Index of Applications in Examples and Activities Act: Activity; Ex: Example Agriculture
College Life
Grape production: Ex 3.6 Strength of bark board: Ex 16.4 Tomato yield and planting density: Ex 15.12, Ex 15.13
Academic success of college sophomores: Ex 14.1 Advantages of multiple SAT scores in college admissions: Act 8.1 Asking questions in seminar class: Ex 6.5 Back-to-college spending: Ex 3.7 College attendance: Ex 10.12 College choice do-over: Ex 1.5 Comparing job offers: Ex 4.18 Detecting plagiarism: Ex 10.9 Enrollments at public universities: Ex 3.15 Gender of college students: Ex 8.7 Graduation rates: Ex 1.10, Ex 13.5, Ex 13.10 Graduation rates and student-related expenditures: Ex 5.1 Graduation rates at small colleges: Ex 14.6, Ex 14.7, Ex 14.8, Ex 14.9 How safe are college campuses? Ex 1.6 Impact of internet and television use on college student reading habits: Ex 9.2 Importance of college education: Ex 9.4 Internet use by college students: Ex 9.1 Math SAT score distribution: Ex 3.14 Misreporting grade point average: Ex 3.17 Money spent on textbooks: Ex 8.1 Predicting graduation rates: Ex 5.13 Roommate satisfaction: Ex 15.7 Students with jumper cables: Ex 7.22 Study habits of college seniors: Ex 3.5 Time required to complete registration: Ex 7.29 Travel distance to college: Ex 3.1 Tuition at public universities: Ex 3.9 Verbal SAT scores: Ex 3.21 Visits to class web site: Ex 4.3, Ex 4.4
Biology Age and flexibility: Act 5.2 Age of a lobster: Ex 5.19 Bee mating behavior: Ex 3.12, Ex 3.13 Black bear habitat selection: Ex 5.9 Calling behavior of Amazonian frogs: Ex 5.22 Cannibalism in wolf spiders: Ex 5.20, Ex 5.21 Charitable behavior of chimpanzees: Ex 9.10, Ex 11.7 Chirp rate for crickets: Ex 10.15 Distance deer mice will travel for food: Ex 5.7, Ex 5.10 Dominant and nondominant hands: Act 3.2 Egg weights: Ex 7.31 Head circumference at birth: Ex 4.19 Loon chick survival factors: Ex 5.17 Predator inspection in guppies: Ex 6.18 Recognizing your roommate’s scent: Ex 7.18 Reflexes with dominant and nondominant hands: Act 11.2 Repertoire size and body characteristics of nightingales: Ex 5.23 Scorpionfly courtship: Ex 8.4 Shark length and jaw width: Ex 13.11, Ex 13.12 Spider phobia: Ex 1.4
Business and Economics Application processing times: Ex 7.8 Cable services: Ex 6.22 Car sales: Ex 7.1 Christmas Price Index: Ex 3.22 Cost of Big Macs: Ex 4.7, Ex 4.8, Ex 4.12 Cost of energy bars: Ex 14.11 Cost of residential air-conditioning: Ex 15.8, Ex 15.9 Credit cards paid in full: Ex 7.21 Daily wasted time at work: Ex 10.14 Education level and income: Ex 3.23 Express mail volume: Ex 7.35 Hybrid car sales: Ex 12.3 Licensing example attempts: Ex 7.9 Mortgage choices: Ex 6.16 Predicting house prices: Ex 14.5 Price of fish: Ex 14.13 Prices of industrial properties: Ex 14.17, Ex 14.19 Resume typos: Ex 3.3 Starting salaries of business school graduates: Ex 16.11, Ex 16.12
Demography and Population Characteristics County population sizes: Ex 4.2 Head circumferences: Act 1.2 Heights and weights of American women: Ex 5.8 Heights of college athletes: Ex 1.1 Heights of mothers: Ex 4.17 Hitchhiker’s thumb: Ex 6.17 Median ages in 2030: Ex 3.10 Newborn birth weights: Ex 7.27 Percentage of population with higher education degrees: Ex 4.9, Ex 4.10 Two-child families: Ex 6.14 Women’s heights and number of siblings: Act 13.1
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
Education and Child Development After-school activities: Ex 6.11 Chess lessons and memory improvement: Ex 11.6 Childcare for preschoolers: Ex 4.15 College plans of high school seniors: Ex 7.4 Combining exam scores: Ex 7.16 Helping hands: Ex 2.7, 2.9 IQ scores: Ex 4.16, Ex 7.28 Predictors of writing competence: Ex 14.4 School enrollment in Northern and Central Africa: Ex 3.11 Standardized test scores: Ex 4.14, Ex 10.16 Students’ knowledge of geography: Act 3.1 Television viewing habits of children: Ex 3.16
Environmental Science Cosmic radiation: Ex 9.7 Lead in tap water: Ex 10.7 Rainfall frequency distributions for Albuquerque: Ex 3.19 River water velocity and distance from shore: Ex 5.16 Soil and sediment characteristics: Ex 14.12, Ex 14.14, Ex 14.16 Water conservation: Ex 10.10 Water quality: Ex 1.2
Food Science Calorie consumption at fast food restaurants: Ex 2.2 Fat content of hot dogs: Ex 8.6 Fish food: Ex 5.15 Pomegranate juice and tumor growth: Ex 5.5 Tannin concentration in wine: Ex 5.2, Ex 5.6
Leisure and Popular Culture Car preferences: Ex 6.1 Do U Txt?: Ex 1.7 iPod shuffles: Ex 7.7 Life insurance for cartoon characters: Ex 2.3 Number of trials required to complete game: Ex 7.2 Probability a Hershey’s Kiss will land on its base: Act 6.1 Selecting cards: Ex 6.20 Selection of contest winners: Ex 6.7 Tossing a coin: Ex 6.8 Twitter words: Act 1.1
Manufacturing and Industry Bottled soda volumes: Ex 8.5 Comprehensive strength of concrete: Ex 7.14 Computer configurations: Ex 6.19 Computer sales: Ex 7.19 Corrosion of underground pipe coatings: Ex 15.14 Durable press rating of cotton fabric: Ex 14.18 DVD player warranties: Ex 6.24 Engineering stress test: Ex 7.3
Ergonomic characteristics of stool designs: Ex 15.10, Ex 15.11 Garbage truck processing times: Ex 7.30 GFI switches: Ex 6.12 Lifetime of compact florescent lightbulbs: Ex 10.2 On-time package delivery: Ex 10.18 Paint flaws: Ex 7.6 Testing for flaws: Ex 7.11, Ex 7.12
Marketing and Consumer Behavior Car choices: Ex 6.10 Energy efficient refrigerators: Ex 7.5 High-pressure sales tactics: Ex 16.13 Impact of food labels: Ex 10.8 Online security: Ex 7.20 Satisfaction with cell phone service: Ex 4.6
Medical Science Apgar scores: Ex 7.10, Ex 7.13 Affect of long work hours on sleep: Ex 11.10 Births and the lunar cycle: Ex 12.1, Ex 12.2 Blood platelet volume: Ex 8.2 Blood pressure and kidney disease: Ex 16.5 Blood test for ovarian cancer: Ex 10.6 Cardiovascular fitness of teens: Ex 10.11 Cerebral volume and ADHD: Ex 11.1 Chronic airflow obstruction: Ex 16.9 Contracting hepatitis from blood transfusion: Ex 8.8, Ex 8.9 Cooling treatment after oxygen deprivation in newborns: Ex 2.5 Diagnosing tuberculosis: Ex 6.15 Drive-through medicine: Ex 9.8 Effect of talking on blood pressure: Ex 11.4 Effects of ethanol on sleep time: Ex 15.6 Evaluating disease treatments: Ex 10.3 Facial expression and self-reported pain level: Ex 12.7 Growth hormone levels and diabetes: Ex 16.10 Hip-to-waist ratio and risk of heart attack: Ex 14.2 Hormones and body fat: Ex 15.4, Ex 15.5 Lead exposure and brain volume: Ex 5.12 Lyme disease: Ex 6.27 Markers for kidney disease: Ex 7.34 Maternal age and baby’s birth weight: Ex 13.2 Medical errors: Ex 6.9 Parental smoking and infant health: Ex 16.2, Ex 16.3 Passive knee extension: Ex 4.1 Platelet volume and heart attack risk: Ex 15.1, Ex 15.2, Ex 15.3 Premature births: Ex 7.36 Sleep duration and blood leptin level: Ex 13.13 Slowing the growth rate of tumors: Ex 10.5 Stroke mortality and education: Ex 12.8 Surviving a heart attack: Ex 6.13 Time perception and nicotine withdrawal: Ex 10.13
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
Treating dyskinesia: Ex 16.8 Treatment for acute mountain sickness: Act 2.5 Ultrasound in treatment of soft-tissue injuries: Ex 11.5, Ex 11.8 Video games and pain management: Act 2.4 Vitamin B12 levels in human blood: Ex 16.7 Waiting time for cardiac procedures in Canada: Ex 9.9 Wart removal methods: Ex 11.9
Physical Sciences Rainfall data: Ex 7.33 Snow cover and temperature: Ex 13.8 Wind chill factor: Ex 14.3
Politics and Public Policy Fair hiring practices: Ex 6.29 Opinions on freedom of speech: Ex 11.11 Predicting election outcomes: Ex 13.3, Ex 13.6, Ex 13.7 Recall petition signatures: Act 9.3 Requests for building permits: Ex 6.31 School board politics: Ex 14.10 Support for affirmative action: Ex 9.1, Ex 9.4
Psychology, Sociology, and Social Issues Benefits of acting out: Ex 1.3 Color and perceived taste: Act 12.2 Estimating sizes: Act 1.3 Extrasensory perception: Ex 6.33 Gender and salary: Ex 11.2 Golden rectangles: Ex 4.11 Hand-holding couples: Ex 6.30 Internet addiction: Ex 6.28 Motivation for revenge: Ex 2.4 One-boy family planning: Ex 6.32 Reading emotions: Ex 11.3 Stroop effect: Act 2.2 Subliminal messages: Ex 2.5 Weight regained proportions for three follow-up methods: Ex 12.6
Public Health and Safety Careless or aggressive driving: Ex 9.5 Effect of cell phone distraction: Ex 2.8 Effects of McDonald’s hamburger sales: Act 2.3
Nicotine content of cigarettes: Ex 10.17 Safety of bicycle helmets: Ex 5.3 Salmonella in restaurant eggs: Act 7.2 Teenage driver citations and traffic school: Ex 6.23
Sports Age and marathon times: Ex 5.4, Ex 5.14 Calling a toss at a football game: Ex 6.6 Concussions in collegiate sports: Ex 12.4, Ex 12.5 Fairness of Euro coin-flipping in European sports: Act 6.2 Helium-filled footballs: Act 11.1 “Hot hand” in basketball: Act 6.3 Losing at golf: Ex 6.2, Ex 6.4 NBA player salaries: Ex 4.5, Ex 4.13 Olympic figure skating: Ex 3.20 Racing starts in competitive swimming: Ex 16.6 Soccer goalie action bias: Ex 6.26 Tennis ball diameters: Ex 10.1 Time to first goal in hockey: Ex 8.3 Treadmill time to exhaustion and ski time of biathletes: Ex 13.4, Ex 13.9 Wrestlers’ weight loss by headstand: Ex 13.1
Surveys and Opinion Polls Are cell phone users different?: Ex 2.1 Collecting and summarizing numerical data: Act 2.2 Designing a sampling plan: Facebook friending: Act 2.1 Selecting a random sample: Ex 2.2
Transportation Accidents by bus drivers: Ex 3.18 Airborne times for San Francisco to Washington D.C. flight: Ex 9.3 Airline luggage weights: Ex 7.17 Airline passenger weights: Act 4.2 Automobile accidents by occupation: Ex 3.8 Comparing gasoline additives: Ex 2.10 Freeway traffic: Ex 7.15 Fuel efficiency of automobiles: Ex 16.1 Lost airline luggage: Ex 6.25 Motorcycle helmets: Ex 1.8, Ex 1.9 On-time airline flights: Ex 10.4 Predicting transit times: Ex 14.15 Turning directions on freeway off-ramp: Ex 6.3
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
This is an electronic version of the print textbook. Due to electronic rights restrictions, some third party content may be suppressed. Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. The publisher reserves the right to remove content from this title at any time if subsequent rights restrictions require it. For valuable information on pricing, previous editions, changes to current editions, and alternate formats, please visit www.cengage.com/highered to search by ISBN#, author, title, or keyword for materials in your areas of interest.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
Statistics: The Exploration and Analysis of Data, Seventh Edition Roxy Peck, Jay L. Devore Publisher: Richard Stratton Senior Sponsoring Editor: Molly Taylor Senior Developmental Editor: Jay Campbell Associate Editor: Daniel Seibert
© 2012, 2008, 2005 Brooks/Cole, Cengage Learning ALL RIGHTS RESERVED. No part of this work covered by the copyright herein may be reproduced, transmitted, stored, or used in any form or by any means, graphic, electronic, or mechanical, including but not limited to photocopying, recording, scanning, digitizing, taping, Web distribution, information networks, or information storage and retrieval systems, except as permitted under Section 107 or 108 of the 1976 United States Copyright Act, without the prior written permission of the publisher.
Senior Editorial Assistant: Shaylin Walsh Associate Media Editor: Andrew Coppola Marketing Manager: Ashley Pickering Marketing Coordinator: Erica O’Connell Marketing Communications Manager: Mary Anne Payumo
For product information and technology assistance, contact us at Cengage Learning Customer & Sales Support, 1-800-354-9706 For permission to use material from this text or product, submit all requests online at www.cengage.com/permissions. Further permissions questions can be emailed to [emailprotected].
Content Project Manager: Susan Miscio Art Director: Linda Helcher
Library of Congress Control Number: 2010937362
Senior Print Buyer: Diane Gibbons
ISBN-13: 978-0-8400-5801-0
Rights Acquisition Specialist: Mandy Groszko
ISBN-10: 0-8400-5801-2
Production Service/Compositor: Graphic World Inc. Text designer: Rokusek Design Cover designer: RHDG
Brooks/Cole 20 Channel Center Street Boston, MA 02210 USA
Cover Image: © Joella Jean Mahoney/Red Stone Gallery Cengage Learning is a leading provider of customized learning solutions with office locations around the globe, including Singapore, the United Kingdom, Australia, Mexico, Brazil, and Japan. Locate your local office at: international.cengage.com/region. Cengage Learning products are represented in Canada by Nelson Education, Ltd. For your course and learning solutions, visit www.cengage.com. Purchase any of our products at your local college store or at our preferred online store www.cengagebrain.com.
Printed in the United States of America 1 2 3 4 5 6 7 14 13 12 11 10
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
Statistics The Exploration and Analysis of Data
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
To Beth Chance and Allan Rossman, whose dedication to improving statistics education is inspirational R. P. To Carol, Allie, and Teri J. D.
About the Cover The cover image is by artist Joella Jean Mahoney, who paints striking abstract landscapes inspired by the American Southwest. In her work, Mahoney is able to beautifully capture the underlying structure of rock formations and canyons. In statistical analyses, we work to capture and learn from the underlying structure we find in data. While the images we create are not nearly as beautiful as Mahoney’s work, in this sense we share a similar goal! Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
EDITION
7
Statistics The Exploration and Analysis of Data Roxy Peck California Polytechnic State University, San Luis Obispo
Jay L. Devore California Polytechnic State University, San Luis Obispo
Australia • Brazil • Japan • Korea • Mexico • Singapore • Spain • United Kingdom • United States
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
About the Authors ROXY PECK is Emerita Associate Dean of the College of Science and Mathematics and Professor of Statistics Emerita at California Polytechnic State University, San Luis Obispo. A faculty member at Cal Poly from 1979 until 2009, Roxy served for 6 years as Chair of the Statistics Department before becoming Associate Dean, a position she held for 13 years. She received an M.S. in Mathematics and a Ph.D. in Applied Statistics from the University of California, Riverside. Roxy is nationally known in the area of statistics education, and she was presented with the Lifetime Achievement Award in Statistics Education at the U.S. Conference on Teaching Statistics in 2009. In 2003 she received the American Statistical Association’s Founder’s Award, recognizing her contributions to K–12 and undergraduate statistics education. She is a Fellow of the American Statistical Association and an elected member of the International Statistics Institute. Roxy served for 5 years as the Chief Reader for the Advanced Placement Statistics Exam and has chaired the American Statistical Association’s Joint Committee with the National Council of Teachers of Mathematics on Curriculum in Statistics and Probability for Grades K–12 and the Section on Statistics Education. In addition to her texts in introductory statistics, Roxy is also co-editor of Statistical Case Studies: A Collaboration Between Academe and Industry and a member of the editorial board for Statistics: A Guide to the Unknown, 4th edition. Outside the classroom, Roxy likes to travel and spends her spare time reading mystery novels. She also collects Navajo rugs and heads to Arizona and New Mexico whenever she can find the time.
JAY L. DEVORE earned his undergraduate degree in Engineering Science from the University of California, Berkeley; spent a year at the University of Sheffield in England; and finished his Ph.D. in statistics at Stanford University. He previously taught at the University of Florida and at Oberlin College and has had visiting appointments at Stanford, Harvard, the University of Washington, New York University, and Columbia. From 1998 to 2006, Jay served as Chair of the Statistics Department at California Polytechnic State University, San Luis Obispo. The Statistics Department at Cal Poly has an international reputation for activities in statistics education. In addition to this book, Jay has written several widely used engineering statistics texts and a book in applied mathematical statistics. He is currently collaborating on a business statistics text, and he also serves as an Associate Editor for Reviews for several statistics journals. He is the recipient of a distinguished teaching award from Cal Poly and is a Fellow of the American Statistical Association. In his spare time, he enjoys reading, cooking and eating good food, playing tennis, and traveling to faraway places. He is especially proud of his wife, Carol, a retired elementary school teacher; his daughter Allison, the executive director of a nonprofit organization in New York City; and his daughter Teresa, an ESL teacher in New York City.
vi Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
Brief Contents 1 CHAPTER 2 CHAPTER 3 CHAPTER 4 CHAPTER 5 CHAPTER 6 CHAPTER 7 CHAPTER 8 CHAPTER 9 CHAPTER 1 0 CHAPTER 1 1 CHAPTER 1 2 CHAPTER 1 3 CHAPTER 1 4 CHAPTER 1 5 CHAPTER 1 6 CHAPTER
The Role of Statistics and the Data Analysis Process 1 Collecting Data Sensibly 31 Graphical Methods for Describing Data 89 Numerical Methods for Describing Data 163 Summarizing Bivariate Data 211 Probability 301 Population Distributions 333 Sampling Variability and Sampling Distributions 385 Estimation Using a Single Example 411 Hypothesis Testing Using a Single ample 457 Comparing Two Populations or Treatments 515 The Analysis of Categorical Data and Goodness-of-Fit Tests 573 Simple Linear Regression and Correlation: Inferential Methods 611 Multiple Regression Analysis 671 Analysis of Variance 703 Nonparametric (Distribution-Free) Statistical Methods 16-1 Appendix A: The Binomial Distribution 729 Appendix B: Statistical Tables 739 Appendix C: References 759 Answers to Selected Odd-Numbered Exercises 763 Index 781 Sections and/or chapter numbers in color can be found at http://www.cengage.com/statistics/peck
vii Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
Contents CHAPTER
1
The Role of Statistics and the Data Analysis Process 1 1.1 1.2 1.3 1.4
CHAPTER
2
Collecting Data Sensibly 31 2.1 2.2 2.3 2.4 2.5 2.6
CHAPTER
3
4
Statistical Studies: Observation and Experimentation 32 Sampling 37 Simple Comparative Experiments 49 More on Experimental Design 65 More on Observational Studies: Designing Surveys (Optional) 70 Interpreting and Communicating the Results of Statistical Analyses 76 Activity 2.1 Facebook Friending 79 Activity 2.2 An Experiment to Test for the Stroop Effect 80 Activity 2.3 McDonald’s and the Next 100 Billion Burgers 81 Activity 2.4 Video Games and Pain Management 81 Activity 2.5 Be Careful with Random Assignment! 82
Graphical Methods for Describing Data 89 3.1 3.2 3.3 3.4 3.5
CHAPTER
Why Study Statistics? 2 The Nature and Role of Variability 3 Statistics and the Data Analysis Process 5 Types of Data and Some Simple Graphical Displays 10 Activity 1.1 Twitter Words 25 Activity 1.2 Head Sizes: Understanding Variability 26 Activity 1.3 Estimating Sizes 26 Activity 1.4 A Meaningful Paragraph 28
Displaying Categorical Data: Comparative Bar Charts and Pie Charts 90 Displaying Numerical Data: Stem-and-Leaf Displays 101 Displaying Numerical Data: Frequency Distributions and Histograms 111 Displaying Bivariate Numerical Data 133 Interpreting and Communicating the Results of Statistical Analyses 142 Activity 3.1 Locating States 152 Activity 3.2 Bean Counters! 152 Cumulative Review Exercises 154
Numerical Methods for Describing Data 163 4.1 4.2 4.3 4.4
Describing the Center of a Data Set 164 Describing Variability in a Data Set 175 Summarizing a Data Set: Boxplots 184 Interpreting Center and Variability: Chebyshev’s Rule, the Empirical Rule, and z Scores 190 4.5 Interpreting and Communicating the Results of Statistical Analyses 199 Activity 4.1 Collecting and Summarizing Numerical Data 204 Activity 4.2 Airline Passenger Weights 204 Activity 4.3 Boxplot Shapes 205 viii Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
Contents
CHAPTER
5
Summarizing Bivariate Data 211 5.1 5.2 5.3 5.4 5.5 5.6
CHAPTER
6
ix
Correlation 212 Linear Regression: Fitting a Line to Bivariate Data 223 Assessing the Fit of a Line 234 Nonlinear Relationships and Transformations 253 Logistic Regression (Optional) 274 Interpreting and Communicating the Results of Statistical Analyses 283 Activity 5.1 Exploring Correlation and Regression Technology Activity (Applets) 290 Activity 5.2 Age and Flexibility 290 Cumulative Review Exercises 295
Probability 301 6.1 Interpreting Probabilities and Basic Probability Rules 302 6.2 Probability as a Basis for Making Decisions 312 6.3 Estimating Probabilities Empirically and by Using Simulation 316 Activity 6.1 Kisses 328 Activity 6.2 A Crisis for European Sports Fans? 328 Activity 6.3 The “Hot Hand” in Basketball 328
CHAPTER
7
Population Distributions 333 7.1 7.2 7.3 7.4
CHAPTER
8
Describing the Distribution of Values in a Population 334 Population Models for Continuous Numerical Variables 342 Normal Distributions 350 Checking for Normality and Normalizing Transformations 367 Activity 7.1 Is It Real? 380 Activity 7.2 Rotten Eggs? 380 Cumulative Review Exercises 383
Sampling Variability and Sampling Distributions 385 8.1 Statistics and Sampling Variability 386 8.2 The Sampling Distribution of a Sample Mean 390 8.3 The Sampling Distribution of a Sample Proportion 401 Activity 8.1 Do Students Who Take the SATs Multiple Times Have an Advantage in College Admissions? 407
CHAPTER
9
Estimation Using a Single Example 411 9.1 9.2 9.3 9.4
CHAPTER
10
Point Estimation 412 Large-Sample Confidence Interval for a Population Proportion 418 Confidence Interval for a Population Mean 431 Interpreting and Communicating the Results of Statistical Analyses 445 Activity 9.1 Getting a Feel for Confidence Level 450 Activity 9.2 An Alternative Confidence Interval for a Population Proportion 452 Activity 9.3 Verifying Signatures on a Recall Petition 452 Activity 9.4 A Meaningful Paragraph 453
Hypothesis Testing Using a Single Sample 457 10.1 10.2 10.3 10.4
Hypotheses and Test Procedures 458 Errors in Hypothesis Testing 462 Large-Sample Hypothesis Tests for a Population Proportion 468 Hypothesis Tests for a Population Mean 482
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
x
Contents
10.5 Power and Probability of Type II Error 493 10.6 Interpreting and Communicating the Results of Statistical Analyses 502 Activity 10.1 Comparing the t and z Distributions 506 Activity 10.2 A Meaningful Paragraph 507 Cumulative Review Exercises 510 CHAPTER
11
Comparing Two Populations or Treatments 515 11.1
Inferences Concerning the Difference Between Two Population or Treatment Means Using Independent Samples 516 11.2 Inferences Concerning the Difference Between Two Population or Treatment Means Using Paired Samples 536 11.3 Large Sample Inferences Concerning a Difference Between Two Population or Treatment Proportions 549 11.4 Interpreting and Communicating the Results of Statistical Analyses 561 Activity 11.1 Helium-Filled Footballs? 565 Activity 11.2 Thinking About Data Collection 565 Activity 11.3 A Meaningful Paragraph 567 CHAPTER
12
The Analysis of Categorical Data and Goodness-of-Fit Tests 573 12.1 Chi-Square Tests for Univariate Data 574 12.2 Tests for Homogeneity and Independence in a Two-way Table 585 12.3 Interpreting and Communicating the Results of Statistical Analyses 601 Activity 12.1 Pick a number, any number... 605 Activity 12.2 Color and Perceived Taste 606
CHAPTER
13
Simple Linear Regression and Correlation: Inferential Methods 611 13.1 13.2 13.3 13.4 13.5 13.6
CHAPTER
14
Simple Linear Regression Model 612 Inferences About the Slope of the Population Regression Line 625 Checking Model Adequacy 635 Inferences Based on the Estimated Regression Line (Optional) 646 Inferences About the Population Correlation Coefficient (Optional) 654 Interpreting and Communicating the Results of Statistical Analyses 658 Activity 13.1 Are Tall Women from “Big” Families? 660 Cumulative Review Exercises 666
Multiple Regression Analysis 671 14.1 Multiple Regression Models 672 14.2 Fitting a Model and Assessing Its Utility 685 Activity 14.1 Exploring the Relationship Between Number of Predictors and Sample Size 701 14.3 Inferences Based on an Estimated Model 14-1 14.4 Other Issues in Multiple Regression 14-13 14.5 Interpreting and Communicating the Results of Statistical Analyses 14-25
CHAPTER
15
Analysis of Variance 703 15.1 15.2 15.3 15.4 15.5
Single-Factor ANOVA and the F Test 704 Multiple Comparisons 717 The F Test for a Randomized Block Experiment 15-1 Two-Factor ANOVA 15-8 Interpreting and Communicating the Results of Statistical Analyses 15-19 Activity 15.1 Exploring Single-Factor ANOVA 725
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
Contents
CHAPTER
16
xi
Nonparametric (Distribution-Free) Statistical Methods 16-1 16.1 Distribution-Free Procedures for Inferences About a Difference Between Two Population or Treatment Means Using Independent Samples (Optional) 16-2 16.2 Distribution-Free Procedures for Inferences About a Difference Between Two Population or Treatment Means Using Paired Samples 16-10 16.3 Distribution-Free ANOVA 16-22 Appendix A: The Binomial Distribution 729 Appendix B: Statistical Tables 739 Appendix C: References 759 Answers to Selected Odd-Numbered Exercises 763 Index 781 Sections and/or chapter numbers in color can be found at http://www.cengage.com/statistics/peck
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
Preface
In a nutshell, statistics is about understanding the role that variability plays in drawing conclusions based on data. Statistics: The Exploration and Analysis of Data, Seventh Edition, develops this crucial understanding of variability through its focus on the data analysis process.
An Organization That Reflects the Data Analysis Process Students are introduced early to the idea that data analysis is a process that begins with careful planning, followed by data collection, data description using graphical and numerical summaries, data analysis, and finally interpretation of results. This process is described in detail in Chapter 1, and the ordering of topics in the first ten chapters of the book mirrors this process: data collection, then data description, then statistical inference. The logical order in the data analysis process can be pictured as shown in the following figure. Step 1: Acknowledging Variability— Collecting Data Sensibly
Step 2: Describing Variability in the Data— Descriptive Statistics
Step 3: Drawing Conclusions in a Way That Recognizes Variability in the Data
Unlike many introductory texts, Statistics: The Exploration and Analysis of Data, Seventh Edition, is organized in a manner consistent with the natural order of the data analysis process: xii Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
Preface
Step 1: Acknowledging Variability— Collecting Data Sensibly
Step 2: Describing Variability in the Data— Descriptive Statistics
Chapters 1–2
Chapters 3–5
Probability Supports the Connection
Chapters 6–7
xiii
Step 3: Drawing Conclusions in a Way That Recognizes Variability in the Data
Chapters 8–15
The Importance of Context and Real Data Statistics is not about numbers; it is about data—numbers in context. It is the context that makes a problem meaningful and something worth considering. For example, exercises that ask students to compute the mean of 10 numbers or to construct a dotplot or boxplot of 20 numbers without context are arithmetic and graphing exercises. They become statistics problems only when a context gives them meaning and allows for interpretation. While this makes for a text that may appear “wordy” when compared to traditional mathematics texts, it is a critical and necessary component of a modern statistics text. Examples and exercises with overly simple settings do not allow students to practice interpreting results in authentic situations or give students the experience necessary to be able to use statistical methods in real settings. We believe that the exercises and examples are a particular strength of this text, and we invite you to compare the examples and exercises with those in other introductory statistics texts. Many students are skeptical of the relevance and importance of statistics. Contrived problem situations and artificial data often reinforce this skepticism. A strategy that we have employed successfully to motivate students is to present examples and exercises that involve data extracted from journal articles, newspapers, and other published sources. Most examples and exercises in the book are of this nature; they cover a very wide range of disciplines and subject areas. These include, but are not limited to, health and fitness, consumer research, psychology and aging, environmental research, law and criminal justice, and entertainment.
A Focus on Interpretation and Communication Most chapters include a section titled “Interpreting and Communicating the Results of Statistical Analyses.” These sections include advice on how to best communicate the results of a statistical analysis and also consider how to interpret statistical sumCopyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
xiv
Preface
maries found in journals and other published sources. A subsection titled “A Word to the Wise” reminds readers of things that must be considered in order to ensure that statistical methods are employed in reasonable and appropriate ways.
Consistent with Recommendations for the Introductory Statistics Course Endorsed by the American Statistical Association In 2005, the American Statistical Association endorsed the report “College Guidelines in Assessment and Instruction for Statistics Education (GAISE Guidelines),” which included the following six recommendations for the introductory statistics course: 1. 2. 3. 4. 5. 6.
Emphasize statistical literacy and develop statistical thinking. Use real data. Stress conceptual understanding rather than mere knowledge of procedures. Foster active learning in the classroom. Use technology for developing conceptual understanding and analyzing data. Use assessments to improve and evaluate student learning.
Statistics: The Exploration and Analysis of Data, Seventh Edition, is consistent with these recommendations and supports the GAISE guidelines in the following ways: 1. Emphasize statistical literacy and develop statistical thinking. Statistical literacy is promoted throughout the text in the many examples and exercises that are drawn from the popular press. In addition, a focus on the role of variability, consistent use of context, and an emphasis on interpreting and communicating results in context work together to help students develop skills in statistical thinking. 2. Use real data. The examples and exercises from Statistics: The Exploration and Analysis of Data, Seventh Edition, are context driven, and the reference sources include the popular press as well as journal articles. 3. Stress conceptual understanding rather than mere knowledge of procedures. Nearly all exercises in Statistics: The Exploration and Analysis of Data, Seventh Edition are multipart and ask students to go beyond just computation. They focus on interpretation and communication, not just in the chapter sections specifically devoted to this topic, but throughout the text. The examples and explanations are designed to promote conceptual understanding. Hands-on activities in each chapter are also constructed to strengthen conceptual understanding. Which brings us to . . . 4. Foster active learning in the classroom. While this recommendation speaks more to pedagogy and classroom practice, Statistics: The Exploration and Analysis of Data, Seventh Edition, provides more than 30 hands-on activities in the text and additional activities in the accompanying instructor resources that can be used in class or assigned to be completed outside of class. In addition, accompanying online materials allow students to assess their understanding and develop a personalized learning plan based on this assessment for each chapter.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
Preface
xv
5. Use technology for developing conceptual understanding and analyzing data. The computer has brought incredible statistical power to the desktop of every investigator. The wide availability of statistical computer packages such as Minitab, S-Plus, JMP, and SPSS, and the graphical capabilities of the modern microcomputer have transformed both the teaching and learning of statistics. To highlight the role of the computer in contemporary statistics, we have included sample output throughout the book. In addition, numerous exercises contain data that can easily be analyzed by computer, though our exposition firmly avoids a presupposition that students have access to a particular statistical package. Technology manuals for specific packages, such as Minitab and SPSS, and for the graphing calculator are available in the online materials that accompany this text. 6. Use assessments to improve and evaluate student learning. Assessment materials in the form of a test bank, quizzes, and chapter exams are available in the instructor resources that accompany this text. The items in the test bank reflect the data-in-context philosophy of the text’s exercises and examples.
Topic Coverage Our book can be used in courses as short as one quarter or as long as one year in duration. Particularly in shorter courses, an instructor will need to be selective in deciding which topics to include and which to set aside. The book divides naturally into four major sections: collecting data and descriptive methods (Chapters 1–5), probability material (Chapters 6–8), the basic one- and two-sample inferential techniques (Chapters 9–12), and more advanced inferential methodology (Chapters 13–16). We include an early chapter (Chapter 5) on descriptive methods for bivariate numerical data. This early exposure raises questions and issues that should stimulate student interest in the subject; it is also advantageous for those teaching courses in which time constraints preclude covering advanced inferential material. However, this chapter can easily be postponed until the basics of inference have been covered, and then combined with Chapter 13 for a unified treatment of regression and correlation. With the possible exception of Chapter 5, Chapters 1–10 should be covered in order. We anticipate that most instructors will then continue with two-sample inference (Chapter 11) and methods for categorical data analysis (Chapter 12), although regression could be covered before either of these topics. Optional portions of Chapter 14 (multiple regression) and Chapter 15 (analysis of variance) and Chapter 16 (nonparametric methods) are included in the online materials that accompany this text.
A Note on Probability This book takes a brief and informal approach to probability, focusing on those concepts needed to understand the inferential methods covered in the later chapters. For those who prefer a more traditional approach to probability, the book Introduction to Statistics and Data Analysis by Roxy Peck, Chris Olsen, and Jay Devore may be a more appropriate choice. Except for the more formal treatment of probability and the inclusion of optional Graphing Calculator Explorations, it parallels the material in this text. Please contact your sales representative for more information about this alternative and other alternative customized options available to you.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
xvi
Preface
In This Edition Look for the following in the Seventh Edition: • More than 50 new examples and more than 270 new exercises that use data
•
•
•
• • •
•
from current newspapers and journals are included. In addition, more of the exercises specifically ask students to write (for example, by requiring students to explain their reasoning, interpret results, and comment on important features of an analysis). Examples and exercises that make use of data sets that can be accessed online from the text website are designated by an icon in the text, as are examples that are further illustrated in the technology manuals (Minitab, SPSS, etc.) that are available in the online materials that accompany this text. Approximately 90 exercises have video solutions, presented by Brian Kotz of Montgomery College, which can be viewed online or downloaded for viewing later. These exercises are designated by an icon in the text. Exercises have been added to the Interpreting and Communicating the Results of Statistical Analyses sections. These exercises give students the chance to practice these important skills. These activities can be used as a chapter capstone or can be integrated at appropriate places as the chapter material is covered in class. Students can now go online with Aplia and CourseMate to further their understanding of the material covered in each chapter. Advanced topics that are often omitted in a one-quarter or one-semester course, such as inference and variable selection methods in multiple regression (Sections 14.3 and 14.4), analysis of variance for randomized block and two-factor designs (Sections 15.3 and 15.4), and distribution-free procedures (Chapter 16), are available in the online materials that accompany this text. Updated materials for instructors are included. In addition to the usual instructor supplements such as a complete solutions manual and a test bank, the following are also available to instructors: • An Instructor’s Resource Binder, which contains additional examples that can be incorporated into classroom presentations and cross-references to resources such as Fathom, Workshop Statistics, and Against All Odds. Of particular interest to those teaching Advanced Placement Statistics, the binder also includes additional data analysis questions of the type encountered on the free response portion of the Advanced Placement exam, as well as a collection of model responses. • For those who use student-response systems in class, a set of “clicker” questions (see JoinIn™ on TurningPoint® under Instructor Resources—Media) for assessing student understanding is available.
Student Resources Digital To access additional course materials and companion resources, please visit www .cengagebrain.com. At the CengageBrain.com home page, search for the ISBN of your title (from the back cover of your book) using the search box at the top of the page. This will take you to the product page where free companion resources can be found. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
Preface
xvii
If your text includes a printed access card, you will have instant access to the following resources referenced throughout your text: • Complete step-by-step instructions for TI-84 Graphing Calculators, Excel,
icon throughout the text. Minitab, SPSS, and JMP indicated by the • Data sets in TI-84, Excel, Minitab, SPSS, SAS, JMP, and ASCII file formats indicated by the icon throughout the text. • Applets used in the Activities found in the text. Also available are other significant online resources: Aplia: Aplia™ is an online interactive learning solution that improves comprehension and outcomes by increasing student effort and engagement. Founded by a professor to enhance his own courses, Aplia provides automatically graded assignments with detailed, immediate explanations for every question, along with innovative teaching materials. Our easy-to-use system has been used by more than 1,000,000 students at over 1800 institutions. Exercises were authored by Aplia content experts and, new for this edition, also taken directly from text. CourseMate: Interested in a simple way to complement your text and course content with study and practice materials? Cengage Learning’s CourseMate brings course concepts to life with interactive learning, study, and exam preparation tools that support the printed textbook. Watch student comprehension soar as your class works with the printed textbook and the textbook-specific website. CourseMate goes beyond the book to deliver what you need! This online component provides a rich array of interactive and supplementary material to accompany the text. Online quizzes, conceptual applets, videos, and a multimedia eBook give students dynamic tools for hands-on learning. An online Activities Manual allows students to take notes and record data with activities from the textbook as well as additional bonus activities for each chapter. Step-by-Step Technology Manuals for Microsoft Excel, Minitab, SPSS, JMP, and TI-84 calculators help students harness the problem-solving power of statistics technology with instruction on how to use these tools, with coverage correlated directly to Examples from the text. Downloadable data sets are also provided for every real-data problem marked in the book in the native file formats for each software type and calculator model covered by the Step-by-Step Manuals. The instructors-only area of CourseMate includes a number of additional classroom aids. Enhanced WebAssign: Exclusively from Cengage Learning, Enhanced WebAssign offers an extensive online program for statistics to encourage the practice that’s so critical for concept mastery. The meticulously crafted pedagogy and exercises in our proven texts become even more effective in Enhanced WebAssign, supplemented by multimedia support and immediate feedback as students complete their assignments. Includes an Enhanced WebAssign Start Smart Guide for Students that helps students get up and running quickly with the program. Key features include: • As many as 1000 homework problems that match the text’s end-of-section exercises • New! Premium eBook with highlighting, note-taking, and search features as well
as links to multimedia resources • Practice Another Version feature on many problems (activated at the instructor’s
discretion), which allows students to attempt the same question with a new set of values until they feel ready to move on • graphPad, which allows students to graph lines, segments, parabolas, and circles as they answer questions Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
xviii
Preface
Print Student Solutions Manual (ISBN: 1-111-57977-6): Contains fully worked-out solutions to all of the odd-numbered exercises in the text, giving students a way to check their answers and ensure that they took the correct steps to arrive at an answer.
Instructor Resources Print Teacher’s Resource Binder (ISBN: 1-111-57474-1): The Teacher’s Resource Binder, prepared by Chris Olsen, is full of wonderful resources for both college professors and AP Statistics teachers. These include: • Additional examples from published sources (with references), classified by chap• •
• • • •
ter in the text. These examples can be used to enrich your classroom discussions. Model responses—examples of responses that can serve as a model for work that would be likely to receive a high mark on the AP exam. A collection of data explorations written by Chris Olsen that can be used throughout the year to help students prepare for the types of questions that they may encounter on the investigative task on the AP Statistics Exam. Advice to AP Statistics teachers on preparing students for the AP Exam, written by Brian Kotz. Activity worksheets, prepared by Carol Marchetti, that can be duplicated and used in class. A list of additional resources for activities, videos, and computer demonstrations, cross-referenced by chapter. A test bank that includes assessment items, quizzes, and chapter exams written by Chris Olsen, Josh Tabor, and Peter Flannigan-Hyde.
Digital • Solution Builder: This online instructor database offers complete worked-out
solutions to all exercises in the text, allowing you to create customized, secure solutions printouts (in PDF format) matched exactly to the problems you assign in class. Sign up for access at www.cengage.com/solutionbuilder. • ExamView® (ISBN: 978-1-111-57423-9): ExamView testing software allows instructors to quickly create, deliver, and customize tests for class in print and online formats, and features automatic grading. Included is a test bank with hundreds of questions customized directly to the text, with all questions also provided in PDF and Microsoft® Word® formats for instructors who opt not to use the software component. ExamView is available within the PowerLecture CD. • PowerLecture (ISBN: 978-1-111-57424-6): This CD-ROM provides the instructor with dynamic media tools for teaching. Create, deliver, and customize tests (both print and online) in minutes with ExamView® Computerized Testing. Easily build solution sets for homework or exams using Solution Builder’s online solutions manual. Microsoft® PowerPoint® lecture slides, JoinIn® assessment material for classroom “clicker” systems, and figures from the book are also included on this CD-ROM. • E-book: This new premium eBook has highlighting, note-taking, and search features as well as links to multimedia resources.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
Preface
xix
• JoinIn™ on TurningPoint® (978-0-495-11881-7): The easiest student class-
room response system to use, JoinIn features instant classroom assessment and learning.
Acknowledgments We are grateful for the thoughtful feedback from the following reviewers that has helped to shape this text over previous editions:
Reviewers for the Seventh Edition Debra Hall Indiana University–Purdue University Indianapolis Hazel Shedd Hinds Community College, Rankin Campus Austin Lampros Colorado State University
Cathleen M. Zucco-Teveloff Rowan University Donna Flint South Dakota State University Douglas A. Noe Miami University Steven T. Garren James Madison University
Rick Gumina Colorado State University
Reviewers of Previous Editions Arun K. Agarwal, Jacob Amidon, Holly Ashton, Barb Barnet, Eddie Bevilacqua, Piotr Bialas, Kelly Black, Jim Bohan, Pat Buchanan, Gabriel Chandler, Andy Chang, Jerry Chen, Richard Chilcoat, Mary Christman, Marvin Creech, Ron Degged, Hemangini Deshmukh, Ann Evans, Guangxiong Fang, Sharon B. Finger, Steven Garren, Mark Glickman, Tyler Haynes, Sonja Hensler, Trish Hutchinson, John Imbrie, Bessie Kirkwood, Jeff Kollath, Christopher Lacke, Michael Leitner, Zia Mahmood, Art Mark, Pam Martin, David Mathiason, Bob Mattson, C. Mark Miller, Megan Mocko, Paul Myers, Kane Nashimoto, Helen Noble, Broderick Oluyede, Elaine Paris, Shelly Ray Parsons, Deanna Payton, Judy Pennington-Price, Michael Phelan, Alan Polansky, Michael Ratliff, David Rauth, Kevin J. Reeves, Lawrence D. Ries, Robb Sinn, Greg Sliwa, Angela Stabley, Jeffery D. Sykes, Yolanda Tra, Joe Ward, Nathan Wetzel, Mark Wilson, Yong Yu, and Toshiyuki Yuasa.
We would also like to express our thanks and gratitude to those whose support made this seventh edition possible: • Molly Taylor, our editor, for her sage advice as she guided us along the way. • Jay Campbell, our developmental editor, who amazed us with his ability to man-
age such a complicated process with kindness and humor, and who kept us on track and moving forward. • Dan Seibert for leading the development of all of the supporting ancillaries. • Mike Ederer, our production editor. • Sandy Brown, our compositor.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
xx
Preface
• • • • •
• • • • • •
Chris Ufer and Rose Boul, who updated the art for this edition. Susan Miscio, our product content manager. Mary Jente for her careful manuscript review and copyediting. Michael Allwood for his heroic work in creating new student and instructor solutions manuals to accompany the text—a huge task he managed beautifully. Kathy Fritz for creating the new interactive PowerPoint presentations that accompany the text and also for sharing her insight in writing the Teaching Tips that accompany each chapter in the annotated instructor editions. Stephen Miller for a masterful job in checking the accuracy of examples and solutions. Brian Kotz for producing the video solutions. Josh Tabor and Peter-Flannagan Hyde for their contributions to the test bank that accompanies the book. Beth Chance and Francisco Garcia for producing the applet used in the confidence interval activities. Gary McClelland for producing the applets from Seeing Statistics used in the regression activities. Carolyn Crockett, our former editor at Cengage, for her support on the previous editions of this book.
And, as always, we thank our families, friends, and colleagues for their continued support. Roxy Peck Jay Devore
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
CHAPTER
1
The Role of Statistics and the Data Analysis Process We encounter data and make conclusions based on data every day. Statistics is the scientific discipline that provides methods to help us make sense of data. Statistical methods, used intelligently, offer a set of powerful tools for gaining insight into the world around us. The widespread use of statistical analyses in diverse fields such as business, medicine, agriculture, social sciences, natural sciences, and engineering has led to increased recognition that statistical literacy—a familiarity with the goals Andresr, 2010/Used under license from Shutterstock.com and methods of statistics—should be a basic component of a well-rounded educational program. The field of statistics teaches us how to make intelligent judgments and informed decisions in the presence of uncertainty and variation. In this chapter, we consider the nature and role of variability in statistical settings, introduce some basic terminology, and look at some simple graphical displays for summarizing data.
Make the most of your study time by accessing everything you need to succeed online with CourseMate. Visit http://www.cengagebrain.com where you will find: • An interactive eBook, which allows you to take notes, highlight, bookmark, search • • • • • •
the text, and use in-context glossary definitions Step-by-step instructions for Minitab, Excel, TI-83/84, SPSS, and JMP Video solutions to selected exercises Data sets available for selected examples and exercises Online quizzes Flashcards Videos
1 Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
2
1.1
Chapter 1 The Role of Statistics and the Data Analysis Process
Why Study Statistics? There is an old saying that “without data, you are just another person with an opinion.” While anecdotes and coincidences may make for interesting stories, you wouldn’t want to make important decisions on the basis of anecdotes alone. For example, just because a friend of a friend ate 16 apricots and then experienced relief from joint pain doesn’t mean that this is all you need to know to help one of your parents choose a treatment for arthritis! Before recommending the apricot treatment, you would definitely want to consider relevant data—that is, data that would allow you to investigate the effectiveness of this treatment. It is difficult to function in today’s world without a basic understanding of statistics. For example, here are a few headlines from articles that draw conclusions based on data that all appeared in a single issue of USA Today (June 29, 2009): •
•
•
•
•
“Infant Colic May Be Linked to Dads” is the headline of an article reporting on a study of the relationship between excessive crying and parents’ depression. The study of more than 7600 babies and their parents concluded that excessive newborn crying is more likely to occur if the father reported symptoms of depression prior to the birth of the baby. The article “Many Adults Can’t Name a Scientist” summarized the results of a survey of 1000 adults. Of those surveyed, 23% were unable to name a single famous scientist. Of those who did come up with a name, Albert Einstein was the scientist of choice, named by 47% of those surveyed. “Few See Themselves as ‘Old’ No Matter What Their Age” is the title of an article that described results from a large survey of 2969 adults. Those surveyed were asked at what age a person would be considered old. The resulting data revealed that there were notable differences in the answer to the question depending on the age of the responder. The average age identified as old by young adults (age 18–29) was 60, while the average was 69 for those who were age 30 to 49, 72 for those age 50 to 64, and 74 for those age 65 and older. The article “Poll Finds Generation Gap Biggest Since Vietnam War” summarized a study that explored opinions regarding social values and political views. Not surprisingly, large behavioral differences between young and old were noted in the use of the Internet, cell phones, and text messaging. The graph titled “If you were given $1000, what would you do?” reported on one aspect of a study of consumer purchasing and saving behavior. Something was definitely amiss in this report, however—the percentages for the response categories (save it, pay off credit card debt, use it for a vacation, etc.) added up to 107%!
To be an informed consumer of reports such as those described above, you must be able to do the following: 1. Extract information from tables, charts, and graphs. 2. Follow numerical arguments. 3. Understand the basics of how data should be gathered, summarized, and analyzed to draw statistical conclusions. Your statistics course will help prepare you to perform these tasks. Studying statistics will also enable you to collect data in a sensible way and then use the data to answer questions of interest. In addition, studying statistics will allow you to critically evaluate the work of others by providing you with the tools you need to make informed judgments. Throughout your personal and professional life, you
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
1.2 The Nature and Role of Variability
3
will need to understand and use data to make decisions. To do this, you must be able to 1. Decide whether existing data is adequate or whether additional information is required. 2. If necessary, collect more information in a reasonable and thoughtful way. 3. Summarize the available data in a useful and informative manner. 4. Analyze the available data. 5. Draw conclusions, make decisions, and assess the risk of an incorrect decision. People informally use these steps when making everyday decisions. Should you go out for a sport that involves the risk of injury? Will your college club do better by trying to raise funds with a benefit concert or with a direct appeal for donations? If you choose a particular major, what are your chances of finding a job when you graduate? How should you select a graduate program based on guidebook ratings that include information on percentage of applicants accepted, time to obtain a degree, and so on? The study of statistics formalizes the process of making decisions based on data and provides the tools for accomplishing the steps listed. We hope that this textbook will help you to understand the logic behind statistical reasoning, prepare you to apply statistical methods appropriately, and enable you to recognize when statistical arguments are faulty.
1.2
The Nature and Role of Variability Statistical methods allow us to collect, describe, analyze and draw conclusions from data. If we lived in a world where all measurements were identical for every individual, these tasks would be simple. Imagine a population consisting of all students at a particular university. Suppose that every student was enrolled in the same number of courses, spent exactly the same amount of money on textbooks this semester, and favored increasing student fees to support expanding library services. For this population, there is no variability in number of courses, amount spent on books, or student opinion on the fee increase. A researcher studying students from this population to draw conclusions about these three variables would have a particularly easy task. It would not matter how many students the researcher studied or how the students were selected. In fact, the researcher could collect information on number of courses, amount spent on books, and opinion on the fee increase by just stopping the next student who happened to walk by the library. Because there is no variability in the population, this one individual would provide complete and accurate information about the population, and the researcher could draw conclusions with no risk of error. The situation just described is obviously unrealistic. Populations with no variability are exceedingly rare, and they are of little statistical interest because they present no challenge! In fact, variability is almost universal. It is variability that makes life (and the life of a statistician, in particular) interesting. We need to understand variability to be able to collect, describe, analyze, and draw conclusions from data in a sensible way. Examples 1.1 and 1.2 illustrate how describing and understanding variability are the keys to learning from data.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
4
Chapter 1 The Role of Statistics and the Data Analysis Process
EXAMPLE 1.1
If the Shoe Fits
The graphs in Figure 1.1 are examples of a type of graph called a histogram. (The construction and interpretation of such graphs is discussed in Chapter 3.) Figure 1.1(a) shows the distribution of the heights of female basketball players who played at a particular university between 2000 and 2008. The height of each bar in the graph indicates how many players’ heights were in the corresponding interval. For example, 40 basketball players had heights between 72 inches and 74 inches, whereas only 2 players had heights between 66 inches and 68 inches Figure 1.1(b) shows the distribution of heights for members of the women’s gymnastics team. Both histograms are based on the heights of 100 women. Frequency
Frequency
40 20
30 20
10 10 0
0 58
60
62
64
66
68 Height
FIGURE 1.1 Histograms of heights (in inches) of female athletes: (a) basketball players; (b) gymnasts.
(a)
70
72
74
76
78
58
60
62
64
66
68 70 Height
72
74
76
78
(b)
The first histogram shows that the heights of female basketball players varied, with most heights falling between 68 inches and 76 inches. In the second histogram we see that the heights of female gymnasts also varied, with most heights in the range of 60 inches to 72 inches It is also clear that there is more variation in the heights of the gymnasts than in the heights of the basketball players, because the gymnast histogram spreads out more about its center than does the basketball histogram. Now suppose that a tall woman (5 feet 11 inches) tells you she is looking for her sister who is practicing with her team at the gym. Would you direct her to where the basketball team is practicing or to where the gymnastics team is practicing? What reasoning would you use to decide? If you found a pair of size 6 shoes left in the locker room, would you first try to return them by checking with members of the basketball team or the gymnastics team? You probably answered that you would send the woman looking for her sister to the basketball practice and that you would try to return the shoes to a gymnastics team member. To reach these conclusions, you informally used statistical reasoning that combined your own knowledge of the relationship between heights of siblings and between shoe size and height with the information about the distributions of heights presented in Figure 1.1. You might have reasoned that heights of siblings tend to be similar and that a height as great as 5 feet 11 inches, although not impossible, would be unusual for a gymnast. On the other hand, a height as tall as 5 feet 11 inches would be a common occurrence for a basketball player. Similarly, you might have reasoned that tall people tend to have bigger feet and that short people tend to have smaller feet. The shoes found were a small size, so it is more likely that they belong to a gymnast than to a basketball player, because small heights and so small feet are usual for gymnasts and unusual for basketball players.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
1.3 Statistics and the Data Analysis Process
© David Chasey/Photodisc/Getty Images
EXAMPLE 1.2
5
Monitoring Water Quality
As part of its regular water quality monitoring efforts, an environmental control board selects five water specimens from a particular well each day. The concentration of contaminants in parts per million (ppm) is measured for each of the five specimens, and then the average of the five measurements is calculated. The histogram in Figure 1.2 summarizes the average contamination values for 200 days. Now suppose that a chemical spill has occurred at a manufacturing plant 1 mile from the well. It is not known whether a spill of this nature would contaminate groundwater in the area of the spill and, if so, whether a spill this distance from the well would affect the quality of well water. One month after the spill, five water specimens are collected from the well, and the average contamination is 15.5 ppm. Considering the variation before the spill, would you interpret this as convincing evidence that the well water was affected by the spill? What if the calculated average was 17.4 ppm? 22.0 ppm? How is your reasoning related to the histogram in Figure 1.2? Frequency
40 30 20 10
FIGURE 1.2 Frequency of average contamination concentration (in parts per million) in well water.
0 10
11
12
13
14
15
16
17
18
19
Average contamination
Before the spill, the average contaminant concentration varied from day to day. An average of 15.5 ppm would not have been an unusual value, so seeing an average of 15.5 ppm after the spill isn’t necessarily an indication that contamination has increased. On the other hand, an average as large as 17.4 ppm is less common, and an average as large as 22.0 ppm is not at all typical of the pre-spill values. In this case, we would probably conclude that the well contamination level has increased.
In these two examples, reaching a conclusion required an understanding of variability. Understanding variability allows us to distinguish between usual and unusual values. The ability to recognize unusual values in the presence of variability is an important aspect of most statistical procedures and is also what enables us to quantify the chance of being incorrect when a conclusion is based on data. These concepts will be developed further in subsequent chapters.
1.3
Statistics and the Data Analysis Process Statistics involves collecting, summarizing, and analyzing data. All three tasks are critical. Without summarization and analysis, raw data are of little value, and even sophisticated analyses can’t produce meaningful information from data that were not collected in a sensible way.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
6
Chapter 1 The Role of Statistics and the Data Analysis Process
Statistical studies are undertaken to answer questions about our world. Is a new flu vaccine effective in preventing illness? Is the use of bicycle helmets on the rise? Are injuries that result from bicycle accidents less severe for riders who wear helmets than for those who do not? How many credit cards do college students have? Do engineering students pay more for textbooks than do psychology students? Data collection and analysis allow researchers to answer such questions. The data analysis process can be viewed as a sequence of steps that lead from planning to data collection to making informed conclusions based on the resulting data. The process can be organized into the following six steps: 1. Understanding the nature of the problem. Effective data analysis requires an understanding of the research problem. We must know the goal of the research and what questions we hope to answer. It is important to have a clear direction before gathering data to ensure that we will be able to answer the questions of interest using the data collected. 2. Deciding what to measure and how to measure it. The next step in the process is deciding what information is needed to answer the questions of interest. In some cases, the choice is obvious (for example, in a study of the relationship between the weight of a Division I football player and position played, you would need to collect data on player weight and position), but in other cases the choice of information is not as straightforward (for example, in a study of the relationship between preferred learning style and intelligence, how would you define learning style and measure it and what measure of intelligence would you use?). It is important to carefully define the variables to be studied and to develop appropriate methods for determining their values. 3. Data collection. The data collection step is crucial. The researcher must first decide whether an existing data source is adequate or whether new data must be collected. Even if a decision is made to use existing data, it is important to understand how the data were collected and for what purpose, so that any resulting limitations are also fully understood and judged to be acceptable. If new data are to be collected, a careful plan must be developed, because the type of analysis that is appropriate and the subsequent conclusions that can be drawn depend on how the data are collected. 4. Data summarization and preliminary analysis. After the data are collected, the next step usually involves a preliminary analysis that includes summarizing the data graphically and numerically. This initial analysis provides insight into important characteristics of the data and can provide guidance in selecting appropriate methods for further analysis. 5. Formal data analysis. The data analysis step requires the researcher to select and apply statistical methods. Much of this textbook is devoted to methods that can be used to carry out this step. 6. Interpretation of results. Several questions should be addressed in this final step. Some examples are: What can we learn from the data? What conclusions can be drawn from the analysis? and How can our results guide future research? The interpretation step often leads to the formulation of new research questions, which, in turn, leads back to the first step. In this way, good data analysis is often an iterative process. For example, the admissions director at a large university might be interested in learning why some applicants who were accepted for the fall 2010 term failed to enroll at the university. The population of interest to the director consists of all accepted applicants who did not enroll in the fall 2010 term. Because this population is large and it may be difficult to contact all the individuals, the director might decide to collect data from only 300 selected students. These 300 students constitute a sample. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
1.3 Statistics and the Data Analysis Process
7
DEFINITION The entire collection of individuals or objects about which information is desired is called the population of interest. A sample is a subset of the population, selected for study. Deciding how to select the 300 students and what data should be collected from each student are steps 2 and 3 in the data analysis process. The next step in the process involves organizing and summarizing data. Methods for organizing and summarizing data, such as the use of tables, graphs, or numerical summaries, make up the branch of statistics called descriptive statistics. The second major branch of statistics, inferential statistics, involves generalizing from a sample to the population from which it was selected. When we generalize in this way, we run the risk of an incorrect conclusion, because a conclusion about the population is based on incomplete information. An important aspect in the development of inferential techniques involves quantifying the chance of an incorrect conclusion.
DEFINITION Descriptive statistics is the branch of statistics that includes methods for organizing and summarizing data. Inferential statistics is the branch of statistics that involves generalizing from a sample to the population from which the sample was selected and assessing the reliability of such generalizations. Example 1.3 illustrates the steps in the data analysis process.
EXAMPLE 1.3
The Benefits of Acting Out
A number of studies have reached the conclusion that stimulating mental activities can lead to improved memory and psychological wellness in older adults. The article “A
Short-Term Intervention to Enhance Cognitive and Affective Functioning in Older Adults” (Journal of Aging and Health [2004]: 562–585) describes a study to investigate whether training in acting has similar benefits. Acting requires a person to consider the goals of the characters in the story, to remember lines of dialogue, to move on stage as scripted, and to do all of this at the same time. The researchers conducting the study wanted to see if participation in this type of complex multitasking would show an improvement in the ability to function independently in daily life. Participants in the study were assigned to one of three groups. One group took part in an acting class for 4 weeks, one group spent a similar amount of time in a class on visual arts, and the third group was a comparison group (called the “no-treatment group”) that did not take either class. A total of 124 adults age 60 to 86 participated in the study. At the beginning of the 4-week study period and again at the end of the 4-week study period, each participant took several tests designed to measure problem solving, memory span, selfesteem, and psychological well-being. After analyzing the data from this study, the researchers concluded that those in the acting group showed greater gains than both the visual arts group and the no-treatment group in both problem solving and psychological well-being. Several new areas of research were suggested in the discussion that followed the analysis. The researchers wondered whether the effect of studying writing or music would be similar to what was observed for acting and described plans to investigate this further. They also noted that the participants in this study were generally well educated and recommended study of a more diverse group before generalizing conclusions about the benefits of studying acting to the larger population of all older adults. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
8
Chapter 1 The Role of Statistics and the Data Analysis Process
This study illustrates the nature of the data analysis process. A clearly defined research question and an appropriate choice of how to measure the variables of interest (the tests used to measure problem solving, memory span, self-esteem, and psychological well-being) preceded the data collection. Assuming that a reasonable method was used to collect the data (we will see how this can be evaluated in Chapter 2) and that appropriate methods of analysis were employed, the investigators reached the conclusion that the study of acting showed promise. However, they recognized the limitations of the study, which in turn led to plans for further research. As is often the case, the data analysis cycle led to new research questions, and the process began again. The six data analysis steps can also be used as a guide for evaluating published research studies. The following questions should be addressed as part of a study evaluation:
Evaluating a Research Study
What were the researchers trying to learn? What questions motivated their research? Was relevant information collected? Were the right things measured? Were the data collected in a sensible way? Were the data summarized in an appropriate way? Was an appropriate method of analysis used, given the type of data and how the data were collected? • Are the conclusions drawn by the researchers supported by the data analysis? • • • • •
Example 1.4 illustrates how these questions can guide an evaluation of a research study.
EXAMPLE 1.4
Afraid of Spiders? You Are Not Alone!
Spider phobia is a common anxiety-producing disorder. In fact, the American Psychiatric Association estimates that between 7% and 15.1% of the population experiences spider phobia. An effective treatment for this condition involves participating in a therapist-led session in which the patient is exposed to live spiders. While this type of treatment has been shown to work for a large proportion of patients, it requires one-onone time with a therapist trained in this technique. The article “Internet-Based Self-
Help versus One-Session Exposure in the Treatment of Spider Phobia” (Cognitive Behaviour Therapy [2009]: 114–120), presented results from a study that compared the effectiveness of online self-help modules to in-person treatment. The article states A total of 30 patients were included following screening on the Internet and a structured clinical interview. The Internet treatment consisted of five weekly text modules, which were presented on a web page, a video in which exposure was modeled, and support provided via Internet. The live-exposure treatment was delivered in a 3-hour session following a brief orientation session. The main outcome measure was the behavioral approach test (BAT), and the authors used questionnaires measuring anxiety symptoms and depression as secondary measures. Results showed that the groups did not differ at post-treatment or follow-up, with the exception of the proportion showing clinically significant change on the BAT. At post-treatment, 46.2% of the Internet group and 85.7% of the live-exposure group achieved this change. At follow-up, the corresponding figures were 66.7% for the Internet group and 72.7% for the live treatment. The researchers concluded that online treatment is a promising new approach for the treatment of spider phobia. The researchers here had a well-defined research question—they wanted to know if online treatment is as effective as in-person exposure treatment. They were interested Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
1.3 Statistics and the Data Analysis Process
9
in this question because online treatment does not require individual time with a therapist, and so, if it works, it might be able to help a larger group of people at a much lower cost. The researchers noted which treatment was received and also recorded results of the BAT and several other measures of anxiety and depression. Participants in the study took these tests prior to beginning treatment, at the end of treatment, and 1 year after the end of treatment. This allowed the researchers to evaluate the immediate and long-term effects of the two treatments and to address the research question. To assess whether the data were collected in a sensible way, it would be useful to know how the participants were selected and how it was determined which of the two treatments a particular participant received. The article indicates that participants were recruited through advertisements and articles in local newspapers and that most were female university students. We will see in Chapter 2 that this may limit our ability to generalize the results of this study. The participants were assigned to one of the two treatments at random, which is a good strategy for ensuring that one treatment does not tend to be favored over the other. The advantages of random assignment in a study of this type are also discussed in Chapter 2. We will also have to delay discussion of the data analysis and the appropriateness of the conclusions because we do not yet have the necessary tools to evaluate these aspects of the study. Many other interesting examples of statistical studies can be found in Statistics: A Guide to the Unknown and in Forty Studies That Changed Psychology: Exploration into the History of Psychological Research (the complete references for these two books can be found in the back of the book).
E X E RC I S E S 1 . 1 - 1 . 1 1 1.1 Give a brief definition of the terms descriptive statistics and inferential statistics. 1.2 Give a brief definition of the terms population and sample. 1.3 Data from a poll conducted by Travelocity led to the following estimates: Approximately 40% of travelers check work e-mail while on vacation, about 33% take cell phones on vacation in order to stay connected with work, and about 25% bring laptop computers on vacation (San Luis Obispo Tribune, December 1, 2005). Are the given percentages population values or were they computed from a sample?
1.4 Based on a study of 2121 children between the ages of 1 and 4, researchers at the Medical College of Wisconsin concluded that there was an association between iron deficiency and the length of time that a child is bottle-fed (Milwaukee Journal Sentinel, November 26, 2005). Describe the sample and the population of interest for this study. Bold exercises answered in back
Data set available online
1.5 The student senate at a university with 15,000 students is interested in the proportion of students who favor a change in the grading system to allow for plus and minus grades (e.g., B1, B, B2, rather than just B). Two hundred students are interviewed to determine their attitude toward this proposed change. What is the population of interest? What group of students constitutes the sample in this problem?
1.6 The increasing popularity of online shopping has many consumers using Internet access at work to browse and shop online. In fact, the Monday after Thanksgiving has been nicknamed “Cyber Monday” because of the large increase in online purchases that occurs on that day. Data from a large-scale survey by a market research firm (Detroit Free Press, November 26, 2005) was used to compute estimates of the percent of men and women who shop online while at work. The resulting estimates probably won’t make most employers happy—42% of the men and 32% of the women in the sample were shopping online at work! Are the estimates given computed using data from a sample or for the entire population? Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
10
Chapter 1 The Role of Statistics and the Data Analysis Process
1.7 The supervisors of a rural county are interested in the proportion of property owners who support the construction of a sewer system. Because it is too costly to contact all 7000 property owners, a survey of 500 owners (selected at random) is undertaken. Describe the population and sample for this problem.
1.8 A consumer group conducts crash tests of new model cars. To determine the severity of damage to 2010 Toyota Camrys resulting from a 10-mph crash into a concrete wall, the research group tests six cars of this type and assesses the amount of damage. Describe the population and sample for this problem.
1.9 A building contractor has a chance to buy an odd lot of 5000 used bricks at an auction. She is interested in determining the proportion of bricks in the lot that are cracked and therefore unusable for her current project, but she does not have enough time to inspect all 5000 bricks. Instead, she checks 100 bricks to determine whether each is cracked. Describe the population and sample for this problem.
1.10 The article “Brain Shunt Tested to Treat Alzheimer’s” (San Francisco Chronicle, October 23, 2002) summarizes the findings of a study that appeared in the journal Neurology. Doctors at Stanford Medical Center were interested in determining whether a new surgical approach to treating Alzheimer’s disease results in improved memory functioning. The surgical procedure involves implanting a thin tube, called a shunt, which is designed to drain toxins from the fluid-filled space that cushions the brain. Eleven patients had shunts implanted and were followed for a year, receiving quarterly tests of memory function. Another sample of Alzheimer’s patients was used as a comparison group. Bold exercises answered in back
1.4
Data set available online
Those in the comparison group received the standard care for Alzheimer’s disease. After analyzing the data from this study, the investigators concluded that the “results suggested the treated patients essentially held their own in the cognitive tests while the patients in the control group steadily declined. However, the study was too small to produce conclusive statistical evidence.” a. What were the researchers trying to learn? What questions motivated their research? b. Do you think that the study was conducted in a reasonable way? What additional information would you want in order to evaluate this study?
1.11 The newspaper article “Spray Away Flu” (Omaha World-Herald, June 8, 1998) reported on a study of the effectiveness of a new flu vaccine that is administered by nasal spray rather than by injection. The article states that the “researchers gave the spray to 1070 healthy children, 15 months to 6 years old, before the flu season two winters ago. One percent developed confirmed influenza, compared with 18% of the 532 children who received a placebo. And only one vaccinated child developed an ear infection after coming down with influenza. . . . Typically 30% to 40% of children with influenza later develop an ear infection.” The researchers concluded that the nasal flu vaccine was effective in reducing the incidence of flu and also in reducing the number of children with flu who subsequently develop ear infections. a. What were the researchers trying to learn? What questions motivated their research? b. Do you think that the study was conducted in a reasonable way? What additional information would you want in order to evaluate this study?
Video Solution available
Types of Data and Some Simple Graphical Displays Every discipline has its own particular way of using common words, and statistics is no exception. You will recognize some of the terminology from previous math and science courses, but much of the language of statistics will be new to you. In this section, you will learn some of the terminology used to describe data.
Types of Data The individuals or objects in any particular population typically possess many characteristics that might be studied. Consider a group of students currently enrolled in a statistics course. One characteristic of the students in the population is the brand of Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
1.4 Types of Data and Some Simple Graphical Displays
11
calculator owned (Casio, Hewlett-Packard, Sharp, Texas Instruments, and so on). Another characteristic is the number of textbooks purchased that semester, and yet another is the distance from the university to each student’s permanent residence. A variable is any characteristic whose value may change from one individual or object to another. For example, calculator brand is a variable, and so are number of textbooks purchased and distance to the university. Data result from making observations either on a single variable or simultaneously on two or more variables. A univariate data set consists of observations on a single variable made on individuals in a sample or population. There are two types of univariate data sets: categorical and numerical. In the previous example, calculator brand is a categorical variable, because each student’s response to the query, “What brand of calculator do you own?” is a category. The collection of responses from all these students forms a categorical data set. The other two variables, number of textbooks purchased and distance to the university, are both numerical in nature. Determining the value of such a numerical variable (by counting or measuring) for each student results in a numerical data set.
DEFINITION A data set consisting of observations on a single characteristic is a univariate data set. A univariate data set is categorical (or qualitative) if the individual observations are categorical responses. A univariate data set is numerical (or quantitative) if each observation is a number.
EXAMPLE 1.5
College Choice Do-Over?
The Higher Education Research Institute at UCLA surveys over 20,000 college seniors each year. One question on the 2008 survey asked seniors the following question: If you could make your college choice over, would you still choose to enroll at your current college? Possible responses were definitely yes (DY), probably yes (PY), probably no (PN), and definitely no (DN). Responses for 20 students were: DY PN DN DY PY PY PN PY PY DY DY PY DY DY PY PY DY DY PN DY (These data are just a small subset of the data from the survey. For a description of the full data set, see Exercise 1.18). Because the response to the question about college choice is categorical, this is a univariate categorical data set. In Example 1.5, the data set consisted of observations on a single variable (college choice response), so this is univariate data. In some studies, attention focuses simultaneously on two different characteristics. For example, both height (in inches) and weight (in pounds) might be recorded for each individual in a group. The resulting data set consists of pairs of numbers, such as (68, 146). This is called a bivariate data set. Multivariate data result from obtaining a category or value for each of two or more attributes (so bivariate data are a special case of multivariate data). For example, multivariate data would result from determining height, weight, pulse rate, and systolic blood pressure for each individual in a group. Example 1.6 illustrates a bivariate data set. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
12
Chapter 1 The Role of Statistics and the Data Analysis Process
EXAMPLE 1.6
How Safe Are College Campuses?
Consider the accompanying data on violent crime on college campuses in Florida during 2005 (http://www.fbi.gov/ucr/05cius/data/table_09.html).
University/College
Student Enrollment
Number of Violent Crimes Reported in 2005
13,067 25,319 5,955 34,865 38,431 692 10,879 13,888 12,775 42,465 47,993 14,533 42,238 9,518
23 4 5 5 29 1 2 3 0 19 17 6 19 1
Florida A&M University Florida Atlantic University Florida Gulf Coast University Florida International University Florida State University New College of Florida Pensacola Junior College Santa Fe Community College Tallahassee Community College University of Central Florida University of Florida University of North Florida University of South Florida University of West Florida
Here two variables—student enrollment and number of violent crimes reported—were recorded for each of the 14 schools. Because this data set consists of values of two variables for each school, it is a bivariate data set. Each of the two variables considered here is numerical (rather than categorical).
Two Types of Numerical Data
Data set available online
There are two different types of numerical data: discrete and continuous. Consider a number line (Figure 1.3) for locating values of the numerical variable being studied. Each possible number (2, 3.125, 8.12976, etc.) corresponds to exactly one point on the number line. Now suppose that the variable of interest is the number of courses in which a student is enrolled. If no student is enrolled in more than eight courses, the possible values are 1, 2, 3, 4, 5, 6, 7, and 8. These values are identified in Figure 1.4(a) by the dots at the points marked 1, 2, 3, 4, 5, 6, 7, and 8. These possible values are isolated from one another on the number line; around any possible value, we can place an interval that is small enough that no other possible value is included in the interval. On the other hand, the line segment in Figure 1.4(b) identifies a plausible set of possible values for the time (in seconds) it takes for the first kernel in a bag of microwave popcorn to pop. Here the possible values make up an entire interval on the number line, and no possible value is isolated from other possible values.
FIGURE 1.3 A number line.
–3
–2
–1
1
2
3
FIGURE 1.4 Possible values of a variable: (a) number of cylinders; (b) quarter-mile time.
2
4 (a)
6
8
10
20
30
40
(b)
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
1.4 Types of Data and Some Simple Graphical Displays
13
DEFINITION A numerical variable results in discrete data if the possible values of the variable correspond to isolated points on the number line. A numerical variable results in continuous data if the set of possible values forms an entire interval on the number line. Discrete data usually arise when observations are determined by counting (for example, the number of roommates a student has or the number of petals on a certain type of flower).
EXAMPLE 1.7
Do U Txt?
The number of text messages sent on a particular day is recorded for each of 12 students. The resulting data set is 23
14
13
15
60
82
40
41
22
Possible values for the variable number of text messages sent are 0, 1, 2, 3. . . . These are isolated points on the number line, so this data set consists of discrete numerical data. Suppose that instead of the number of text messages sent, the time spent texting had been recorded. Even though time spent may have been reported rounded to the nearest minute, the actual time spent could have been 6 minutes, 6.2 minutes, 6.28 minutes, or any other value in an entire interval. So, recording values of time spent texting would result in continuous data. In general, data are continuous when observations involve making measurements, as opposed to counting. In practice, measuring instruments do not have infinite accuracy, so possible measured values, strictly speaking, do not form a continuum on the number line. However, any number in the continuum could be a value of the variable. The distinction between discrete and continuous data will be important in our discussion of probability models.
Frequency Distributions and Bar Charts for Categorical Data Data set available online
An appropriate graphical or tabular display of data can be an effective way to summarize and communicate information. When the data set is categorical, a common way to present the data is in the form of a table, called a frequency distribution.
A frequency distribution for categorical data is a table that displays the possible categories along with the associated frequencies and/or relative frequencies. The frequency for a particular category is the number of times the category appears in the data set. The relative frequency for a particular category is calculated as relative frequency 5
frequency number of obervations in the data set
The relative frequency for a particular category is the proporton of the observations that belong to that category. If the table includes relative frequencies, it is sometimes referred to as a relative frequency distribution.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
14
Chapter 1 The Role of Statistics and the Data Analysis Process
EXAMPLE 1.8
Motorcycle Helmets—Can You See Those Ears?
The U.S. Department of Transportation established standards for motorcycle helmets. To ensure a certain degree of safety, helmets should reach the bottom of the motorcyclist’s ears. The report “Motorcycle Helmet Use in 2005—Overall Results” (National Highway Traffic Safety Administration, August 2005) summarized data collected in June of 2005 by observing 1700 motorcyclists nationwide at selected roadway locations. Each time a motorcyclist passed by, the observer noted whether the rider was wearing no helmet, a noncompliant helmet, or a compliant helmet. Using the coding NH 5 noncompliant helmet CH 5 compliant helmet N 5 no helmet a few of the observations were CH
N
CH
NH
N
CH
CH
CH
N
N
There were also 1690 additional observations, which we didn’t reproduce here! In total, there were 731 riders who wore no helmet, 153 who wore a noncompliant helmet, and 816 who wore a compliant helmet. The corresponding frequency distribution is given in Table 1.1.
T A B L E 1.1
Frequency Distribution for Helmet Use
Helmet Use Category
Frequency
Relative Frequency
No helmet Noncompliant helmet Compliant helmet
731 153 816 1700
0.430 0.090 0.480 1.000 Total number of observations
731/1700 153/1700
Should total 1, but in some cases may be slightly off due to rounding
From the frequency distribution, we can see that a large number of riders (43%) were not wearing a helmet, but most of those who wore a helmet were wearing one that met the Department of Transportation safety standard. A frequency distribution gives a tabular display of a data set. It is also common to display categorical data graphically. A bar chart is one of the most widely used types of graphical displays for categorical data.
Bar Charts A bar chart is a graph of a frequency distribution of categorical data. Each category in the frequency distribution is represented by a bar or rectangle, and the picture is constructed in such a way that the area of each bar is proportional to the corresponding frequency or relative frequency.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
1.4 Types of Data and Some Simple Graphical Displays
15
Bar Charts When to Use Categorical data. How to Construct 1. Draw a horizontal axis, and write the category names or labels below the line at regularly spaced intervals. 2. Draw a vertical axis, and label the scale using either frequency or relative frequency. 3. Place a rectangular bar above each category label. The height is determined by the category’s frequency or relative frequency, and all bars should have the same width. With the same width, both the height and the area of the bar are proportional to frequency and relative frequency.
What to Look For • Frequently and infrequently occurring categories.
EXAMPLE 1.9
Revisiting Motorcycle Helmets
Example 1.8 used data on helmet use from a sample of 1700 motorcyclists to construct a frequency distribution (Table 1.1). Figure 1.5 shows the bar chart corresponding to this frequency distribution. Frequency 900 800 700 600 500 400 300 200 100
FIGURE 1.5 Bar chart of helmet use.
Step-by-Step technology instructions available online
No helmet
Noncompliant helmet
Compliant helmet
Helmet Use Category
The bar chart provides a visual representation of the information in the frequency distribution. From the bar chart, it is easy to see that the compliant helmet use category occurred most often in the data set. The bar for compliant helmets is about five times as tall (and therefore has five times the area) as the bar for noncompliant helmets because approximately five times as many motorcyclists wore compliant helmets than wore noncompliant helmets.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
16
Chapter 1
The Role of Statistics and the Data Analysis Process
Dotplots for Numerical Data A dotplot is a simple way to display numerical data when the data set is reasonably small. Each observation is represented by a dot above the location corresponding to its value on a horizontal measurement scale. When a value occurs more than once, there is a dot for each occurrence and these dots are stacked vertically.
Dotplots When to Use Small numerical data sets. How to Construct 1. Draw a horizontal line and mark it with an appropriate measurement scale. 2. Locate each value in the data set along the measurement scale, and represent it by a dot. If there are two or more observations with the same value, stack the dots vertically.
What to Look For • • • •
Dotplots convey information about:
A representative or typical value in the data set. The extent to which the data values spread out. The nature of the distribution of values along the number line. The presence of unusual values in the data set.
E X A M P L E 1 . 1 0 Making It to Graduation . . . The article “Keeping Score When It Counts: Graduation Rates and Academic Progress Rates for 2009 NCAA Men’s Division I Basketball Tournament Teams” (The Institute for Diversity and Ethics in Sport, University of Central Florida, March 2009) compared graduation rates of basketball players to those of all student
Data set available online
FIGURE 1.6 Minitab dotplot of graduation rates.
athletes for the universities and colleges that sent teams to the 2009 Division I playoffs. The graduation rates in the accompanying table represent the percentage of athletes who started college in 2002 who had graduated by the end of 2008. Also shown are the differences between the graduation rate for all student athletes and the graduation rate for basketball student athletes. (Note: Teams from 63 schools made it to the playoffs, but two of them—Cornell and North Dakota State—did not report graduation rates.) Minitab, a computer software package for statistical analysis, was used to construct a dotplot of the 61 graduation rates for basketball players (see Figure 1.6). From this dotplot, we see that basketball graduation rates varied a great deal from school to school, ranging from a low of 8% to a high of 100%. We can also see that the graduation rates seem to cluster in several groups, denoted by the colored ovals that have been added to the dotplot. There are several schools with graduation rates of 100% (excellent!) and another group of 13 schools with graduation rates that are
10
20
30
40 50 60 70 Graduation rates for basketball players (%)
80
90
100
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
1.4 Types of Data and Some Simple Graphical Displays
17
Graduation Rates (%) Basketball
All Athletes
Difference (All - BB)
63 56 31 20 38 100 70 91 92 30 8 34 29 71 33 89 89 60 100 67 80 64 40 42 100 10 55 46 60 36 53 36
75 57 86 64 69 85 96 79 89 76 56 53 82 83 81 96 97 67 80 89 86 70 69 75 94 79 72 83 79 72 78 71
12 1 55 44 31 215 26 212 23 46 48 19 53 12 48 7 8 7 220 22 6 6 29 33 26 69 17 37 19 36 25 35
Basketball
All Athletes
57 45 86 67 53 55 92 69 17 77 80 100 86 37 42 50 57 38 31 47 46 67 100 89 53 100 50 41 100 86 82
70 57 85 81 78 69 75 84 48 79 91 95 94 69 63 83 71 78 72 72 79 75 82 95 71 92 83 68 80 79 92
Difference (All - BB) 13 12 21 14 25 14 217 15 31 2 11 25 8 32 21 33 14 40 41 25 33 8 218 6 18 28 33 27 220 27 10
higher than most. The majority of schools are in the large cluster with graduation rates from about 30% to about 72%. And then there is that bottom group of four schools with embarrassingly low graduation rates for basketball players: Northridge (8%), Maryland (10%), Portland State (17%), and Arizona (20%). Figure 1.7 shows two dotplots of graduation rates—one for basketball players and one for all student athletes. There are some striking differences that are easy to
Basketball
FIGURE 1.7 MINITAB dotplot of graduation rates for basketball players and for all athletes.
All athletes 10
20
30
40
50 60 Graduation rates (%)
70
80
90
100
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
18
Chapter 1
The Role of Statistics and the Data Analysis Process
see when the data is displayed in this way. The graduation rates for all student athletes tend to be higher and to vary less from school to school than the graduation rates for basketball players. The dotplots in Figure 1.7 are informative, but we can do even better. The data given here are an example of paired data. Each basketball graduation rate is paired with a graduation rate for all student athletes from the same school. When data are paired in this way, it is usually more informative to look at differences—in this case, the difference between the graduation rate for all student athletes and for basketball players for each school. These differences (all ⫺ basketball) are also shown in the data table. Figure 1.8 gives a dotplot of the 61 differences. Notice that one difference is equal to 0. This corresponded to a school for which the basketball graduation rate is equal to the graduation rate of all student athletes. There are 11 schools for which the difference is negative. Negative differences correspond to schools that have a graduation rate for basketball players that is higher than the graduation rate for all student athletes. The most interesting features of the difference dotplot are the very large number of positive differences and the wide spread. Positive differences correspond to schools that have a lower graduation rate for basketball players. There is a lot of variability in the graduation rate difference from school to school, and three schools have differences that are noticeably higher than the rest. (In case you were wondering, these schools were Clemson with a difference of 53%, American University with a difference of 55%, and Maryland with a difference of 69%.)
Difference negative
Difference positive
FIGURE 1.8 Dotplot of graduation rate differences (ALL ⫺ BB)
–10
10 20 30 40 Difference in graduation rate % (ALL – BB)
50
60
EX E RC I S E S 1 . 1 2 - 1 . 3 1 1.12 Classify each of the following variables as either
1.13 Classify each of the following variables as either
categorical or numerical. For those that are numerical, determine whether they are discrete or continuous. a. Number of students in a class of 35 who turn in a term paper before the due date b. Gender of the next baby born at a particular hospital c. Amount of fluid (in ounces) dispensed by a machine used to fill bottles with soda pop d. Thickness of the gelatin coating of a vitamin E capsule e. Birth order classification (only child, firstborn, middle child, lastborn) of a math major
categorical or numerical. For those that are numerical, determine whether they are discrete or continuous. a. Brand of computer purchased by a customer b. State of birth for someone born in the United States c. Price of a textbook d. Concentration of a contaminant (micrograms per cubic centimeter) in a water sample e. Zip code (Think carefully about this one.) f. Actual weight of coffee in a 1-pound can
Bold exercises answered in back
Data set available online
1.14 For the following numerical variables, state whether each is discrete or continuous. Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
1.4 Types of Data and Some Simple Graphical Displays
a. The number of insufficient-funds checks received by a grocery store during a given month b. The amount by which a 1-pound package of ground beef decreases in weight (because of moisture loss) before purchase c. The number of New York Yankees during a given year who will not play for the Yankees the next year d. The number of students in a class of 35 who have purchased a used copy of the textbook
1.18 The report “Findings from the 2008 Administration of the College Senior Survey” (Higher Education Research Institute, UCLA, June 2009) gave the following relative frequency distribution summarizing student responses to the question “If you could make your college choice over, would you still choose to enroll at your current college?” Response
1.16 For each of the following situations, give a set of possible data values that might arise from making the observations described. a. The manufacturer for each of the next 10 automobiles to pass through a given intersection is noted. b. The grade point average for each of the 15 seniors in a statistics class is determined. c. The number of gas pumps in use at each of 20 gas stations at a particular time is determined. d. The actual net weight of each of 12 bags of fertilizer having a labeled weight of 50 pounds is determined. e. Fifteen different radio stations are monitored during a 1-hour period, and the amount of time devoted to commercials is determined for each.
1.17 In a survey of 100 people who had recently purchased motorcycles, data on the following variables were recorded: Gender of purchaser Brand of motorcycle purchased Number of previous motorcycles owned by purchaser Telephone area code of purchaser Weight of motorcycle as equipped at purchase a. Which of these variables are categorical? b. Which of these variables are discrete numerical? c. Which type of graphical display would be an appropriate choice for summarizing the gender data, a bar chart or a dotplot? d. Which type of graphical display would be an appropriate choice for summarizing the weight data, a bar chart or a dotplot? Bold exercises answered in back
Data set available online
Relative Frequency
Definitely yes Probably yes Probably no Definitely no
1.15 For the following numerical variables, state whether each is discrete or continuous. a. The length of a 1-year-old rattlesnake b. The altitude of a location in California selected randomly by throwing a dart at a map of the state c. The distance from the left edge at which a 12-inch plastic ruler snaps when bent sufficiently to break d. The price per gallon paid by the next customer to buy gas at a particular station
19
.447 .373 .134 .046
a. Use this information to construct a bar chart for the response data. b. If you were going to use the response data and the bar chart from Part (a) as the basis for an article for your student paper, what would be a good headline for your article?
1.19 The article “Feasting on Protein” (AARP Bulletin, September 2009) gave the cost per gram of protein for 19 common food sources of protein. Cost (cents per gram of protein)
Food Chicken Salmon Turkey Soybeans Roast beef Cottage cheese Ground beef Ham Lentils Beans Yogurt Milk Peas Tofu Cheddar cheese Nuts Eggs Peanut butter Ice cream
1.8 5.8 1.5 3.1 2.7 3.1 2.3 2.1 3.3 2.9 5.0 2.5 5.2 6.9 3.6 5.2 5.7 1.8 5.3
a. Construct a dotplot of the cost-per-gram data. b. Locate the cost per gram for meat and poultry items in your dotplot and highlight them in a different color. Based on the dotplot, do meat and poultry items appear to be a good value? That is, do they appear to be relatively low cost compared to other sources of protein? Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
20
Chapter 1 The Role of Statistics and the Data Analysis Process
Box Office Mojo (www.boxofficemojo.com) tracks movie ticket sales. Ticket sales (in millions of dollars) for each of the top 20 movies in 2007 and 2008 are shown in the accompanying table.
1.20
Movie (2007) Spider-Man 3 Shrek the Third Transformers Pirates of the Caribbean: At World’s End Harry Potter and the Order of the Phoenix I Am Legend The Bourne Ultimatum National Treasure: Book of Secrets Alvin and the Chipmunks 300 Ratatouille The Simpsons Movie Wild Hogs Knocked Up Juno Rush Hour 3 Live Free or Die Hard Fantastic Four: Rise of the Silver Surfer American Gangster Enchanted
Movie (2008) The Dark Knight Iron Man Indiana Jones and the Kingdom of the Crystal Skull Hancock WALL-E Kung Fu Panda Twilight Madagascar: Escape 2 Africa Quantum of Solace Dr. Suess’ Horton Hears a Who!
2007 Sales (millions of dollars) 336.5 322.7 319.2 309.4 292.0 256.4 227.5 220.0
Sex and the City Gran Torino Mamma Mia! Marley and Me The Chronicles of Narnia: Prince Caspian Slumdog Millionaire The Incredible Hulk Wanted Get Smart The Curious Case of Benjamin Button
152.6 148.1 144.1 143.2 141.6 141.3 134.8 134.5 130.3 127.5
a. Construct a dotplot of the 2008 ticket sales data. Comment on any interesting features of the dotplot. b. Construct a dotplot of the 2007 ticket sales data. Comment on any interesting features of the dotplot. In what ways are the distributions of the 2007 and 2008 ticket sales observations similar? In what ways are they different?
217.3 210.6 206.4 183.1 168.3 148.8 143.5 140.1 134.5 131.9
About 38,000 students attend Grant MacEwan College in Edmonton, Canada. In 2004, the college surveyed non-returning students to find out why they did not complete their degree (Grant MacEwan College Early Leaver Survey Report, 2004). Sixty-three students gave a personal (rather than an academic) reason for leaving. The accompanying frequency distribution summarizes primary reason for leaving for these 63 students.
1.21
130.2 127.8
2008 Sales (millions of dollars) 533.3 318.4 317.1 227.9 223.8 215.4 192.8 180.0 168.4 154.5 (continued)
Bold exercises answered in back
Movie (2008)
2008 Sales (millions of dollars)
Data set available online
Primary Reason for Leaving Financial Health Employment Family issues Wanted to take a break Moving Travel Other personal reasons
Frequency 19 12 8 6 4 2 2 10
Summarize the reason for leaving data using a bar chart and write a few sentences commenting on the most common reasons for leaving.
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
1.4 Types of Data and Some Simple Graphical Displays
1.22 Figure EX-1.22 is a graph that appeared in USA Unique Visitors
Total Visits
Visits per Unique Visitor
68,557,534 58,555,800 5,979,052 7,645,423 11,274,160 4,448,915 17,296,524 3,312,898 4,720,720 9,047,491 13,704,990 5,673,549 1,530,329 2,997,929 2,398,323 1,317,551 1,647,336 1,568,439 1,831,376 1,499,057 494,464 329,041 452,090 81,245 96,155
1,191,373,339 810,153,536 54,218,731 53,389,974 42,744,438 39,630,927 35,219,210 33,121,821 25,221,354 22,993,608 20,278,100 19,511,682 10,173,342 9,849,137 9,416,265 9,358,966 8,586,261 7,279,050 7,009,577 5,199,702 5,081,235 2,961,250 2,170,315 1,118,245 109,492
17.3777 13.8356 9.0681 6.9833 3.7914 8.9080 2.0362 9.9978 5.3427 2.5414 1.4796 3.4391 6.6478 3.2853 3.9262 7.1033 5.2122 4.6410 3.8275 3.4686 10.2762 8.9996 4.8006 13.7639 1.1387
Today (June 29, 2009). This graph is meant to be a bar graph of responses to the question shown in the graph. a. Is response to the question a categorical or numerical variable? b. Explain why a bar chart rather than a dotplot was used to display the response data. c. There must have been an error made in constructing this graph. How can you tell that the graph is not a correct representation of the response data?
Site facebook.com myspace.com twitter.com fixter.com linkedin.com tagged.com classmates.com myyearbook.com livejournal.com imeem.com reunion.com ning.com blackplanet.com bebo.com hi5.com yuku.com cafemom.com friendster.com xanga.com 360.yahoo.com orkut.com urbanchat.com fubar.com asiantown.net tickle.com
Image not available due to copyright restrictions
1.23
a. A dotplot of the total visits data is shown in Figure EX-1.23a. What are the most obvious features of the dotplot? What does it tell you about the online social networking sites? b. A dotplot for the number of unique visitors is shown in Figure EX-1.23b. In what way is this dotplot different from the dotplot for total visits in Part (a)?
The online article “Social Networks: Facebook
Takes Over Top Spot, Twitter Climbs” (Compete.com, February 9, 2009) included the accompanying data on number of unique visitors and total number of visits for January 2009 for the top 25 online social network sites. The data on total visits and unique visitors were used to compute the values in the final column of the data table, in which visits per unique visitor 5
total visits number of unique visitors
200,000,000
400,000,000
FIGURE EX-1.23a
Bold exercises answered in back
21
Data set available online
600,000,000 Total visits
800,000,000
1,000,000,000 1,200,000,000
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
22
Chapter 1 The Role of Statistics and the Data Analysis Process
10,000,000
20,000,000
FIGURE EX-1.23b
5
FIGURE EX-1.23c
What does this tell you about the online social networking sites? c. A dotplot for the visits per unique visitor data is shown in Figure EX-1.23c. What new information about the online social networks is provided by this dotplot?
1.24 Heal the Bay is an environmental organization that releases an annual beach report card based on water quality (Heal the Bay Beach Report Card, May 2009). The 2009 ratings for 14 beaches in San Francisco County during wet weather were: A1 C B A A1 A1 A A1 B D C D F F a. Would it be appropriate to display the ratings data using a dotplot? Explain why or why not. b. Summarize the wet weather ratings by constructing a relative frequency distribution and a bar chart. c. The dry weather ratings for these same beaches were: A B B A1 A F A A A A A A B A Construct a bar graph for the dry weather ratings. d. Do the bar graphs from parts (b) and (c) support the statement that beach water quality tends to be better in dry weather conditions? Explain. The article “Going Wireless” (AARP Bulletin, June 2009) reported the estimated percentage of households with only wireless phone service (no landline) for the 50 states and the District of Columbia. In the accompanying data table, each state was also classified into one of three geographical regions—West (W), Middle states (M), and East (E).
1.25
Wireless %
Region
State
13.9
M
AL
Bold exercises answered in back
Data set available online
30,000,000 40,000,000 Unique visitors
50,000,000
10 Visits per unique visitor
Wireless % 11.7 18.9 22.6 9.0 16.7 5.6 5.7 20.0 16.8 16.5 8.0 22.1 16.5 13.8 22.2 16.8 21.4 15.0 13.4 10.8 9.3 16.3 17.4 19.1 9.9 9.2 23.2 10.8 16.9 11.6 8.0 21.1 11.4 16.3 14.0 23.2
15
Region
60,000,000
70,000,000
20
State
W AK W AZ M AR W CA W CO E CN E DE E DC E FL E GA W HI W ID M IL M IN M IA M KA M KY M LA E ME E MD E MA M MI M MN M MS M MO W MT M NE W NV M ND E NH E NJ W NM E NY E NC E OH M OK (data continued on following page)
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
1.4 Types of Data and Some Simple Graphical Displays
Wireless %
Region
State
17.7 10.8 7.9 20.6 6.4 20.3 20.9 25.5 10.8 5.1 16.3 11.6 15.2 11.4
W E E E M M M W E E W E M W
OR PA RI SC SD TN TX UT VA VT WA WV WI WY
a. Display the data graphically in a way that makes it possible to compare wireless percent for the three geographical regions. b. Does the graphical display in Part (a) reveal any striking differences in wireless percent for the three geographical regions or are the distributions of wireless percent observations similar for the three regions? Example 1.6 gave the accompanying data on violent crime on college campuses in Florida during 2005 (from the FBI web site):
1.26
University/College
23
Student Enrollment
Number of Violent Crimes Reported in 2005
13,067 25,319
23 4
5,955
5
34,865
5
38,431 692 10,879 13,888
29 1 2 3
12,775
42,465
19
47,993 14,533
17 6
42,238
19
9,518
1
Florida A&M University Florida Atlantic University Florida Gulf Coast University Florida International University Florida State University New College of Florida Pensacola Junior College Santa Fe Community College Tallahassee Community College University of Central Florida University of Florida University of North Florida University of South Florida University of West Florida
a. Construct a dotplot using the 14 observations on number of violent crimes reported. Which schools stand out from the rest? b. One of the Florida schools only has 692 students and a few of the schools are quite a bit larger than the rest. Because of this, it might make more sense to consider a crime rate by calculating the number of violent crimes reported per 1000 students. For example, for Florida A&M University the violent crime rate would be 23 110002 5 1.00182 110002 5 1.8 13067 Calculate the violent crime rate for the other 13 schools and then use those values to construct a dotplot. Do the same schools stand out as unusual in this dotplot? c. Based on your answers from parts (a) and (b), write a couple of sentences commenting on violent crimes reported at Florida universities and colleges in 2005.
Bold exercises answered in back
Data set available online
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
24
Chapter 1 The Role of Statistics and the Data Analysis Process
The article “Fliers Trapped on Tarmac Push for Rules on Release” (USA Today, July 28, 2009) gave the following data for 17 airlines on number of flights that were delayed on the tarmac for at least 3 hours for the period from October 2008 to May 2009:
Airline
Number of Delays
Rate per 10,000 Flights
93 72 81 29 44 46 18 48 24 17 29 5 29 13 11 7 11
4.9 4.1 2.8 2.7 1.6 1.6 1.4 1.3 1.2 1.1 1.1 0.9 0.8 0.7 0.6 0.4 0.1
ExpressJet Continental Delta Comair American Eagle US Airways JetBlue American Northwest Mesa United Frontier SkyWest Pinnacle Atlantic Southeast AirTran Southwest
sumers” (San Luis Obispo Tribune, February 2, 2005) included the accompanying breakdown of identity theft complaints by type. Type of Complaint
Percent of All Complaints
Credit card fraud Phone or utilities fraud Bank fraud Employment fraud Other
28% 19% 18% 13% 22%
Construct a bar chart for these data and write a sentence or two commenting on the most common types of identity theft complaints.
1.29 A 2005 AP-IPSOS poll found that 21% of American adults surveyed said their child was heavier than doctors recommend. The reasons given as the most important contributing factor to the child’s weight problem are summarized in the accompanying table. Lack of exercise Easy access to junk food Genetics Eating unhealthy food Medical condition Overeating
Figure EX-1.27 shows two dotplots: one displays the number of delays data, and one displays the rate per 10,000 flights data. a. If you were going to rank airlines based on flights delayed on the tarmac for at least three hours, would you use the total number of flights data or the rate per 10,000 flights data? Explain the reason for your choice. b. Write a short paragraph that could be used as part of a newspaper article on flight delays that could accompany the dotplot of the rate per 10,000 flights data.
10
20
30
40
0.0
0.5
1.0
1.5
2.0 2.5 3.0 Rate per 10,000 flights
Data set available online
38% 23% 12% 9% 8% 7%
a. Construct a bar chart for the data on the most important contributing factor. b. Do you think that it would be reasonable to combine some of these contributing factors into a single category? If so, which categories would you combine and why?
FIGURE EX-1.27
Bold exercises answered in back
The article “Fraud, Identity Theft Afflict Con-
1.28
1.27
50 60 Number of delays
70
80
90
3.5
4.0
4.5
100
5.0
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
1.4 Types of Data and Some Simple Graphical Displays
The article “Americans Drowsy on the Job and the Road” (Associated Press, March 28, 2001)
1.30
summarized data from the 2001 Sleep in America poll. Each individual in a sample of 1004 adults was asked questions about his or her sleep habits. The article states that “40 percent of those surveyed say they get sleepy on the job and their work suffers at least a few days each month, while 22 percent said the problems occur a few days each week. And 7 percent say sleepiness on the job is a daily occurrence.” Assuming that everyone else reported that sleepiness on the job was not a problem, summarize the given information by constructing a relative frequency bar chart.
1.31 “Ozzie and Harriet Don’t Live Here Anymore” (San Luis Obispo Tribune, February 26, 2002) is the title of an article that looked at the changing makeup of America’s suburbs. The article states that nonfamily households (for example, homes headed by a single proBold exercises answered in back
A C TI V I T Y 1 . 1
Data set available online
25
fessional or an elderly widow) now outnumber married couples with children in suburbs of the nation’s largest metropolitan areas. The article goes on to state: In the nation’s 102 largest metropolitan areas, “nonfamilies” comprised 29 percent of households in 2000, up from 27 percent in 1990. While the number of married-with-children homes grew too, the share did not keep pace. It declined from 28 percent to 27 percent. Married couples without children at home live in another 29 percent of suburban households. The remaining 15 percent are single-parent homes. Use the given information on type of household in 2000 to construct a frequency distribution and a bar chart. (Be careful to extract the 2000 percentages from the given information).
Video Solution available
Twitter Words
This activity requires Internet access. TweetVolume is a web site that allows you to enter up to three words to produce a bar chart based on how often those words appear on Twitter. 1. Go to www.tweetvolume.com and spend a few minutes experimenting with different words to see how the site works. For example, in July 2010, the words statistics, sample and population resulted in the following bar chart.
2. Find a set of three words that result in a bar chart in which all three bars are approximately the same height. 3. Find a set of three words that satisfy the following: i. One word begins with the letter a, one word begins with the letter b, and one word begins with the letter c. ii. The word that begins with the letter a is more common on Twitter (has the highest bar in the bar graph) than the other two words. iii. The word that begins with the letter b is more common on Twitter than the word that begins with the letter c.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
26
Chapter 1 The Role of Statistics and the Data Analysis Process
AC TI V I TY 1 . 2
Head Sizes: Understanding Variability
Materials needed: Each team will need a measuring tape. For this activity, you will work in teams of 6 to 10 people. 1. Designate a team leader for your team by choosing the person on your team who celebrated his or her last birthday most recently. 2. The team leader should measure and record the head size (measured as the circumference at the widest part of the forehead) of each of the other members of his or her team. 3. Record the head sizes for the individuals on your team as measured by the team leader. 4. Next, each individual on the team should measure the head size of the team leader. Do not share your measurement with the other team members until all team members have measured the team leader’s head size. 5. After all team members have measured the team leader’s head, record the different team leader head size measurements obtained by the individuals on your team. 6. Using the data from Step 3, construct a dotplot of the team leader’s measurements of team head sizes. Then, using the same scale, construct a separate dotplot of the different measurements of the team leader’s head size (from Step 5).
AC TI V I TY 1 . 3
Now use the available information to answer the following questions: 7. Do you think the team leader’s head size changed in between measurements? If not, explain why the measurements of the team leader’s head size are not all the same. 8. Which data set was more variable—head size measurements of the different individuals on your team or the different measurements of the team leader’s head size? Explain the basis for your choice. 9. Consider the following scheme (you don’t actually have to carry this out): Suppose that a group of 10 people measured head sizes by first assigning each person in the group a number between 1 and 10. Then person 1 measured person 2’s head size, person 2 measured person 3’s head size, and so on, with person 10 finally measuring person 1’s head size. Do you think that the resulting head size measurements would be more variable, less variable, or show about the same amount of variability as a set of 10 measurements resulting from a single individual measuring the head size of all 10 people in the group? Explain.
Estimating Sizes
1. Construct an activity sheet that consists of a table that has 6 columns and 10 rows. Label the columns of the table with the following six headings: (1) Shape, (2) Estimated Size, (3) Actual Size, (4) Difference (Estimated 2 Actual), (5) Absolute Difference, and (6) Squared Difference. Enter the numbers from 1 to 10 in the “Shape” column. 2. Next you will be visually estimating the sizes of the shapes in Figure 1.9. Size will be described as the number of squares of this size
that would fit in the shape. For example, the shape
would be size 3, as illustrated by
You should now quickly visually estimate the sizes of the shapes in Figure 1.9. Do not draw on the figure—these are to be quick visual estimates. Record your estimates in the “Estimated Size” column of the activity sheet. 3. Your instructor will provide the actual sizes for the 10 shapes, which should be entered into the “Actual Size” column of the activity sheet. Now complete the “Difference” column by subtracting the actual value from your estimate for each of the 10 shapes. 4. What would cause a difference to be negative? What would cause a difference to be positive?
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
Activities
27
FIGURE 1.9 Shapes for Activity 1.3.
3
2
1
6
5 4
9
8
7
10
5. Would the sum of the differences tell you if the estimates and actual values were in close agreement? Does a sum of 0 for the differences indicate that all the estimates were equal to the actual value? Explain. 6. Compare your estimates with those of another person in the class by comparing the sum of the absolute values of the differences between estimates and corresponding actual values. Who was better at estimating shape sizes? How can you tell? 7. Use the last column of the activity sheet to record the squared differences (for example, if the difference for shape 1 was 23, the squared difference
would be (23)2 5 9. Explain why the sum of the squared differences can also be used to assess how accurate your shape estimates were. 8. For this step, work with three or four other students from your class. For each of the 10 shapes, form a new size estimate by computing the average of the size estimates for that shape made by the individuals in your group. Is this new set of estimates more accurate than your own individual estimates were? How can you tell? 9. Does your answer from Step 8 surprise you? Explain why or why not.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
28
Chapter 1 The Role of Statistics and the Data Analysis Process
AC TI V I TY 1 . 4
A Meaningful Paragraph
Write a meaningful paragraph that includes the following six terms: sample, population, descriptive statistics, bar chart, numerical variable, and dotplot. A “meaningful paragraph” is a coherent piece of writing in an appropriate context that uses all of the listed words. The paragraph should show that you un-
derstand the meanings of the terms and their relationships to one another. A sequence of sentences that just define the terms is not a meaningful paragraph. When choosing a context, think carefully about the terms you need to use. Choosing a good context will make writing a meaningful paragraph easier.
Summary of Key Concepts and Formulas TERM OR FORMULA
COMMENT
Population
The entire collection of individuals or measurements about which information is desired.
Sample
A part of the population selected for study.
Descriptive statistics
Numerical, graphical, and tabular methods for organizing and summarizing data.
Inferential statistics
Methods for generalizing from a sample to a population.
Categorical data
Individual observations are categorical responses (nonnumerical).
Numerical data
Individual observations are numerical (quantitative) in nature.
Discrete numerical data
Possible values are isolated points along the number line.
Continuous numerical data
Possible values form an entire interval along the number line.
Univariate, bivariate and multivariate data
Each observation consists of one (univariate), two (bivariate), or two or more (multivariate) responses or values.
Frequency distribution for categorical data
A table that displays frequencies, and sometimes relative frequencies, for each of the possible values of a categorical variable.
Bar chart
A graph of a frequency distribution for a categorical data set. Each category is represented by a bar, and the area of the bar is proportional to the corresponding frequency or relative frequency.
Dotplot
A graph of numerical data in which each observation is represented by a dot on or above a horizontal measurement scale.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
Chapter Review Exercises
29
Chapter Review Exercises 1.32 - 1.37 The report “Testing the Waters 2009” (www .nrdc.org) included information on the water quality at
1.32
the 82 most popular swimming beaches in California. Thirty-eight of these beaches are in Los Angeles County. For each beach, water quality was tested weekly and the data below are the percent of the tests in 2008 that failed to meet water quality standards. Los Angeles County 32 4 6 4 19 13 11 19 33 12 29 3 17 26 17 20 Other Counties 0 15 1 0 10
0 8 0 8 40
0 1 2 8 3
2 5 7 8
4 9 11 10
7 11 6 6
4 16 22 14
27 23 18 11
19 19 31
23 16 43
3 0 0 0
7 5 2 0
5 4 2 17
11 1 3 4
5 0 5 3
7 1 3 7
a. Construct a dotplot of the percent of tests failing to meet water quality standards for the Los Angeles County beaches. Write a few sentences describing any interesting features of the dotplot. b. Construct a dotplot of the percent of tests failing to meet water quality standards for the beaches in other counties. Write a few sentences describing any interesting features of the dotplot. c. Based on the two dotplots from Parts (a) and (b), describe how the percent of tests that fail to meet water quality standards for beaches in Los Angeles county differs from those of other counties.
1.33 The U.S. Department of Education reported that 14% of adults were classified as being below a basic literacy level, 29% were classified as being at a basic literacy level, 44% were classified as being at an intermediate literacy level, and 13% were classified as being at a proficient level (2003 National Assessment of Adult Literacy). a. Is the variable literacy level categorical or numerical? b. Would it be appropriate to display the given information using a dotplot? Explain why or why not. c. Construct a bar chart to display the given data on literacy level.
Bold exercises answered in back
Data set available online
1.34 The Computer Assisted Assessment Center at the University of Luton published a report titled “Technical Review of Plagiarism Detection Software.” The authors of this report asked faculty at academic institutions about the extent to which they agreed with the statement “Plagiarism is a significant problem in academic institutions.” The responses are summarized in the accompanying table. Construct a bar chart for these data. Response Strongly disagree Disagree Not sure Agree Strongly agree
Frequency 5 48 90 140 39
1.35 The article “Just How Safe Is That Jet?” (USA Today, March 13, 2000) gave the following relative frequency distribution that summarized data on the type of violation for fines imposed on airlines by the Federal Aviation Administration: Type of Violation
Relative Frequency
Security Maintenance Flight operations Hazardous materials Other
.43 .39 .06 .03 .09
Use this information to construct a bar chart for type of violation, and then write a sentence or two commenting on the relative occurrence of the various types of violation. Each year, U.S. News and World Report publishes a ranking of U.S. business schools. The following data give the acceptance rates (percentage of applicants admitted) for the best 25 programs in a recent survey:
1.36
16.3 12.0 25.1 20.3 31.9 20.7 30.1 19.5 36.2 46.9 25.8 36.7 33.8 24.2 21.5 35.1 37.6 23.9 17.0 38.4 31.2 43.8 28.9 31.4 48.9 Construct a dotplot, and comment on the interesting features of the plot.
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
30
Chapter 1 The Role of Statistics and the Data Analysis Process
1.37 Many adolescent boys aspire to be professional athletes. The paper “Why Adolescent Boys Dream of Becoming Professional Athletes” (Psychological Reports [1999]:1075–1085) examined some of the reasons. Each boy in a sample of teenage boys was asked the following question: “Previous studies have shown that more teenage boys say that they are considering becoming professional athletes than any other occupation. In your opinion, why do these boys want to become professional athletes?” The resulting data are shown in the following table: Bold exercises answered in back
Data set available online
Response
Frequency
Fame and celebrity Money Attract women Like sports Easy life Don’t need an education Other
94 56 29 27 24 19 19
Construct a bar chart to display these data. Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
CHAPTER
2
Collecting Data Sensibly A primary goal of statistical studies is to collect data that can then be used to make informed decisions. It should come as no surprise that the ability to make good decisions depends on the quality of the information available. The data collection step is critical to obtaining reliable information; both the type of analysis that is appropriate and the conclusions that can be drawn depend on how the data are collected. In this chapter, we first consider two types of statistical studies and then focus on two widely used methods of data collection: sampling and experimentation.
Purestock/Kwame Zikomo/SuperStock
Make the most of your study time by accessing everything you need to succeed online with CourseMate. Visit http://www.cengagebrain.com where you will find: • An interactive eBook, which allows you to take notes, highlight, bookmark, search • • • • • •
the text, and use in-context glossary definitions Step-by-step instructions for Minitab, Excel, TI-83/84, SPSS, and JMP Video solutions to selected exercises Data sets available for selected examples and exercises Online quizzes Flashcards Videos
31 Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
32
2.1
Chapter 2 Collecting Data Sensibly
Statistical Studies: Observation and Experimentation On September 25, 2009, results from a study of the relationship between spanking and IQ were reported by a number of different news media. Some of the headlines that appeared that day were:
“Spanking lowers a child’s IQ” (Los Angeles Times) “Do you spank? Studies indicate it could lower your kid’s IQ” (SciGuy, Houston Chronicle) “Spanking can lower IQ” (NBC4i, Columbus, Ohio) “Smacking hits kids’ IQ” (newscientist.com) In the study that these headlines refer to, the investigators followed 806 kids age 2 to 4 and 704 kids age 5 to 9 for 4 years. IQ was measured at the beginning of the study and again 4 years later. The researchers found that at the end of the study, the average IQ of kids who were not spanked was 5 points higher than that of kids who were spanked among the kids who were 2 to 4 years old when the study began, and 2.8 points higher among the kids who were 5 to 9 years old when the study began. These headlines all imply that spanking was the cause of the observed difference in IQ. Is this conclusion reasonable? The answer depends in a critical way on the study design. We’ll return to these headlines and decide if they are on target after first considering some important aspects of study design.
Observation and Experimentation Data collection is an important step in the data analysis process. When we set out to collect information, it is important to keep in mind the questions we hope to answer on the basis of the resulting data. Sometimes we are interested in answering questions about characteristics of a single existing population or in comparing two or more well-defined populations. To accomplish this, we select a sample from each population under consideration and use the sample information to gain insight into characteristics of those populations. For example, an ecologist might be interested in estimating the average shell thickness of bald eagle eggs. A social scientist studying a rural community may want to determine whether gender and attitude toward abortion are related. These are examples of studies that are observational in nature. In these studies, we want to observe characteristics of members of an existing population or of several populations, and then use the resulting information to draw conclusions. In an observational study, it is important to obtain a sample that is representative of the corresponding population. Sometimes the questions we are trying to answer deal with the effect of certain explanatory variables on some response and cannot be answered using data from an observational study. Such questions are often of the form, “What happens when ... ?” or, “What is the effect of ... ?” For example, an educator may wonder what would happen to test scores if the required lab time for a chemistry course were increased from 3 hours to 6 hours per week. To answer such questions, the researcher conducts an experiment to collect relevant data. The value of some response variable (test score in the chemistry example) is recorded under different experimental conditions (3-hour lab and 6-hour lab). In an experiment, the researcher manipulates one or more explanatory variables, also sometimes called factors, to create the experimental conditions.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
2.1
Statistical Studies: Observation and Experimentation
33
DEFINITION A study is an observational study if the investigator observes characteristics of a sample selected from one or more existing populations. The goal of an observational study is usually to draw conclusions about the corresponding population or about differences between two or more populations. In a welldesigned observational study, the sample is selected in a way that is designed to produce a sample that is respresentative of the population. A study is an experiment if the investigator observes how a response variable behaves when one or more explanatory variables, also called factors, are manipulated. The usual goal of an experiment is to determine the effect of the manipulated explanatory variables (factors) on the response variable. In a welldesigned experiment, the composition of the groups that will be exposed to different experimental conditions is determined by random assignment. The type of conclusion that can be drawn from a statistical study depends on the study design. Both observational studies and experiments can be used to compare groups, but in an experiment the researcher controls who is in which group, whereas this is not the case in an observational study. This seemingly small difference is critical when it comes to drawing conclusions based on data from the study. A well-designed experiment can result in data that provide evidence for a causeand-effect relationship. This is an important difference between an observational study and an experiment. In an observational study, it is impossible to draw clear cause-and-effect conclusions because we cannot rule out the possibility that the observed effect is due to some variable other than the explanatory variable being studied. Such variables are called confounding variables.
DEFINITION A confounding variable is one that is related to both group membership and the response variable of interest in a research study. Consider the role of confounding variables in the following three studies: • The article
“Panel Can’t Determine the Value of Daily Vitamins” (San Luis Obispo Tribune, July 1, 2003) summarized the conclusions of a government
advisory panel that investigated the benefits of vitamin use. The panel looked at a large number of studies on vitamin use and concluded that the results were “inadequate or conflicting.” A major concern was that many of the studies were observational in nature and the panel worried that people who take vitamins might be healthier just because they tend to take better care of themselves in general. This potential confounding variable prevented the panel from concluding that taking vitamins is the cause of observed better health among those who take vitamins. • Studies have shown that people over age 65 who get a flu shot are less likely than those who do not get a flu shot to die from a flu-related illness during the following year. However, recent research has shown that people over age 65 who get a flu shot are also less likely than those who don’t to die from any cause during the following year (International Journal of Epidemiology, December 21, 2005).
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
34
Chapter 2 Collecting Data Sensibly
This has lead to the speculation that those over age 65 who get flu shots are healthier as a group than those who do not get flu shots. If this is the case, observational studies that compare two groups—those who get flu shots and those who do not—may overestimate the effectiveness of the flu vaccine because general health differs in the two groups. General health is a possible confounding variable in such studies. • The article “Heartfelt Thanks to Fido” (San Luis Obispo Tribune, July 5, 2003) summarized a study that appeared in the American Journal of Cardiology (March 15, 2003). In this study researchers measured heart rate variability (a measure of the heart’s ability to handle stress) in patients who had recovered from a heart attack. They found that heart rate variability was higher (which is good and means the heart can handle stress better) for those who owned a dog than for those who did not. Should someone who suffers a heart attack immediately go out and get a dog? Well, maybe not yet. The American Heart Association recommends additional studies to determine if the improved heart rate variability is attributable to dog ownership or due to the fact that dog owners get more exercise. If in fact dog owners do tend to get more exercise than nonowners, level of exercise is a confounding variable that would prevent us from concluding that owning a dog is the cause of improved heart rate variability. Each of the three studies described above illustrates why potential confounding variables make it unreasonable to draw a cause-and-effect conclusion from an observational study. Let’s return to the study on spanking and IQ described at the beginning of this section. Is this study an observational study or an experiment? Two groups were compared (children who were spanked and children who were not spanked), but the researchers did not randomly assign children to the spanking or no-spanking groups. The study is observational, and so cause-and-effect conclusions such as “spanking lowers IQ” are not justified based on the observed data. What we can say is that there is evidence that, as a group, children who are spanked tend to have a lower IQ than children who are not spanked. What we cannot say is that spanking is the cause of the lower IQ. It is possible that other variables—such as home or school environment, socio-economic status, or parents’ education—are related to both IQ and whether or not a child was spanked. These are examples of possible confounding variables. Fortunately, not everyone made the same mistake as the writers of the headlines given earlier in this section. Some examples of headlines that got it right are:
“Lower IQ’s measured in spanked children” (world-science.net) “Children who get spanked have lower IQs” (livescience.com) “Research suggests an association between spanking and lower IQ in children” (CBSnews.com)
Drawing Conclusions from Statistical Studies In this section, two different types of conclusions have been described. One type involves generalizing from what we have seen in a sample to some larger population, and the other involves reaching a cause-and-effect conclusion about the effect of an explanatory variable on a response. When is it reasonable to draw such conclusions? The answer depends on the way that the data were collected. Table 2.1 summarizes the types of conclusions that can be made with different study designs. As you can see from Table 2.1, it is important to think carefully about the objectives of a statistical study before planning how the data will be collected. Both Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
2.1
Statistical Studies: Observation and Experimentation
35
T A B L E 2.1 Drawing Conclusions from Statistical Studies Reasonable to Generalize Conclusions about Group Characteristics to the Population?
Study Description
Reasonable to Draw Causeand-Effect Conclusion?
Observational study with sample selected at random from population of interest
Yes
No
Observational study based on convenience or voluntary response sample (poorly designed sampling plan)
No
No
No
Yes
Yes
Yes
No
No
Experiment with groups formed by random assignment of individuals or objects to experimental conditions Individuals or objects used in study are volunteers or not randomly selected from some population of interest Individuals or objects used in study are randomly selected from some population of interest Experiment with groups not formed by random assignment to experimental conditions (poorly designed experiment)
observational studies and experiments must be carefully designed if the resulting data are to be useful. The common sampling procedures used in observational studies are considered in Section 2.2. In Sections 2.3 and 2.4, we consider experimentation and explore what constitutes good practice in the design of simple experiments.
E X E RC I S E S 2 . 1 - 2 . 1 2 2.1
The article “Television’s Value to Kids: It’s All in How They Use It” (Seattle Times, July 6, 2005)
2.2 The article “Acupuncture for Bad Backs: Even Sham Therapy Works” (Time, May 12, 2009) summa-
described a study in which researchers analyzed standardized test results and television viewing habits of 1700 children. They found that children who averaged more than 2 hours of television viewing per day when they were younger than 3 tended to score lower on measures of reading ability and short-term memory. a. Is the study described an observational study or an experiment? b. Is it reasonable to conclude that watching two or more hours of television is the cause of lower reading scores? Explain.
rized a study conducted by researchers at the Group Health Center for Health Studies in Seattle. In this study, 638 adults with back pain were randomly assigned to one of four groups. People in group 1 received the usual care for back pain. People in group 2 received acupuncture at a set of points tailored specifically for each individual. People in group 3 received acupuncture at a standard set of points typically used in the treatment of back pain. Those in group 4 received fake acupuncture—they were poked with a toothpick at the same set of points used for the people in group 3! Two notable conclusions from the study were: (1) patients receiving real or fake acupuncture
Bold exercises answered in back
Data set available online
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
36
Chapter 2 Collecting Data Sensibly
experienced a greater reduction in pain than those receiving usual care; and (2) there was no significant difference in pain reduction for those who received acupuncture (at individualized or the standard set of points) and those who received fake acupuncture toothpick pokes. a. Is this study an observational study or an experiment? Explain. b. Is it reasonable to conclude that receiving either real or fake acupuncture was the cause of the observed reduction in pain in those groups compared to the usual care group? What aspect of this study supports your answer?
2.5 Consider the following graphical display that appeared in the New York Times:
Image not available due to copyright restrictions
2.3 The article “Display of Health Risk Behaviors on MySpace by Adolescents” (Archives of Pediatrics and Adolescent Medicine [2009]:27–34) described a study in which researchers looked at a random sample of 500 publicly accessible MySpace web profiles posted by 18-year-olds. The content of each profile was analyzed. One of the conclusions reported was that displaying sport or hobby involvement was associated with decreased references to risky behavior (sexual references or references to substance abuse or violence). a. Is the study described an observational study or an experiment? b. Is it reasonable to generalize the stated conclusion to all 18-year-olds with a publicly accessible MySpace web profile? What aspect of the study supports your answer? c. Not all MySpace users have a publicly accessible profile. Is it reasonable to generalize the stated conclusion to all 18-year-old MySpace users? Explain. d. Is it reasonable to generalize the stated conclusion to all MySpace users with a publicly accessible profile? Explain.
2.4 Can choosing the right music make wine taste better? This question was investigated by a researcher at a university in Edinburgh (www.decanter.com/news). Each of 250 volunteers was assigned at random to one of five rooms where they were asked to taste and rate a glass of wine. In one of the rooms, no music was playing and a different style of music was playing in each of the other four rooms. The researchers concluded that cabernet sauvignon is perceived as being richer and more robust when bold music is played than when no music is heard. a. Is the study described an observational study or an experiment? b. Can a case be made for the researcher’s conclusion that the music played was the cause for the higher rating? Explain. Bold exercises answered in back
Data set available online
Based on the data summarized in the graph, we can see that students who have a high school GPA or 3.5 or higher and a combined SAT score of over 1200 have an 89% graduation rate when they attend a “most selective” college, but only a 59% graduation rate when they attend a “least selective” college. Give an example of a potential confounding variable that might explain why the following statement is not reasonable: If all the students that have a GPA of 3.5 or higher and a combined SAT score of 1200 or higher and that were admitted to a “least selective” college were moved to a “most selective” college, the graduation rate for these students would be approximately 89%.
2.6 “Fruit Juice May Be Fueling Pudgy Preschoolers, Study Says” is the title of an article that appeared in the San Luis Obispo Tribune (February 27, 2005). This article describes a study that found that for 3- and 4-year-olds, drinking something sweet once or twice a day doubled the risk of being seriously overweight one year later. The authors of the study state Total energy may be a confounder if consumption of sweet drinks is a marker for other dietary factors associated with overweight (Pediatrics, November 2005). Give an example of a dietary factor that might be one of the potentially confounding variables the study authors are worried about.
2.7 The article “Americans are ‘Getting the Wrong Idea’ on Alcohol and Health” (Associated Press, April 19, 2005) reported that observational studies in recent years that have concluded that moderate drinking is associated with a reduction in the risk of heart disease may be misleading. The article refers to a study conducted by Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
2.2
the Centers for Disease Control and Prevention that showed that moderate drinkers, as a group, tended to be better educated, wealthier, and more active than nondrinkers. Explain why the existence of these potentially confounding variables prevents drawing the conclusion that moderate drinking is the cause of reduced risk of heart disease.
2.8 An article titled “Guard Your Kids Against Allergies: Get Them a Pet” (San Luis Obispo Tribune, August 28, 2002) described a study that led researchers to conclude that “babies raised with two or more animals were about half as likely to have allergies by the time they turned six.” a. Do you think this study was an observational study or an experiment? Explain. b. Describe a potential confounding variable that illustrates why it is unreasonable to conclude that being raised with two or more animals is the cause of the observed lower allergy rate.
2.9 Researchers at the Hospital for Sick Children in Toronto compared babies born to mothers with diabetes to babies born to mothers without diabetes (“Conditioning and Hyperanalgesia in Newborns Exposed to Repeated Heel Lances,” Journal of the American Medical Association [2002]: 857–861). Babies born to mothers with diabetes have their heels pricked numerous times during the first 36 hours of life in order to obtain blood samples to monitor blood sugar level. The researchers noted that the babies born to diabetic mothers were more likely to grimace or cry when having blood drawn than the babies born to mothers without diabetes. This led the researchers to conclude that babies who experience pain early in life become highly sensitive to pain. Comment on the appropriateness of this conclusion. Bold exercises answered in back
2.2
Data set available online
Sampling
37
2.10 Based on a survey conducted on the DietSmart .com web site, investigators concluded that women who regularly watched Oprah were only one-seventh as likely to crave fattening foods as those who watched other daytime talk shows (San Luis Obispo Tribune, October 14, 2000). a. Is it reasonable to conclude that watching Oprah causes a decrease in cravings for fattening foods? Explain. b. Is it reasonable to generalize the results of this survey to all women in the United States? To all women who watch daytime talk shows? Explain why or why not. A survey of affluent Americans (those with incomes of $75,000 or more) indicated that 57% would rather have more time than more money (USA Today,
2.11
January 29, 2003). a. What condition on how the data were collected would make the generalization from the sample to the population of affluent Americans reasonable? b. Would it be reasonable to generalize from the sample to say that 57% of all Americans would rather have more time than more money? Explain.
2.12 Does living in the South cause high blood pressure? Data from a group of 6278 whites and blacks questioned in the Third National Health and Nutritional Examination Survey between 1988 and 1994 (see CNN.com web site article of January 6, 2000, titled “High Blood Pressure Greater Risk in U.S. South, Study Says”) indicates that a greater percentage of Southerners have high blood pressure than do people in any other region of the United States. This difference in rate of high blood pressure was found in every ethnic group, gender, and age category studied. List at least two possible reasons we cannot conclude that living in the South causes high blood pressure. Video Solution available
Sampling Many studies are conducted in order to generalize from a sample to the corresponding population. As a result, it is important that the sample be representative of the population. To be reasonably sure of this, we must carefully consider the way in which the sample is selected. It is sometimes tempting to take the easy way out and gather data in a haphazard way; but if a sample is chosen on the basis of convenience alone, it becomes impossible to interpret the resulting data with confidence. For example, it might be easy to use the students in your statistics class as a sample of students at your university. However, not all majors include a statistics course in their curriculum, and most students take statistics in their sophomore or junior year. The difficulty is that it is not clear whether or how these factors (and others that we might not be aware of) affect any conclusions based on information from such a sample.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
38
Chapter 2 Collecting Data Sensibly
There is no way to tell just by looking at a sample whether it is representative of the population from which it was drawn. Our only assurance comes from the method used to select the sample.
There are many reasons for selecting a sample rather than obtaining information from an entire population (a census). Sometimes the process of measuring the characteristics of interest is destructive, as with measuring the lifetime of flashlight batteries or the sugar content of oranges, and it would be foolish to study the entire population. But the most common reason for selecting a sample is limited resources. Restrictions on available time or money usually prohibit observation of an entire population.
Bias in Sampling Bias in sampling is the tendency for samples to differ from the corresponding population in some systematic way. Bias can result from the way in which the sample is selected or from the way in which information is obtained once the sample has been chosen. The most common types of bias encountered in sampling situations are selection bias, measurement or response bias, and nonresponse bias. Selection bias (sometimes also called undercoverage) is introduced when the way the sample is selected systematically excludes some part of the population of interest. For example, a researcher may wish to generalize from the results of a study to the population consisting of all residents of a particular city, but the method of selecting individuals may exclude the homeless or those without telephones. If those who are excluded from the sampling process differ in some systematic way from those who are included, the sample is virtually guaranteed to be unrepresentative of the population. If this difference between the included and the excluded occurs on a variable that is important to the study, conclusions based on the sample data may not be valid for the population of interest. Selection bias also occurs if only volunteers or self-selected individuals are used in a study, because those who choose to participate (for example, in a call-in telephone poll) may well differ from those who choose not to participate. Measurement or response bias occurs when the method of observation tends to produce values that systematically differ from the true value in some way. This might happen if an improperly calibrated scale is used to weigh items or if questions on a survey are worded in a way that tends to influence the response. For example, a Gallup survey sponsored by the American Paper Institute (Wall Street Journal, May 17, 1994) included the following question: “It is estimated that disposable diapers account for less than 2 percent of the trash in today’s landfills. In contrast, beverage containers, third-class mail and yard waste are estimated to account for about 21 percent of trash in landfills. Given this, in your opinion, would it be fair to tax or ban disposable diapers?” It is likely that the wording of this question prompted people to respond in a particular way. Other things that might contribute to response bias are the appearance or behavior of the person asking the question, the group or organization conducting the study, and the tendency for people not to be completely honest when asked about illegal behavior or unpopular beliefs. Although the terms measurement bias and response bias are often used interchangeably, the term measurement bias is usually used to describe systematic deviation from the true value as a result of a faulty measurement instrument (as with the improperly calibrated scale). Nonresponse bias occurs when responses are not obtained from all individuals selected for inclusion in the sample. As with selection bias, nonresponse bias can distort Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
2.2
Sampling
39
results if those who respond differ in important ways from those who do not respond. Although some level of nonresponse is unavoidable in most surveys, the biasing effect on the resulting sample is lowest when the response rate is high. To minimize nonresponse bias, it is critical that a serious effort be made to follow up with individuals who do not respond to an initial request for information. The nonresponse rate for surveys or opinion polls varies dramatically, depending on how the data are collected. Surveys are commonly conducted by mail, by phone, and by personal interview. Mail surveys are inexpensive but often have high nonresponse rates. Telephone surveys can also be inexpensive and can be implemented quickly, but they work well only for short surveys and they can also have high nonresponse rates. Personal interviews are generally expensive but tend to have better response rates. Some of the many challenges of conducting surveys are discussed in Section 2.5.
Types of Bias Selection Bias Tendency for samples to differ from the corresponding population as a result of systematic exclusion of some part of the population. Measurement or Response Bias Tendency for samples to differ from the corresponding population because the method of observation tends to produce values that differ from the true value. Nonresponse Bias Tendency for samples to differ from the corresponding population because data are not obtained from all individuals selected for inclusion in the sample. It is important to note that bias is introduced by the way in which a sample is selected or by the way in which the data are collected from the sample. Increasing the size of the sample, although possibly desirable for other reasons, does nothing to reduce bias if the method of selecting the sample is flawed or if the nonresponse rate remains high. A good discussion of types of bias appears in the sampling book by Lohr listed in the references in the back of the book. Potential sources of bias are illustrated in the following examples.
EXAMPLE 2.1
Are Cell Phone Users Different?
Many surveys are conducted by telephone and participants are often selected from phone books that include only landline telephones. For many years, it was thought that this was not a serious problem because most cell phone users also had a landline phone and so they still had a chance of being included in the survey. But the number of people with only cell phones is growing, and this trend is a concern for survey organizations. The article “Omitting Cell Phone Users May Affect Polls” (Associated Press, September 25, 2008) described a study that examined whether people who only have a cell phone are different that those who have landline phones. One finding from the study was that for people under the age of 30 with only a cell phone, 28% were Republicans compared to 36% of landline users. This suggests that researchers who use telephone surveys need to worry about how selection bias might influence the ability to generalize the results of a survey if only landlines are used.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
40
Chapter 2 Collecting Data Sensibly
E X A M P L E 2 . 2 Think Before You Order That Burger! The article “What People Buy from Fast-Food Restaurants: Caloric Content and Menu Item Selection” (Obesity [2009]: 1369–1374) reported that the average number of calories consumed at lunch in New York City fast food restaurants was 827. The researchers selected 267 fast food locations at random. The paper states that at each of these locations “adult customers were approached as they entered the restaurant and asked to provide their food receipt when exiting and to complete a brief survey.” Approaching customers as they entered the restaurant and before they ordered may have influenced what they purchased. This introduces the potential for response bias. In addition, some people chose not to participate when approached. If those who chose not to participate differed from those who did participate, the researchers also need to be concerned about nonresponse bias. Both of these potential sources of bias limit the researchers’ ability to generalize conclusions based on data from this study.
Random Sampling Most of the inferential methods introduced in this text are based on the idea of random selection. The most straightforward sampling method is called simple random sampling. A simple random sample is a sample chosen using a method that ensures that each different possible sample of the desired size has an equal chance of being the one chosen. For example, suppose that we want a simple random sample of 10 employees chosen from all those who work at a large design firm. For the sample to be a simple random sample, each of the many different subsets of 10 employees must be equally likely to be the one selected. A sample taken from only full-time employees would not be a simple random sample of all employees, because someone who works part-time has no chance of being selected. Although a simple random sample may, by chance, include only full-time employees, it must be selected in such a way that each possible sample, and therefore every employee, has the same chance of inclusion in the sample. It is the selection process, not the final sample, which determines whether the sample is a simple random sample. The letter n is used to denote sample size; it is the number of individuals or objects in the sample. For the design firm scenario of the previous paragraph, n510.
DEFINITION A simple random sample of size n is a sample that is selected from a population in a way that ensures that every different possible sample of the desired size has the same chance of being selected.
The definition of a simple random sample implies that every individual member of the population has an equal chance of being selected. However, the fact that every individual has an equal chance of selection, by itself, is not enough to guarantee that the sample is a simple random sample. For example, suppose that a class is made up of 100 students, 60 of whom are female. A researcher decides to select 6 of the female students by writing all 60 names on slips of paper, mixing the slips, and then picking 6. She then selects 4 male students from the class using a similar procedure. Even though every student in the class has an equal chance of being included in the sample (6 of 60 females Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
2.2
Sampling
41
are selected and 4 of 40 males are chosen), the resulting sample is not a simple random sample because not all different possible samples of 10 students from the class have the same chance of selection. Many possible samples of 10 students—for example, a sample of 7 females and 3 males or a sample of all females—have no chance of being selected. The sample selection method described here is not necessarily a bad choice (in fact, it is an example of stratified sampling, to be discussed in more detail shortly), but it does not produce a simple random sample, and this must be considered when a method is chosen for analyzing data resulting from such a sampling method.
Selecting a Simple Random Sample A number of different methods can be used to select a simple random sample. One way is to put the name or number of each member of the population on different but identical slips of paper. The process of thoroughly mixing the slips and then selecting n slips one by one yields a random sample of size n. This method is easy to understand, but it has obvious drawbacks. The mixing must be adequate, and producing the necessary slips of paper can be extremely tedious, even for relatively small populations. A commonly used method for selecting a random sample is to first create a list, called a sampling frame, of the objects or individuals in the population. Each item on the list can then be identified by a number, and a table of random digits or a random number generator can be used to select the sample. A random number generator is a procedure that produces a sequence of numbers that satisfies properties associated with the notion of randomness. Most statistics software packages include a random number generator, as do many calculators. A small table of random digits can be found in Appendix A, Table 1. For example, suppose a list containing the names of the 427 customers who purchased a new car during 2009 at a large dealership is available. The owner of the dealership wants to interview a sample of these customers to learn about customer satisfaction. She plans to select a simple random sample of 20 customers. Because it would be tedious to write all 427 names on slips of paper, random numbers can be used to select the sample. To do this, we can use three-digit numbers, starting with 001 and ending with 427, to represent the individuals on the list. The random digits from rows 6 and 7 of Appendix A, Table 1 are shown here: 09387679956256584264 41010220475119479751 We can use blocks of three digits from this list (underlined in the lists above) to identify the individuals who should be included in the sample. The first block of three digits is 093, so the 93rd person on the list will be included in the sample. The next five blocks of three digits (876, 799, 562, 565, and 842) do not correspond to anyone on the list, so we ignore them. The next block that corresponds to a person on the list is 410, so that person is included in the sample. This process would continue until 20 people have been selected for the sample. We would ignore any three-digit repeats since any particular person should only be selected once for the sample. Another way to select the sample would be to use computer software or a graphing calculator to generate 20 random numbers. For example, Minitab produced the following when 20 random numbers between 1 and 427 were requested. 289 10
67 203
29 346
26 186
205 232
214 410
422 43
31 293
233 98 25 371
These numbers could be used to determine which 20 customers to include in the sample. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
42
Chapter 2 Collecting Data Sensibly
When selecting a random sample, researchers can choose to do the sampling with or without replacement. Sampling with replacement means that after each successive item is selected for the sample, the item is “replaced” back into the population and may therefore be selected again at a later stage. In practice, sampling with replacement is rarely used. Instead, the more common method is to not allow the same item to be included in the sample more than once. After being included in the sample, an individual or object would not be considered for further selection. Sampling in this manner is called sampling without replacement.
DEFINITION Sampling without replacement: Once an individual from the population is selected for inclusion in the sample, it may not be selected again in the sampling process. A sample selected without replacement includes n distinct individuals from the population. Sampling with replacement: After an individual from the population is selected for inclusion in the sample and the corresponding data are recorded, the individual is placed back in the population and can be selected again in the sampling process. A sample selected with replacement might include any particular individual from the population more than once. Although these two forms of sampling are different, when the sample size n is small relative to the population size, as is often the case, there is little practical difference between them. In practice, the two can be viewed as equivalent if the sample size is at most 10% of the population size.
EXAMPLE 2.3
Selecting a Random Sample of Glass Soda Bottles
Breaking strength is an important characteristic of glass soda bottles. Suppose that we want to measure the breaking strength of each bottle in a random sample of size n53 selected from four crates containing a total of 100 bottles (the population). Each crate contains five rows of five bottles each. We can identify each bottle with a number from 1 to 100 by numbering across the rows in each crate, starting with the top row of crate 1, as pictured: Crate 1 1 2 3 6
4
5
Crate 4 76 77 ...
Crate 2 26 27 28 ...
...
© BananaStock/Alamy Images
... 100
Using a random number generator from a calculator or statistical software package, we could generate three random numbers between 1 and 100 to determine which bottles would be included in our sample. This might result in bottles 15 (row 3 column 5 of crate 1), 89 (row 3 column 4 of crate 4), and 60 (row 2 column 5 of crate 3) being selected.
Step-by-Step technology instructions available online Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
2.2
Sampling
43
The goal of random sampling is to produce a sample that is likely to be representative of the population. Although random sampling does not guarantee that the sample will be representative, it does allow us to assess the risk of an unrepresentative sample. It is the ability to quantify this risk that will enable us to generalize with confidence from a random sample to the corresponding population.
An Important Note Concerning Sample Size It is a common misconception that if the size of a sample is relatively small compared to the population size, the sample cannot possibly accurately reflect the population. Critics of polls often make statements such as, “There are 14.6 million registered voters in California. How can a sample of 1000 registered voters possibly reflect public opinion when only about 1 in every 14,000 people is included in the sample?” These critics do not understand the power of random selection! Consider a population consisting of 5000 applicants to a state university, and suppose that we are interested in math SAT scores for this population. A dotplot of the values in this population is shown in Figure 2.1(a). Figure 2.1(b) shows dotplots of the math SAT scores for individuals in five different random samples from the population, ranging in sample size from n550 to n51000. Notice that the samples tend to reflect the distribution of scores in the population. If we were interested in using the
300 400 500 Each dot represents up to 3 observations.
600
700
800
n = 1000
n = 500 n = 250 n = 100
FIGURE 2.1 (a) Dotplot of math SAT scores for the entire population. (b) Dotplots of math SAT scores for random samples of sizes 50, 100, 250, 500, and 1000.
n = 50 300 400 Each dot represents up to 3 observations.
500
600
700
800
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
44
Chapter 2 Collecting Data Sensibly
sample to estimate the population average or to say something about the variability in SAT scores, even the smallest of the samples (n550) pictured would provide reliable information. Although it is possible to obtain a simple random sample that does not do a reasonable job of representing the population, this is likely only when the sample size is very small, and unless the population itself is small, this risk does not depend on what fraction of the population is sampled. The random selection process allows us to be confident that the resulting sample adequately reflects the population, even when the sample consists of only a small fraction of the population.
Other Sampling Methods Simple random sampling provides researchers with a sampling method that is objective and free of selection bias. In some settings, however, alternative sampling methods may be less costly, easier to implement, and sometimes even more accurate.
Stratified Random Sampling When the entire population can be divided into a set of nonoverlapping subgroups, a method known as stratified sampling often proves easier to implement and more cost-effective than simple random sampling. In stratified random sampling, separate simple random samples are independently selected from each subgroup. For example, to estimate the average cost of malpractice insurance, a researcher might find it convenient to view the population of all doctors practicing in a particular metropolitan area as being made up of four subpopulations: (1) surgeons, (2) internists and family practitioners, (3) obstetricians, and (4) a group that includes all other areas of specialization. Rather than taking a random simple sample from the population of all doctors, the researcher could take four separate simple random samples—one from the group of surgeons, another from the internists and family practitioners, and so on. These four samples would provide information about the four subgroups as well as information about the overall population of doctors. When the population is divided in this way, the subgroups are called strata and each individual subgroup is called a stratum (the singular of strata). Stratified sampling entails selecting a separate simple random sample from each stratum. Stratified sampling can be used instead of simple random sampling if it is important to obtain information about characteristics of the individual strata as well as of the entire population, although a stratified sample is not required to do this—subgroup estimates can also be obtained by using an appropriate subset of data from a simple random sample. The real advantage of stratified sampling is that it often allows us to make more accurate inferences about a population than does simple random sampling. In general, it is much easier to produce relatively accurate estimates of characteristics of a homogeneous group than of a heterogeneous group. For example, even with a small sample, it is possible to obtain an accurate estimate of the average grade point average (GPA) of students graduating with high honors from a university. The individual GPAs of these students are all quite similar (a homogeneous group), and even a sample of three or four individuals from this subpopulation should be representative. On the other hand, producing a reasonably accurate estimate of the average GPA of all seniors at the university, a much more diverse group of GPAs, is a more difficult task. Thus, if a varied population can be divided into strata, with each stratum being much more homogeneous than the population with respect to the characteristic of interest, then a stratified random sample can produce more accurate estimates of population characteristics than a simple random sample of the same size.
Cluster Sampling Sometimes it is easier to select groups of individuals from a population than it is to select individuals themselves. Cluster sampling involves dividing the population of interest into nonoverlapping subgroups, called clusters. Clusters Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
2.2
Sampling
45
are then selected at random, and then all individuals in the selected clusters are included in the sample. For example, suppose that a large urban high school has 600 senior students, all of whom are enrolled in a first period homeroom. There are 24 senior homerooms, each with approximately 25 students. If school administrators wanted to select a sample of roughly 75 seniors to participate in an evaluation of the college and career placement advising available to students, they might find it much easier to select three of the senior homerooms at random and then include all the students in the selected homerooms in the sample. In this way, an evaluation survey could be administered to all students in the selected homerooms at the same time—certainly easier logistically than randomly selecting 75 students and then administering the survey to those individual seniors. Because whole clusters are selected, the ideal situation for cluster sampling is when each cluster mirrors the characteristics of the population. When this is the case, a small number of clusters results in a sample that is representative of the population. If it is not reasonable to think that the variability present in the population is reflected in each cluster, as is often the case when the cluster sizes are small, then it becomes important to ensure that a large number of clusters are included in the sample. Be careful not to confuse clustering and stratification. Even though both of these sampling strategies involve dividing the population into subgroups, both the way in which the subgroups are sampled and the optimal strategy for creating the subgroups are different. In stratified sampling, we sample from every stratum, whereas in cluster sampling, we include only selected whole clusters in the sample. Because of this difference, to increase the chance of obtaining a sample that is representative of the population, we want to create homogeneous groups for strata and heterogeneous (reflecting the variability in the population) groups for clusters.
Systematic Sampling Systematic sampling is a procedure that can be used when it is possible to view the population of interest as consisting of a list or some other sequential arrangement. A value k is specified (for example, k550 or k5200). Then one of the first k individuals is selected at random, after which every kth individual in the sequence is included in the sample. A sample selected in this way is called a 1 in k systematic sample. For example, a sample of faculty members at a university might be selected from the faculty phone directory. One of the first k520 faculty members listed could be selected at random, and then every 20th faculty member after that on the list would also be included in the sample. This would result in a 1 in 20 systematic sample. The value of k for a 1 in k systematic sample is generally chosen to achieve a desired sample size. For example, in the faculty directory scenario just described, if there were 900 faculty members at the university, the 1 in 20 systematic sample described would result in a sample size of 45. If a sample size of 100 was desired, a 1 in 9 systematic sample could be used (because 900/10059). As long as there are no repeating patterns in the population list, systematic sampling works reasonably well. However, if there are such patterns, systematic sampling can result in an unrepresentative sample. For example, suppose that workers at the entry station of a state park have recorded the number of visitors to the park each day for the past 10 years. In a 1 in 70 systematic sample of days from this list, we would pick one of the first 70 days at random and then every 70th day after that. But if the first day selected happened to be a Wednesday, every day selected in the entire sample would also be a Wednesday (because there are 7 days a week and 70 is a multiple of 7). It is unlikely that such a sample would be representative of the entire collection of days. The number of visitors is likely to be higher on weekend days, and no Saturdays or Sundays would be included in the sample. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
46
Chapter 2 Collecting Data Sensibly
It is often tempting to resort to “convenience” sampling—that is, using an easily available or convenient group to form a sample. This is a recipe for disaster! Results from such samples are rarely informative, and it is a mistake to try to generalize from a convenience sample to any larger population. One common form of convenience sampling is sometimes called voluntary response sampling. Such samples rely entirely on individuals who volunteer to be a part of the sample, often by responding to an advertisement, calling a publicized telephone number to register an opinion, or logging on to an Internet site to complete a survey. It is extremely unlikely that individuals participating in such voluntary response surveys are representative of any larger population of interest.
Convenience Sampling: Don’t Go There!
EX E RC I S E S 2 . 1 3 - 2 . 3 2 2.13 As part of a curriculum review, the psychology
2.17 Suppose that a group of 1000 orange trees is laid
department would like to select a simple random sample of 20 of last year’s 140 graduates to obtain information on how graduates perceived the value of the curriculum. Describe two different methods that might be used to select the sample.
out in 40 rows of 25 trees each. To determine the sugar content of fruit from a sample of 30 trees, researcher A suggests randomly selecting five rows and then randomly selecting six trees from each sampled row. Researcher B suggests numbering each tree on a map of the trees from 1 to 1000 and using random numbers to select 30 of the trees. Which selection method is preferred? Explain.
2.14 A petition with 500 signatures is submitted to a university’s student council. The council president would like to determine the proportion of those who signed the petition who are actually registered students at the university. There is not enough time to check all 500 names with the registrar, so the council president decides to select a simple random sample of 30 signatures. Describe how this might be done.
2.15 During the previous calendar year, a county’s small claims court processed 870 cases. Describe how a simple random sample of size n 5 50 might be selected from the case files to obtain information regarding the average award in such cases. 2.16 The financial aid advisor of a university plans to use a stratified random sample to estimate the average amount of money that students spend on textbooks each term. For each of the following proposed stratification schemes, discuss whether it would be worthwhile to stratify the university students in this manner. a. Strata corresponding to class standing (freshman, sophomore, junior, senior, graduate student) b. Strata corresponding to field of study, using the following categories: engineering, architecture, business, other c. Strata corresponding to the first letter of the last name: A–E, F–K, etc.
Bold exercises answered in back
Data set available online
2.18 For each of the situations described, state whether the sampling procedure is simple random sampling, stratified random sampling, cluster sampling, systematic sampling, or convenience sampling. a. All first-year students at a university are enrolled in 1 of 30 sections of a seminar course. To select a sample of freshmen at this university, a researcher selects four sections of the seminar course at random from the 30 sections and all students in the four selected sections are included in the sample. b. To obtain a sample of students, faculty, and staff at a university, a researcher randomly selects 50 faculty members from a list of faculty, 100 students from a list of students, and 30 staff members from a list of staff. c. A university researcher obtains a sample of students at his university by using the 85 students enrolled in his Psychology 101 class. d. To obtain a sample of the seniors at a particular high school, a researcher writes the name of each senior on a slip of paper, places the slips in a box and mixes them, and then selects 10 slips. The students whose names are on the selected slips of paper are included in the sample. e. To obtain a sample of those attending a basketball game, a researcher selects the 24th person through the door. Then, every 50th person after that is also included in the sample. Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
2.2
2.19 Of the 6500 students enrolled at a community college, 3000 are part time and the other 3500 are full time. The college can provide a list of students that is sorted so that all full-time students are listed first, followed by the part-time students. a. Describe a procedure for selecting a stratified random sample that uses full-time and part-time students as the two strata and that includes 10 students from each stratum. b. Does every student at this community college have the same chance of being selected for inclusion in the sample? Explain.
2.20 Briefly explain why it is advisable to avoid the use of convenience samples.
2.21 A sample of pages from this book is to be obtained, and the number of words on each selected page will be determined. For the purposes of this exercise, equations are not counted as words and a number is counted as a word only if it is spelled out—that is, ten is counted as a word, but 10 is not. a. Describe a sampling procedure that would result in a simple random sample of pages from this book. b. Describe a sampling procedure that would result in a stratified random sample. Explain why you chose the specific strata used in your sampling plan. c. Describe a sampling procedure that would result in a systematic sample. d. Describe a sampling procedure that would result in a cluster sample. e. Using the process you gave in Part (a), select a simple random sample of at least 20 pages, and record the number of words on each of the selected pages. Construct a dotplot of the resulting sample values, and write a sentence or two commenting on what it reveals about the number of words on a page. f. Using the process you gave in Part (b), select a stratified random sample that includes a total of at least 20 selected pages, and record the number of words on each of the selected pages. Construct a dotplot of the resulting sample values, and write a sentence or two commenting on what it reveals about the number of words on a page.
2.22 In 2000, the chairman of a California ballot initiative campaign to add “none of the above” to the list of ballot options in all candidate races was quite critical of a Field poll that showed his measure trailing by 10 percentage points. The poll was based on a random sample of 1000 registered voters in California. He is quoted by Bold exercises answered in back
Data set available online
Sampling
47
the Associated Press (January 30, 2000) as saying, “Field’s sample in that poll equates to one out of 17,505 voters,” and he added that this was so dishonest that Field should get out of the polling business! If you worked on the Field poll, how would you respond to this criticism?
2.23 The authors of the paper “Digital Inequality: Differences in Young Adults’ Use of the Internet” (Communication Research [2008]: 602–621) were interested in determining if people with higher levels of education use the Internet in different ways than those who do not have as much formal education. To answer this question, they used data from a national telephone survey. Approximately 1300 households were selected for the survey, and 270 of them completed the interview. What type of bias should the researchers be concerned about and why?
2.24 The authors of the paper “Illicit Use of Psychostimulants among College Students” (Psychology, Health & Medicine [2002]: 283–287) surveyed college students about their use of legal and illegal stimulants. The sample of students surveyed consisted of students enrolled in a psychology class at a small, competitive college in the United States. a. Was this sample a simple random sample, a stratified sample, a systematic sample, or a convenience sample? Explain. b. Give two reasons why the estimate of the proportion of students who reported using illegal stimulants based on data from this survey should not be generalized to all U.S. college students.
2.25 The paper “Deception and Design: The Impact of Communication Technology on Lying Behavior” (Computer-Human Interaction [2009]: 130–136) describes an investigation into whether lying is less common in face-to-face communication than in other forms of communication such as phone conversations or e-mail. Participants in this study were 30 students in an upperdivision communications course at Cornell University who received course credit for participation. Participants were asked to record all of their social interactions for a week, making note of any lies told. Based on data from these records, the authors of the paper concluded that students lie more often in phone conversations than in face-to-face conversations and more often in face-to-face conversations than in e-mail. Discuss the limitations of this study, commenting on the way the sample was selected and potential sources of bias. Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
48
Chapter 2 Collecting Data Sensibly
2.26 The authors of the paper “Popular Video Games: Quantifying the Presentation of Violence and its Context” (Journal of Broadcasting & Electronic Media [2003]: 58–76) investigated the relationship between video game rating—suitable for everyone (E), suitable for 13 years of age and older (T), and suitable for 17 years of age and older (M)—and the number of violent interactions per minute of play. The sample of games examined consisted of 60 video games—the 20 most popular (by sales) for each of three game systems. The researchers concluded that video games rated for older children had significantly more violent interactions per minute than did those games rated for more general audiences. a. Do you think that the sample of 60 games was selected in a way that makes it reasonable to think it is representative of the population of all video games? b. Is it reasonable to generalize the researchers’ conclusion to all video games? Explain why or why not.
2.27 Participants in a study of honesty in online dating profiles were recruited through print and online advertisements in the Village Voice, one of New York City’s most prominent weekly newspapers, and on Craigslist New York City (“The Truth About Lying in Online
Dating Profiles,” Computer-Human Interaction [2007]: 1–4). The actual height, weight, and age of the participants were compared to what appeared in their online dating profiles. The resulting data was then used to draw conclusions about how common deception was in online dating profiles. What concerns do you have about generalizing conclusions based on data from this study to the population of all people who have an online dating profile? Be sure to address at least two concerns and give the reason for your concern.
2.28 The report “Undergraduate Students and Credit Cards in 2004: An Analysis of Usage Rates and Trends” (Nellie Mae, May 2005) estimated that 21% of undergraduates with credit cards pay them off each month and that the average outstanding balance on undergraduates’ credit cards is $2169. These estimates were based on an online survey that was sent to 1260 students. Responses were received from 132 of these students. Is it reasonable to generalize the reported estimates to the population of all undergraduate students? Address at least two possible sources of bias in your answer.
2.29 Suppose that you were asked to help design a survey of adult city residents in order to estimate the proportion that would support a sales tax increase. The
Bold exercises answered in back
Data set available online
plan is to use a stratified random sample, and three stratification schemes have been proposed. Scheme 1: Stratify adult residents into four strata based on the first letter of their last name (A–G, H–N, O–T, U–Z). Scheme 2: Stratify adult residents into three strata: college students, nonstudents who work full time, nonstudents who do not work full time. Scheme 3: Stratify adult residents into five strata by randomly assigning residents into one of the five strata. Which of the three stratification schemes would be best in this situation? Explain.
2.30 The article “High Levels of Mercury Are Found in Californians” (Los Angeles Times, February 9, 2006) describes a study in which hair samples were tested for mercury. The hair samples were obtained from more than 6000 people who voluntarily sent hair samples to researchers at Greenpeace and The Sierra Club. The researchers found that nearly one-third of those tested had mercury levels that exceeded the concentration thought to be safe. Is it reasonable to generalize this result to the larger population of U.S. adults? Explain why or why not. Whether or not to continue a Mardi Gras Parade through downtown San Luis Obispo, CA, is a hotly debated topic. The parade is popular with students and many residents, but some celebrations have led to complaints and a call to eliminate the parade. The local newspaper conducted online and telephone surveys of its readers and was surprised by the results. The survey web site received more than 400 responses, with more than 60% favoring continuing the parade, while the telephone response line received more than 120 calls, with more than 90% favoring banning the parade (San Luis Obispo Tribune, March 3, 2004). What factors may have contributed to these very different results?
2.31
2.32 The article “Gene’s Role in Cancer May Be Overstated” (San Luis Obispo Tribune, August 21, 2002) states that “early studies that evaluated breast cancer risk among gene mutation carriers selected women in families where sisters, mothers, and grandmothers all had breast cancer. This created a statistical bias that skewed risk estimates for women in the general population.” Is the bias described here selection bias, measurement bias, or nonresponse bias? Explain.
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
2.3
2.3
Simple Comparative Experiments
49
Simple Comparative Experiments Sometimes the questions we are trying to answer deal with the effect of certain explanatory variables on some response. Such questions are often of the form, “What happens when . . . ?” or “What is the effect of . . . ?” For example, an industrial engineer may be considering two different workstation designs and might want to know whether the choice of design affects work performance. A medical researcher may want to determine how a proposed treatment for a disease compares to a standard treatment. Experiments provide a way to collect data to answer these types of questions.
DEFINITION An experiment is a study in which one or more explanatory variables are manipulated in order to observe the effect on a response variable. The explanatory variables are those variables that have values that are controlled by the experimenter. Explanatory variables are also called factors. The response variable is a variable that is not controlled by the experimenter and that is measured as part of the experiment. An experimental condition is any particular combination of values for the explanatory variables. Experimental conditions are also called treatments.
Suppose we are interested in determining the effect of room temperature on performance on a first-year calculus exam. In this case, the explanatory variable is room temperature (it can be manipulated by the experimenter). The response variable is exam performance (the variable that is not controlled by the experimenter and that will be measured). In general, we can identify the explanatory variables and the response variable easily if we can describe the purpose of the experiment in the following terms: The purpose is to assess the effect of on . explanatory response variable variable Let’s return to the example of an experiment to assess the effect of room temperature on exam performance. We might decide to use two room temperature settings, 65° and 75°. This would result in an experiment with two experimental conditions (or equivalently, two treatments) corresponding to the two temperature settings. Suppose that there are 10 sections of first-semester calculus that have agreed to participate in our study. We might design an experiment in this way: Set the room temperature (in degrees Fahrenheit) to 65° in five of the rooms and to 75° in the other five rooms on test day, and then compare the exam scores for the 65° group and the 75° group. Suppose that the average exam score for the students in the 65° group was noticeably higher than the average for the 75° group. Could we conclude that the increased temperature resulted in a lower average score? Based on the information given, the answer is no because many other factors might be related to exam score. Were the sections at different times of the day? Did they have the same instructor? Different textbooks? Did the sections differ with respect to the abilities of the students? Any of these other factors could provide a plausible explanation (having nothing to do with room temperature) for why the average test score was different for the two groups. It is not possible to separate the effect of temperature from the effects of Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
50
Chapter 2 Collecting Data Sensibly
these other factors. As a consequence, simply setting the room temperatures as described makes for a poorly designed experiment.
A well-designed experiment requires more than just manipulating the explanatory variables; the design must also eliminate other possible explanations or the experimental results will not be conclusive.
The goal is to design an experiment that will allow us to determine the effects of the explanatory variables on the chosen response variable. To do this, we must take into consideration any extraneous variables that, although not of interest in the current study, might also affect the response variable.
DEFINITION An extraneous variable is one that is not one of the explanatory variables in the study but is thought to affect the response variable.
A well-designed experiment copes with the potential effects of extraneous variables by using random assignment to experimental conditions and sometimes also by incorporating direct control and/or blocking into the design of the experiment. Each of these strategies—random assignment, direct control, and blocking—is described in the paragraphs that follow. A researcher can directly control some extraneous variables. In the calculus test example, the textbook used is an extraneous variable because part of the differences in test results might be attributed to this variable. We could control this variable directly, by requiring that all sections use the same textbook. Then any observed differences between temperature groups could not be explained by the use of different textbooks. The extraneous variable time of day might also be directly controlled in this way by having all sections meet at the same time. The effects of some extraneous variables can be filtered out by a process known as blocking. Extraneous variables that are addressed through blocking are called blocking variables. Blocking creates groups (called blocks) that are similar with respect to blocking variables; then all treatments are tried in each block. In our example, we might use instructor as a blocking variable. If five instructors are each teaching two sections of calculus, we would make sure that for each instructor, one section was part of the 65° group and the other section was part of the 75° group. With this design, if we see a difference in exam scores for the two temperature groups, the extraneous variable instructor can be ruled out as a possible explanation, because all five instructors’ students were present in each temperature group. (Had we controlled the instructor variable by choosing to have only one instructor, that would be an example of direct control. Of course we can’t directly control both time of day and instructor.) If one instructor taught all the 65° sections and another taught all the 75° sections, we would be unable to distinguish the effect of temperature from the effect of the instructor. In this situation, the two variables (temperature and instructor) are said to be confounded.
Two variables are confounded if their effects on the response variable cannot be distinguished from one another.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
2.3
51
Simple Comparative Experiments
If an extraneous variable is confounded with the explanatory variables (which define the treatments), it is not possible to draw an unambiguous conclusion about the effect of the treatment on the response. Both direct control and blocking are effective in ensuring that the controlled variables and blocking variables are not confounded with the variables that define the treatments. We can directly control some extraneous variables by holding them constant, and we can use blocking to create groups that are similar to essentially filter out the effect of others. But what about variables, such as student ability in our calculus test example, which cannot be controlled by the experimenter and which would be difficult to use as blocking variables? These extraneous variables are handled by the use of random assignment to experimental groups. Random assignment ensures that our experiment does not systematically favor one experimental condition over any other and attempts to create experimental groups that are as much alike as possible. For example, if the students requesting calculus could be assigned to one of the ten available sections using a random mechanism, we would expect the resulting groups to be similar with respect to student ability as well as with respect to other extraneous variables that are not directly controlled or used as a basis for blocking. Note that random assignment in an experiment is different from random selection of subjects. The ideal situation would be to have both random selection of subjects and random assignment of subjects to experimental conditions, as this would allow conclusions from the experiment to be generalized to a larger population. For many experiments the random selection of subjects is not possible. As long as subjects are assigned at random to experimental conditions, it is still possible to assess treatment effects. To get a sense of how random assignment tends to create similar groups, suppose that 50 college freshmen are available to participate as subjects in an experiment to investigate whether completing an online review of course material before an exam improves exam performance. The 50 subjects vary quite a bit with respect to achievement, which is reflected in their math and verbal SAT scores, as shown in Figure 2.2.
Math
FIGURE 2.2 Dotplots of math and verbal SAT scores for 50 freshmen.
Verbal 500
600
700
800
If these 50 students are to be assigned to the two experimental groups (one that will complete the online review and one that will not), we want to make sure that the assignment of students to groups does not favor one group over the other by tending to assign the higher achieving students to one group and the lower achieving students to the other. Creating groups of students with similar achievement levels in a way that considers both verbal and math SAT scores simultaneously would be difficult, so we rely on random assignment. Figure 2.3(a) shows the math SAT scores of the students assigned to each of the two experimental groups (one shown in orange and one shown in blue) for each of three different random assignments of students to groups. Figure 2.3(b) shows the verbal SAT scores for the two experimental groups for each of the same three random assignments. Notice that each of the three random assignments produced groups that are similar with respect to both verbal and math SAT scores. So, if any of these three assignments were used and the two groups differed on exam performance, we could rule out differences in math or verbal SAT scores as possible competing explanations for the difference. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
52
Chapter 2 Collecting Data Sensibly
(a) 500
600
700
800
500
600
700
800
FIGURE 2.3 Dotplots for three different random assignments to two groups, one shown in orange and one shown in blue: (a) math SAT score; (b) verbal SAT score.
(b)
Not only will random assignment tend to create groups that are similar with respect to verbal and math SAT scores, but it will also tend to even out the groups with respect to other extraneous variables. As long as the number of subjects is not too small, we can rely on the random assignment to produce comparable experimental groups. This is the reason that random assignment is a part of all well-designed experiments. Not all experiments require the use of human subjects. For example, a researcher interested in comparing three different gasoline additives with respect to gasoline mileage might conduct an experiment using a single car with an empty tank. One gallon of gas with one of the additives will be put in the tank, and the car will be driven along a standard route at a constant speed until it runs out of gas. The total distance traveled on the gallon of gas could then be recorded. This could be repeated a number of times—10, for example—with each additive. The experiment just described can be viewed as consisting of a sequence of trials. Because a number of extraneous variables (such as variations in environmental conditions like wind speed or humidity and small variations in the condition of the car) might have an effect on gas mileage, it would not be a good idea to use additive 1 for the first 10 trials, additive 2 for the next 10 trials, and so on. A better approach would be to randomly assign additive 1 to 10 of the 30 planned trials, and then randomly assign additive 2 to 10 of the remaining 20 trials. The resulting plan for carrying out the experiment might look as follows: Trial Additive
1 2
2 2
3 3
4 3
5 2
6 1
7 2
... ...
30 1
When an experiment can be viewed as a sequence of trials, random assignment involves the random assignment of treatments to trials. Remember that random assignment— either of subjects to treatments or of treatments to trials—is a critical component of a good experiment. Random assignment can be effective only if the number of subjects or observations in each experimental condition (treatment) is large enough for each experimental group to reliably reflect variability in the population. For example, if there were only 20 students requesting calculus, it is unlikely that we would get equivalent groups for comparison, even with random assignment to the ten sections. Replication is the design strategy of making multiple observations for each experimental condition. Together, replication and random assignment allow the researcher to be reasonably confident of comparable experimental groups. To illustrate the design of a simple experiment, consider the dilemma of Anna, a waitress in a local restaurant. She would like to increase the amount of her tips, and her strategy is simple: She will write “Thank you” on the back of some of the checks before giving them to the patrons and on others she will write nothing. She plans to calculate the percentage of the tip as her measure of success (for example, a 15% tip is common). She will compare the average percentage of the tips calculated from Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
2.3
Simple Comparative Experiments
53
Key Concepts in Experimental Design Random Assignment Random assignment (of subjects to treatments or of treatments to trials) to ensure that the experiment does not systematically favor one experimental condition (treatment) over another. Blocking Using extraneous variables to create groups (blocks) that are similar. All experimental conditions (treatments) are then tried in each block. Direct Control Holding extraneous variables constant so that their effects are not confounded with those of the experimental conditions (treatments). Replication Ensuring that there is an adequate number of observations for each experimental condition.
checks with and without the handwritten “Thank you.” If writing “Thank you” does not produce higher tips, she may try a different strategy. Anna is untrained in the art of planning experiments, but already she has taken some common sense steps in the right direction to answer her question—Will writing “Thank you” produce the desired outcome of higher tips? Anna has defined a manageable problem, and collecting the appropriate data is feasible. It should be easy to gather data as a normal part of her work. Anna wonders whether writing “Thank you” on the customers’ bills will have an effect on the amount of her tip. In the language of experimentation, we would refer to the writing of “Thank you” and the not writing of “Thank you” as treatments (the two experimental conditions to be compared in the experiment). The two treatments together are the possible values of the explanatory variable. The tipping percentage is the response variable. The idea behind this terminology is that the tipping percentage is a response to the treatments writing “Thank you” or not writing “Thank you.” Anna’s experiment may be thought of as an attempt to explain the variability in the response variable in terms of its presumed cause, the variability in the explanatory variable. That is, as she manipulates the explanatory variable, she expects the response by her customers to vary. Anna has a good start, but now she must consider the four fundamental design principles. Replication. Anna cannot run a successful experiment by gathering tipping information on only one person for each treatment. There is no reason to believe that any single tipping incident is representative of what would happen in other incidents, and therefore it would be impossible to evaluate the two treatments with only two subjects. To interpret the effects of a particular treatment, she must replicate each treatment in the experiment. Blocking. Suppose that Anna works on both Thursdays and Fridays. Because day of the week might affect tipping behavior, Anna should block on day of the week and make sure that observations for both treatments are made on each of the two days. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
54
Chapter 2 Collecting Data Sensibly
Direct Control and Random Assignment. There are a number of extraneous variables that might have an effect on the size of tip. Some restaurant patrons will be seated near the window with a nice view; some will have to wait for a table, whereas others may be seated immediately; and some may be on a fixed income and cannot afford a large tip. Some of these variables can be directly controlled. For example, Anna may choose to use only window tables in her experiment, thus eliminating table location as a potential confounding variable. Other variables, such as length of wait and customer income, cannot be easily controlled. As a result, it is important that Anna use random assignment to decide which of the window tables will be in the “Thank you” group and which will be in the “No thank you” group. She might do this by flipping a coin as she prepares the check for each window table. If the coin lands with the head side up, she could write “Thank you” on the bill, omitting the “Thank you” when a tail is observed. The accompanying box summarizes how experimental designs deal with extraneous variables.
Taking Extraneous Variables into Account Extraneous variables are variables other than the explanatory variables in an experiment that may also have an effect on the response variable. There are several strategies for dealing with extraneous variables in order to avoid confounding. Extraneous variables that we know about and choose to incorporate into the experimental design: Strategies Direct control—holds extraneous variables fixed so that they can’t affect the response variable Blocking—allows for valid comparisons because each treatment is tried in each block Extraneous variables that we don’t know about or choose not to incorporate into the experimental design through direct control or blocking: Strategy Random assignment Extraneous variables that are not incorporated into the design of the experiment are sometimes called lurking variables.* *For more on lurking variables, see “Lurking Variables: Some Examples” (The American Statistician [1981]: 227–233).
A Note on Random Assignment There are several strategies that can be used to perform random assignment of subjects to treatments or treatments to trials. Two common strategies are: • Write the name of each subject or a unique number that corresponds to a subject
on a slip of paper. Place all of the slips in a container and mix well. Then draw out the desired number of slips to determine those that will be assigned to the first treatment group. This process of drawing slips of paper then continues until all treatment groups have been determined. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
2.3
Simple Comparative Experiments
55
• Assign each subject a unique number from 1 to n, where n represents the total
number of subjects. Use a random number generator or table of random numbers to obtain numbers that will identify which subjects will be assigned to the first treatment group. This process would be repeated, ignoring any random numbers generated that correspond to subjects that have already been assigned to a treatment group, until all treatment groups have been formed. The two strategies above work well and can be used for experiments in which the desired number of subjects in each treatment group has been predetermined. Another strategy that is sometimes employed is to use a random mechanism (such as tossing a coin or rolling a die) to determine which treatment will be assigned to a particular subject. For example, in an experiment with two treatments, you might toss a coin to determine if the first subject is assigned to treatment 1 or treatment 2. This could continue for each subject—if the coin lands H, the subject is assigned to treatment 1, and if the coin lands T, the subject is assigned to treatment 2. This strategy is fine, but may result in treatment groups of unequal size. For example, in an experiment with 100 subjects, 53 might be assigned to treatment 1 and 47 to treatment 2. If this is acceptable, the coin flip strategy is a reasonable way to assign subjects to treatments. But, suppose you want to ensure that there is an equal number of subjects in each treatment group. Is it acceptable to use the coin flip strategy until one treatment group is complete and then just assign all of the remaining subjects to groups that are not yet full? The answer to this question is that it is probably not acceptable. For example, suppose a list of 20 subjects is in order by age from youngest to oldest and that we want to form two treatment groups each consisting of 10 subjects. Tossing a coin to make the assignments might result in the following (based on using the first row of random digits in Appendix A, Table 1, with an even number representing H and an odd number representing T):
Subject
Random Number
Coin Toss Equivalent
Treatment Group
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
4 5 1 8 5 0 3 3 7 1 2 8 4 5 1
H T T H T H T T T T H H H T T
1 2 2 1 2 1 2 2 2 2 1 1 1 2 2 1 1 1 1 1
Treatment group 2 filled. Assign all others to treatment group 1.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
56
Chapter 2 Collecting Data Sensibly
If the list of subjects was ordered by age, treatment group 1 would end up with a disproportionate number of older people. This strategy usually results in one treatment group drawing disproportionately from the end of the list. So, the only time the strategy of assigning at random until groups fill up and then assigning the remaining subjects to the group that is not full is reasonable is if you can be sure that the list is in random order with respect to all variables that might be related to the response variable. Because of this, it is best to avoid this strategy. Activity 2.5 investigates potential difficulties with this type of strategy. On the other hand, if the number of subjects is large, it may not be important that every treatment group has exactly the same number of subjects. If this is the case, it is reasonable to use a coin flip strategy (or other strategies of this type) that does not involve stopping assignment of subjects to a group that becomes full.
Evaluating an Experimental Design The key concepts of experimental design provide a framework for evaluating an experimental design, as illustrated in the following examples.
E X A M P L E 2 . 4 Revenge Is Sweet The article “The Neural Basis of Altruistic Punishment” (Science, August 27, 2004) described a study that examined motivation for revenge. Subjects in the study were all healthy, right-handed men. Subjects played a game with another player in which they could both earn money by trusting each other or one player could double-cross the other player and keep all of the money. In some cases the double cross was required by the rules of the game in certain circumstances, while in other cases the double cross was the result of a deliberate choice. The victim of a double cross was then given the opportunity to retaliate by imposing a fine, but sometimes the victim had to spend some of his own money in order to impose the fine. This study was an experiment with four experimental conditions or treatments: 1. double cross not deliberate (double cross dictated by the rules of the game) and no cost to the victim to retaliate 2. double cross deliberate and no cost to the victim to retaliate 3. double cross not deliberate and a cost to the victim to retaliate 4. double cross deliberate and a cost to the victim to retaliate All subjects chose revenge (imposed a fine on the double-crosser) when the double cross was deliberate and retaliation was free, and 86% of the subjects chose revenge when the double cross was deliberate, even if it cost them money. Only 21% imposed a fine if the double cross was dictated by the rules of the game and was not deliberate. Assuming that the researchers randomly assigned the subjects to the four experimental conditions, this study is an experiment that incorporated random assignment, direct control (controlled sex, health, and handedness by using only healthy, righthanded males as subjects), and replication (many subjects assigned to each experimental condition).
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
2.3
Simple Comparative Experiments
57
E X A M P L E 2 . 5 Subliminal Messages The article “The Most Powerful Manipulative Messages Are Hiding in Plain Sight” (Chronicle of Higher Education, January 29, 1999) reported the results of an interesting experiment on priming—the effect of subliminal messages on how we behave. In the experiment, subjects completed a language test in which they were asked to construct a sentence using each word in a list of words. One group of subjects received a list of words related to politeness, and a second group was given a list of words related to rudeness. Subjects were told to complete the language test and then come into the hall and find the researcher so that he could explain the next part of the test. When each subject came into the hall, he or she found the researcher engaged in conversation. The researcher wanted to see whether the subject would interrupt the conversation. The researcher found that 63% of those primed with words related to rudeness interrupted the conversation, whereas only 17% of those primed with words related to politeness interrupted. If we assume that the researcher randomly assigned the subjects to the two groups, then this study is an experiment that compares two treatments (primed with words related to rudeness and primed with words related to politeness). The response variable, politeness, has the values interrupted conversation and did not interrupt conversation. The experiment uses replication (many subjects in each treatment group) and random assignment to control for extraneous variables that might affect the response.
Many experiments compare a group that receives a particular treatment to a control group that receives no treatment.
E X A M P L E 2 . 6 Chilling Newborns? Then You Need a Control Group... Researchers for the National Institute of Child Health and Human Development studied 208 infants whose brains were temporarily deprived of oxygen as a result of complications at birth (The New England Journal of Medicine, October 13, 2005). These babies were subjects in an experiment to determine if reducing body temperature for three days after birth improved their chances of surviving without brain damage. The experiment was summarized in a paper that stated “infants were randomly assigned to usual care (control group) or whole-body cooling.” Including a control group in the experiment provided a basis for comparison of death and disability rates for the proposed cooling treatment and those for usual care. Some extraneous variables that might also affect death and disability rates, such as the duration of oxygen deprivation, could not be directly controlled, so to ensure that the experiment did not unintentionally favor one experimental condition over the other, random assignment of the infants to the two groups was critical. Because this was a well-designed experiment, the researchers were able to use the resulting data and statistical methods that you will see in Chapter 11 to conclude that cooling did reduce the risk of death and disability for infants deprived of oxygen at birth.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
58
Chapter 2 Collecting Data Sensibly
Visualizing the Underlying Structure of Some Common Experimental Designs Simple diagrams are sometimes used to highlight important features of some common experimental designs. The structure of an experiment that is based on random assignment of experimental units (the units to which treatments are assigned, usually subjects or trials) to one of two treatments is displayed in Figure 2.4. The diagram can be easily adapted for an experiment with more than two treatments. In any particular setting, we would also want to customize the diagram by indicating what the treatments are and what response will be measured. This is illustrated in Example 2.7.
Treatment A
Measure response for Treatment A
Random assignment Experimental units/ subjects
Compare treatments Random assignment Treatment B
Measure response for Treatment B
FIGURE 2.4 Diagram of an experiment with random assignment of experimental units to two treatments.
EXAMPLE 2.7
A Helping Hand
Can moving their hands help children learn math? This is the question investigated by the authors of the paper “Gesturing Gives Children New Ideas about Math” (Psychological Science [2009]: 267–272). An experiment was conducted to compare two different methods for teaching children how to solve math problems of the form 312185____18. One method involved having students point to the 312 on the left side of the equal sign with one hand and then point to the blank on the right side of the equal sign before filling in the blank to complete the equation. The other method did not involve using these hand gestures. The paper states that the study used children ages 9 and 10 who were given a pretest containing six problems of the type described above. Only children who answered all six questions incorrectly became subjects in the experiment. There were a total of 128 subjects. To compare the two methods, the 128 children were assigned at random to the two experimental conditions. Children assigned to one experimental condition were taught a method that used hand gestures and children assigned to the other experimental condition were taught a similar strategy that did not involve using hand gestures. Each child then took a test with six problems and the number correct was determined for each child. The researchers used the resulting data to reach the conclusion that the average number correct for children who used the method that incorporated hand gestures was significantly higher than the average number correct for children who were taught the method that did not use hand gestures. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
2.3
Method with hand gestures
Simple Comparative Experiments
59
Measure number correct on test
Random assignment Compare number correct for those who used hand gestures and those who did not
128 children Random assignment Method without hand gestures
Measure number correct on test
FIGURE 2.5 Diagram for the experiment of Example 2.7.
The basic structure of this experiment can be diagramed as shown in Figure 2.5. This type of diagram provides a nice summary of the experiment, but notice that several important characteristics of the experiment are not captured in the diagram. For example, the diagram does not show that some extraneous variables were considered by the researchers and directly controlled. In this example, both age and prior math knowledge were directly controlled by using only children who were 9 and 10 years old and who were not able to solve any of the questions on the pretest correctly. So, be aware that while a diagram of an experiment may be a useful tool, it usually cannot stand alone in describing an experimental design.
Some experiments consist of a sequence of trials, and treatments are assigned at random to the trials. The diagram in Figure 2.6 illustrates the underlying structure of such an experiment. Example 2.8 shows how this diagram can be customized to describe a particular experiment.
Trials for Treatment A
Measure response for Treatment A
Random assignment Experimental trials
Compare treatments Random assignment Trials for Treatment B
Measure response for Treatment B
FIGURE 2.6 Diagram of an experiment with random assignment of treatments to trials. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
60
Chapter 2 Collecting Data Sensibly
E X A M P L E 2 . 8 Distracted? Watch Out for Those Cars! The paper “Effect of Cell Phone Distraction on Pediatric Pedestrian Injury Risk” (Pediatrics [2009]: e179–e185) describes an experiment to investigate whether pedestrians who are talking on a cell phone are at greater risk of an accident when crossing the street than when not talking on a cell phone. No children were harmed in this experiment—a virtual interactive pedestrian environment was used! One possible way of conducting such an experiment would be to have a person cross 20 streets in this virtual environment. The person would talk on a cell phone for some crossings and would not use the cell phone for others. It would be important to randomly assign the two treatments (talking on the phone, not talking on the phone) to the 20 trials (the 20 simulated street crossings). This would result in a design that did not favor one treatment over the other because the pedestrian became more careful with experience or more tired and, therefore, easily distracted over time. The basic structure of this experiment is diagramed in Figure 2.7.
Crossings with cell phone
Accident or close call: YES or NO
Random assignment Compare accident risk for cell phone vs. no cell phone
20 street crossings Random assignment Crossings without cell phone
FIGURE 2.7 Diagram for the experiment of Example 2.8 with random assignment to trials.
Accident or close call: YES or NO
The actual experiment conducted by the authors of the paper was a bit more sophisticated than the one just described. In this experiment, 77 children age 10 and 11 each performed simulated crossings with and without a cell phone. Random assignment was used to decide which children would cross first with the cell phone followed by no cell phone and which children could cross first with no cell phone. The structure of this experiment is diagramed in Figure 2.8.
Cell phone first, then no cell phone
Accident or close call with cell phone? Accident or close call without cell phone?
Compare accident risk for cell phone vs. no cell phone for those in the cell phone first group
Random assignment Compare accident risk for cell phone vs. no cell phone
77 children Random assignment No cell phone first, then cell phone
Accident or close call without cell phone? Accident or close call with cell phone?
Compare accident risk for cell phone vs. no cell phone for those in the no cell phone first group
FIGURE 2.8 Diagram for the Experiment of Example 2.8 with 77 children. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
2.3
Simple Comparative Experiments
61
As was the case in Example 2.7, note that while the diagram is informative, by itself, it does not capture all of the important aspects of the design. In particular, it does not capture the direct control of age (only children age 10 and 11 were used as subjects in the experiment). Experimental designs in which experimental units are assigned at random to treatments or in which treatments are assigned at random to trials (like those of the experiments in Examples 2.7 and 2.8) are called completely randomized designs. Diagrams are also useful for highlighting the structure of experiments that use blocking. This is illustrated in Example 2.9.
E X A M P L E 2 . 9 A Helping Hand Revisited
FIGURE 2.9 Diagram for the experiment of Example 2.9 using gender to form blocks.
Create blocks
Random assignment
81 girls
Let’s return to the experiment described in Example 2.7. Take a minute to go back and re-read that example. The experiment described in Example 2.7, a completely randomized design with 128 subjects, was used to compare two different methods for teaching kids how to solve a particular type of math problem. Age and prior math knowledge were extraneous variables that the researchers thought might be related to performance on the math test given at the end of the lesson, so the researchers chose to directly control these variables. The 128 children were assigned at random to the two experimental conditions (treatments). The researchers relied on random assignment to create treatment groups that would be roughly equivalent with respect to other extraneous variables. But suppose that we were worried that gender might also be related to performance on the math test. One possibility would be to use direct control of gender— that is, we might use only boys or only girls as subjects in the experiment. Then if we saw a difference in test performance for the two teaching methods, it could not be due to one experimental group containing more boys and fewer girls than the other group. The downside to this strategy is that if we use only boys in the experiment, there is no basis for also generalizing any conclusions from the experiment to girls. Another strategy for dealing with extraneous variables is to incorporate blocking into the design. In the case of gender, we could create two blocks, one consisting of girls and one consisting of boys. Then, once the blocks are formed, we would randomly assign the girls to the two treatments and randomly assign the boys to the two treatments. In the actual study, the group of 128 children included 81 girls and 47 boys. A diagram that shows the structure of an experiment that includes blocking using gender is shown in Figure 2.9. Method with hand gestures
Measure number correct on test
Method without hand gestures
Measure number correct on test
Compare number correct for girls who used hand gestures and girls who did not Compare number correct for those who used hand gestures and those who did not
128 children
47 boys
Random assignment
Create blocks
Method with hand gestures
Measure number correct on test
Method without hand gestures
Measure number correct on test
Compare number correct for boys who used hand gestures and boys who did not
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
62
Chapter 2 Collecting Data Sensibly
When blocking is used, the design is called a randomized block design. Note that one difference between the diagram that describes the experiment in which blocking is used (Figure 2.9) and the diagram of the original experiment (Figure 2.5) is at what point the random assignment occurs. When blocking is incorporated in an experiment, the random assignment to treatments occurs after the blocks have been formed and is done separately for each block.
Before proceeding with an experiment, you should be able to give a satisfactory answer to each of the following 10 questions. 1. What is the research question that data from the experiment will be used to answer? 2. What is the response variable? 3. How will the values of the response variable be determined? 4. What are the explanatory variables for the experiment? 5. For each explanatory variable, how many different values are there, and what are these values? 6. What are the treatments for the experiment? 7. What extraneous variables might influence the response? 8. How does the design incorporate random assignment of subjects to treatments (or treatments to subjects) or random assignment of treatments to trials? 9. For each extraneous variable listed in Question 7, how does the design protect against its potential influence on the response through blocking, direct control, or random assignment? 10. Will you be able to answer the research question using the data collected in this experiment?
EX E RC I S E S 2 . 3 3 - 2 . 4 7 2.33 The head of the quality control department at a printing company would like to carry out an experiment to determine which of three different glues results in the greatest binding strength. Although they are not of interest in the current investigation, other factors thought to affect binding strength are the number of pages in the book and whether the book is being bound as a paperback or a hardback. a. What is the response variable in this experiment? b. What explanatory variable will determine the experimental conditions? c. What two extraneous variables are mentioned in the problem description? Are there other extraneous variables that should be considered?
Bold exercises answered in back
Data set available online
2.34 A study of college students showed a temporary gain of up to 9 IQ points after listening to a Mozart piano sonata. This conclusion, dubbed the Mozart effect, has since been criticized by a number of researchers who have been unable to confirm the result in similar studies. Suppose that you wanted to see whether there is a Mozart effect for students at your school. a. Describe how you might design an experiment for this purpose. b. Does your experimental design include direct control of any extraneous variables? Explain. c. Does your experimental design use blocking? Explain why you did or did not include blocking in your design. d. What role does random assignment play in your design?
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
2.3
Simple Comparative Experiments
63
2.35 The following is from an article titled “After the
2.37 The Institute of Psychiatry at Kings College Lon-
Workout, Got Chocolate Milk?” that appeared in the Chicago Tribune (January 18, 2005):
don found that dealing with “infomania” has a temporary, but significant derogatory effect on IQ (Discover, November 2005). In this experiment, researchers divided volunteers into two groups. Each subject took an IQ test. One group had to check e-mail and respond to instant messages while taking the test, and the second group took the test without any distraction. The distracted group had an average score that was 10 points lower than the average for the control group. Explain why it is important that the researchers created the two experimental groups in this study by using random assignment.
The article is not explicit about this, but in order for this to have been a well-designed experiment, it must have incorporated random assignment. Briefly explain where the researcher would have needed to use random assign in order for the conclusion of the experiment to be valid.
2.36 The report “Comparative Study of Two Computer Mouse Designs” (Cornell Human Factors Laboratory Technical Report RP7992) included the following description of the subjects used in an experiment: Twenty-four Cornell University students and staff (12 males and 12 females) volunteered to participate in the study. Three groups of 4 men and 4 women were selected by their stature to represent the 5th percentile (female 152.160.3 cm, male 164.160.4 cm), 50th percentile (female 162.460.1 cm, male 174.160.7 cm), and 95th percentile (female 171.960.2 cm, male 185.760.6 cm) ranges . . . All subjects reported using their right hand to operate a computer mouse. This experimental design incorporated direct control and blocking. a. Are the potential effects of the extraneous variable stature (height) addressed by blocking or direct control? b. Whether the right or left hand is used to operate the mouse was considered to be an extraneous variable. Are the potential effects of this variable addressed by blocking or direct control?
Bold exercises answered in back
Data set available online
2.38 In an experiment to compare two different surgical procedures for hernia repair (“A Single-Blinded, Ran-
domized Comparison of Laparoscopic Versus Open Hernia Repair in Children,” Pediatrics [2009]: 332– 336), 89 children were assigned at random to one of the two surgical methods. The researchers relied on the random assignment of subjects to treatments to create comparable groups with respect to extraneous variables that they did not control. One such extraneous variable was age. After random assignment to treatments, the researchers looked at the age distribution of the children in each of the two experimental groups (laparoscopic repair (LR) and open repair (OR)). The accompanying figure is from the paper. 16 14 12 10 Age, y
Researchers at Indiana University at Bloomington have found that chocolate milk effectively helps athletes recover from an intense workout. They had nine cyclists bike, rest four hours, then bike again, three separate times. After each workout, the cyclists downed chocolate milk or energy drinks Gatorade or Endurox (two to three glasses per hour); then, in the second workout of each set, they cycled to exhaustion. When they drank chocolate milk, the amount of time they could cycle until they were exhausted was similar to when they drank Gatorade and longer than when they drank Endurox.
8 6 4 2 0 LR
OR
Based on this figure, has the random assignment of subjects to experimental groups been successful in creating groups that are similar with respect to the ages of the children in the groups? Explain. Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
64
Chapter 2 Collecting Data Sensibly
2.39 In many digital environments, users are allowed to choose how they are represented visually online. Does how people are represented online affect online behavior? This question was examined by the authors of the paper “The Proteus Effect: The Effect of Transformed
Self-Representation on Behavior” (Human Communication Research [2007]: 271–290). Participants were randomly assigned either an attractive avatar (a graphical image that represents a person) to represent them or an unattractive avatar. a. The researchers concluded that when interacting with a person of the opposite gender in an online virtual environment, those assigned an attractive avatar moved significantly closer to the other person than those who had been assigned an unattractive avatar. This difference was attributed to the attractiveness of the avatar. Explain why the researchers would not have been able to reach this conclusion if participants had been allowed to choose one of the two avatars (attractive, unattractive) to represent them online. b. Construct a diagram to represent the underlying structure of this experiment.
2.40 To examine the effect of exercise on body composition, healthy women age 35 to 50 were classified as either active (9 hours or more of physical activity per week) or sedentary (“Effects of Habitual Physical Activ-
ity on the Resting Metabolic Rates and Body Composition of Women aged 35 to 50 Years,” Journal of the American Dietetic Association [2001]: 1181–1191). Percent body fat was measured and the researchers found that percent body fat was significantly lower for women who were active than for sedentary women. a. Is the study described an experiment? If so, what are the explanatory variable and the response variable? If not, explain why it is not an experiment. b. From this study alone, is it reasonable to conclude that physical activity is the cause of the observed difference in body fat percentage? Justify your answer.
2.41 Does playing action video games provide more than just entertainment? The authors of the paper “Action-Video-Game Experience Alters the Spatial Resolution of Vision” (Psychological Science [2007]: 88–94) concluded that spatial resolution, an important aspect of vision, is improved by playing action video games. They based this conclusion on data from an experiment in which 32 volunteers who had not played action video games were “equally and randomly divided between the experimental and control groups.” Subjects in each group Bold exercises answered in back
Data set available online
played a video game for 30 hours over a period of 6 weeks. Those in the experimental group played Unreal Tournament 2004, an action video game. Those in the control group played the game Tetris, a game that does not require the user to process multiple objects at once. Explain why the random assignment to the two groups is an important aspect of this experiment.
2.42 Construct a diagram to represent the subliminal messages experiment of Example 2.5.
2.43 Construct a diagram to represent the gasoline additive experiment described on page 52.
2.44 An advertisement for a sweatshirt that appeared in SkyMall Magazine (a catalog distributed by some airlines) stated the following: “This is not your ordinary hoody! Why? Fact: Research shows that written words on containers of water can influence the water’s structure for better or worse depending on the nature and intent of the word. Fact: The human body is 70% water. What if positive words were printed on the inside of your clothing?” For only $79, you could purchase a hooded sweatshirt that had over 200 positive words (such as hope, gratitude, courage and love) in 15 different languages printed on the inside of the sweatshirt so that you could benefit from being surrounded by these positive words. The reference to the “fact” that written words on containers of water can influence the water’s structure appears to be based on the work of Dr. Masaru Emoto who typed words on paper, pasted the words on bottles of water, and observed how the water reacted to the words by seeing what kind of crystals were formed in the water. He describes several of his experiments in his selfpublished book, The Message from Water. If you were going to interview Dr. Emoto, what questions would you want to ask him about his experiment? 2.45 An experiment was carried out to assess the effect of Sweet Talk, a text messaging support system for patients with diabetes (“A Randomized Controlled Trial of
Sweet Talk,” Diabetic Medicine [2006]: 1332–1338). Participants in the experiment were 92 patients, age 8 to 18, with type I diabetes who had been on conventional insulin treatment for at least one year. Participants were assigned at random to one of three experimental groups: Group 1: continued conventional insulin therapy Group 2: continued conventional insulin therapy with Sweet Talk support Group 3: followed a new intensive insulin therapy with Sweet Talk support Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
2.4
One response variable was a measure of glucose concentration in the blood. There was no significant difference in glucose concentration between groups 1 and 2, but group 3 showed a significant improvement in this measure compared to groups 1 and 2. a. Explain why it is not reasonable to attribute the observed improvement in group 3 compared to group 1 to the use of Sweet Talk, even though subjects were randomly assigned to the three experimental groups. b. How would you modify this experiment so that you could tell if improvement in glucose concentration was attributable to the intensive insulin therapy, the use of Sweet Talk, or a combination of the two? c. Draw a diagram showing the structure of the modified experiment from Part (b).
2.46 The Pew Research Center conducted a study of gender bias. The report “Men or Women: Who is the
Better Leader? A Paradox in Public Attitudes” (www. pewsocialtrends.org, August 28, 2008) describes how the study was conducted: In the experiment, two separate random samples of more than 1000 registered voters were asked to read a profile sent to them online of a hypothetical candidate for U.S. Congress in their district. One random sample of 1161 respondents read a profile of Ann Clark, described as a lawyer, a churchgoer, a member of the local Chamber of Commerce, an environmentalist and a member of the same party as the survey respondent. They were then asked what they liked and didn’t like about her, whether they considered her qualified and whether they Bold exercises answered in back
2.4
Data set available online
More on Experimental Design
65
were inclined to vote for her. There was no indication that this was a survey about gender or gender bias. A second random sample of 1139 registered voters was asked to read a profile of Andrew Clark, who—except for his gender—was identical in every way to Ann Clark. These respondents were then asked the same questions. a. What are the two treatments in this experiment? b. What are the response variables in this experiment? c. Explain why “taking two separate random samples” has the same benefits as random assignment to the two treatments in this experiment.
2.47 Red wine contains flavonol, an antioxidant thought to have beneficial health effects. But to have an effect, the antioxidant must be absorbed into the blood. The article “Red Wine is a Poor Source of Bioavailable
Flavonols in Men” (The Journal of Nutrition [2001]: 745–748) describes a study to investigate three sources of dietary flavonol—red wine, yellow onions, and black tea—to determine the effect of source on absorption. The article included the following statement: We recruited subjects via posters and local newspapers. To ensure that subjects could tolerate the alcohol in the wine, we only allowed men with a consumption of at least seven drinks per week to participate... Throughout the study, the subjects consumed a diet that was low in flavonols. a. What are the three treatments in this experiment? b. What is the response variable? c. What are three extraneous variables that the researchers chose to control in the experiment? Video Solution available
More on Experimental Design The previous section covered basic principles for designing simple comparative experiments—control, blocking, random assignment, and replication. The goal of an experimental design is to provide a method of data collection that (1) minimizes extraneous sources of variability in the response so that any differences in response for various experimental conditions can be more easily assessed and (2) creates experimental groups that are similar with respect to extraneous variables that cannot be controlled either directly or through blocking. In this section, we look at some additional considerations that you may need to think about when planning an experiment.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
66
Chapter 2 Collecting Data Sensibly
Use of a Control Group If the purpose of an experiment is to determine whether some treatment has an effect, it is important to include an experimental group that does not receive the treatment. Such a group is called a control group. The use of a control group allows the experimenter to assess how the response variable behaves when the treatment is not used. This provides a baseline against which the treatment groups can be compared to determine whether the treatment had an effect.
Stockbyte/Getty Images
E X A M P L E 2 . 1 0 Comparing Gasoline Additives Suppose that an engineer wants to know whether a gasoline additive increases fuel efficiency (miles per gallon). Such an experiment might use a single car (to eliminate car-to-car variability) and a sequence of trials in which 1 gallon of gas is put in an empty tank, the car is driven around a racetrack at a constant speed, and the distance traveled on the gallon of gas is recorded. To determine whether the additive increases gas mileage, it would be necessary to include a control group of trials in which distance traveled was measured when gasoline without the additive was used. The trials would be assigned at random to one of the two experimental conditions (additive or no additive). Even though this experiment consists of a sequence of trials all with the same car, random assignment of trials to experimental conditions is still important because there will always be uncontrolled variability. For example, temperature or other environmental conditions might change over the sequence of trials, the physical condition of the car might change slightly from one trial to another, and so on. Random assignment of experimental conditions to trials will tend to even out the effects of these uncontrollable factors.
Although we usually think of a control group as one that receives no treatment, in experiments designed to compare a new treatment to an existing standard treatment, the term control group is sometimes also used to describe the group that receives the current standard treatment. Not all experiments require the use of a control group. For example, many experiments are designed to compare two or more conditions—an experiment to compare density for three different formulations of bar soap or an experiment to determine how oven temperature affects the cooking time of a particular type of cake. However, sometimes a control group is included even when the ultimate goal is to compare two or more different treatments. An experiment with two treatments and no control group might allow us to determine whether there is a difference between the two treatments and even to assess the magnitude of the difference if one exists, but it would not allow us to assess the individual effect of either treatment. For example, without a control group, we might be able to say that there is no difference in the increase in mileage for two different gasoline additives, but we would not be able to tell if this was because both additives increased gas mileage by a similar amount or because neither additive had any effect on gas mileage.
Use of a Placebo In experiments that use human subjects, use of a control group may not be enough to determine whether a treatment really does have an effect. People sometimes respond merely to the power of suggestion! For example, suppose a study designed to determine Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
2.4
More on Experimental Design
67
whether a particular herbal supplement is effective in promoting weight loss uses an experimental group that takes the herbal supplement and a control group that takes nothing. It is possible that those who take the herbal supplement and believe that they are taking something that will help them to lose weight may be more motivated and may unconsciously change their eating behavior or activity level, resulting in weight loss. Although there is debate about the degree to which people respond, many studies have shown that people sometimes respond to treatments with no active ingredients and that they often report that such “treatments” relieve pain or reduce symptoms. So, if an experiment is to enable researchers to determine whether a treatment really has an effect, comparing a treatment group to a control group may not be enough. To address the problem, many experiments use what is called a placebo.
DEFINITION A placebo is something that is identical (in appearance, taste, feel, etc.) to the treatment received by the treatment group, except that it contains no active ingredients. For example, in the herbal supplement experiment, rather than using a control group that received no treatment, the researchers might want to include a placebo group. Individuals in the placebo group would take a pill that looked just like the herbal supplement but did not contain the herb or any other active ingredient. As long as the subjects did not know whether they were taking the herb or the placebo, the placebo group would provide a better basis for comparison and would allow the researchers to determine whether the herbal supplement had any real effect over and above the “placebo effect.”
Single-Blind and Double-Blind Experiments Because people often have their own personal beliefs about the effectiveness of various treatments, it is desirable to conduct experiments in such a way that subjects do not know what treatment they are receiving. For example, in an experiment comparing four different doses of a medication for relief of headache pain, someone who knows that he is receiving the medication at its highest dose may be subconsciously influenced to report a greater degree of headache pain reduction. By ensuring that subjects are not aware of which treatment they receive, we can prevent the subjects’ personal perceptions from influencing the response. An experiment in which subjects do not know what treatment they have received is described as single-blind. Of course, not all experiments can be made single-blind. For example, in an experiment to compare the effect of two different types of exercise on blood pressure, it is not possible for participants to be unaware of whether they are in the swimming group or the jogging group! However, when it is possible, “blinding” the subjects in an experiment is generally a good strategy. In some experiments, someone other than the subject is responsible for measuring the response. To ensure that the person measuring the response does not let personal beliefs influence the way in which the response is recorded, the researchers should make sure that the measurer does not know which treatment was given to any particular individual. For example, in a medical experiment to determine whether a new vaccine reduces the risk of getting the flu, doctors must decide whether a particular individual who is not feeling well actually has the flu or some other unrelated illness. If the doctor knew that a participant with flu-like symptoms had received the new flu vaccine, she might be less likely to determine that the participant had the flu and more likely to interpret the symptoms as being the result of some other illness. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
68
Chapter 2 Collecting Data Sensibly
There are two ways in which blinding might occur in an experiment. One involves blinding the subjects, and the other involves blinding the individuals who measure the response. If subjects do not know which treatment was received and those measuring the response do not know which treatment was given to which subject, the experiment is described as double-blind. If only one of the two types of blinding is present, the experiment is single-blind.
DEFINITION A double-blind experiment is one in which neither the subjects nor the individuals who measure the response know which treatment was received. A single-blind experiment is one in which the subjects do not know which treatment was received but the individuals measuring the response do know which treatment was received, or one in which the subjects do know which treatment was received but the individuals measuring the response do not know which treatment was received.
Experimental Units and Replication An experimental unit is the smallest unit to which a treatment is applied. In the language of experimental design, treatments are assigned at random to experimental units, and replication means that each treatment is applied to more than one experimental unit. Replication is necessary for random assignment to be an effective way to create similar experimental groups and to get a sense of the variability in the values of the response for individuals who receive the same treatment. As we will see in Chapters 9–15, this enables us to use statistical methods to decide whether differences in the responses in different treatment groups can be attributed to the treatment received or whether they can be explained by chance variation (the natural variability seen in the responses to a single treatment). Be careful when designing an experiment to ensure that there is replication. For example, suppose that children in two third-grade classes are available to participate in an experiment to compare two different methods for teaching arithmetic. It might at first seem reasonable to select one class at random to use one method and then assign the other method to the remaining class. But what are the experimental units here? If treatments are randomly assigned to classes, classes are the experimental units. Because only one class is assigned to each treatment, this is an experiment with no replication, even though there are many children in each class. We would not be able to determine whether there was a difference between the two methods based on data from this experiment, because we would have only one observation per treatment. One last note on replication: Do not confuse replication in an experimental design with replicating an experiment. Replicating an experiment means conducting a new experiment using the same experimental design as a previous experiment; it is a way of confirming conclusions based on a previous experiment, but it does not eliminate the need for replication in each of the individual experiments themselves.
Using Volunteers as Subjects in an Experiment Although the use of volunteers in a study that involves collecting data through sampling is never a good idea, it is a common practice to use volunteers as subjects in an experiment. Even though the use of volunteers limits the researcher’s ability to generalize to a larger population, random assignment of the volunteers to treatments should result in comparable groups, and so treatment effects can still be assessed. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
2.4
More on Experimental Design
69
E X E RC I S E S 2 . 4 8 - 2 . 5 9 2.48 Explain why some studies include both a control group and a placebo treatment. What additional comparisons are possible if both a control group and a placebo group are included?
2.49 Explain why blinding is a reasonable strategy in many experiments. 2.50 Give an example of an experiment for each of the following: a. Single-blind experiment with the subjects blinded b. Single-blind experiment with the individuals measuring the response blinded c. Double-blind experiment d. An experiment for which it is not possible to blind the subjects Swedish researchers concluded that viewing and discussing art soothes the soul and helps relieve medical conditions such as high blood pressure and constipation (AFP International News Agency, October 14, 2005). This conclusion was based on a study in which 20 elderly women gathered once a week to discuss different works of art. The study also included a control group of 20 elderly women who met once a week to discuss their hobbies and interests. At the end of 4 months, the art discussion group was found to have a more positive attitude, to have lower blood pressure, and to use fewer laxatives than the control group. a. Why would it be important to determine if the researchers assigned the women participating in the study at random to one of the two groups? b. Explain why you think that the researchers included a control group in this study.
2.51
2.52 In an experiment to compare two different surgical procedures for hernia repair (“A Single-Blinded, Ran-
domized Comparison of Laparoscopic Versus Open Hernia Repair in Children,” Pediatrics [2009]: 332– 336), 89 children were assigned at random to one of the two surgical methods. The methods studied were laparoscopic repair and open repair. In laparoscopic repair, three small incisions are made and the surgeon works through these incisions with the aid of a small camera that is inserted through one of the incisions. In the open repair, a larger incision is used to open the abdomen. One of the response variables in this study was the amount of medication that was given after the surgery for the control of pain and nausea. The paper states “For Bold exercises answered in back
Data set available online
postoperative pain, rescue fentanyl (1 mg/kg) and for nausea, ondansetron (0.1 mg/kg) were given as judged necessary by the attending nurse blinded to the operative approach.” a. Why do you think it was important that the nurse who administered the medications did not know which type of surgery was performed? b. Explain why it was not possible for this experiment to be double-blind.
2.53 The article “Placebos Are Getting More Effective. Drug Makers Are Desperate to Know Why.” (Wired Magazine, August 8, 2009) states that “according to research, the color of a tablet can boost the effectiveness even of genuine meds—or help convince a patient that a placebo is a potent remedy.” Describe how you would design an experiment to investigate if adding color to Tylenol tablets would result in greater perceived pain relief. Be sure to address how you would select subjects, how you would measure pain relief, what colors you would use, and whether or not you would include a control group in your experiment.
2.54 A novel alternative medical treatment for heart attacks seeds the damaged heart muscle with cells from the patient’s thigh muscle (“Doctors Mend Damaged
Hearts with Cells from Muscles,” San Luis Obispo Tribune, November 18, 2002). Doctor Dib from the Arizona Heart Institute evaluated the approach on 16 patients with severe heart failure. The article states that “ordinarily, the heart pushes out more than half its blood with each beat. Dib’s patients had such severe heart failure that their hearts pumped just 23 percent. After bypass surgery and cell injections, this improved to 36 percent, although it was impossible to say how much, if any, of the new strength resulted from the extra cells.” a. Explain why it is not reasonable to generalize to the population of all heart attack victims based on the data from these 16 patients. b. Explain why it is not possible to say whether any of the observed improvement was due to the cell injections, based on the results of this study. c. Describe a design for an experiment that would allow researchers to determine whether bypass surgery plus cell injections was more effective than bypass surgery alone.
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
70
Chapter 2 Collecting Data Sensibly
2.55
The article “Doctor Dogs Diagnose Cancer by
Sniffing It Out” (Knight Ridder Newspapers, January 9, 2006) reports the results of an experiment described in the journal Integrative Cancer Therapies. In this experiment, dogs were trained to distinguish between people with breast and lung cancer and people without cancer by sniffing exhaled breath. Dogs were trained to lay down if they detected cancer in a breath sample. After training, dogs’ ability to detect cancer was tested using breath samples from people whose breath had not been used in training the dogs. The paper states “The researchers blinded both the dog handlers and the experimental observers to the identity of the breath samples.” Explain why this blinding is an important aspect of the design of this experiment.
2.56 An experiment to evaluate whether vitamins can help prevent recurrence of blocked arteries in patients who have had surgery to clear blocked arteries was described in the article “Vitamins Found to Help Prevent Blocked Arteries” (Associated Press, September 1, 2002). The study involved 205 patients who were given either a treatment consisting of a combination of folic acid, vitamin B12, and vitamin B6 or a placebo for 6 months. a. Explain why a placebo group was used in this experiment. b. Explain why it would be important for the researchers to have assigned the 205 subjects to the two groups (vitamin and placebo) at random. c. Do you think it is appropriate to generalize the results of this experiment to the population of all patients who have undergone surgery to clear blocked arteries? Explain.
2.57 Pismo Beach, California, has an annual clam festival that includes a clam chowder contest. Judges rate clam chowders from local restaurants, and the judging is done in such a way that the judges are not aware of which chowder is from which restaurant. One year, much to the dismay of the seafood restaurants on the waterfront, Denny’s chowder was declared the winner! (When asked what the ingredients were, the cook at Bold exercises answered in back
2.5
Data set available online
Denny’s said he wasn’t sure—he just had to add the right amount of nondairy creamer to the soup stock that he got from Denny’s distribution center!) a. Do you think that Denny’s chowder would have won the contest if the judging had not been “blind?” Explain. b. Although this was not an experiment, your answer to Part (a) helps to explain why those measuring the response in an experiment are often blinded. Using your answer in Part (a), explain why experiments are often blinded in this way.
2.58 The San Luis Obispo Tribune (May 7, 2002) reported that “a new analysis has found that in the majority of trials conducted by drug companies in recent decades, sugar pills have done as well as—or better than—antidepressants.” What effect is being described here? What does this imply about the design of experiments with a goal of evaluating the effectiveness of a new medication?
2.59 The article “A Debate in the Dentist’s Chair” (San Luis Obispo Tribune, January 28, 2000) described an ongoing debate over whether newer resin fillings are a better alternative to the more traditional silver amalgam fillings. Because amalgam fillings contain mercury, there is concern that they could be mildly toxic and prove to be a health risk to those with some types of immune and kidney disorders. One experiment described in the article used sheep as subjects and reported that sheep treated with amalgam fillings had impaired kidney function. a. In the experiment, a control group of sheep that received no fillings was used but there was no placebo group. Explain why it is not necessary to have a placebo group in this experiment. b. The experiment compared only an amalgam filling treatment group to a control group. What would be the benefit of also including a resin filling treatment group in the experiment? c. Why do you think the experimenters used sheep rather than human subjects? Video Solution available
More on Observational Studies: Designing Surveys (Optional) Designing an observational study to compare two populations on the basis of some easily measured characteristic is relatively straightforward, with attention focusing on choosing a reasonable method of sample selection. However, many observational
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
2.5 More on Observational Studies: Designing Surveys (Optional)
71
studies attempt to measure personal opinion or attitudes using responses to a survey. In such studies, both the sampling method and the design of the survey itself are critical to obtaining reliable information. At first glance it might seem that a survey is a simple method for acquiring information. However, it turns out that designing and administering a survey is not an easy task. Great care must be taken in order to obtain good information from a survey.
Survey Basics A survey is a voluntary encounter between strangers in which an interviewer seeks information from a respondent by engaging in a special type of conversation. This conversation might take place in person, over the telephone, or even in the form of a written questionnaire, and it is quite different from usual social conversations. Both the interviewer and the respondent have certain roles and responsibilities. The interviewer gets to decide what is relevant to the conversation and may ask questions— possibly personal or even embarrassing questions. The respondent, in turn, may refuse to participate in the conversation and may refuse to answer any particular question. But having agreed to participate in the survey, the respondent is responsible for answering the questions truthfully. Let’s consider the situation of the respondent.
The Respondent’s Tasks Understanding of the survey process has been improved in the past two decades by contributions from the field of psychology, but there is still much uncertainty about how people respond to survey questions. Survey researchers and psychologists generally agree that the respondent is confronted with a sequence of tasks when asked a question: comprehension of the question, retrieval of information from memory, and reporting the response.
Task 1: Comprehension Comprehension is the single most important task facing the respondent, and fortunately it is the characteristic of a survey question that is most easily controlled by the question writer. Understandable directions and questions are characterized by (1) a vocabulary appropriate to the population of interest, (2) simple sentence structure, and (3) little or no ambiguity. Vocabulary is often a problem. As a rule, it is best to use the simplest possible word that can be used without sacrificing clear meaning. Simple sentence structure also makes it easier for the respondent to understand the question. A famous example of difficult syntax occurred in 1993 when the Roper organization created a survey related to the Holocaust. One question in this survey was
“Does it seem possible or does it seem impossible to you that the Nazi extermination of the Jews never happened?” The question has a complicated structure and a double negative—“impossible . . . never happened”—that could lead respondents to give an answer opposite to what they actually believed. The question was rewritten and given a year later in an otherwise unchanged survey: “Does it seem possible to you that the Nazi extermination of the Jews never happened, or do you feel certain that it happened?” This question wording is much clearer, and in fact the respondents’ answers were quite different, as shown in the following table (the “unsure” and “no opinion” percentages have been omitted): Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
72
Chapter 2 Collecting Data Sensibly
Original Roper Poll Impossible Possible
65% 12%
Revised Roper Poll Certain it happened Possible it never happened
91% 1%
It is also important to filter out ambiguity in questions. Even the most innocent and seemingly clear questions can have a number of possible interpretations. For example, suppose that you are asked, “When did you move to Cedar Rapids?” This would seem to be an unambiguous question, but some possible answers might be (1) “In 1971,” (2) “When I was 23,” and (3) “In the summer.” The respondent must decide which of these three answers, if any, is the appropriate response. It may be possible to lessen the ambiguity with more precise questions: 1. In what year did you move to Cedar Rapids? 2. How old were you when you moved to Cedar Rapids? 3. In what season of the year did you move to Cedar Rapids? One way to find out whether or not a question is ambiguous is to field-test the question and to ask the respondents if they were unsure how to answer a question. Ambiguity can also arise from the placement of questions as well as from their phrasing. Here is an example of ambiguity uncovered when the order of two questions differed in two versions of a survey on happiness. The questions were 1. Taken altogether, how would you say things are these days: Would you say that you are very happy, pretty happy, or not too happy? 2. Taking things altogether, how would you describe your marriage: Would you say that your marriage is very happy, pretty happy, or not too happy?
The proportions of responses to the general happiness question differed for the different question orders, as follows:
Response to General Happiness Question
Very happy Pretty happy Not too happy
General Asked First
General Asked Second
52.4% 44.2% 3.4%
38.1% 52.8% 9.1%
If the goal in this survey was to estimate the proportion of the population that is generally happy, these numbers are quite troubling—they cannot both be right! What seems to have happened is that Question 1 was interpreted differently depending on whether it was asked first or second. When the general happiness question was asked after the marital happiness question, the respondents apparently interpreted it to be asking about their happiness in all aspects of their lives except their marriage. This was a reasonable interpretation, given that they had just been asked about their marital happiness, but it is a different interpretation than when the general happiness question was asked first. The troubling lesson here is that even carefully worded questions can have different interpretations in the context of the rest of the survey. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
2.5 More on Observational Studies: Designing Surveys (Optional)
73
Task 2: Retrieval from Memory Retrieving relevant information from memory to answer the question is not always an easy task, and it is not a problem limited to questions of fact. For example, consider this seemingly elementary “factual” question: How many times in the past 5 years did you visit your dentist’s office? a. b. c. d. e.
0 times 1–5 times 6–10 times 11–15 times more than 15 times
It is unlikely that many people will remember with clarity every single visit to the dentist in the past 5 years. But generally, people will respond to such a question with answers consistent with the memories and facts they are able to reconstruct given the time they have to respond to the question. An individual may, for example, have a sense that he usually makes about two trips a year to the dentist’s office, so he may extrapolate the typical year and get 10 times in 5 years. Then there may be three particularly memorable visits, say, for a root canal in the middle of winter. Thus, the best recollection is now 13, and the respondent will choose Answer (d), 11–15 times. Perhaps not exactly correct, but the best that can be reported under the circumstances. What are the implications of this relatively fuzzy memory for those who construct surveys about facts? First, the investigator should understand that most factual answers are going to be approximations of the truth. Second, events closer to the time of a survey are easier to recall. Attitude and opinion questions can also be affected in significant ways by the respondent’s memory of recently asked questions. For example, one study contained a survey question asking respondents their opinion about how much they followed politics. When that question was preceded by a factual question asking whether they knew the name of the congressional representative from their district, the percentage who reported they follow politics “now and then” or “hardly ever” jumped from 21% to 39%! Respondents apparently concluded that, because they didn’t know the answer to the previous knowledge question, they must not follow politics as much as they might have thought otherwise. In a survey that asks for an opinion about the degree to which the respondent believes drilling for oil should be permitted in national parks, the response might be different if the question is preceded by questions about the high price of gasoline than if the question is preceded by questions about the environment.
Task 3: Reporting the Response The task of formulating and reporting a response can be influenced by the social aspects of the survey conversation. In general, if a respondent agrees to take a survey, he or she will be motivated to answer truthfully. Therefore, if the questions are not too difficult (taxing the respondent’s knowledge or memory) and if there are not too many questions (taxing the respondent’s patience), the answers to questions will be reasonably accurate. However, it is also true that the respondents often wish to present themselves in a favorable light. This desire leads to what is known as a social desirability bias. Sometimes this bias is a response to the particular wording in a question. In 1941, the following questions were analyzed in two different forms of a survey (emphasis added): 1. Do you think the United States should forbid public speeches against democracy? 2. Do you think the United States should allow public speeches against democracy?
It would seem logical that these questions are opposites and that the proportion who would not allow public speeches against democracy should be equal to the proportion who would forbid public speeches against democracy. But only 45% of those respondents offering an opinion on Question 1 thought the United States should “forbid,” Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
74
Chapter 2 Collecting Data Sensibly
whereas 75% of the respondents offering an opinion on Question 2 thought the United States should “not allow” public speeches against democracy. Most likely, respondents reacted negatively to the word forbid, as forbidding something sounds much harsher than not allowing it. Some survey questions may be sensitive or threatening, such as questions about sex, drugs, or potentially illegal behavior. In this situation, a respondent not only will want to present a positive image but also will certainly think twice about admitting illegal behavior! In such cases, the respondent may shade the actual truth or may even lie about particular activities and behaviors. In addition, the tendency toward positive presentation is not limited to obviously sensitive questions. For example, consider the question about general happiness previously described. Several investigators have reported higher happiness scores in face-to-face interviews than in responses to a mailed questionnaire. Presumably, a happy face presents a more positive image of the respondent to the interviewer. On the other hand, if the interviewer was a clearly unhappy person, a respondent might shade answers to the less happy side of the scale, perhaps thinking that it is inappropriate to report happiness in such a situation. It is clear that constructing surveys and writing survey questions can be a daunting task. Keep in mind the following three things: 1. Questions should be understandable by the individuals in the population being surveyed. Vocabulary should be at an appropriate level, and sentence structure should be simple. 2. Questions should, as much as possible, recognize that human memory is fickle. Questions that are specific will aid the respondent by providing better memory cues. The limitations of memory should be kept in mind when interpreting the respondent’s answers. 3. As much as possible, questions should not create opportunities for the respondent to feel threatened or embarrassed. In such cases respondents may introduce a social desirability bias, the degree of which is unknown to the interviewer. This can compromise conclusions drawn from the survey data.
Constructing good surveys is a difficult task, and we have given only a brief introduction to this topic. For a more comprehensive treatment, we recommend the book by Sudman and Bradburn listed in the references in the back of the book.
EX E RC I S E S 2 . 6 0 - 2 . 6 5 2.60 A tropical forest survey conducted by Conservation International included the following statements in the material that accompanied the survey: “A massive change is burning its way through the earth’s environment.” “The band of tropical forests that encircle the earth is being cut and burned to the ground at an alarming rate.” “Never in history has mankind inflicted such sweeping changes on our planet as the clearing of rain forest taking place right now!”
Bold exercises answered in back
Data set available online
The survey that followed included the questions given in Parts (a)–(d) below. For each of these questions, identify a word or phrase that might affect the response and possibly bias the results of any analysis of the responses. a. “Did you know that the world’s tropical forests are being destroyed at the rate of 80 acres per minute?” b. “Considering what you know about vanishing tropical forests, how would you rate the problem?” c. “Do you think we have an obligation to prevent the man-made extinction of animal and plant species?” d. “Based on what you know now, do you think there is a link between the destruction of tropical forests and changes in the earth’s atmosphere?” Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
2.5 More on Observational Studies: Designing Surveys (Optional)
2.61 Fast-paced lifestyles, in which students balance the requirements of school, after-school activities, and jobs, are thought by some to lead to reduced sleep. Suppose that you are assigned the task of designing a survey that will provide answers to the accompanying questions. Write a set of survey questions that might be used. In some cases, you may need to write more than one question to adequately address a particular issue. For example, responses might be different for weekends and school nights. You may also have to define some terms to make the questions understandable to the target audience, which is adolescents. Topics to be addressed: How much sleep do the respondents get? Is this enough sleep? Does sleepiness interfere with schoolwork? If they could change the starting and ending times of the school day, what would they suggest? (Sorry, they cannot reduce the total time spent in school during the day!)
2.62 Asthma is a chronic lung condition characterized by difficulty in breathing. Some studies have suggested that asthma may be related to childhood exposure to some animals, especially dogs and cats, during the first year of life (“Exposure to Dogs and
Cats in the First Year of Life and Risk of Allergic Sensitization at 6 to 7 Years of Age,” Journal of the American Medical Association [2002]: 963–972). Some environmental factors that trigger an asthmatic response are (1) cold air, (2) dust, (3) strong fumes, and (4) inhaled irritants. a. Write a set of questions that could be used in a survey to be given to parents of young children suffering from asthma. The survey should include questions about the presence of pets in the first year of the child’s life as well as questions about the presence of pets today. Also, the survey should include questions that address the four mentioned household environmental factors. b. It is generally thought that low-income persons, who tend to be less well educated, have homes in environments where the four environmental factors are present. Mindful of the importance of comprehension, can you improve the questions in Part (a) by making your vocabulary simpler or by changing the wording of the questions?
Bold exercises answered in back
Data set available online
75
c. One problem with the pet-related questions is the reliance on memory. That is, parents may not actually remember when they got their pets. How might you check the parents’ memories about these pets?
2.63 In national surveys, parents consistently point to school safety as an important concern. One source of violence in junior high schools is fighting (“Self-Reported Characterization of Seventh-Grade Student Fights,” Journal of Adolescent Health [1998]: 103–109). To construct a knowledge base about student fights, a school administrator wants to give two surveys to students after fights are broken up. One of the surveys is to be given to the participants, and the other is to be given to students who witnessed the fight. The type of information desired includes (1) the cause of the fight, (2) whether or not the fight was a continuation of a previous fight, (3) whether drugs or alcohol was a factor, (4) whether or not the fight was gang related, and (5) the role of bystanders. a. Write a set of questions that could be used in the two surveys. Each question should include a set of possible responses. For each question, indicate whether it would be used on both surveys or just on one of the two. b. How might the tendency toward positive selfpresentation affect the responses of the fighter to the survey questions you wrote for Part (a)? c. How might the tendency toward positive selfpresentation affect the responses of a bystander to the survey questions you wrote for Part (a)?
2.64 Doctors have expressed concern about young women drinking large amounts of soda and about their decreased consumption of milk (“Teenaged Girls, Car-
bonated Beverage Consumption, and Bone Fractures,” Archives of Pediatric and Adolescent Medicine [2000]: 610–613). In parts (a)–(d), construct two questions that might be included in a survey of teenage girls. Each question should include possible responses from which the respondent can select. (Note: The questions as written are vague. Your task is to clarify the questions for use in a survey, not just to change the syntax!) a. How much “cola” beverage does the respondent consume? b. How much milk (and milk products) is consumed by the respondent? c. How physically active is the respondent? d. What is the respondent’s history of bone fractures?
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
76
Chapter 2 Collecting Data Sensibly
2.65 A survey described in the paper “The Adolescent Health Review: A Brief Multidimensional Screening Instrument” (Journal of Adolescent Health [2001]:131–139) attempted to address psychosocial factors thought to be of importance in preventive health care for adolescents. For each risk area in the following list, construct a question that would be comprehensible to students in grades 9–12 and that would provide information about the risk factor. Bold exercises answered in back
2.6
Data set available online
Make your questions multiple-choice, and provide possible responses. a. Lack of exercise b. Poor nutrition c. Emotional distress d. Sexual activity e. Cigarette smoking f. Alcohol use Video Solution available
Interpreting and Communicating the Results of Statistical Analyses Statistical studies are conducted to allow investigators to answer questions about characteristics of some population of interest or about the effect of some treatment. Such questions are answered on the basis of data, and how the data are obtained determines the quality of information available and the type of conclusions that can be drawn. As a consequence, when describing a study you have conducted (or when evaluating a published study), you must consider how the data were collected. The description of the data collection process should make it clear whether the study is an observational study or an experiment. For observational studies, some of the issues that should be addressed are: 1. What is the population of interest? What is the sampled population? Are these two populations the same? If the sampled population is only a subset of the population of interest, undercoverage limits our ability to generalize to the population of interest. For example, if the population of interest is all students at a particular university, but the sample is selected from only those students who choose to list their phone number in the campus directory, undercoverage may be a problem. We would need to think carefully about whether it is reasonable to consider the sample as representative of the population of all students at the university. Overcoverage results when the sampled population is actually larger than the population of interest. This would be the case if we were interested in the population of all high schools that offer Advanced Placement (AP) Statistics but sampled from a list of all schools that offered an AP class in any subject. Both undercoverage and overcoverage can be problematic. 2. How were the individuals or objects in the sample actually selected? A description of the sampling method helps the reader to make judgments about whether the sample can reasonably be viewed as representative of the population of interest. 3. What are potential sources of bias, and is it likely that any of these will have a substantial effect on the observed results? When describing an observational study, you should acknowledge that you are aware of potential sources of bias and explain any steps that were taken to minimize their effect. For example, in a mail survey, nonresponse can be a problem, but the sampling plan may seek to minimize its effect by offering incentives for participation and by following up one or more times with those who do not respond to the first request. A common misperception is that increasing the sample size is a way to reduce bias in observational studies, but this is not the case. For example, if measurement bias is
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
2.6 Interpreting and Communicating the Results of Statistical Analyses
77
present, as in the case of a scale that is not correctly calibrated and tends to weigh too high, taking 1000 measurements rather than 100 measurements cannot correct for the fact that the measured weights will be too large. Similarly, a larger sample size cannot compensate for response bias introduced by a poorly worded question. For experiments, some of the issues that should be addressed are: 1. What is the role of random assignment? All good experiments use random assignment as a means of coping with the effects of potentially confounding variables that cannot easily be directly controlled. When describing an experimental design, you should be clear about how random assignment (subjects to treatments, treatments to subjects, or treatments to trials) was incorporated into the design. 2. Were any extraneous variables directly controlled by holding them at fixed values throughout the experiment? If so, which ones and at which values? 3. Was blocking used? If so, how were the blocks created? If an experiment uses blocking to create groups of homogeneous experimental units, you should describe the criteria used to create the blocks and their rationale. For example, you might say something like “Subjects were divided into two blocks—those who exercise regularly and those who do not exercise regularly—because it was believed that exercise status might affect the responses to the diets.” Because each treatment appears at least once in each block, the block size must be at least as large as the number of treatments. Ideally, the block sizes should be equal to the number of treatments, because this presumably would allow the experimenter to create small groups of extremely homogeneous experimental units. For example, in an experiment to compare two methods for teaching calculus to first-year college students, we may want to block on previous mathematics knowledge by using math SAT scores. If 100 students are available as subjects for this experiment, rather than creating two large groups (above-average math SAT score and below-average math SAT score), we might want to create 50 blocks of two students each, the first consisting of the two students with the highest math SAT scores, the second containing the two students with the next highest scores, and so on. We would then select one student in each block at random and assign that student to teaching method 1. The other student in the block would be assigned to teaching method 2.
A Word to the Wise: Cautions and Limitations It is a big mistake to begin collecting data before thinking carefully about research objectives and developing a plan. A poorly designed plan for data collection may result in data that do not enable the researcher to answer key questions of interest or to generalize conclusions based on the data to the desired populations of interest. Clearly defining the objectives at the outset enables the investigator to determine whether an experiment or an observational study is the best way to proceed. Watch out for the following inappropriate actions: 1. Drawing a cause-and-effect conclusion from an observational study. Don’t do this, and don’t believe it when others do it! 2. Generalizing results of an experiment that uses volunteers as subjects to a larger population. This is not sensible without a convincing argument that the group of volunteers can reasonably be considered a representative sample from the population. 3. Generalizing conclusions based on data from a sample to some population of interest. This is sometimes a sensible thing to do, but on other occasions it is not Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
78
Chapter 2 Collecting Data Sensibly
reasonable. Generalizing from a sample to a population is justified only when there is reason to believe that the sample is likely to be representative of the population. This would be the case if the sample was a random sample from the population and there were no major potential sources of bias. If the sample was not selected at random or if potential sources of bias were present, these issues would have to be addressed before a judgment could be made regarding the appropriateness of generalizing the study results. For example, the Associated Press (January 25, 2003) reported on the high cost of housing in California. The median home price was given for each of the 10 counties in California with the highest home prices. Although these 10 counties are a sample of the counties in California, they were not randomly selected and (because they are the 10 counties with the highest home prices) it would not be reasonable to generalize to all California counties based on data from this sample. 4. Generalizing conclusions based on an observational study that used voluntary response or convenience sampling to a larger population. This is almost never reasonable.
EX E RC I S E S 2 . 6 6 - 2 . 6 9 2.66 The following paragraph appeared in USA Today (August 6, 2009): Cement doesn’t hold up to scrutiny A common treatment that uses medical cement to fix cracks in the spinal bones of elderly people worked no better than a sham treatment, the first rigorous studies of a popular procedure reveal. Pain and disability were virtually the same up to six months later, whether patients had a real treatment or a fake one, shows the research in today’s New England Journal of Medicine. Tens of thousands of Americans each year are treated with bone cement, especially older women with osteoporosis. The researchers said it is yet another example of a procedure coming into wide use before proven safe and effective. Medicare pays $1,500 to $2,100 for the outpatient procedure. The paper referenced in this paragraph is “A Randomized Trial of Vertebroplasty for Painful Osteoporotic Vertebral Fractures” (New England Journal of Medicine [2009]: 557–568). Obtain a copy of this paper through your university library or your instructor. Read the following sections of the paper: the abstract on page 557; the study design section on page 558; the participants section on pages 558–559; the outcome assessment section on pages 559–560; and the discussion section that begins on page 564. Bold exercises answered in back
Data set available online
The summary of this study that appeared in USA Today consisted of just one paragraph. If the newspaper had allowed four paragraphs, other important aspects of the study could have been included. Write a fourparagraph summary that the paper could have used. Remember—you are writing for the USA Today audience, not for the readers of the New England Journal of Medicine!
2.67 The article “Effects of Too Much TV Can Be Undone” (USA Today, October 1, 2007) included the following paragraph: Researchers at Johns Hopkins Bloomberg School of Public Health report that it’s not only how many hours children spend in front of the TV, but at what age they watch that matters. They analyzed data from a national survey in which parents of 2707 children were interviewed first when the children were 30–33 months old and again when they were 5 12, about their TV viewing and their behavior. a. Is the study described an observational study or an experiment? b. The article says that data from a sample of 2707 parents were used in the study. What other information about the sample would you want in order to evaluate the study?
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
Activities
c. The actual paper referred to by the USA Today article was “Children’s Television Exposure and Behavioral
and Social Outcomes at 5.5 years: Does Timing of Exposure Matter?” (Pediatrics [2007]: 762–769). The paper describes the sample as follows: The study sample included 2707 children whose mothers completed telephone interviews at both 30 to 33 months and 5.5 years and reported television exposure at both time points. Of those completing both interviewers, 41 children (1%) were excluded because of missing data on television exposure at one or both time points. Compared with those enrolled in the HS clinical trial, parents in the study sample were disproportionately older, white, more educate, and married. The “HS clinical trial” referred to in the excerpt from the paper was a nationally representative sample used in the Healthy Steps for Young Children national evaluation. Based on the above description of the study sample, do you think that it is reasonable to regard the sample as representative of parents of all children at age 5.5 years? Explain. d. The USA Today article also includes the following summary paragraph: The study did not examine what the children watched and can’t show TV was the cause of later problems, but it does “tell parents that even if kids are watching TV early in life, and they stop, it could reduce the risk for behavioral and social problems later,” Mistry says. Bold exercises answered in back
A C TI V I T Y 2 . 1
79
What potentially confounding variable is identified in this passage? e. The passage in Part (d) says that the study cannot show that TV was the cause of later problems. Is the quote from Kamila Mistry (one of the study authors) in the passage consistent with the statement about cause? Explain.
2.68 The short article “Developing Science-Based Food and Nutrition Information” (Journal of the American Dietetic Association [2001]: 1144–1145) includes some guidelines for evaluating a research paper. Obtain a copy of this paper through your university library or your instructor. Read this article and make a list of questions that can be used to evaluate a research study.
2.69 An article titled “I Said, Not While You Study: Science Suggests Kids Can’t Study and Groove at the Same Time” appeared in the Washington Post (September 5, 2006). This provides an example of a reporter summarizing the result of a scientific study in a way that is designed to make it accessible to the newspaper’s readers. You can find the newspaper article online by searching on the title or by going to http://www.washingtonpost.com/ wp-dyn/content/article/2006/09/03/AR2006090300592 .html. The study referenced in the newspaper article was published in the Proceedings of the National Academies of Science and can be found at http://www.pnas.org/ content/103/31/11778.full. Read the newspaper article and then take a look at the published paper. Comment on whether you think that the author was successful in communicating the findings of the study to the intended audience.
Data set available online
Video Solution available
Facebook Friending
Background: The article “Professors Prefer Face Time to Facebook” appeared in the student newspaper at Cal Poly, San Luis Obispo (Mustang Daily, August 27, 2009). The article examines how professors and students felt about using Facebook as a means of faculty-student communication. The student who wrote this article got mixed opinions when she interviewed students to ask whether they wanted to become Facebook friends with their professors. Two student comments included in the article were “I think the younger the professor is, the more you can relate to them and the less awkward it would be if you were to become friends on Facebook. The
older the professor, you just would have to wonder, ‘Why are they friending me?’” and “I think becoming friends with professors on Facebook is really awkward. I don’t want them being able to see into my personal life, and frankly, I am not really interested in what my professors do in their free time.” Even if the students interviewed had expressed a consistent opinion, it would still be unreasonable to think this represented general student opinion on this issue because only four students were interviewed and it is not clear from the article how these students were selected.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
80
Chapter 2 Collecting Data Sensibly
In this activity, you will work with a partner to develop a plan to assess student opinion about being Facebook friends with professors at your school. 1. Suppose you will select a sample of 50 students at your school to participate in a survey. Write one or more questions that you would ask each student in the sample. 2. Discuss with your partner whether you think it would be easy or difficult to obtain a simple random sample of 50 students at your school and to obtain the desired information from all the students selected for the sample. Write a summary of your discussion. 3. With your partner, decide how you might go about selecting a sample of 50 students from your school
AC TI V I TY 2 . 2
An Experiment to Test for the Stroop Effect
Background: In 1935, John Stroop published the results of his research into how people respond when presented with conflicting signals. Stroop noted that most people are able to read words quickly and that they cannot easily ignore them and focus on other attributes of a printed word, such as text color. For example, consider the following list of words: green
blue
red
blue
yellow
red
It is easy to quickly read this list of words. It is also easy to read the words even if the words are printed in color, and even if the text color is different from the color of the word. For example, people can read the words in the list green
blue
that reasonably could be considered representative of the population of interest even if it may not be a simple random sample. Write a brief description of your sampling plan, and point out the aspects of your plan that you think make it reasonable to argue that it will be representative. 4. Explain your plan to another pair of students. Ask them to critique your plan. Write a brief summary of the comments you received. Now reverse roles, and provide a critique of the plan devised by the other pair. 5. Based on the feedback you received in Step 4, would you modify your original sampling plan? If not, explain why this is not necessary. If so, describe how the plan would be modified.
red
blue
yellow
red
as quickly as they can read the list that isn’t printed in color. However, Stroop found that if people are asked to name the text colors of the words in the list (red, yellow, blue, green, red, green), it takes them longer. Psychologists believe that this is because the reader has to inhibit a natural response (reading the word) and produce a different response (naming the color of the text).
If Stroop is correct, people should be able to name colors more quickly if they do not have to inhibit the word response, as would be the case if they were shown the following:
1. Design an experiment to compare times to identify colors when they appear as text to times to identify colors when there is no need to inhibit a word response. Indicate how random assignment is incorporated into your design. What is your response variable? How will you measure it? How many subjects will you use in your experiment, and how will they be chosen? 2. When you are satisfied with your experimental design, carry out the experiment. You will need to construct your list of colored words and a corresponding list of colored bars to use in the experiment. You will also need to think about how you will implement the random assignment scheme. 3. Summarize the resulting data in a brief report that explains whether your findings are consistent with the Stroop effect.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
Activities
A C TI V I T Y 2 . 3
McDonald’s and the Next 100 Billion Burgers
Background: The article “Potential Effects of the Next
100 Billion Hamburgers Sold by McDonald’s” (American Journal of Preventative Medicine [2005]: 379–381) estimated that 992.25 million pounds of saturated fat would be consumed as McDonald’s sells its next 100 billion hamburgers. This estimate was based on the assumption that the average weight of a burger sold would be 2.4 oz. This is the average of the weight of a regular hamburger (1.6 oz.) and a Big Mac (3.2 oz.). The authors took this approach because McDonald’s does not publish sales and profits of individual items. Thus, it is not possible to estimate how many of McDonald’s first 100 billion beef burgers sold were 1.6 oz hamburgers, 3.2 oz. Big Macs (introduced in 1968), 4.0 oz. Quarter Pounders (introduced in 1973), or other sandwiches.
A C TI V I T Y 2 . 4
81
This activity can be completed as an individual or as a team. Your instructor will specify which approach (individual or team) you should use. 1. The authors of the article believe that the use of 2.4 oz. as the average size of a burger sold at McDonald’s is “conservative,” which would result in the estimate of 992.25 million pounds of saturated fat being lower than the actual amount that would be consumed. Explain why the authors’ belief might be justified. 2. Do you think it would be possible to collect data that could lead to an estimate of the average burger size that would be better than 2.4 oz.? If so, explain how you would recommend collecting such data. If not, explain why you think it is not possible.
Video Games and Pain Management
Background: Video games have been used for pain management by doctors and therapists who believe that the attention required to play a video game can distract the player and thereby decrease the sensation of pain. The paper “Video Games and Health” (British Medical Journal [2005]:122–123) states However, there has been no long term follow-up and no robust randomized controlled trials of such interventions. Whether patients eventually tire of such games is also unclear. Furthermore, it is not known whether any distracting effect depends simply on concentrating on an interactive task or whether the content of games is also an important factor as there have been no controlled trials comparing video games with other distracters. Further research should examine factors within games such as novelty, users’ preferences, and relative levels of challenge and should compare video games with other potentially distracting activities.
1. Working with a partner, select one of the areas of potential research suggested in the passage from the paper and formulate a specific question that could be addressed by performing an experiment. 2. Propose an experiment that would provide data to address the question from Step 1. Be specific about how subjects might be selected, what the experimental conditions (treatments) would be, and what response would be measured. 3. At the end of Section 2.3 there are 10 questions that can be used to evaluate an experimental design. Answer these 10 questions for the design proposed in Step 2. 4. After evaluating your proposed design, are there any changes you would like to make to your design? Explain.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
82
Chapter 2 Collecting Data Sensibly
AC TI V I TY 2 . 5
Be Careful with Random Assignment!
When individuals climb to high altitudes, a condition known as acute mountain sickness (AMS) may occur. AMS is brought about by a combination of reduced air pressure and lower oxygen concentration that occurs at high altitudes. Two standard treatments for AMS are a medication, acetazolamide (which stimulates breathing and reduces mild symptoms) and the use of portable hyperbaric chambers. With increasing numbers of younger inexperienced mountaineers, it is important to re-evaluate these treatments for the 12 to 14 year age group. An experimental plan under consideration is to study the first 18 youngsters diagnosed with AMS at a high altitude park ranger station whose parents consent to participation in the experiment. Equal numbers of each treatment are desired and the researchers are considering the following strategy for random assignment of treatments: Assign the treatments using a coin flip until one treatment has been assigned nine times; then assign the other treatment to the remaining subjects. The table below presents data on the first 18 young climbers whose parents consented to participation in the experiment. Order
Gender
Age (yr)
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
male female male male male male female female male female female male female female male female female male
12.90 13.34 12.39 13.95 13.63 13.62 12.55 13.54 12.34 13.74 13.78 14.05 14.22 13.91 14.39 13.54 13.85 14.11
1. Describe how you would implement a strategy equivalent to the one proposed by the researchers. Your plan should assign the treatments M (medicine) and H (hyperbaric chamber) to these climbers as they appear at the ranger station. 2. Implement your strategy in Step (1), assigning treatments to climbers 1–18. 3. Looking at which climbers were assigned to each of the two groups, do you feel that this method worked well? Why or why not? 4. Compute the proportion of females in the medicine group. How does this proportion compare to the proportion of females in the entire group of 18 subjects? 5. Construct two dotplots—one of the ages of those assigned to the medicine treatment and one of the ages of those assigned to the hyperbaric chamber treatment. Are the age distributions for the two groups similar? 6. Compute the average age of those assigned to the medicine group. How does it compare to the average age for the other treatment group? 7. Record the proportion of females in the medicine group, the average age of those assigned to the medicine group, and the average age of those assigned to the hyperbaric chamber group obtained by each student on your class. 8. Using the values from Step (6), construct a dotplot of each of the following: the proportion of females in the medicine group, the average age of those assigned to the medicine group, and the average age of those assigned to the hyperbaric chamber group. 9. Using the results of the previous steps, evaluate the success of this random assignment strategy. Write a short paragraph explaining to the researchers whether or not they should use the proposed strategy for random assignment and why.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
Summary of Key Concepts and Formulas
83
Summary of Key Concepts and Formulas TERM OR FORMULA
COMMENT
Observational study
A study that observes characteristics of an existing population.
Simple random sample
A sample selected in a way that gives every different sample of size n an equal chance of being selected.
Stratified sampling
Dividing a population into subgroups (strata) and then taking a separate random sample from each stratum.
Cluster sampling
Dividing a population into subgroups (clusters) and forming a sample by randomly selecting clusters and including all individuals or objects in the selected clusters in the sample.
1 in k systematic sampling
A sample selected from an ordered arrangement of a population by choosing a starting point at random from the first k individuals on the list and then selecting every kth individual thereafter.
Confounding variable
A variable that is related both to group membership and to the response variable.
Measurement or response bias
The tendency for samples to differ from the population because the method of observation tends to produce values that differ from the true value.
Selection bias
The tendency for samples to differ from the population because of systematic exclusion of some part of the population.
Nonresponse bias
The tendency for samples to differ from the population because measurements are not obtained from all individuals selected for inclusion in the sample.
Experiment
A procedure for investigating the effect of experimental conditions (treatments) on a response variable.
Treatments
The experimental conditions imposed by the experimenter.
Extraneous variable
A variable that is not an explanatory variable in the study but is thought to affect the response variable.
Direct control
Holding extraneous variables constant so that their effects are not confounded with those of the experimental conditions.
Blocking
Using extraneous variables to create groups that are similar with respect to those variables and then assigning treatments at random within each block, thereby filtering out the effect of the blocking variables.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
84
Chapter 2 Collecting Data Sensibly
TERM OR FORMULA
COMMENT
Random assignment
Assigning experimental units to treatments or treatments to trials at random.
Replication
A strategy for ensuring that there is an adequate number of observations on each experimental treatment.
Placebo treatment
A treatment that resembles the other treatments in an experiment in all apparent ways but that has no active ingredients.
Control group
A group that receives no treatment.
Single-blind experiment
An experiment in which the subjects do not know which treatment they received but the individuals measuring the response do know which treatment was received, or an experiment in which the subjects do know which treatment they received but the individuals measuring the response do not know which treatment was received.
Double-blind experiment
An experiment in which neither the subjects nor the individuals who measure the response know which treatment was received.
Chapter Review Exercises 2.70 - 2.85 2.70 A pollster for the Public Policy Institute of California explains how the Institute selects a sample of California adults (“It’s About Quality, Not Quantity,” San Luis Obispo Tribune, January 21, 2000): That is done by using computer-generated random residential telephone numbers with all California prefixes, and when there are no answers, calling back repeatedly to the original numbers selected to avoid a bias against hard-to-reach people. Once a call is completed, a second random selection is made by asking for the adult in the household who had the most recent birthday. It is as important to randomize who you speak to in the household as it is to randomize the household you select. If you didn’t, you’d primarily get women and older people. Comment on this approach to selecting a sample. How does the sampling procedure attempt to minimize certain types of bias? Are there sources of bias that may still be a concern? Bold exercises answered in back
Data set available online
Based on a survey of 4113 U.S. adults, researchers at Stanford University concluded that Internet use leads to increased social isolation. The survey was conducted by an Internet-based polling company that selected its samples from a pool of 35,000 potential respondents, all of whom had been given free Internet access and WebTV hardware in exchange for agreeing to regularly participate in surveys conducted by the polling company. Two criticisms of this study were expressed in an article that appeared in the San Luis Obispo Tribune (February 28, 2000). The first criticism was that increased social isolation was measured by asking respondents if they were talking less to family and friends on the phone. The second criticism was that the sample was selected only from a group that was induced to participate by the offer of free Internet service, yet the results were generalized to all U.S. adults. For each criticism, indicate what type of bias is being described and why it might make you question the conclusion drawn by the researchers.
2.71
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
Chapter Review Exercises
2.72 The article “I’d Like to Buy a Vowel, Drivers Say” (USA Today, August 7, 2001) speculates that young people prefer automobile names that consist of just numbers and/or letters that do not form a word (such as Hyundai’s XG300, Mazda’s 626, and BMW’s 325i). The article goes on to state that Hyundai had planned to identify the car that was eventually marketed as the XG300 with the name Concerto, until they determined that consumers hated it and that they thought XG300 sounded more “technical” and deserving of a higher price. Do the students at your school feel the same way? Describe how you would go about selecting a sample to answer this question.
2.73 A study in Florida is examining whether health literacy classes and using simple medical instructions that include pictures and avoid big words and technical terms can keep Medicaid patients healthier (San Luis Obispo Tribune, October 16, 2002). Twenty-seven community health centers are participating in the study. For 2 years, half of the centers will administer standard care. The other centers will have patients attend classes and will provide special health materials that are easy to understand. Explain why it is important for the researchers to assign the 27 centers to the two groups (standard care and classes with simple health literature) at random. Is status related to a student’s understanding of science? The article “From Here to Equity: The Influence
2.74
of Status on Student Access to and Understanding of Science” (Culture and Comparative Studies [1999]: 577– 602) described a study on the effect of group discussions on learning biology concepts. An analysis of the relationship between status and “rate of talk” (the number of on-task speech acts per minute) during group work included gender as a blocking variable. Do you think that gender is a useful blocking variable? Explain. The article “Tots’ TV-Watching May Spur Attention Problems” (San Luis Obispo Tribune, April 5, 2004) describes a study that appeared in the journal
2.75
Pediatrics. In this study, researchers looked at records of 2500 children who were participating in a long-term health study. They found that 10% of these children had attention disorders at age 7 and that hours of television watched at age 1 and age 3 was associated with an increased risk of having an attention disorder at age 7. a. Is the study described an observational study or an experiment? Bold exercises answered in back
Data set available online
85
b. Give an example of a potentially confounding variable that would make it unwise to draw the conclusion that hours of television watched at a young age is the cause of the increased risk of attention disorder.
2.76 A study of more than 50,000 U.S. nurses found that those who drank just one soda or fruit punch a day tended to gain much more weight and had an 80% increased risk in developing diabetes compared to those who drank less than one a month. (The Washington Post, August 25, 2004). “The message is clear. . . . Anyone who cares about their health or the health of their family would not consume these beverages,” said Walter Willett of the Harvard School of Public Health, who helped conduct the study. The sugar and beverage industries said that the study was fundamentally flawed. “These allegations are inflammatory. Women who drink a lot of soda may simply have generally unhealthy lifestyles,” said Richard Adamson of the American Beverage Association. a. Do you think that the study described was an observational study or an experiment? b. Is it reasonable to conclude that drinking soda or fruit punch causes the observed increased risk of diabetes? Why or why not?
2.77 “Crime Finds the Never Married” is the conclusion drawn in an article from USA Today (June 29, 2001). This conclusion is based on data from the Justice Department’s National Crime Victimization Survey, which estimated the number of violent crimes per 1000 people, 12 years of age or older, to be 51 for the never married, 42 for the divorced or separated, 13 for married individuals, and 8 for the widowed. Does being single cause an increased risk of violent crime? Describe a potential confounding variable that illustrates why it is unreasonable to conclude that a change in marital status causes a change in crime risk. The article “Workers Grow More Dissatisfied” in the San Luis Obispo Tribune (August 22, 2002) states that “a survey of 5000 people found that while most Americans continue to find their jobs interesting, and are even satisfied with their commutes, a bare majority like their jobs.” This statement was based on the fact that only 51 percent of those responding to a mail survey indicated that they were satisfied with their jobs. Describe any potential sources of bias that might limit the researcher’s ability to draw conclusions about working Americans based on the data collected in this survey.
2.78
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
86
Chapter 2 Collecting Data Sensibly
2.79 According to the article “Effect of Preparation Methods on Total Fat Content, Moisture Content, and Sensory Characteristics of Breaded Chicken Nuggets and Beef Steak Fingers” (Family and Consumer Sciences Research Journal [1999]: 18–27), sensory tests were conducted using 40 college student volunteers at Texas Women’s University. Give three reasons, apart from the relatively small sample size, why this sample may not be ideal as the basis for generalizing to the population of all college students.
2.80 Do ethnic group and gender influence the type of care that a heart patient receives? The following passage is from the article “Heart Care Reflects Race and Sex, Not Symptoms” (USA Today, February 25, 1999, reprinted with permission): Previous research suggested blacks and women were less likely than whites and men to get cardiac catheterization or coronary bypass surgery for chest pain or a heart attack. Scientists blamed differences in illness severity, insurance coverage, patient preference, and health care access. The researchers eliminated those differences by videotaping actors—two black men, two black women, two white men, and two white women—describing chest pain from identical scripts. They wore identical gowns, used identical gestures, and were taped from the same position. Researchers asked 720 primary care doctors at meetings of the American College of Physicians or the American Academy of Family Physicians to watch a tape and recommend care. The doctors thought the study focused on clinical decision making. Evaluate this experimental design. Do you think this is a good design or a poor design, and why? If you were designing such a study, what, if anything, would you propose to do differently?
2.81 An article in the San Luis Obispo Tribune (September 7, 1999) described an experiment designed to investigate the effect of creatine supplements on the development of muscle fibers. The article states that the researchers “looked at 19 men, all about 25 years of age and similar in weight, lean body mass, and capacity to lift weights. Ten were given creatine—25 grams a day for the first week, followed by 5 grams a day for the rest of the study. The rest were given a fake preparation. No one was told what he was getting. All the men worked out under the guidance of the same trainer. The response variable measured was gain in fat-free mass (in percent).” Bold exercises answered in back
Data set available online
a. What extraneous variables are identified in the given statement, and what strategy did the researchers use to deal with them? b. Do you think it was important that the men participating in the experiment were not told whether they were receiving creatine or the placebo? Explain. c. This experiment was not conducted in a double-blind manner. Do you think it would have been a good idea to make this a double-blind experiment? Explain. Researchers at the University of Houston decided to test the hypothesis that restaurant servers who squat to the level of their customers would receive a larger tip
2.82
(“Effect of Server Posture on Restaurant Tipping,” Journal of Applied Social Psychology [1993]: 678–685). In the experiment, the waiter would flip a coin to determine whether he would stand or squat next to the table. The waiter would record the amount of the bill and of the tip and whether he stood or squatted. a. Describe the treatments and the response variable. b. Discuss possible extraneous variables and how they could be controlled. c. Discuss whether blocking would be necessary. d. Identify possible confounding variables. e. Discuss the role of random assignment in this experiment.
2.83 You have been asked to determine on what types of grasslands two species of birds, northern harriers and short-eared owls, build nests. The types of grasslands to be used include undisturbed native grasses, managed native grasses, undisturbed nonnative grasses, and managed nonnative grasses. You are allowed a plot of land 500 meters square to study. Explain how you would determine where to plant the four types of grasses. What role would random assignment play in this determination? Identify any confounding variables. Would this study be considered an observational study or an experiment? (Based on the article “Response of Northern Harriers and Short-Eared Owls to Grassland Management in Illinois,” Journal of Wildlife Management [1999]: 517–523.) A manufacturer of clay roofing tiles would like to investigate the effect of clay type on the proportion of tiles that crack in the kiln during firing. Two different types of clay are to be considered. One hundred tiles can be placed in the kiln at any one time. Firing temperature varies slightly at different locations in the kiln, and firing temperature may also affect cracking. Discuss the design of an experiment to collect information that could be
2.84
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
Chapter Review Exercises
used to decide between the two clay types. How does your proposed design deal with the extraneous variable temperature?
2.85 A mortgage lender routinely places advertisements in a local newspaper. The advertisements are of three different types: one focusing on low interest rates, one featuring low fees for first-time buyers, and one appealing to people who may want to refinance their homes. The lender would like to determine which adver-
87
tisement format is most successful in attracting customers to call for more information. Describe an experiment that would provide the information needed to make this determination. Be sure to consider extraneous variables, such as the day of the week that the advertisement appears in the paper, the section of the paper in which the advertisement appears, or daily fluctuations in the interest rate. What role does random assignment play in your design?
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
CHAPTER
3
Graphical Methods for Describing Data Most college students (and their parents) are concerned about the cost of a college education. The Chronicle of Higher Education (August 2008) reported the average tuition and fees for 4-year public institutions in each of the 50 U.S. states for the 2006-2007 academic year. Average tuition and fees (in dollars) are given for each state:
Florin Tirlea/iStockphoto
4712 3930 7629 3943 5077
4422 4155 7504 5022 5009
4669 8038 7392 4038 5114
4937 6284 4457 5471 3757
4452 6019 6320 9010 9783
4634 4966 5378 4176 6447
7151 5821 5181 5598 5636
7417 3778 2844 9092 4063
3050 6557 9003 6698 6048
3851 7106 9333 7914 2951
Make the most of your study time by accessing everything you need to succeed online with CourseMate. Visit http://www.cengagebrain.com where you will find: • An interactive eBook, which allows you to take notes, highlight, bookmark, search • • • • • •
the text, and use in-context glossary definitions Step-by-step instructions for Minitab, Excel, TI-83/84, SPSS, and JMP Video solutions to selected exercises Data sets available for selected examples and exercises Online quizzes Flashcards Videos
89 Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
90
Chapter 3 Graphical Methods for Describing Data
Several questions could be posed about these data. What is a typical value of average tuition and fees for the 50 states? Are observations concentrated near the typical value, or does average tuition and fees differ quite a bit from state to state? Are there any states whose average tuition and fees are somehow unusual compared to the rest? What proportion of the states have average tuition and fees exceeding $6000? Exceeding $8000? Questions such as these are most easily answered if the data can be organized in a sensible manner. In this chapter, we introduce some techniques for organizing and describing data using tables and graphs.
3.1
Displaying Categorical Data: Comparative Bar Charts and Pie Charts Comparative Bar Charts In Chapter 1 we saw that categorical data could be summarized in a frequency distribution and displayed graphically using a bar chart. Bar charts can also be used to give a visual comparison of two or more groups. This is accomplished by constructing two or more bar charts that use the same set of horizontal and vertical axes, as illustrated in Example 3.1.
EXAMPLE 3.1
How Far Is Far Enough
Each year The Princeton Review conducts a survey of high school students who are applying to college and parents of college applicants. The report “2009 College Hopes
& Worries Survey Findings” (www.princetonreview.com/uploadedFiles/Test_ Preparation/Hopes_and_Worries/colleg_hopes_worries_details.pdf) included a summary of how 12,715 high school students responded to the question “Ideally how far from home would you like the college you attend to be?” Also included was a summary of how 3007 parents of students applying to college responded to the question “How far from home would you like the college your child attends to be?” The accompanying relative frequency table summarized the student and parent responses.
FREQUENCY
Ideal Distance Less than 250 miles 250 to 500 miles 500 to 1000 miles More than 1000 miles
Step-by-Step technology instructions available online
RELATIVE FREQUENCY
Students
Parents
Students
Parents
4450 3942 2416 1907
1594 902 331 180
.35 .31 .19 .15
.53 .30 .11 .06
When constructing a comparative bar chart we use the relative frequency rather than the frequency to construct the scale on the vertical axis so that we can make meaningful comparisons even if the sample sizes are not the same. The comparative bar chart for these data is shown in Figure 3.1. It is easy to see the differences between students and parents. A higher proportion of parents prefer a college close to home, and a higher
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
3.1 Displaying Categorical Data: Comparative Bar Charts and Pie Charts
91
proportion of students than parents believe that the ideal distance from home would be more than 500 miles. To see why it is important to use relative frequencies rather than frequencies to compare groups of different sizes, consider the incorrect bar chart constructed using the frequencies rather than the relative frequencies (Figure 3.2). The incorrect bar chart conveys a very different and misleading impression of the differences between students and parents.
Relative frequency Students Parents
0.6 0.5 0.4 0.3 0.2 0.1
FIGURE 3.1
0 <250 miles
Comparative bar chart of ideal distance from home.
250–500 miles 500–1000 miles
>1000 miles
Ideal distance
Frequency Students Parents
5000 4000 3000 2000 1000
FIGURE 3.2
An incorrect comparative bar chart for the data of Example 3.1.
0 <250 miles
250–500 miles 500–1000 miles
>1000 miles
Ideal distance
Pie Charts A categorical data set can also be summarized using a pie chart. In a pie chart, a circle is used to represent the whole data set, with “slices” of the pie representing the possible categories. The size of the slice for a particular category is proportional to the corresponding frequency or relative frequency. Pie charts are most effective for summarizing data sets when there are not too many different categories.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
92
Chapter 3 Graphical Methods for Describing Data
EXAMPLE 3.2
Life Insurance for Cartoon Characters??
The article “Fred Flintstone, Check Your Policy” (The Washington Post, October 2, 2005) summarized the results of a survey of 1014 adults conducted by the Life and Health Insurance Foundation for Education. Each person surveyed was asked to select which of five fictional characters, Spider-Man, Batman, Fred Flintstone, Harry Potter, and Marge Simpson, he or she thought had the greatest need for life insurance. The resulting data are summarized in the pie chart of Figure 3.3. Don’t know 12.0% Marge Simpson 11.0%
Spider-Man 28.0%
Harry Potter 15.0% Batman 18.0%
FIGURE 3.3 Pie chart of data on which fictional character most needs life insurance.
Fred Flintstone 16.0%
The survey results were quite different from an insurance expert’s assessment. His opinion was that Fred Flintstone, a married father with a young child, was by far the one with the greatest need for life insurance. Spider-Man, unmarried with an elderly aunt, would need life insurance only if his aunt relied on him to supplement her income. Batman, a wealthy bachelor with no dependents, doesn’t need life insurance in spite of his dangerous job!
Pie Chart for Categorical Data When to Use Categorical data with a relatively small number of possible categories. Pie charts are most useful for illustrating proportions of the whole data set for various categories. How to Construct 1. Draw a circle to represent the entire data set. 2. For each category, calculate the “slice” size. Because there are 360 degrees in a circle slice size 5 360 ? (category relative frequency) 3. Draw a slice of appropriate size for each category. This can be tricky, so most pie charts are generated using a graphing calculator or a statistical software package.
What to Look For • Categories that form large and small proportions of the data set.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
3.1 Displaying Categorical Data: Comparative Bar Charts and Pie Charts
EXAMPLE 3.3
93
Watch Those Typos
Typos on a résumé do not make a very good impression when applying for a job. Senior executives were asked how many typos in a résumé would make them not consider a job candidate (“Job Seekers Need a Keen Eye,” USA Today, August 3, 2009). The resulting data are summarized in the accompanying relative frequency distribution.
Number of Typos
Frequency
Relative Frequency
60 54 21 10 5
.40 .36 .14 .07 .03
1 2 3 4 or more Don’t know
To draw a pie chart by hand, we would first compute the slice size for each category. For the one typo category, the slice size would be slice size 5 (.40)(360) 5 144 degrees 144 degrees, to represent first attempt category
We would then draw a circle and use a protractor to mark off a slice corresponding to about 144°, as illustrated here in the figure shown in the margin. Continuing to add slices in this way leads to a completed pie chart. It is much easier to use a statistical software package to create pie charts than to construct them by hand. A pie chart for the typo data, created with the statistical software package Minitab, is shown in Figure 3.4.
4 or more 7.0% 3 14.0%
Don’t know 3.0%
1 40.0%
FIGURE 3.4 Pie chart for the typo data of Example 3.3.
Step-by-Step technology instructions available online
2 36.0%
Pie charts can be used effectively to summarize a single categorical data set if there are not too many different categories. However, pie charts are not usually the best tool if the goal is to compare groups on the basis of a categorical variable. This is illustrated in Example 3.4.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
94
Chapter 3 Graphical Methods for Describing Data
EXAMPLE 3.4
Scientists and Nonscientists Do Not See Eye-to-Eye
Scientists and nonscientists were asked to indicate if they agreed or disagreed with the following statement: “When something is run by the government, it is usually inefficient and wasteful.” The resulting data (from “Scientists, Public Differ in Outlooks,” USA Today, July 10, 2009) were used to create the two pie charts in Figure 3.5.
Scientists
Nonscientists
Don’t know
Don’t know
Agree Disgree Agree Disgree
(a)
(b)
FIGURE 3.5 Pie charts for Example 3.4: (a) scientist data; (b) nonscientist data.
Although differences between scientists and nonscientists can be seen by comparing the pie charts of Figure 3.5, it can be difficult to compare category proportions using pie charts. A comparative bar chart (Figure 3.6) makes this type of comparison easier.
Relative frequency Scientists Nonscientists
0.7 0.6 0.5 0.4 0.3 0.2 0.1
FIGURE 3.6 Comparative bar chart for the scientist and nonscientist data.
0 Agree
Disagree
Don’t know
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
3.1 Displaying Categorical Data: Comparative Bar Charts and Pie Charts
95
A Different Type of “Pie” Chart: Segmented Bar Graphs A pie chart can be difficult to construct by hand, and the circular shape sometimes makes it difficult to compare areas for different categories, particularly when the relative frequencies for categories are similar. The segmented bar graph (also sometimes called a stacked bar graph) avoids these difficulties by using a rectangular bar rather than a circle to represent the entire data set. The bar is divided into segments, with different segments representing different categories. As with pie charts, the area of the segment for a particular category is proportional to the relative frequency for that category. Example 3.5 illustrates the construction of a segmented bar graph.
EXAMPLE 3.5
How College Seniors Spend Their Time
Each year, the Higher Education Research Institute conducts a survey of college seniors. In 2008, approximately 23,000 seniors participated in the survey (“Findings
from the 2008 Administration of the College Senior Survey,” Higher Education Research Institute, June 2009). The accompanying relative frequency table summarizes student response to the question: “During the past year, how much time did you spend studying and doing homework in a typical week?”
STUDYING/HOMEWORK
Amount of Time 2 hours or less 3 to 5 hours 6 to 10 hours 11 to 15 hours 16 to 20 hours Over 20 hours
Relative Frequency .074 .227 .285 .181 .122 .111
To construct a segmented bar graph for these data, first draw a bar of any fixed width and length, and then add a scale that ranges from 0 to 1, as shown.
1.00 0.90 0.80 0.70 0.60 0.50 0.40 0.30 0.20 0.10 0.00
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
96
Chapter 3
Graphical Methods for Describing Data
Then divide the bar into six segments, corresponding to the six possible time categories in this example. The first segment, corresponding to the 2 hours or less category, ranges from 0 to .074. The second segment, corresponding to 3 to 5 hours, ranges from .074 to .301 (for a length of .227, the relative frequency for this category), and so on. The segmented bar graph is shown in Figure 3.7.
1.00 0.90 0.80 Over 20 hours 16 to 20 hours 11 to 15 hours 6 to 10 hours 3 to 5 hours 2 hours or less
0.70 0.60 0.50 0.40 0.30
FIGURE 3.7 Segmented bar graph for the study time data of Example 3.5.
0.20 0.10 0.00
The same report also gave data on amount of time spent on exercise or sports in a typical week. Figure 3.8 shows horizontal segmented bar graphs (segmented bar graphs can be displayed either vertically or horizontally) for both time spent studying and time spent exercising. Viewing these graphs side by side makes it easy to see how students differ with respect to time spent on these two types of activities.
Amount of Time >20 hours 16–20 hours 11–15 hours 6–10 hours 3–5 hours 2 hours or less
Exercise/sport
Studying/homework
FIGURE 3.8 Segmented bar graphs for time spent studying and time spent exercising.
0.0
0.2
0.4
0.6
0.8
1.0
Data
Other Uses of Bar Charts and Pie Charts As we have seen in previous examples, bar charts and pie charts can be used to summarize categorical data sets. However, they are occasionally used for other purposes, as illustrated in Examples 3.6 and 3.7.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
3.1 Displaying Categorical Data: Comparative Bar Charts and Pie Charts
97
© PhotoLink/Photodisc/Getty Images
E X A M P L E 3 . 6 Grape Production The 2008 Grape Crush Report for California gave the following information on grape production for each of four different types of grapes (California Department of Food and Agriculture, March 10, 2009): Type of Grape
Tons Produced
Red Wine Grapes White Wine Grapes Raisin Grapes Table Grapes Total
1,715,000 1,346,000 494,000 117,000 3,672,000
Although this table is not a frequency distribution, it is common to represent information of this type graphically using a pie chart, as shown in Figure 3.9. The pie represents the total grape production, and the slices show the proportion of the total production for each of the four types of grapes. Table Raisin
Red wine
FIGURE 3.9
White wine
Pie chart for grape production data.
EXAMPLE 3.7
Back-to-College Spending
The National Retail Federation’s 2008 Back to College Consumer Intentions and Actions Survey (www.nrf.com) asked each person in a sample of college students how much they planned to spend in various categories during the upcoming academic year. The average amounts of money (in dollars) that men and women planned to spend for five different types of purchases are shown in the accompanying table.
Type of Purchase Clothing and Accessories Shoes School Supplies Electronics and Computers Dorm or Apartment Furnishings
Average for Men
Average for Women
$207.46 $107.22 $86.85 $533.17 $266.69
$198.15 $88.65 $81.56 $344.90 $266.98
Data set available online
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
98
Chapter 3 Graphical Methods for Describing Data
Even though this table is not a frequency distribution, this type of information is often represented graphically in the form of a bar chart, as illustrated in Figure 3.10. From the bar chart, we can see that the average amount of money that men and women plan to spend is similar for all of the types of purchases except for electronics and computers, in which the average for men is quite a bit higher than the average for women. Amount plan to spend
600 Men Women
500 400 300 200 100
FIGURE 3.10 Comparative bar chart for the backto-college spending data of men and women.
0 Clothing
Shoes
School supplies
Electronics and computers
Dorm or apartment furnishings
EX E RC I S E S 3 . 1 - 3 . 1 4 3.1 Each person in a nationally representative sample of 1252 young adults age 23 to 28 years old was asked how they viewed their “financial physique” (“2009 Young
Adults & Money Survey Findings,” Charles Schwab, 2009). “Toned and fit” was chosen by 18% of the respondents, while 55% responded “a little bit flabby,” and 27% responded “seriously out of shape.” Summarize this information in a pie chart.
Image not available due to copyright restrictions
3.2 The accompanying graphical display appeared in USA Today (October 22, 2009). It summarizes survey responses to a question about whether visiting social networking sites is allowed at work. Which of the graph types introduced in this section is used to display the responses? (USA Today frequently adds artwork and text to their graphs to try to make them look more interesting.)
3.3 The survey referenced in the previous exercise was conducted by Robert Half Technology. This company issued a press release (“Whistle—But Don’t Tweet—
While You Work,” www.roberthalftechnology.com, October 6, 2009) that provided more detail than in the USA Today snapshot graph. The actual question asked Bold exercises answered in back
Data set available online
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
3.1 Displaying Categorical Data: Comparative Bar Charts and Pie Charts
99
was “Which of the following most closely describes your company’s policy on visiting social networking sites, such as Facebook, MySpace and Twitter, while at work?” The responses are summarized in the following table:
Construct an appropriate graph to summarize the information in the table. Explain why you chose this particular type of graph.
Relative Frequency (expressed as percent)
ated school cafeterias in 20 school districts across the United States. Each district was assigned a numerical score on the basis of rigor of food codes, frequency of food safety inspections, access to inspection information, and the results of cafeteria inspections. Based on the score assigned, each district was also assigned one of four grades. The scores and grades are summarized in the accompanying table, which appears in the report “Making the Grade: An
Response Category Prohibited completely Permitted for business purposes only Permitted for limited personal use Permitted for any type of personal use Don’t know/no answer
54% 19% 16% 10% 1%
a. Explain how the survey response categories and corresponding relative frequencies were used or modified to produce the graphical display in Exercise 3.2. b. Using the original data in the table, construct a segmented bar graph. c. What are two other types of graphical displays that would be appropriate for summarizing these data?
3.4 The National Confectioners Association asked 1006 adults the following question: “Do you set aside a personal stash of Halloween candy?” Fifty-five percent of those surveyed responded no, 41% responded yes, and 4% either did not answer the question or said they did not know (USA Today, October 22, 2009). Use the given information to construct a pie chart. 3.5 The report “Communicating to Teens (Aged 12–17)” (U.S. Department of Health and Human Services, www.cdc.gov) suggests that teens can be classified into five groups based on attitude, behavior, and conformity. The report also includes estimates of the percentage of teens who fall into each of these groups. The groups are described in the accompanying table. Group and Description Explorer: creative, independent, and differs from the norm. Visible: well known and popular because of looks, personality or athletic ability Status Quo: display traditional values of moderation and achievement, seek mainstream acceptance Non-Teen: behave more like adults or young children because of lack of social skills or indifference to teen culture and style Isolator: psychologically isolated from both peers and adults
Bold exercises answered in back
Percentage of Teens in This Group 10% 30% 38%
14%
8%
Data set available online
3.6 The Center for Science in the Public Interest evalu-
Analysis of Food Safety in School Cafeterias” (cspi.us/ new/pdf/makingthegrade.pdf, 2007). Top of the Class
Passing
Barely Passing
Jurisdiction
Failing
Overall Score (out of 100)
City of Fort Worth, TX King County, WA City of Houston, TX Maricopa County, AZ City and County of Denver, CO
80 79 78 77 75
Dekalb County, GA Farmington Valley Health District, CT State of Virginia
73 72 72
Fulton County, GA City of Dallas, TX City of Philadelphia, PA City of Chicago, IL City and County of San Francisco, CA Montgomery County, MD
68 67 67 65 64 63
Hillsborough County, FL City of Minneapolis, MN Dade County, FL State of Rhode Island District of Columbia City of Hartford, CT
60 60 59 54 46 37
a. Two variables are summarized in the figure, grade and overall score. Is overall score a numerical or categorical variable? Is grade (indicated by the different colors in the figure) a numerical or categorical variable? b. Explain how the figure is equivalent to a segmented bar graph of the grade data. c. Construct a dotplot of the overall score data. Based on the dotplot, suggest an alternate assignment of grades (top of class, passing, etc.) to the 20 school districts. Explain the reasoning you used to make your assignment. Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
100
Chapter 3 Graphical Methods for Describing Data
3.7 The article “Housework around the World” (USA Today, September 15, 2009) included the percentage of women who say their spouses never help with household chores for five different countries. Country
3.9 The article “Rinse Out Your Mouth” (Associated
Percentage
Japan France United Kingdom United States Canada
74% 44% 40% 34% 31%
a. Display the information in the accompanying table in a bar chart. b. The article did not state how the author arrived at the given percentages. What are two questions that you would want to ask the author about how the data used to compute the percentages were collected? c. Assuming that the data that were used to compute these percentages were collected in a reasonable way, write a few sentences describing how the five countries differ in terms of spouses helping their wives with housework.
3.8 The report “Findings from the 2008 Administration of the College Senior Survey” (Higher Education Research Institute, 2009) asked a large number of college seniors how they would rate themselves compared to the average person of their age with respect to physical health. The accompanying relative frequency table summarizes the responses for men and women. Relative Frequency
Rating of Physical Health
Men
Women
Highest 10% Above average Average Below average Lowest 10%
.220 .399 .309 .066 .005
.101 .359 .449 .086 .005
a. Construct a comparative bar graph of the responses that allows you to compare the responses of men and women. b. There were 8110 men and 15,260 women who responded to the survey. Explain why it is important that the comparative bar graph be constructed using the relative frequencies rather than the actual numbers of people (the frequencies) responding in each category.
Bold exercises answered in back
c. Write a few sentences commenting on how college seniors perceive themselves with respect to physical health and how men and women differ in their perceptions.
Data set available online
Press, March 29, 2006) summarized results from a survey of 1001 adults on the use of profanity. When asked “How many times do you use swear words in conversations?” 46% responded a few or more times per week, 32% responded a few times a month or less, and 21% responded never. Use the given information to construct a segmented bar chart.
3.10 The article “The Need to Be Plugged In” (Associated Press, December 22, 2005) described the results of a survey of 1006 adults who were asked about various technologies, including personal computers, cell phones, and DVD players. The accompanying table summarizes the responses to questions about how essential these technologies were. Relative Frequency Response
Personal Computer
Cell Phone
DVD Player
.46
.41
.19
.28
.25
.35
.26
.34
.46
Cannot imagine living without Would miss but could do without Could definitely live without
Construct a comparative bar chart that shows the distribution of responses for the three different technologies.
3.11 Poor fitness in adolescents and adults increases the risk of cardiovascular disease. In a study of 3110 adolescents and 2205 adults ( Journal of the American Medical Association, December 21, 2005), researchers found 33.6% of adolescents and 13.9% of adults were unfit; the percentage was similar in adolescent males (32.9%) and females (34.4%), but was higher in adult females (16.2%) than in adult males (11.8%). a. Summarize this information using a comparative bar graph that shows differences between males and females within the two different age groups. b. Comment on the interesting features of your graphical display. 3.12 A survey of 1001 adults taken by Associated Press–Ipsos asked “How accurate are the weather fore-
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
3.2 Displaying Numerical Data: Stem-and-Leaf Displays
casts in your area?” (San Luis Obispo Tribune, June 15, 2005). The responses are summarized in the table below. Extremely Very Somewhat Not too Not at all Not sure
4% 27% 53% 11% 4% 1%
101
a. Do you think this is an effective use of a pie chart? Why or why not? b. Construct a bar chart to show the distribution of deaths by object struck. Is this display more effective than the pie chart in summarizing this data set? Explain.
3.14 The article “Death in Roadwork Zones at Record High” (San Luis Obispo Tribune, July 25, 2001) included a bar chart similar to this one:
a. Construct a pie chart to summarize these data. b. Construct a bar chart to summarize these data. c. Which of these charts—a pie chart or a bar chart— best summarizes the important information? Explain.
Number of deaths 900 800
3.13 In a discussion of accidental deaths involving
700
roadside hazards, the web site highwaysafety.com included a pie chart like the one shown:
600 500 400
Embankment (11.0%)
Tree (28.0%)
300 200 100
Guardrail (9.0%)
’91
’92
’93
’94
’95
’96
’97
’98
’99
Year Utility pole (9.0%)
Other (11.0%)
Ditch (8.0%) Curb (6.0%) Sign or post (6.0%)
Bold exercises answered in back
3.2
Bridge rail (1.0%) Concrete bar (2.0%) Fence (4.0%) Culvert (5.0%)
Data set available online
a. Comment on the trend over time in the number of people killed in highway work zones. b. Would a pie chart have also been an effective way to summarize these data? Explain why or why not.
Video Solution available
Displaying Numerical Data: Stem-and-Leaf Displays A stem-and-leaf display is an effective and compact way to summarize univariate numerical data. Each number in the data set is broken into two pieces, a stem and a leaf. The stem is the first part of the number and consists of the beginning digit(s). The leaf is the last part of the number and consists of the final digit(s). For example, the number 213 might be split into a stem of 2 and a leaf of 13 or a stem of 21 and a leaf of 3. The resulting stems and leaves are then used to construct the display.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
102
Chapter 3 Graphical Methods for Describing Data
E X A M P L E 3 . 8 Should Doctors Get Auto Insurance Discounts? Many auto insurance companies give job-related discounts of between 5 and 15%. The article “Auto-Rate Discounts Seem to Defy Data” (San Luis Obispo Tribune, June 19, 2004) included the accompanying data on the number of automobile accidents per year for every 1000 people in 40 occupations.
Occupation Student Physician Lawyer Architect Real estate broker Enlisted military Social worker Manual laborer Analyst Engineer Consultant Sales Military officer Nurse School administrator Skilled labor Librarian Creative arts Executive Insurance agent
4 5 6 7 8 9 10 11 12 13 14 15
3 7 56667889 44567788999 000013445689 2569
2
Stem: Tens Leaf: Ones
FIGURE 3.11 Stem-and-leaf display for accident rate per 1000 for forty occupations
Step-by-step technology instructions available online
Accidents per 1000 152 109 106 105 102 199 198 196 195 194 194 193 191 190 190 190 190 190 189 189
Occupation Banking-finance Customer service Manager Medical support Computer-related Dentist Pharmacist Proprietor Teacher, professor Accountant Law enforcement Physical therapist Veterinarian Clerical, secretary Clergy Homemaker Politician Pilot Firefighter Farmer
Accidents per 1000 89 88 88 87 87 86 85 84 84 84 79 78 78 77 76 76 76 75 67 43
Figure 3.11 shows a stem-and-leaf display for the accident rate data. The numbers in the vertical column on the left of the display are the stems. Each number to the right of the vertical line is a leaf corresponding to one of the observations in the data set. The legend Stem: Leaf:
Tens Ones
tells us that the observation that had a stem of 4 and a leaf of 3 corresponds to an occupation with an accident rate of 43 per 1000 (as opposed to 4.3 or 0.43). Similarly, the observation with the stem of 10 and leaf of 2 corresponds to 102 accidents per 1000 (the leaf of 2 is the ones digit) and the observation with the stem of 15 and leaf of 2 corresponds to 152 accidents per 1000. The display in Figure 3.11 suggests that a typical or representative value is in the stem 8 or 9 row, perhaps around 90. The observations are mostly concentrated in the 75 to 109 range, but there are a couple of values that stand out on the low end (43 and 67) and one observation (152) that is far removed from the rest of the data on the high end.
Data set available online Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
3.2 Displaying Numerical Data: Stem-and-Leaf Displays
103
From the point of view of an auto insurance company it might make sense to offer discounts to occupations with low accident rates—maybe farmers (43 auto accidents per 1000 farmers) or firefighters (67 accidents per 1000 firefighters) or even some of the occupations with accident rates in the 70s. The “discounts seem to defy data” in the title of the article refers to the fact that some insurers provide discounts to doctors and engineers, but not to homemakers, politicians, and other occupations with lower accident rates. Two possible explanations were offered for this apparent discrepancy. One is that it is possible that while some occupations have higher accident rates, they also have lower average cost per claim. Accident rates alone may not reflect the actual cost to the insurance company. Another possible explanation is that the insurance companies may offer the discounted auto insurance in order to attract people who would then also purchase other types of insurance such as malpractice or liability insurance.
The leaves on each line of the display in Figure 3.11 have been arranged in order from smallest to largest. Most statistical software packages order the leaves this way, but it is not necessary to do so to get an informative display that still shows many of the important characteristics of the data set, such as shape and spread. Stem-and-leaf displays can be useful to get a sense of a typical value for the data set, as well as a sense of how spread out the values in the data set are. It is also easy to spot data values that are unusually far from the rest of the values in the data set. Such values are called outliers. The stem-and-leaf display of the accident rate data (Figure 3.11) shows an outlier on the low end (43) and an outlier on the high end (152).
DEFINITION An outlier is an unusually small or large data value. A precise rule for deciding when an observation is an outlier is given in Chapter 4.
Stem-and-Leaf Displays When to Use Numerical data sets with a small to moderate number of observations (does not work well for very large data sets) How to Construct 1. Select one or more leading digits for the stem values. The trailing digits (or sometimes just the first one of the trailing digits) become the leaves. 2. List possible stem values in a vertical column. 3. Record the leaf for every observation beside the corresponding stem value. 4. Indicate the units for stems and leaves someplace in the display.
What to Look For The display conveys information about • • • • •
a representative or typical value in the data set the extent of spread about a typical value the presence of any gaps in the data the extent of symmetry in the distribution of values the number and location of peaks
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
104
Chapter 3
Graphical Methods for Describing Data
E X A M P L E 3 . 9 Tuition at Public Universities The introduction to this chapter gave data on average tuition and fees at public institutions in the year 2007 for the 50 U.S. states. The observations ranged from a low value of 2844 to a high value of 9783. The data are reproduced here: 4712 4422 4669 4937 4452 4634 7151 7417 3050 3851 3930 4155 8038 6284 6019 4966 5821 3778 6557 7106 7629 7504 7392 4457 6320 5378 5181 2844 9003 9333 3943 5022 4038 5471 9010 4176 5598 9092 6698 7914 5077 5009 5114 3757 9783 6447 5636 4063 6048 2951 A natural choice for the stem is the leading (thousands) digit. This would result in a display with 7 stems (2, 3, 4, 5, 6, 7, 8, and 9). Using the first two digits of a number as the stem would result in 69 stems (28, 29, . . . , 97). A stem-and-leaf display with 56 stems would not be an effective summary of the data. In general, stemand-leaf displays that use between 5 and 20 stems tend to work well. If we choose the thousands digit as the stem, the remaining three digits (the hundreds, tens, and ones) would form the leaf. For example, for the first few values in the first column of data, we would have 4712 S stem ⫽ 4, leaf ⫽ 712 3930 S stem ⫽ 3, leaf ⫽ 930 7629 S stem ⫽ 7, leaf ⫽ 629
Data set available online
FIGURE 3.12 Stem-and-leaf display of average tuition and fees.
The leaves have been entered in the display of Figure 3.12 in the order they are encountered in the data set. Commas are used to separate the leaves only when each leaf has two or more digits. Figure 3.12 shows that most states had average tuition and fees in the $4000 to $7000 range and that the typical average tuition and fees is around $6000. A few states have average tuition and fees at public four-year institutions that are quite a bit higher than most other states (the five states with the highest values were Vermont, New Jersey, Pennsylvania, Ohio, and New Hampshire). 2 3 4 5 6 7 8 9
844, 951 050, 851, 930, 778, 943, 757 712, 422, 669, 937, 452, 634, 155, 966, 457, 038, 176, 063 821, 378, 181, 022, 471, 598, 077, 009, 114, 636 284, 019, 557, 320, 698, 447, 048 151, 417, 106, 629, 504, 392, 914 Stem: Thousands 038 Leaf: Ones 003, 333, 010, 092, 783
An alternative display (Figure 3.13) results from dropping all but the first digit of the leaf. This is what most statistical computer packages do when generating a display; little information about typical value, spread, or shape is lost in this truncation and the display is simpler and more compact.
FIGURE 3.13 Stem-and-leaf display of the average tuition and fees data using truncated stems.
2 3 4 5 6 7 8 9
89 089797 746946194010 8310450016 2053640 1416539 0 03007
Stem: Thousands Leaf: Hundreds
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
3.2 Displaying Numerical Data: Stem-and-Leaf Displays
105
Repeated Stems to Stretch a Display Sometimes a natural choice of stems gives a display in which too many observations are concentrated on just a few stems. A more informative picture can be obtained by dividing the leaves at any given stem into two groups: those that begin with 0, 1, 2, 3, or 4 (the “low” leaves) and those that begin with 5, 6, 7, 8, or 9 (the “high” leaves). Then each stem value is listed twice when constructing the display, once for the low leaves and once again for the high leaves. It is also possible to repeat a stem more than twice. For example, each stem might be repeated five times, once for each of the leaf groupings {0, 1}, {2, 3}, {4, 5}, {6, 7}, and {8, 9}.
E X A M P L E 3 . 1 0 Median Ages in 2030 The accompanying data on the Census Bureau’s projected median age in 2030 for the 50 U.S. states and Washington D.C. appeared in the article “2030 Forecast: Mostly Gray” (USA Today, April 21, 2005). The median age for a state is the age that divides the state’s residents so that half are younger than the median age and half are older than the median age. Projected Median Age 41.0 32.9 39.3 29.3 39.2 37.8 37.7 42.0 41.1 39.6 46.0 38.4 37.9 39.1 42.1 40.7 46.7 41.6 46.4
37.4 39.1 39.4 41.3
35.6 40.0 42.1 41.5
41.1 38.8 40.8 38.3
43.6 46.9 44.8 34.6
33.7 37.5 39.9 30.4
45.4 40.2 36.8 43.9
35.6 40.2 43.2 37.8
38.7 39.0 40.2 38.5
The ages in the data set range from 29.3 to 46.9. Using the first two digits of each data value for the stem results in a large number of stems, while using only the first digit results in a stem-and-leaf display with only three stems. The stem-and-leaf display using single digit stems and leaves truncated to a single digit is shown in Figure 3.14. A stem-and-leaf display that uses repeated stems is shown in Figure 3.15. Here each stem is listed twice, once for the low leaves (those beginning with 0, 1, 2, 3, 4) and once for the high leaves (those beginning with 5, 6, 7, 8, 9). This display is more informative than the one in Figure 3.14, but is much more compact than a display based on two-digit stems.
FIGURE 3.14 Stem-and-leaf display for the projected median age data.
FIGURE 3.15 Stem-and-leaf display for the projected median age data using repeated stems.
2 3 4
2H 3L 3H 4L 4H
9 02345567777778888899999999 000000111111222333456666 Stem: Tens Leaf: Ones
9 0234 5567777778888899999999 0000001111112223334 56666 Stem: Tens Leaf: Ones
Data set available online Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
106
Chapter 3 Graphical Methods for Describing Data
Comparative Stem-and-Leaf Displays Frequently an analyst wishes to see whether two groups of data differ in some fundamental way. A comparative stem-and-leaf display, in which the leaves for one group are listed to the right of the stem values and the leaves for the second group are listed to the left, can provide preliminary visual impressions and insights.
E X A M P L E 3 . 1 1 Progress for Children The report “Progress for Children” (UNICEF, April 2005) included the accompanying data on the percentage of primary-school-age children who were enrolled in school for 19 countries in Northern Africa and for 23 countries in Central Africa. Northern Africa 54.6 34.3 48.9 98.8 91.6 97.8 Central Africa 58.3 34.6 35.5 63.4 58.4 61.9 98.9
77.8 96.1
59.6 92.2
88.5 94.9
97.4 98.6
92.5 86.6
83.9
96.9
88.9
45.4 40.9
38.6 73.9
63.8 34.8
53.9 74.4
61.9 97.4
69.9 61.0
43.0 66.7
85.0 79.6
We will construct a comparative stem-and-leaf display using the first digit of each observation as the stem and the remaining two digits as the leaf. To keep the display simple the leaves will be truncated to one digit. For example, the observation 54.6 would be processed as 54.6 S stem ⫽ 5, leaf ⫽ 4 (truncated from 4.6) and the observation 34.3 would be processed as 34.3 S stem ⫽ 3, leaf ⫽ 4 (truncated from 4.3) The resulting comparative stem-and-leaf display is shown in Figure 3.16. Central Africa
FIGURE 3.16 Comparative stem-and-leaf display for percentage of children enrolled in primary school.
4854 035 838 6113913 943 5 87
Northern Africa 3 4 5 6 7 8 9
4 8 49 76 8386 7268176248
Stem: Tens Leaf: Ones
From the comparative stem-and-leaf display you can see that there is quite a bit of variability in the percentage enrolled in school for both Northern and Central African countries and that the shapes of the two data distributions are quite different. The percentage enrolled in school tends to be higher in Northern African countries than in Central African countries, although the smallest value in each of the two data sets is about the same. For Northern African countries the distribution of values has a single peak in the 90s with the number of observations declining as we move toward the stems corresponding to lower percentages enrolled in school. For Central African countries the distribution is more symmetric, with a typical value in the mid 60s. Data set available online Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
3.2 Displaying Numerical Data: Stem-and-Leaf Displays
107
E X E RC I S E S 3 . 1 5 - 3 . 2 1 The U.S. Department of Health and Human Services provided the data in the accompanying table in the report “Births: Preliminary Data for 2007” (National Vital Statistics Reports, March 18, 2009). Entries
3.15
in the table are the birth rates (births per 1,000 of population) for the year 2007. State Alabama Alaska Arizona Arkansas California Colorado Connecticut Delaware District of Columbia Florida Georgia Hawaii Idaho Illinois Indiana Iowa Kansas Kentucky Louisiana Maine Maryland Massachusetts Michigan Minnesota Mississippi Missouri Montana Nebraska Nevada New Hampshire New Jersey New Mexico New York North Carolina North Dakota Ohio Oklahoma Oregon Pennsylvania Rhode Island South Carolina
Births per 1,000 of Population 14.0 16.2 16.2 14.6 15.5 14.6 11.9 14.1 15.1 13.1 15.9 14.9 16.7 14.1 14.2 13.7 15.1 14.0 15.4 10.7 13.9 12.1 12.4 14.2 15.9 13.9 13.0 15.2 16.1 10.8 13.4 15.5 13.1 14.5 13.8 13.2 15.2 13.2 12.1 11.7 14.3
State South Dakota Tennessee Texas Utah Vermont Virginia Washington West Virginia Wisconsin Wyoming
15.4 14.1 17.1 20.8 10.5 14.1 13.8 12.1 13.0 15.1
Construct a stem-and-leaf display using stems 10, 11...20. Comment on the interesting features of the display.
3.16 The National Survey on Drug Use and Health, conducted in 2006 and 2007 by the Office of Applied Studies, led to the following state estimates of the total number of people ages 12 and older who had used a tobacco product within the last month. State Alabama Alaska Arizona Arkansas California Colorado Connecticut Delaware District of Columbia Florida Georgia Hawaii Idaho Illinois Indiana Iowa Kansas Kentucky Louisiana Maine Maryland Massachusetts Michigan Minnesota
Data set available online
Number of People (in thousands) 1,307 161 1,452 819 6,751 1,171 766 200 141 4,392 2,341 239 305 3,149 1,740 755 726 1,294 1,138 347 1,206 1,427 2,561 1,324 (continued)
(continued)
Bold exercises answered in back
Births per 1,000 of Population
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
108
Chapter 3
Graphical Methods for Describing Data
State Mississippi Missouri Montana Nebraska Nevada New Hampshire New Jersey New Mexico New York North Carolina North Dakota Ohio Oklahoma Oregon Pennsylvania Rhode Island South Carolina South Dakota Tennessee Texas Utah Vermont Virginia Washington West Virginia Wisconsin Wyoming
Number of People (in thousands) 763 1,627 246 429 612 301 1,870 452 4,107 2,263 162 3,256 1,057 857 3,170 268 1,201 202 1,795 5,533 402 158 1,771 1,436 582 1,504 157
a. Construct a stem-and-leaf display using thousands (of thousands) as the stems and truncating the leaves to the tens (of thousands) digit. b. Write a few sentences describing the shape of the distribution and any unusual observations. c. The four largest values were for California, Texas, Florida, and New York. Does this indicate that tobacco use is more of a problem in these states than elsewhere? Explain. d. If you wanted to compare states on the basis of the extent of tobacco use, would you use the data in the given table? If yes, explain why this would be reasonable. If no, what would you use instead as the basis for the comparison? The article “Going Wireless” (AARP Bulletin, June 2009) reported the estimated percentage of house-
3.17
holds with only wireless phone service (no land line) for the 50 U.S. states and the District of Columbia. In the accompanying data table, each state was also classified into one of three geographical regions—West (W), Middle states (M), and East (E). Bold exercises answered in back
Data set available online
Wireless %
Region
State
13.9 11.7 18.9 22.6 9.0 16.7 5.6 5.7 20.0 16.8 16.5 8.0 22.1 16.5 13.8 22.2 16.8 21.4 15.0 13.4 10.8 9.3 16.3 17.4 19.1 9.9 9.2 23.2 10.8 16.9 11.6 8.0 21.1 11.4 16.3 14.0 23.2 17.7 10.8 7.9 20.6 6.4 20.3 20.9 25.5 10.8 5.1 16.3 11.6 15.2 11.4
M W W M W W E E E E E W W M M M M M M E E E M M M M W M W M E E W E E E M W E E E M M M W E E W E M W
AL AK AZ AR CA CO CN DE DC FL GA HI ID IL IN IA KA KY LA ME MD MA MI MN MS MO MT NE NV ND NH NJ NM NY NC OH OK OR PA RI SC SD TN TX UT VA VT WA WV WI WY
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
3.2 Displaying Numerical Data: Stem-and-Leaf Displays
a. Construct a stem-and-leaf display for the wireless percentage using the data from all 50 states and the District of Columbia. What is a typical value for this data set? b. Construct a back-to-back stem-and-leaf display for the wireless percentage of the states in the West and the states in the East. How do the distributions of wireless percentages compare for states in the East and states in the West?
3.18 The article “Economy Low, Generosity High” (USA Today, July 28, 2009) noted that despite a weak economy in 2008, more Americans volunteered in their communities than in previous years. Based on census data (www.volunteeringinamerica.gov), the top and bottom five states in terms of percentage of the population who volunteered in 2008 were identified. The top five states were Utah (43.5%), Nebraska (38.9%), Minnesota (38.4%), Alaska (38.0%), and Iowa (37.1%). The bottom five states were New York (18.5%), Nevada (18.8%), Florida (19.6%), Louisiana (20.1%), and Mississippi (20.9%). a. For the data set that includes the percentage who volunteered in 2008 for each of the 50 states, what is the largest value? What is the smallest value? b. If you were going to construct a stem-and-leaf display for the data set consisting of the percentage who volunteered in 2008 for the 50 states, what stems would you use to construct the display? Explain your choice. The article “Frost Belt Feels Labor Drain” (USA Today, May 1, 2008) points out that even though
3.19
total population is increasing, the pool of young workers is shrinking in many states. This observation was prompted by the data in the accompanying table. Entries in the table are the percent change in the population of 25- to 44-year-olds over the period from 2000 to 2007. A negative percent change corresponds to a state that had fewer 25- to 44-year-olds in 2007 than in 2000 (a decrease in the pool of young workers). State Alabama Alaska Arizona Arkansas California Colorado Connecticut Delaware
% Change ⫺4.1 ⫺2.5 17.8 0.9 ⫺0.4 4.1 ⫺9.9 ⫺2.2 (continued)
Bold exercises answered in back
Data set available online
State District of Columbia Florida Georgia Hawaii Idaho Illinois Indiana Iowa Kansas Kentucky Louisiana Maine Maryland Massachusetts Michigan Minnesota Mississippi Missouri Montana Nebraska Nevada New Hampshire New Jersey New Mexico New York North Carolina North Dakota Ohio Oklahoma Oregon Pennsylvania Rhode Island South Carolina South Dakota Tennessee Texas Utah Vermont Virginia Washington West Virginia Wisconsin Wyoming
109
% Change 1.8 5.8 7.2 ⫺1.3 11.1 ⫺4.6 ⫺3.1 ⫺6.5 ⫺5.3 ⫺1.7 ⫺11.9 ⫺8.7 ⫺5.7 ⫺9.6 ⫺9.1 ⫺4.5 ⫺5.2 ⫺2.9 ⫺3.7 ⫺5.6 22.0 ⫺7.5 ⫺7.8 0.6 ⫺8.0 2.4 ⫺10.9 ⫺8.2 ⫺1.6 4.4 ⫺9.1 ⫺8.8 0.1 ⫺4.1 0.6 7.3 19.6 ⫺10.4 ⫺1.1 1.6 ⫺5.4 ⫺5.0 ⫺2.3
a. The smallest value in the data set is ⫺11.9 and the largest value is 22.0. One possible choice of stems for a stem-and-leaf display would be to use the tens digit, resulting in stems of ⫺1, ⫺0, 0, 1, and 2. Notice that because there are both negative and positive values in the data set, we would want to use two 0 stems—one where we can enter leaves for the Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
110
Chapter 3 Graphical Methods for Describing Data
negative percent changes that are between 0 and ⫺9.9, and one where we could enter leaves for the positive percent changes that are between 0 and 9.9. Construct a stem-and-leaf plot using these five stems. (Hint: Think of each data value as having two digits before the decimal place, so 4.1 would be regarded as 04.1.) b. Using two-digit stems would result in more than 30 stems, which is more than we would usually want for a stem-and-leaf display. Describe a strategy for using repeated stems that would result in a stemand-leaf display with about 10 stems. c. The article described “the frost belt” as the cold part of the country—the Northeast and Midwest— noting that states in the frost belt generally showed a decline in the number of people in the 25- to 44-year-old age group. How would you describe the group of states that saw a marked increase in the number of 25- to 44-year-olds? A report from Texas Transportation Institute (Texas A&M University System, 2005) titled “Congestion Reduction Strategies” included the accompanying data on extra travel time for peak travel time in hours per year per traveler for different sized urban areas.
3.20
Very Large Urban Areas Los Angeles, CA San Francisco, CA Washington DC, VA, MD Atlanta, GA Houston, TX Dallas, Fort Worth, TX Chicago, IL-IN Detroit, MI Miami, FL Boston, MA, NH, RI New York, NY-NJ-CT Phoenix, AZ Philadelphia, PA-NJ-DE-MD
Large Urban Areas Riverside, CA Orlando, FL San Jose, CA San Diego, CA
Extra Hours per Year per Traveler 93 72 69 67 63 60 58 57 51 51 49 49 38
Extra Hours per Year per Traveler 55 55 53 52 (continued)
Large Urban Areas
Extra Hours per Year per Traveler
Denver, CO Baltimore, MD Seattle, WA Tampa, FL Minneapolis, St Paul, MN Sacramento, CA Portland, OR, WA Indianapolis, IN St Louis, MO-IL San Antonio, TX Providence, RI, MA Las Vegas, NV Cincinnati, OH-KY-IN Columbus, OH Virginia Beach, VA Milwaukee, WI New Orleans, LA Kansas City, MO-KS Pittsburgh, PA Buffalo, NY Oklahoma City, OK Cleveland, OH
51 50 46 46 43 40 39 38 35 33 33 30 30 29 26 23 18 17 14 13 12 10
a. Construct a comparative stem-and-leaf plot for annual delay per traveler for each of the two different sizes of urban areas. b. Is the following statement consistent with the display constructed in Part (a)? Explain. The larger the urban area, the greater the extra travel time during peak period travel. High school dropout rates (percentages) for 2008 for the 50 states were given in the 2008 Kids Count Data Book (www.aecf.org) and are shown in the following table:
3.21
State
Rate
Alabama Alaska Arizona Arkansas California Colorado Connecticut Delaware Florida Georgia Hawaii Idaho
8% 10% 9% 9% 6% 8% 5% 7% 7% 8% 8% 6% (continued)
Bold exercises answered in back
Data set available online
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
3.3 Displaying Numerical Data: Frequency Distributions and Histograms
State
Rate
State
Illinois Indiana Iowa Kansas Kentucky Louisiana Maine Maryland Massachusetts Michigan Minnesota Mississippi Missouri Montana Nebraska Nevada New Hampshire New Jersey New Mexico New York North Carolina North Dakota Ohio Oklahoma Oregon Pennsylvania Rhode Island South Carolina South Dakota
6% 8% 3% 5% 7% 10% 6% 6% 4% 6% 3% 7% 7% 9% 4% 10% 3% 4% 10% 5% 8% 7% 5% 8% 6% 5% 6% 7% 6%
Tennessee Texas Utah Vermont Virginia Washington West Virginia Wisconsin Wyoming
(continued)
Bold exercises answered in back
3.3
Data set available online
111
Rate 7% 7% 7% 4% 4% 7% 8% 4% 6%
Note that dropout rates range from a low of 3% to a high of 10%. In constructing a stem-and-leaf display for these data, if we regard each dropout rate as a two-digit number and use the first digit for the stem, then there are only two possible stems, 0 and 1. One solution is to use repeated stems. Consider a scheme that divides the leaf range into five parts: 0 and 1, 2 and 3, 4 and 5, 6 and 7, and 8 and 9. Then, for example, stem 0 could be repeated as 0 0t 0f 0s 0*
with leaves 0 and 1 with leaves 2 and 3 with leaves 4 and 5 with leaves 6 and 7 with leaves 8 and 9
Construct a stem-and-leaf display for this data set that uses stems 0t, 0f, 0s, 0*, and 1. Comment on the important features of the display.
Video Solution available
Displaying Numerical Data: Frequency Distributions and Histograms A stem-and-leaf display is not always an effective way to summarize data; it is unwieldy when the data set contains a large number of observations. Frequency distributions and histograms are displays that work well for large data sets.
Frequency Distributions and Histograms for Discrete Numerical Data Discrete numerical data almost always result from counting. In such cases, each observation is a whole number. As in the case of categorical data, a frequency distribution for discrete numerical data lists each possible value (either individually or grouped into intervals), the associated frequency, and sometimes the corresponding relative frequency. Recall that relative frequency is calculated by dividing the frequency by the total number of observations in the data set. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
112
Chapter 3 Graphical Methods for Describing Data
E X A M P L E 3 . 1 2 Promiscuous Queen Bees Queen honey bees mate shortly after they become adults. During a mating flight, the queen usually takes multiple partners, collecting sperm that she will store and use throughout the rest of her life. The authors of the paper “The Curious Promiscuity of Queen Honey Bees” (Annals of Zoology [2001]: 255–265) studied the behavior of 30 queen honey bees to learn about the length of mating flights and the number of partners a queen takes during a mating flight. The accompanying data on number of partners were generated to be consistent with summary values and graphs given in the paper. Number of Partners 12 2 4 8 3 5 9 7 5
6 6 4
6 7 7
7 10 4
8 1 6
7 9 7
8 7 8
11 6 10
The corresponding relative frequency distribution is given in Table 3.1. The smallest value in the data set is 1 and the largest is 12, so the possible values from 1 to 12 are listed in the table, along with the corresponding frequency and relative frequency.
T A B L E 3.1 Relative Frequency Distribution for Number of Partners Number of Partners
Frequency
Relative Frequency
1 2 3 4 5 6 7 8 9 10 11 12 Total
1 1 1 3 2 5 7 4 2 2 1 1 30
.033 .033 .033 .100 .067 .167 .233 .133 .067 .067 .033 .033 .999
1 5 30
3 .03
1 rom ding f s n fer ou Dif to r e du
From the relative frequency distribution, we can see that five of the queen bees had six partners during their mating flight. The corresponding relative frequency, 5 30 5 .167, tells us that the proportion of queens with six partners is .167, or equivalently 16.7% of the queens had six partners. Adding the relative frequencies for the values 10, 11, and 12 gives .067 1 .033 1 .033 5 .133 indicating that 13.3% of the queens had 10 or more partners.
Data set available online Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
3.3 Displaying Numerical Data: Frequency Distributions and Histograms
113
It is possible to create a more compact frequency distribution by grouping some of the possible values into intervals. For example, we might group together 1, 2, and 3 partners to form an interval of 1–3, with a corresponding frequency of 3. The grouping of other values in a similar way results in the relative frequency distribution shown in Table 3.2.
T A B L E 3.2 Relative Frequency Distribution of Number of Partners Using Intervals Number of Partners
Frequency
Relative Frequency
1–3 4–6 7–9 10–12
3 10 13 4
.100 .333 .433 .133
A histogram for discrete numerical data is a graph of the frequency or relative frequency distribution, and it is similar to the bar chart for categorical data. Each frequency or relative frequency is represented by a rectangle centered over the corresponding value (or range of values) and the area of the rectangle is proportional to the corresponding frequency or relative frequency.
Histogram for Discrete Numerical Data When to Use Discrete numerical data. Works well, even for large data sets. How to Construct 1. Draw a horizontal scale, and mark the possible values of the variable. 2. Draw a vertical scale, and mark it with either frequency or relative frequency. 3. Above each possible value, draw a rectangle centered at that value (so that the rectangle for 1 is centered at 1, the rectangle for 5 is centered at 5, and so on). The height of each rectangle is determined by the corresponding frequency or relative frequency. Often possible values are consecutive whole numbers, in which case the base width for each rectangle is 1.
What to Look For • • • • •
Center or typical value Extent of spread or variability General shape Location and number of peaks Presence of gaps and outliers
E X A M P L E 3 . 1 3 Revisiting Promiscuous Queen Bees The queen bee data of Example 3.12 were summarized in a frequency distribution. The corresponding histogram is shown in Figure 3.17. Note that each rectangle in the histogram is centered over the corresponding value. When relative frequency instead of frequency is used for the vertical scale, the scale on the vertical axis is different but all essential characteristics of the graph (shape, location, spread) are unchanged. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
114
Chapter 3 Graphical Methods for Describing Data
Frequency
Relative frequency .25
7 6
.20
5 .15
4 3
.10
2 .05 1 0
0 1
2
3
4
9 5 6 7 8 Number of partners
FIGURE 3.17 Histogram and relative frequency histogram of queen bee data.
10
11
12
1
2
3
4
9 5 6 7 8 Number of partners
10
11
12
A histogram based on the grouped frequency distribution of Table 3.2 can be constructed in a similar fashion, and is shown in Figure 3.18. A rectangle represents the frequency or relative frequency for each interval. For the interval 1–3, the rectangle extends from .5 to 3.5 so that there are no gaps between the rectangles of the histogram. Frequency 14 12 10 8 6 4 2
FIGURE 3.18 Histogram of queen bee data using intervals.
0 1
2
3
4
9 5 6 7 8 Number of partners
10
11
12
Sometimes a discrete numerical data set contains a large number of possible values and perhaps also has a few large or small values that are far away from most of the data. In this case, rather than forming a frequency distribution with a very long list of possible values, it is common to group the observed values into intervals or ranges. This is illustrated in Example 3.14.
E X A M P L E 3 . 1 4 Math SAT Score Distribution Each of the 1,530,128 students who took the math portion of the SAT exam in 2009 received a score between 200 and 800. The score distribution was summarized in a frequency distribution table that appeared in the College Board report titled “2009 College Bound Seniors.” A relative frequency distribution is given in Table 3.3 and Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
3.3 Displaying Numerical Data: Frequency Distributions and Histograms
115
T A B L E 3.3 Relative Frequency Distribution of Math SAT Score Math SAT Score
Frequency
Relative Frequency
200–299 300–399 400–499 500–599 600–699 700–800
97,296 295,693 449,238 454,497 197,741 35,663
0.064 0.193 0.294 0.297 0.129 0.023
Relative frequency .30 .25 .20 .15 .10 .05
FIGURE 3.19 Relative frequency histogram for the math SAT data.
0 199.5
299.5
399.5 499.5 599.5 Math SAT score
699.5
799.5
the corresponding relative frequency histogram is shown in Figure 3.19. Notice that rather than list each possible individual score value between 200 and 800, the scores are grouped into intervals (200 to 299, 300 to 399, etc.). This results in a much more compact table that still communicates the important features of the data set. Also, notice that because the data set is so large, the frequencies are also large numbers. Because of these large frequencies, it is easier to focus on the relative frequencies in our interpretation. From the relative frequency distribution and histogram, we can see that while there is a lot of variability in individual math SAT scores, the majority were in the 400 to 600 range and a typical value for math SAT looks to be something in the low 500s. Before leaving this example, take a second look at the relative frequency histogram of Figure 3.19. Notice that there is one rectangle for each score interval in the relative frequency distribution. For simplicity we have chosen to treat the very last interval, 700 to 800, as if it were 700 to 799 so that all of the score ranges in the frequency distribution are the same width. Also note that the rectangle representing the score range 400 to 499 actually extends from 399.5 to 499.5 on the score scale. This is similar to what happens in histograms for discrete numerical data where there is no grouping. For example, in Figure 3.17 the rectangle representing 2 is centered at 2 but extends from 1.5 to 2.5 on the number of partners scale.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
116
Chapter 3
Graphical Methods for Describing Data
Frequency Distributions and Histograms for Continuous Numerical Data The difficulty in constructing tabular or graphical displays with continuous data, such as observations on reaction time (in seconds) or weight of airline passenger carry-on luggage (in pounds), is that there are no natural categories. The way out of this dilemma is to define our own categories. For carry-on luggage weight, we might expect weights up to about 30 pounds. One way to group the weights into 5-pound intervals is shown in Figure 3.20. Then each observed data value could be classified into one of these intervals. The intervals used are sometimes called class intervals. The class intervals play the same role that the categories or individual values played in frequency distributions for categorical or discrete numerical data. FIGURE 3.20 Suitable class intervals for carry-on luggage weight data.
5
10
15
20
25
30
There is one further difficulty we need to address. Where should we place an observation such as 20, which falls on a boundary between classes? Our convention is to define intervals so that such an observation is placed in the upper rather than the lower class interval. Thus, in a frequency distribution, one class might be 15 to ⬍20, where the symbol ⬍ is a substitute for the phrase less than. This class will contain all observations that are greater than or equal to 15 and less than 20. The observation 20 would then fall in the class 20 to ⬍25.
E X A M P L E 3 . 1 5 Enrollments at Public Universities States differ widely in the percentage of college students who are enrolled in public institutions. The National Center for Education Statistics provided the accompanying data on this percentage for the 50 U.S. states for fall 2007. Percentage of College Students Enrolled in Public Institutions 96 86 81 84 77 90 73 53 90 93 76 86 78 76 88 86 87 64 89 86 80 66 70 90 89 82 73 72 56 55 75 77 82 83 79 75 43 50 64 80 82 75
96 60 81 59
73 58 73 59
The smallest observation is 46 (Massachusetts) and the largest is 96 (Alaska and Wyoming). It is reasonable to start the first class interval at 40 and let each interval have a width of 10. This gives class intervals of 40 to ⬍50, 50 to ⬍60, 60 to ⬍70, 70 to ⬍80, 80 to ⬍90, and 90 to ⬍100. Table 3.4 displays the resulting frequency distribution, along with the relative frequencies.
Data set available online Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
3.3 Displaying Numerical Data: Frequency Distributions and Histograms
117
T A B L E 3.4 Frequency Distribution for Percentage of College Students Enrolled in Public Institutions Class Interval
Frequency
Relative Frequency
40 to ⬍50 50 to ⬍60 60 to ⬍70 70 to ⬍80 80 to ⬍90 90 to ⬍100
1 7 4 15 17 6 50
.02 .14 .08 .30 .34 .12 1.00
Various relative frequencies can be combined to yield other interesting information. For example, proportion of states proportion in 40 proportion in 50 ° with percent in public ¢ 5 a b1a b to ,50 class to ,60 class institutions less than 60 5 .02 1 .14 5 .16 116%2 and proportion of states proportion proportion proportion with percent in ± ≤ 5 ° in 60 to ¢ 1 ° in 70 to ¢ 1 ° in 80 to ¢ public institutions ,70 class ,80 class ,90 class between 60 and 90 5 .08 1 .30 1 .34 5 .72 172%2
There are no set rules for selecting either the number of class intervals or the length of the intervals. Using a few relatively wide intervals will bunch the data, whereas using a great many relatively narrow intervals may spread the data over too many intervals, so that no interval contains more than a few observations. Neither type of distribution will give an informative picture of how values are distributed over the range of measurement, and interesting features of the data set may be missed. In general, with a small amount of data, relatively few intervals, perhaps between 5 and 10, should be used. With a large amount of data, a distribution based on 15 to 20 (or even more) intervals is often recommended. The quantity "number of observations is often used as an estimate of an appropriate number of intervals: 5 intervals for 25 observations, 10 intervals when the number of observations is 100, and so on. Two people making reasonable and similar choices for the number of intervals, their width, and the starting point of the first interval will usually obtain similar histograms of the data.
Histograms for Continuous Numerical Data When the class intervals in a frequency distribution are all of equal width, it is easy to construct a histogram using the information in a frequency distribution. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
118
Chapter 3
Graphical Methods for Describing Data
Histogram for Continuous Numerical Data When the Class Interval Widths are Equal When to Use Continuous numerical data. Works well, even for large data sets. How to Construct 1. Mark the boundaries of the class intervals on a horizontal axis. 2. Use either frequency or relative frequency on the vertical axis. 3. Draw a rectangle for each class directly above the corresponding interval (so that the edges are at the class interval boundaries). The height of each rectangle is the frequency or relative frequency of the corresponding class interval.
What to Look For • • • • •
Center or typical value Extent of spread, variability General shape Location and number of peaks Presence of gaps and outliers
E X A M P L E 3 . 1 6 TV Viewing Habits of Children The article “Early Television Exposure and Subsequent Attention Problems in Children” (Pediatrics, April 2004) investigated the television viewing habits of children in the United States. Table 3.5 gives approximate relative frequencies (read from graphs that appeared in the article) for the number of hours spent watching TV per day for a sample of children at age 1 year and a sample of children at age 3 years. The data summarized in the article were obtained as part of a large scale national survey.
T A B L E 3.5 Relative Frequency Distribution for Number of Hours Spent Watching TV per Day
Step-by-step technology instructions available online
TV Hours per Day
Age 1 Year Relative Frequency
Age 3 Years Relative Frequency
0 to ⬍2 2 to ⬍4 4 to ⬍6 6 to ⬍8 8 to ⬍10 10 to ⬍12 12 to ⬍14 14 to ⬍16
.270 .390 .190 .085 .030 .020 .010 .005
.630 .195 .100 .025 .020 .015 .010 .005
Figure 3.21(a) is the relative frequency histogram for the 1-year-old children and Figure 3.21(b) is the relative frequency histogram for 3-year-old children. Notice that both histograms have a single peak with the majority of children in both age groups concentrated in the smaller TV hours intervals. Both histograms are quite stretched out at the upper end, indicating some young children watch a lot of TV. The big difference between the two histograms is at the low end, with a much higher proportion of 3-year-old children falling in the 0 to 2 TV hours interval than is the case for 1-year-old children. A typical number of TV hours per day for 1-yearold children would be somewhere between 2 and 4 hours, whereas a typical number of TV hours for 3-year-old children is in the 0 to 2 hours interval.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
3.3 Displaying Numerical Data: Frequency Distributions and Histograms
Relative frequency
119
Relative frequency
.7
.7
.6
.6
.5
.5
.4
.4
.3
.3
.2
.2
.1
.1
0 0
2
4
6 8 10 12 Hours of TV per day
14
16
(a)
2
4
6 8 10 12 Hours of TV per day
14
16
(b)
FIGURE 3.21 Histogram of TV hours per day: (a) 1-year-old children; (b) 3-year-old children.
FIGURE 3.22 Three choices of class intervals for a data set with outliers: (a) many short intervals of equal width; (b) a few wide intervals of equal width; (c) intervals of unequal width.
Figure 3.22 shows a data set in which a great many observations are concentrated at the center of the data set, with only a few unusual, or stray, values both below and above the main body of data. If a frequency distribution is based on short intervals of equal width, a great many intervals will be required to capture all observations, and many of them will contain no observations, as shown in Figure 3.22(a). On the other hand, only a few wide intervals will capture all values, but then most of the observations will be grouped into a few intervals, as shown in Figure 3.22(b). In such situations, it is best to use a combination of wide class intervals where there are few data points and shorter intervals where there are many data points, as shown in Figure 3.22(c).
Class Intervals of Unequal Widths
(a) (b) (c)
Constructing a Histogram for Continuous Data When Class Interval Widths are Unequal When class intervals are not of equal width, frequencies or relative frequencies should not be used on the vertical axis. Instead, the height of each rectangle, called the density for the class interval, is given by density 5 rectangle height 5
relative frequency of class interval class interval width
The vertical axis is called the density scale. The use of the density scale to construct the histogram ensures that the area of each rectangle in the histogram will be proportional to the corresponding relative frequency. The formula for density can also be used when class widths are equal. However, when the intervals are of equal width, the extra arithmetic required to obtain the densities is unnecessary. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
120
Chapter 3 Graphical Methods for Describing Data
E X A M P L E 3 . 1 7 Misreporting Grade Point Average When people are asked for the values of characteristics such as age or weight, they sometimes shade the truth in their responses. The article “Self-Reports of Academic Performance” (Social Methods and Research [November 1981]: 165–185) focused on such characteristics as SAT scores and grade point average (GPA). For each student in a sample, the difference in GPA (reported – actual) was determined. Positive differences resulted from individuals reporting GPAs larger than the correct values. Most differences were close to 0, but there were some rather large errors. Because of this, the frequency distribution based on unequal class widths shown in Table 3.6 gives an informative yet concise summary.
T A B L E 3.6 Frequency Distribution for Errors in Reported GPA Class Interval
Relative Frequency
Width
Density
⫺2.0 to ⬍⫺0.4 ⫺0.4 to ⬍⫺0.2 ⫺0.2 to ⬍⫺0.1 ⫺0.1 to ⬍0 0 to ⬍0.1 0.1 to ⬍0.2 0.2 to ⬍0.4 0.4 to ⬍2.0
.023 .055 .097 .210 .189 .139 .116 .171
1.6 .2 .1 .1 .1 .1 .2 1.6
0.014 0.275 0.970 2.100 1.890 1.390 0.580 0.107
Density
2.0 1.5 1.0
ph correct gra
0.5
–2.0
–0.4
0.4
2.0
(a) Relative frequency
.20 .15
aph incorrect gr
.10
FIGURE 3.23 Histograms for errors in reporting GPA: (a) a correct histogram (height ⫽ density); (b) an incorrect histogram (height ⫽ relative frequency).
.05
–2.0
–0.4
0.4
2.0
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
3.3 Displaying Numerical Data: Frequency Distributions and Histograms
121
Figure 3.23 displays two histograms based on this frequency distribution. The histogram in Figure 3.23(a) is correctly drawn, with density used to determine the height of each bar. The histogram in Figure 3.23(b) has height equal to relative frequency and is therefore not correct. In particular, this second histogram considerably exaggerates the incidence of grossly overreported and underreported values—the areas of the two most extreme rectangles are much too large. The eye is naturally drawn to large areas, so it is important that the areas correctly represent the relative frequencies.
Histogram Shapes General shape is an important characteristic of a histogram. In describing various shapes it is convenient to approximate the histogram itself with a smooth curve (called a smoothed histogram). This is illustrated in Figure 3.24.
FIGURE 3.24 Approximating a histogram with a smooth curve.
One description of general shape relates to the number of peaks, or modes.
DEFINITION A histogram is said to be unimodal if it has a single peak, bimodal if it has two peaks, and multimodal if it has more than two peaks. These shapes are illustrated in Figure 3.25.
FIGURE 3.25 Smoothed histograms with various numbers of modes: (a) unimodal; (b) bimodal; (c) multimodal.
(a)
(b)
(c)
Bimodality sometimes occurs when the data set consists of observations on two quite different kinds of individuals or objects. For example, consider a large data set consisting of driving times for automobiles traveling between San Luis Obispo, California, and Monterey, California. This histogram would show two peaks, one for those cars that took the inland route (roughly 2.5 hours) and another for those cars traveling up the coast highway (3.5–4 hours). However, bimodality does not automatically follow in such situations. Bimodality will occur in the histogram of the combined groups only if the centers of the two separate histograms are far apart relative to the variability in the two data sets. Thus, a large data set consisting of heights of college students would probably not produce a bimodal histogram because the typical height for males (about 69 in.) and the typical height for females (about 66 in.) are not very far apart. Many histograms encountered in practice are unimodal, and multimodality is not as common. Unimodal histograms come in a variety of shapes. A unimodal histogram is symmetric if there is a vertical line of symmetry such that the part of the histogram to the left of the line is a mirror image of the part to the right. (Bimodal and multimodal Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
122
Chapter 3 Graphical Methods for Describing Data
FIGURE 3.26 Several symmetric unimodal smoothed histograms.
histograms can also be symmetric in this way.) Several different symmetric smoothed histograms are shown in Figure 3.26. Proceeding to the right from the peak of a unimodal histogram, we move into what is called the upper tail of the histogram. Going in the opposite direction moves us into the lower tail.
DEFINITION A unimodal histogram that is not symmetric is said to be skewed. If the upper tail of the histogram stretches out much farther than the lower tail, then the distribution of values is positively skewed or right skewed. If, on the other hand, the lower tail is much longer than the upper tail, the histogram is negatively skewed or left skewed. These two types of skewness are illustrated in Figure 3.27. Positive skewness is much more frequently encountered than is negative skewness. An example of positive skewness occurs in the distribution of single-family home prices in Los Angeles County; most homes are moderately priced (at least for California), whereas the relatively few homes in Beverly Hills and Malibu have much higher price tags.
FIGURE 3.27 Two examples of skewed smoothed histograms: (a) positive skew; (b) negative skew.
(a)
(b)
One rather specific shape, a normal curve, arises more frequently than any other in statistical applications. Many histograms can be well approximated by a normal curve (for example, characteristics such as arm span and the weight of an apple). Here we briefly mention several of the most important qualitative properties of normal curves, postponing a more detailed discussion until Chapter 7. A normal curve is both symmetric and bell-shaped; it looks like the curve in Figure 3.28(a). However, not all bell-shaped curves are normal. In a normal curve, starting from the top of the bell the height of the curve decreases at a well-defined rate when moving toward either tail. (This rate of decrease is specified by a certain mathematical function.)
FIGURE 3.28 Three examples of bell-shaped histograms: (a) normal; (b) heavy-tailed; (c) light-tailed.
(a)
(b)
(c)
A curve with tails that do not decline as rapidly as the tails of a normal curve is called heavy-tailed (compared to the normal curve). Similarly, a curve with tails that decrease more rapidly than the normal tails is called light-tailed. Figures 3.28(b) and 3.28(c) illustrate these possibilities. The reason that we are concerned about the tails Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
3.3 Displaying Numerical Data: Frequency Distributions and Histograms
123
in a distribution is that many inferential procedures that work well (i.e., they result in accurate conclusions) when the population distribution is approximately normal perform poorly when the population distribution is heavy-tailed.
Do Sample Histograms Resemble Population Histograms? Sample data are usually collected to make inferences about a population. The resulting conclusions may be in error if the sample is unrepresentative of the population. So how similar might a histogram of sample data be to the histogram of all population values? Will the two histograms be centered at roughly the same place and spread out to about the same extent? Will they have the same number of peaks, and will the peaks occur at approximately the same places? A related issue concerns the extent to which histograms based on different samples from the same population resemble one another. If two different sample histograms can be expected to differ from one another in obvious ways, then at least one of them might differ substantially from the population histogram. If the sample differs substantially from the population, conclusions about the population based on the sample are likely to be incorrect. Sampling variability—the extent to which samples differ from one another and from the population—is a central idea in statistics. Example 3.18 illustrates sampling variability in histogram shapes.
E X A M P L E 3 . 1 8 What You Should Know About Bus Drivers . . . A sample of 708 bus drivers employed by public corporations was selected, and the number of traffic accidents in which each bus driver was involved during a 4-year period was determined (“Application of Discrete Distribution Theory to the Study of Non-
communicable Events in Medical Epidemiology,” in Random Counts in Biomedical and Social Sciences, G. P. Patil, ed. [University Park, PA: Pennsylvania State University Press, 1970]). A listing of the 708 sample observations might look like this: 306002141...602 The frequency distribution (Table 3.7) shows that 117 of the 708 drivers had no accidents, a relative frequency of 117/708 ⫽ .165 (or 16.5%). Similarly, the proportion
T A B L E 3.7 Frequency Distribution for Number of Accidents by Bus Drivers
Data set available online
Number of Accidents
Frequency
Relative Frequency
0 1 2 3 4 5 6 7 8 9 10 11
117 157 158 115 78 44 21 7 6 1 3 1 708
.165 .222 .223 .162 .110 .062 .030 .010 .008 .001 .004 .001 .998
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
124
Chapter 3
Graphical Methods for Describing Data
of sampled drivers who had 1 accident is .222 (or 22.2%). The largest sample observation was 11. Although the 708 observations actually constituted a sample from the population of all bus drivers, we will regard the 708 observations as constituting the entire population. The first histogram in Figure 3.29, then, represents the population histogram. The other four histograms in Figure 3.29 are based on four different samples of 50 observations each selected at random from this population. The five histograms
Relative frequency
.25 Population histogram
.20 .15 .10 .05
1
2
3 4 5 6 7 Number of accidents
Relative frequency
8
9
10 11
Relative frequency
.25 Sample histogram 1
.20
Sample histogram 2
.15 .10 .05 0 1 2 3 4 5 6 7 8 9 10 11 Relative frequency
0 1 2 3 4 5 6 7 8 9 10 11 Relative frequency
.35 .30 .25 .20
Sample histogram 3
Sample histogram 4
.15 .10
FIGURE 3.29 Comparison of population and sample histograms for number of accidents.
.05 0 1 2 3 4 5 6 7 8 9 10 11
0 1 2 3 4 5 6 7 8 9 10 11
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
3.3 Displaying Numerical Data: Frequency Distributions and Histograms
125
certainly resemble one another in a general way, but some dissimilarities are also obvious. The population histogram rises to a peak and then declines smoothly, whereas the sample histograms tend to have more peaks, valleys, and gaps. Although the population data set contained an observation of 11, none of the four samples did. In fact, in the first two samples, the largest observations were 7 and 8, respectively. In Chapters 8–15 we will see how sampling variability can be described and taken into account when we use sample data to draw conclusions about a population.
Cumulative Relative Frequencies and Cumulative Relative Frequency Plots Rather than wanting to know what proportion of the data fall in a particular class, we often wish to determine the proportion falling below a specified value. This is easily done when the value is a class boundary. Consider the following intervals and relative frequencies for carry-on luggage weight for passengers on flights between Phoenix and New York City during October 2009: Class Relative frequency
0 to 5 .05
5 to ⬍10 .10
10 to ⬍15 .18
15 to ⬍20 .25
... ...
Then proportion of passengers with carry-on luggage weight less than 15 lbs. ⫽ proportion in one of the first three classes ⫽ .05 ⫹ .10 ⫹ .18 ⫽ .33 Similarly, proportion of passengers with carry-on luggage weight less than 20 lbs. ⫽ .05 ⫹ .10 ⫹ .18 ⫹ .25 ⫽ .33 ⫹ .25 ⫽ .58 Each such sum of relative frequencies is called a cumulative relative frequency. Notice that the cumulative relative frequency .58 is the sum of the previous cumulative relative frequency .33 and the “current” relative frequency .25. The use of cumulative relative frequencies is illustrated in Example 3.19.
E X A M P L E 3 . 1 9 Albuquerque Rainfall The National Climatic Data Center has been collecting weather data for many years. Annual rainfall totals for Albuquerque, New Mexico, from 1950 to 2008 (www .ncdc.noaa.gov/oa/climate/research/cag3/city.html) were used to construct the relative frequency distribution shown in Table 3.8. The table also contains a column of cumulative relative frequencies. The proportion of years with annual rainfall less than 10 inches is .585, the cumulative relative frequency for the 9 to ⬍10 interval. What about the proportion of years with annual rainfall less than 8.5 inches? Because 8.5 is not the endpoint of one of the intervals in the frequency distribution, we can only estimate this from the information given. The value 8.5 is halfway between the endpoints of the 8 to 9 inter-
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
126
Chapter 3
Graphical Methods for Describing Data
T A B L E 3.8 Relative Frequency distribution for Albuquerque Rainfall Data with Cumulative Relative Frequencies Annual Rainfall (inches)
Frequency
Relative Frequency
Cumulative Relative Frequency
4 to ⬍5 5 to ⬍6 6 to ⬍7
3 6 5
0.052 0.103 0.086
7 to ⬍8 8 to ⬍9 9 to ⬍10 10 to ⬍11 11 to ⬍12 12 to ⬍13 13 to ⬍14
6 10 4 12 6 3 3
0.103 0.172 0.069 0.207 0.103 0.052 0.052
0.052 0.155 ⫽ .052 ⫹ .103 0.241 ⫽ .052 ⫹ .103 ⫹ .086 or .155 ⫹ .086 0.344 0.516 0.585 0.792 0.895 0.947 0.999
val, so it is reasonable to estimate that half of the relative frequency of .172 for this interval belongs in the 8 to 8.5 range. Then estimate of proportion of 1 ° years with rainfall less ¢ 5 .052 1 .103 1 .086 1 .103 1 1.1722 5 .430 2 than 8.5 inches This proportion could also have been computed using the cumulative relative frequencies as estimate of proportion of 1 ° years with rainfall less ¢ 5 .344 1 1.1722 5 .430 2 than 8.5 inches Similarly, since 11.25 is one-fourth of the way between 11 and 12, estimate of proportion of 1 ° years with rainfall less ¢ 5 .792 1 1.1032 5 .818 4 than 11.25 inches
A cumulative relative frequency plot is just a graph of the cumulative relative frequencies against the upper endpoint of the corresponding interval. The pairs (upper endpoint of interval, cumulative relative frequency) are plotted as points on a rectangular coordinate system, and successive points in the plot are connected by a line segment. For the rainfall data of Example 3.19, the plotted points would be (5, .052) (10, .585)
(6, .155) (11, .792)
(7, .241) (12, .895)
(8, .344) (13, .947)
(9, .516) (14, .999)
One additional point, the pair (lower endpoint of first interval, 0), is also included in the plot (for the rainfall data, this would be the point (4 0)), and then points are connected by line segments. Figure 3.30 shows the cumulative relative Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
3.3 Displaying Numerical Data: Frequency Distributions and Histograms
127
frequency plot for the rainfall data. The cumulative relative frequency plot can be used to obtain approximate answers to questions such as What proportion of the observations is smaller than a particular value? and What value separates the smallest p percent from the larger values? Cumulative relative frequency 1.0
0.8
0.6
0.4
0.2
FIGURE 3.30
0.0
Cumulative relative frequency plot for the rainfall data of Example 3.19.
3
4
5
6
7
8 9 10 11 12 13 14 Rainfall
For example, to determine the approximate proportion of years with annual rainfall less than 9.5 inches, we would follow a vertical line up from 9.5 on the x-axis and then read across to the y-axis to obtain the corresponding relative frequency, as illustrated in Figure 3.31(a). Approximately .55, or 55%, of the years had annual rainfall less than 9.5 inches. Cumulative relative frequency
Cumulative relative frequency 1.0
1.0
0.9
0.9
0.8
0.8
0.7
0.7
0.6
0.6
0.5
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0.0
3
4
5
6
7
8 9 10 11 12 13 14 Rainfall (a)
0.0
3
4
5
6
7
8 9 10 11 12 13 14 Rainfall (b)
FIGURE 3.31 Using the cumulative relative frequency plot. (a) Determining the approximate proportion of years with annual rainfall less than 9.5 inches. (b) Finding the amount of rainfall that separates the 30% of years with the lowest rainfall from the 70% with higher rainfall. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
128
Chapter 3 Graphical Methods for Describing Data
Similarly, to find the amount of rainfall that separates the 30% of years with the smallest annual rainfall from years with higher rainfall, start at .30 on the cumulative relative frequency axis and move across and then down to find the corresponding rainfall amount, as shown in Figure 3.31(b). Approximately 30% of the years had annual rainfall of less than 7.6 inches.
EX E RC I S E S 3 . 2 2 - 3 . 3 7 The article “Americans on the Move” (USA Today, November 30, 2007) included the data in the
3.22
accompanying table. Entries in the table are the percentage of state residents who had moved during 2006.
State Alabama Alaska Arizona Arkansas California Colorado Connecticut Delaware District of Columbia Florida Georgia Hawaii Idaho Illinois Indiana Iowa Kansas Kentucky Louisiana Maine Maryland Massachusetts Michigan Minnesota Mississippi Missouri Montana Nebraska Nevada New Hampshire New Jersey New Mexico New York North Carolina
Percentage of Residents Who Moved During 2006 16.1 21.2 20.2 18.9 15.9 19.6 13.1 14.0 18.8 17.4 18.8 14.5 21.0 15.0 16.8 17.0 18.7 16.8 18.9 14.4 14.5 13.6 14.2 14.2 17.2 17.5 17.5 18.0 22.0 13.7 11.1 16.8 11.5 17.5
Percentage of Residents Who Moved During 2006
State North Dakota Ohio Oklahoma Oregon Pennsylvania Rhode Island South Carolina South Dakota Tennessee Texas Utah Vermont Virginia Washington West Virginia Wisconsin Wyoming
17.2 15.7 19.2 20.2 12.7 13.4 16.6 16.7 16.6 19.1 20.7 14.5 16.3 19.5 12.7 15.3 18.8
Construct a histogram of these data using class intervals of 10 to ⬍12, 12 to ⬍14, 14 to ⬍16, and so on. Write a few sentences to describe the shape, center, and spread of the distribution.
3.23 The accompanying data on annual maximum wind speed (in meters per second) in Hong Kong for each year in a 45-year period were given in an article that appeared in the journal Renewable Energy (March, 2007). Use the annual maximum wind speed data to construct a histogram. Is the histogram approximately symmetric, positively skewed, or negatively skewed? Would you describe the histogram as unimodal, bimodal, or multimodal? 30.3 27.2 37.0 28.3 37.5
39.0 52.9 34.4 39.1 31.5
33.9 45.8 35.5 55.0 32.0
38.6 63.3 62.2 35.0 35.5
44.6 36.0 30.3 28.8 37.5
31.4 64.0 40.0 25.7 41.0
26.7 31.4 36.0 62.7 37.5
51.9 42.2 39.4 32.4 48.6
31.9 41.1 34.4 31.9 28.1
(continued) Bold exercises answered in back
Data set available online
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
3.3 Displaying Numerical Data: Frequency Distributions and Histograms
The accompanying relative frequency table is based on data from the 2007 College Bound Seniors Report for California (College Board, 2008).
% of Workers who Belong to a Union
3.24
Score on SAT Reasoning Exam
Relative Frequency for Males
Relative Frequency for Females
200 to ⬍250 250 to ⬍ 300 300 to ⬍350 350 to ⬍ 400 400 to ⬍ 450 450 to ⬍ 500 500 to ⬍ 550 550 to ⬍600 600 to ⬍ 650 650 to ⬍ 700 700 to ⬍ 750 750 to ⬍800
.0404 .0546 .1076 .1213 .1465 .1556 .1400 .1126 .0689 .0331 .0122 .0072
.0183 .0299 .0700 .0896 .1286 .1540 .1667 .1550 .1050 .0529 .0194 .0105
a. Construct a relative frequency histogram for SAT reasoning score for males. b. Construct a relative frequency histogram for SAT reasoning score for females. c. Based on the histograms from Parts (a) and (b), write a few sentences commenting on the similarities and differences in the distribution of SAT reasoning scores for males and females. The data in the accompanying table represents the percentage of workers who are members of a union for each U.S. state and the District of Columbia (AARP Bulletin, September 2009).
3.25
State Alabama Alaska Arizona Arkansas California Colorado Connecticut Delaware District of Columbia Florida Georgia Hawaii Idaho Illinois Indiana
% of Workers who Belong to a Union 9.8 23.5 8.8 5.9 18.4 8.0 16.9 12.2 13.4 6.4 3.7 24.3 7.1 16.6 12.4 (continued)
Bold exercises answered in back
Data set available online
129
State Iowa Kansas Kentucky Louisiana Maine Maryland Massachusetts Michigan Minnesota Mississippi Missouri Montana Nebraska Nevada New Hampshire New Jersey New Mexico New York North Carolina North Dakota Ohio Oklahoma Oregon Pennsylvania Rhode Island South Carolina South Dakota Tennessee Texas Utah Vermont Virginia Washington West Virginia Wisconsin Wyoming
10.6 7.0 8.6 4.6 12.3 15.7 12.6 18.8 16.1 5.3 11.2 12.2 8.3 16.7 3.5 6.1 10.6 18.3 7.2 24.9 14.2 6.6 16.6 15.4 16.5 3.9 5.0 5.5 4.5 5.8 4.1 10.4 19.8 13.8 15.0 7.7
a. Construct a histogram of these data using class intervals of 0 to ⬍5, 5 to ⬍10, 10 to ⬍15, 15 to ⬍20, and 20 to ⬍25. b. Construct a dotplot of these data. Comment on the interesting features of the plot. c. For this data set, which is a more informative graphical display—the dotplot from Part (b) or the histogram constructed in Part (a)? Explain. d. Construct a histogram using about twice as many class intervals as the histogram in Part (a). Use 2.5 to ⬍5 as the first class interval. Write a few sentences that explain why this histogram does a better job of displaying this data set than does the histogram in Part (a). Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
130
Chapter 3 Graphical Methods for Describing Data
Medicare’s new medical plans offer a wide range of variations and choices for seniors when picking a drug plan (San Luis Obispo Tribune, November 25, 2005). The monthly cost for a stand-alone drug plan varies from plan to plan and from state to state. The accompanying table gives the premium for the plan with the lowest cost for each state.
3.26
State Alabama Alaska Arizona Arkansas California Colorado Connecticut Delaware District of Columbia Florida Georgia Hawaii Idaho Illinois Indiana Iowa Kansas Kentucky Louisiana Maine Maryland Massachusetts Michigan Minnesota Mississippi Missouri Montana Nebraska Nevada New Hampshire New Jersey New Mexico New York North Carolina North Dakota Ohio Oklahoma Oregon Pennsylvania Rhode Island South Carolina South Dakota Tennessee Bold exercises answered in back
State Texas Utah Vermont Virginia Washington West Virginia Wisconsin Wyoming
Cost per Month (dollars) 14.08 20.05 6.14 10.31 5.41 8.62 7.32 6.44 6.44 10.35 17.91 17.18 6.33 13.32 12.30 1.87 9.48 12.30 17.06 19.60 6.44 7.32 13.75 1.87 11.60 10.29 1.87 1.87 6.42 19.60 4.43 10.65 4.10 13.27 1.87 14.43 10.07 6.93 10.14 7.32 16.57 1.87 14.08 (continued) Data set available online
Cost per Month (dollars) 10.31 6.33 7.32 8.81 6.93 10.14 11.42 1.87
a. Use class intervals of $0 to ⬍$3, $3 to ⬍$6, $6 to ⬍$9, etc., to create a relative frequency distribution for these data. b. Construct a histogram and comment on its shape. c. Using the relative frequency distribution or the histogram, estimate the proportion of the states that have a minimum monthly plan of less than $13.00 a month. The following two relative frequency distributions were constructed using data that appeared in the report “Undergraduate Students and Credit Cards in 2004” (Nellie Mae, May 2005). One relative frequency distribution is based on credit bureau data for a random sample of 1413 college students, while the other is based on the result of a survey completed by 132 of the 1260 college students who received the survey.
3.27
Credit Card Balance (dollars)— Credit Bureau Data
Relative Frequency
0 to ⬍100 100 to ⬍500 500 to ⬍1000 1000 to ⬍2000 2000 to ⬍3000 3000 to ⬍7000 7000 or more
.18 .19 .14 .16 .10 .16 .07
Credit Card Balance (dollars)— Survey Data
Relative Frequency
0 to ⬍100 100 to ⬍500 500 to ⬍1000 1000 to ⬍2000 2000 to ⬍3000 3000 to ⬍7000 7000 or more
.18 .22 .17 .22 .07 .14 .00
a. Construct a histogram for the credit bureau data. For purposes of constructing the histogram, assume that none of the students in the sample had a balance Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
3.3 Displaying Numerical Data: Frequency Distributions and Histograms
higher than $15,000 and that the last interval can be regarded as 7000 to ⬍15,000. Be sure to use the density scale when constructing the histogram. b. Construct a histogram for the survey data. Use the same scale that you used for the histogram in Part (a) so that it will be easy to compare the two histograms. c. Comment on the similarities and differences in the histograms from Parts (a) and (b). d. Do you think the high nonresponse rate for the survey may have contributed to the observed differences in the two histograms? Explain. U.S. Census data for San Luis Obispo County, California, were used to construct the following frequency distribution for commute time (in minutes) of working adults (the given frequencies were read from a graph that appeared in the San Luis Obispo Tribune [September 1, 2002] and so are only approximate):
3.28
Commute Time
Frequency
0 to ⬍5 5 to ⬍10 10 to ⬍15 15 to ⬍20 20 to ⬍25 25 to ⬍30 30 to ⬍35 35 to ⬍40 40 to ⬍45 45 to ⬍60 60 to ⬍90 90 to ⬍120
5,200 18,200 19,600 15,400 13,800 5,700 10,200 2,000 2,000 4,000 2,100 2,200
a. Notice that not all intervals in the frequency distribution are equal in width. Why do you think that unequal width intervals were used? b. Construct a table that adds a relative frequency and a density column to the given frequency distribution (see Example 3.17). c. Use the densities computed in Part (b) to construct a histogram for this data set. (Note: The newspaper displayed an incorrectly drawn histogram based on frequencies rather than densities!) Write a few sentences commenting on the important features of the histogram. d. Compute the cumulative relative frequencies, and construct a cumulative relative frequency plot. e. Use the cumulative relative frequency plot constructed in Part (d) to answer the following questions. Bold exercises answered in back
Data set available online
131
i. Approximately what proportion of commute times were less than 50 minutes? ii. Approximately what proportion of commute times were greater than 22 minutes? iii. What is the approximate commute time value that separates the shortest 50% of commute times from the longest 50%?
3.29 Student loans can add up, especially for those attending professional schools to study in such areas as medicine, law, or dentistry. Researchers at the University of Washington studied medical students and gave the following information on the educational debt of medical students on completion of their residencies (Annals of Internal Medicine [March 2002]: 384–398): Educational Debt (dollars)
Relative Frequency
0 to ⬍5000 5000 to ⬍20,000 20,000 to ⬍50,000 50,000 to ⬍100,000 100,000 or more
.427 .046 .109 .232 .186
a. What are two reasons that you could not use the given information to construct a histogram with the educational debt intervals on the horizontal axis and relative frequency on the y-axis? b. Suppose that no student had an educational debt of $150,000 or more upon completion of his or her residency, so that the last class in the relative frequency distribution would be 100,000 to ⬍150,000. Summarize this distribution graphically by constructing a histogram of the educational debt data. (Don’t forget to use the density scale for the heights of the bars in the histogram, because the interval widths aren’t all the same.) c. Based on the histogram of Part (b), write a few sentences describing the educational debt of medical students completing their residencies.
3.30 An exam is given to students in an introductory statistics course. What is likely to be true of the shape of the histogram of scores if: a. the exam is quite easy? b. the exam is quite difficult? c. half the students in the class have had calculus, the other half have had no prior college math courses, and the exam emphasizes mathematical manipulation? Explain your reasoning in each case. Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
132
Chapter 3 Graphical Methods for Describing Data
3.31 The accompanying frequency distribution summarizes data on the number of times smokers who had successfully quit smoking attempted to quit before their final successful attempt (“Demographic Variables,
Smoking Variables, and Outcome Across Five Studies,” Health Psychology [2007]: 278–287). Number of Attempts
Frequency
0 1 2 3–4 5 or more
778 306 274 221 238
Assume that no one had made more than 10 unsuccessful attempts, so that the last entry in the frequency distribution can be regarded as 5–10 attempts. Summarize this data set using a histogram. Be careful—the class intervals are not all the same width, so you will need to use a density scale for the histogram. Also remember that for a discrete variable, the bar for 1 will extend from 0.5 to 1.5. Think about what this will mean for the bars for the 3–4 group and the 5–10 group.
3.32 Example 3.19 used annual rainfall data for Albuquerque, New Mexico, to construct a relative frequency distribution and cumulative relative frequency plot. The National Climate Data Center also gave the accompanying annual rainfall (in inches) for Medford, Oregon, from 1950 to 2008. 28.84 20.85 23.43 16.05 13.71 18.81
20.15 19.86 19.55 22.08 14.68 15.15
18.88 23.34 20.82 19.44 15.16 18.16
25.72 19.08 19.04 30.38 16.77 19.99
16.42 29.23 18.77 18.79 12.33 19.00
20.18 18.32 19.63 10.89 21.93 23.97
28.96 21.27 12.39 17.25 31.57 21.99
20.72 18.93 22.39 14.95 18.13 17.25
23.58 15.47 15.95 13.86 28.87 14.07
10.62 20.68 20.46 15.30 16.69
a. Construct a relative frequency distribution for the Medford rainfall data. b. Use the relative frequency distribution of Part (a) to construct a histogram. Describe the shape of the histogram. c. Construct a cumulative relative frequency plot for the Medford rainfall data. d. Use the cumulative relative frequency plot of Part (c) to answer the following questions: i. Approximately what proportion of years had annual rainfall less than 15.5 inches? ii. Approximately what proportion of years had annual rainfall less than 25 inches? Bold exercises answered in back
Data set available online
iii. Approximately what proportion of years had annual rainfall between 17.5 and 25 inches?
3.33 The National Climate Data Center referenced in the previous exercise and Example 3.19 also gives rainfall data for a number of other U.S. cities. Go to the web site www.ncdc.noaa.gov/oa/climate/research/cag3/city.html and select one of the other cities. Use the data from 1950 to the most recent year for which data is available for the city you have selected to construct a relative frequency distribution and histogram. Write a few sentences comparing the distribution of annual rainfall values for the city you selected to the rainfall distribution for Medford, Oregon. (Use the histogram for Medford constructed in Exercise 3.32.) 3.34 The authors of the paper “Myeloma in Patients Younger than Age 50 Years Presents with More Favorable Features and Shows Better Survival” (Blood [2008]: 4039–4047) studied patients who had been diagnosed with stage 2 multiple myeloma prior to the age of 50. For each patient who received high dose chemotherapy, the number of years that the patient lived after the therapy (survival time) was recorded. The cumulative relative frequencies in the accompanying table were approximated from survival graphs that appeared in the paper.
Years Survived
Cumulative Relative Frequency
0 to ⬍2 2 to ⬍4 4 to ⬍6 6 to ⬍8 8 to ⬍10 10 to ⬍12 12 to ⬍14 14 to ⬍16
.10 .52 .54 .64 .68 .70 .72 1.00
a. Use the given information to construct a cumulative relative frequency plot. b. Use the cumulative relative frequency plot from Part (a) to answer the following questions: i. What is the approximate proportion of patients who lived fewer than 5 years after treatment? ii. What is the approximate proportion of patients who lived fewer than 7.5 years after treatment? iii. What is the approximate proportion of patients who lived more than 10 years after treatment?
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
3.4 Displaying Bivariate Numerical Data
ing table, and state whether each histogram is symmetric, bimodal, positively skewed, or negatively skewed:
3.35 a. Use the cumulative relative frequencies given in the previous exercise to compute the relative frequencies for each class interval and construct a relative frequency distribution. b. Summarize the survival time data with a histogram. c. Based on the histogram, write a few sentences describing survival time of the stage 2 myeloma patients in this study. d. What additional information would you need in order to decide if it is reasonable to generalize conclusions about survival time from the group of patients in the study to all patients younger than 50 years old who are diagnosed with multiple myeloma and who receive high dose chemotherapy?
Frequency I
II
III
IV
V
0 to ⬍10 10 to ⬍20 20 to ⬍30 30 to ⬍40 40 to ⬍50 50 to ⬍60 60 to ⬍70
5 10 20 30 20 10 5
40 25 10 8 7 5 5
30 10 8 7 7 8 30
15 25 8 7 20 25 10
6 5 6 9 9 23 42
140, . . . , 180 to 200, devise a frequency distribution based on 70 observations whose histogram could be described as follows: a. symmetric c. positively skewed b. bimodal d. negatively skewed
the five frequency distributions, I–V, given in the follow-
3.4
Class Interval
3.37 Using the five class intervals 100 to 120, 120 to
3.36 Construct a histogram corresponding to each of
Bold exercises answered in back
133
Data set available online
Video Solution available
Displaying Bivariate Numerical Data A bivariate data set consists of measurements or observations on two variables, x and y. For example, x might be the distance from a highway and y the lead content of soil at that distance. When both x and y are numerical variables, each observation consists of a pair of numbers, such as (14, 5.2) or (27.63, 18.9). The first number in a pair is the value of x, and the second number is the value of y. An unorganized list of bivariate data provides little information about the distribution of either the x values or the y values separately and even less information about how the two variables are related to one another. Just as graphical displays can be used to summarize univariate data, they can also help with bivariate data. The most important graph based on bivariate numerical data is a scatterplot. In a scatterplot each observation (pair of numbers) is represented by a point on a rectangular coordinate system, as shown in Figure 3.32(a). The horizontal axis is identified with values of x and is scaled so that any x value can be easily located. Similarly, the vertical or y-axis is marked for easy location of y values. The point corresponding to any particular (x, y) pair is placed where a vertical line from the value on y
y
40
40
30
30
20
FIGURE 3.32 Constructing a scatterplot: (a) rectangular coordinate system; (b) point corresponding to (4.5, 15).
y =15
10
Point corresponding to (4.5, 15)
20 10
1
2
3 (a)
4
5
x
1
2
3 (b)
4
5
x
x = 4.5
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
134
Chapter 3 Graphical Methods for Describing Data
the x-axis meets a horizontal line from the value on the y-axis. Figure 3.32(b) shows the point representing the observation (4.5, 15); it is above 4.5 on the horizontal axis and to the right of 15 on the vertical axis.
E X A M P L E 3 . 2 0 Olympic Figure Skating Do tall skaters have an advantage when it comes to earning high artistic scores in figure skating competitions? Data on x ⫽ height (in cm) and y ⫽ artistic score in the free skate for both male and female singles skaters at the 2006 Winter Olympics are shown in the accompanying table. (Data set courtesy of John Walker.)
Name
Data set available online
PLUSHENKO Yevgeny BUTTLE Jeffrey LYSACEK Evan LAMBIEL Stephane SAVOIE Matt WEIR Johnny JOUBERT Brian VAN DER PERREN Kevin TAKAHASHI Daisuke KLIMKIN Ilia ZHANG Min SAWYER Shawn LI Chengjiang SANDHU Emanuel VERNER Tomas DAVYDOV Sergei CHIPER Gheorghe DINEV Ivan DAMBIER Frederic LINDEMANN Stefan KOVALEVSKI Anton BERNTSSON Kristoffer PFEIFER Viktor TOTH Zoltan ARAKAWA Shizuka COHEN Sasha SLUTSKAYA Irina SUGURI Fumie ROCHETTE Joannie MEISSNER Kimmie HUGHES Emily MEIER Sarah KOSTNER Carolina SOKOLOVA Yelena YAN Liu LEUNG Mira GEDEVANISHVILI Elene KORPI Kiira POYKIO Susanna
Gender
Height
Artistic
M M M M M M M M M M M M M M M M M M M M M M M M F F F F F F F F F F F F F F F
178 173 177 176 175 172 179 177 165 170 176 163 170 183 180 159 176 174 163 163 171 175 180 185 166 157 160 157 157 160 165 164 168 162 164 168 159 166 159
41.2100 39.2500 37.1700 38.1400 35.8600 37.6800 36.7900 33.0100 36.6500 32.6100 31.8600 34.2500 28.4700 35.1100 28.6100 30.4700 32.1500 29.2500 31.2500 31.0000 28.7500 28.0400 28.7200 25.1000 39.3750 39.0063 38.6688 37.0313 35.0813 33.4625 31.8563 32.0313 34.9313 31.4250 28.1625 26.7000 31.2250 27.2000 31.2125
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
3.4 Displaying Bivariate Numerical Data
Name
135
Gender
Height
Artistic
F F F F F F F F F
162 163 160 166 164 165 158 168 160
31.5688 26.5125 28.5750 25.5375 28.6375 23.0000 26.3938 23.6688 24.5438
ANDO Miki EFREMENKO Galina LIASHENKO Elena HEGEL Idora SEBESTYEN Julia KARADEMIR Tugba FONTANA Silvia PAVUK Viktoria MAXWELL Fleur
Figure 3.33(a) gives a scatterplot of the data. Looking at the data and the scatterplot, we can see that
40
40
35
35 Artistic
Artistic
1. Several observations have identical x values but different y values (for example, x ⫽ 176 cm for both Stephane Lambiel and Min Zhang, but Lambiel’s artistic score was 38.1400 and Zhang’s artistic score was 31.8600). Thus, the value of y is not determined solely by the value of x but by various other factors as well.
30
25
Gender F M
30
25 160
165
170 175 Height
180
185
160
165
180
185
(b)
40
40
35
35 Artistic
Artistic
(a)
170 175 Height
30
30
25
25 160
165
170 175 Height
180
(c)
185 160
165 Height
170
175
(d)
FIGURE 3.33 Scatterplots for the data of Example 3.20: (a) scatterplot of data; (b) scatterplot of data with observations for males and females distinguished by color; (c) scatterplot for male skaters; (d) scatterplot for female skaters. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
136
Chapter 3 Graphical Methods for Describing Data
2. At any given height there is quite a bit of variability in artistic score. For example, for those skaters with height 160 cm, artistic scores ranged from a low of about 24.5 to a high of about 39. 3. There is no noticeable tendency for artistic score to increase as height increases. There does not appear to be a strong relationship between height and artistic score. The data set used to construct the scatter plot included data for both male and female skaters. Figure 3.33(b) shows a scatterplot of the (height, artistic score) pairs with observations for male skaters shown in blue and observations for female skaters shown in orange. Not surprisingly, the female skaters tend to be shorter than the male skaters (the observations for females tend to be concentrated toward the left side of the scatterplot). Careful examination of this plot shows that while there was no apparent pattern in the combined (male and female) data set, there may be a relationship between height and artistic score for female skaters. Figures 3.33(c) and 3.33(d) show separate scatterplots for the male and female skaters, respectively. It is interesting to note that it appears that for female skaters, higher artistic scores seem to be associated with smaller height values, but for men there does not appear to be a relationship between height and artistic score. The relationship between height and artistic score for women is not evident in the scatterplot of the combined data.
The horizontal and vertical axes in the scatterplots of Figure 3.33 do not intersect at the point (0, 0). In many data sets, the values of x or of y or of both variables differ considerably from 0 relative to the ranges of the values in the data set. For example, a study of how air conditioner efficiency is related to maximum daily outdoor temperature might involve observations at temperatures of 80°, 82°, . . . , 98°, 100°. In such cases, the plot will be more informative if the axes intersect at some point other than (0, 0) and are marked accordingly. This is illustrated in Example 3.21.
E X A M P L E 3 . 2 1 Taking Those “Hard” Classes Pays Off The report titled “2007 College Bound Seniors” (College Board, 2007) included the accompanying table showing the average score on the writing and math sections of the SAT for groups of high school seniors completing different numbers of years of study in six core academic subjects (arts and music, English, foreign languages, mathematics, natural sciences, and social sciences and history). Figure 3.34(a) and (b) show two scatterplots of x ⫽ total number of years of study and y ⫽ average writing SAT score. The scatterplots were produced by the statistical computer package Minitab. In Figure 3.34(a), we let Minitab select the scale for both axes. Figure 3.34(b) was obtained by specifying that the axes would intersect at the point (0, 0). The second plot does not make effective use of space. It is more crowded than the first plot, and such crowding can make it more difficult to see the general nature of any relationship. For example, it can be more difficult to spot curvature in a crowded plot. Step-by-step technology instructions available online Data set available online Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
3.4 Displaying Bivariate Numerical Data
600 Average SAT writing score
550 Average SAT writing score
137
525 500 475
500 400 300 200 100
450
0 15
16
17 18 Years of study
19
20
5
10 Years of study
(a)
15
20
(b)
Average SAT score
550 525 Variable Average SAT—writing Average SAT—math
500 475 450
FIGURE 3.34 Minitab scatterplots of data in Example 3.21: (a) scale for both axes selected by Minitab; (b) axes intersect at the point (0, 0); (c) math and writing on same plot.
15
16
17 18 Years of study
19
20
(c)
Years of Study
Average Writing Score
Average Math Score
15 16 17 18 19 20
442 447 454 469 486 534
461 466 473 490 507 551
The scatterplot for average writing SAT score exhibits a fairly strong curved pattern, indicating that there is a strong relationship between average writing SAT score and the total number of years of study in the six core academic subjects. Although the pattern in the plot is curved rather than linear, it is still easy to see that the average writing SAT score increases as the number of years of study increases. Figure 3.34(c) shows a scatterplot with the average writing SAT scores represented by blue squares and the average math SAT scores represented by orange dots. From this plot we can see that while the average math SAT scores tend to be higher than the average writing scores at all of the values of total number of years of study, the general curved form of the relationship is similar.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
138
Chapter 3 Graphical Methods for Describing Data
In Chapter 5, methods for summarizing bivariate data when the scatterplot reveals a pattern are introduced. Linear patterns are relatively easy to work with. A curved pattern, such as the one in Example 3.21, is a bit more complicated to analyze, and methods for summarizing such nonlinear relationships are developed in Section 5.4.
Time Series Plots Data sets often consist of measurements collected over time at regular intervals so that we can learn about change over time. For example, stock prices, sales figures, and other socio-economic indicators might be recorded on a weekly or monthly basis. A time-series plot (sometimes also called a time plot) is a simple graph of data collected over time that can be invaluable in identifying trends or patterns that might be of interest. A time-series plot can be constructed by thinking of the data set as a bivariate data set, where y is the variable observed and x is the time at which the observation was made. These (x, y) pairs are plotted as in a scatterplot. Consecutive observations are then connected by a line segment; this aids in spotting trends over time.
E X A M P L E 3 . 2 2 The Cost of Christmas The Christmas Price Index is computed each year by PNC Advisors, and it is a humorous look at the cost of the giving all of the gifts described in the popular Christmas song “The Twelve Days of Christmas.” The year 2008 was the most costly year since the index began in 1984, with the “cost of Christmas” at $21,080. A plot of the Christmas Price Index over time appears on the PNC web site (www .pncchristmaspriceindex.com) and the data given there were used to construct the time-series plot of Figure 3.35. The plot shows an upward trend in the index from
Price of Christmas 21000 20000 19000 18000 17000 16000 15000 14000 13000
FIGURE 3.35 Time-series plot for the Christmas Price Index data of Example 3.22.
12000 1984
1986
1988
1990
1992
1994
1996 Year
1998
2000
2002
2004
2006
2008
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
3.4 Displaying Bivariate Numerical Data
139
1984 until 1993. A dramatic drop in the cost occurred between 1993 and 1995, but there has been a clear upward trend in the index since then. You can visit the web site to see individual time-series plots for each of the twelve gifts that are used to determine the Christmas Price Index (a partridge in a pear tree, two turtle doves, etc.). See if you can figure out what caused the dramatic decline in 1995.
E X A M P L E 3 . 2 3 Education Level and Income—Stay in School! The time-series plot shown in Figure 3.36 appears on the U.S. Census Bureau web site. It shows the average earnings of workers by educational level as a proportion of the average earnings of a high school graduate over time. For example, we can see from this plot that in 1993 the average earnings for people with bachelor’s degrees was about 1.5 times the average for high school graduates. In that same year, the average earnings for those who were not high school graduates was only about 75% (a proportion of .75) of the average for high school graduates. The time-series plot also shows that the gap between the average earnings for high school graduates and those with a bachelor’s degree or an advanced degree widened during the 1990s.
Average earnings as a proportion of high school graduates’ earnings 3.0
2.5
Advanced degree
2.0 Bachelor’s degree 1.5 Some college or associate’s degree 1.0
High school graduate Not high school graduate
FIGURE 3.36 Time-series plot for average earnings as a proportion of the average earnings of high school graduates.
0.5 1975 1977 1979 1981 1983 1985 1987 1989 1991 1993 1995 1997 1999 Year
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
140
Chapter 3 Graphical Methods for Describing Data
EX E RC I S E S 3 . 3 8 - 3 . 4 5 Consumer Reports Health (www.consumer reports.org) gave the accompanying data on saturated fat
3.38
(in grams), sodium (in mg), and calories for 36 fast-food items. Fat
Sodium
Calories
2 5 3 2 1 6 4.5 5 3.5 1 2 3 6 3 2 5 3.5 2.5 0 2.5 1 3 1 4 3 1.5 3 9 1 1.5 2.5 3 0 0 2.5 3
1042 921 250 770 635 440 490 1160 970 1120 350 450 800 1190 1090 570 1215 1160 520 1120 240 650 1620 660 840 1050 1440 750 500 1200 1200 1250 1040 760 780 500
268 303 260 660 180 290 290 360 300 315 160 200 320 420 120 290 285 390 140 330 120 180 340 380 300 490 380 560 230 370 330 330 220 260 220 230
a. Construct a scatterplot using y ⫽ calories and x ⫽ fat. Does it look like there is a relationship between fat and calories? Is the relationship what you expected? Explain. b. Construct a scatterplot using y ⫽ calories and x ⫽ sodium. Write a few sentences commenting on the difference between the relationship of calories to fat and calories to sodium. Bold exercises answered in back
Data set available online
c. Construct a scatterplot using y ⫽ sodium and x ⫽ fat. Does there appear to be a relationship between fat and sodium? d. Add a vertical line at x ⫽ 3 and a horizontal line at y ⫽ 900 to the scatterplot in Part (c). This divides the scatterplot into four regions, with some of the points in the scatterplot falling into each of the four regions. Which of the four regions corresponds to healthier fast-food choices? Explain.
3.39 The report “Wireless Substitution: Early Release of Estimates from the National Health Interview Survey” (Center for Disease Control, 2009) gave the following estimates of the percentage of homes in the United States that had only wireless phone service at 6-month intervals from June 2005 to December 2008. Percent with Only Wireless Phone Service
Date June 2005 December 2005 June 2006 December 2006 June 2007 December 2007 June 2008 December 2008
7.3 8.4 10.5 12.8 13.6 15.8 17.5 20.2
Construct a time-series plot for these data and describe the trend in the percent of homes with only wireless phone service over time. Has the percent increased at a fairly steady rate? The accompanying table gives the cost and an overall quality rating for 15 different brands of bike helmets (www.consumerreports.org).
3.40
Cost 35 20 30 40 50 23 30 18 40 28 20
Rating 65 61 60 55 54 47 47 43 42 41 40 (continued)
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
3.4 Displaying Bivariate Numerical Data
Cost
Rating
25 30 30 40
32 63 63 53
The accompanying table gives the cost and an overall quality rating for 10 different brands of men’s athletic shoes and nine different brands of women’s athletic shoes (www.consumerreports.org).
3.41
Rating
65 45 45 80 110 110 30 80 110 70 65 70 85 80 45 70 55 110 70
71 70 62 59 58 57 56 52 51 51 71 70 66 66 65 62 61 60 59
Type Men’s Men’s Men’s Men’s Men’s Men’s Men’s Men’s Men’s Men’s Women’s Women’s Women’s Women’s Women’s Women’s Women’s Women’s Women’s
a. Using the data for all 19 shoes, construct a scatterplot using y ⫽ quality rating and x ⫽ cost. Write a sentence describing the relationship between quality rating and cost. b. Construct a scatterplot of the 19 data points that uses different colors or different symbols to distinguish the points that correspond to men’s shoes from those that correspond to women’s shoes. How do men’s and women’s athletic shoes differ with respect to cost and quality rating? Are the relationships between cost and quality rating the same for men and women? If not, how do the relationships differ? Bold exercises answered in back
The article “Medicine Cabinet is a Big Killer” (The Salt Lake Tribune, August 1, 2007) looked at the number of prescription-drug-overdose deaths in Utah over the period from 1991 to 2006. Construct a timeseries plot for these data and describe the trend over time. Has the number of overdose deaths increased at a fairly steady rate?
3.42
a. Construct a scatterplot using y ⫽ quality rating and x ⫽ cost. b. Based on the scatterplot from Part (a), does there appear to be a relationship between cost and quality rating? Does the scatterplot support the statement that the more expensive bike helmets tended to receive higher quality ratings?
Cost
141
Data set available online
Year
Number of Overdose Deaths
1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006
32 52 73 61 68 64 85 89 88 109 153 201 237 232 308 307
3.43 The article “Cities Trying to Rejuvenate Recycling Efforts” (USA Today, October 27, 2006) states that the amount of waste collected for recycling has grown slowly in recent years. This statement was supported by the data in the accompanying table. Use these data to construct a time-series plot. Explain how the plot is or is not consistent with the given statement. Year
Recycled Waste (in millions of tons)
1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005
29.7 32.9 36.0 37.9 43.5 46.1 46.4 47.3 48.0 50.1 52.7 52.8 53.7 55.8 57.2 58.4
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
Chapter 3 Graphical Methods for Describing Data
Some days of the week are more dangerous than others, according to Traffic Safety Facts produced by the National Highway Traffic Safety Administration. The average number of fatalities per day for each day of the week are shown in the accompanying table.
3.44
3.45 The accompanying time-series plot of movie box office totals (in millions of dollars) over 18 weeks of summer for both 2001 and 2002 appeared in USA Today (September 3, 2002): USA TODAY. September 03, 2002. Reprinted with permission.
142
Average Fatalities per Day (day of the week) 1978–1982 1983–1987 1988–1992 1993–1997 1998–2002 Total
Mon Tue Wed Thurs
Fri
Sat Sun
103 98 97 97 99 99
156 140 139 129 129 138
201 174 168 148 149 168
101 96 94 93 96 96
107 99 97 96 98 100
116 108 106 102 104 107
159 140 135 127 130 138
a. Using the midpoint of each year range (e.g., 1980 for the 1978–1982 range), construct a time-series plot that shows the average fatalities over time for each day of the week. Be sure to label each line clearly as to which day of the week it represents. b. Write a sentence or two commenting on the difference in average number of fatalities for the days of the week. What is one possible reason for the differences? c. Write a sentence or two commenting on the change in average number of fatalities over time. What is one possible reason for the change? Bold exercises answered in back
3.5
Data set available online
Patterns that tend to repeat on a regular basis over time are called seasonal patterns. Describe any seasonal patterns that you see in the summer box office data. Hint: Look for patterns that seem to be consistent from year to year.
Video Solution available
Interpreting and Communicating the Results of Statistical Analyses A graphical display, when used appropriately, can be a powerful tool for organizing and summarizing data. By sacrificing some of the detail of a complete listing of a data set, important features of the data distribution are more easily seen and more easily communicated to others.
Communicating the Results of Statistical Analyses When reporting the results of a data analysis, a good place to start is with a graphical display of the data. A well-constructed graphical display is often the best way to highlight the essential characteristics of the data distribution, such as shape and spread for numerical data sets or the nature of the relationship between the two variables in a bivariate numerical data set. For effective communication with graphical displays, some things to remember are • Be sure to select a display that is appropriate for the given type of data. • Be sure to include scales and labels on the axes of graphical displays. • In comparative plots, be sure to include labels or a legend so that it is clear which
parts of the display correspond to which samples or groups in the data set. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
3.5 Interpreting and Communicating the Results of Statistical Analyses
143
• Although it is sometimes a good idea to have axes that do not cross at (0, 0) in a
•
•
•
•
• •
scatterplot, the vertical axis in a bar chart or a histogram should always start at 0 (see the cautions and limitations later in this section for more about this). Keep your graphs simple. A simple graphical display is much more effective than one that has a lot of extra “junk.” Most people will not spend a great deal of time studying a graphical display, so its message should be clear and straightforward. Keep your graphical displays honest. People tend to look quickly at graphical displays, so it is important that a graph’s first impression is an accurate and honest portrayal of the data distribution. In addition to the graphical display itself, data analysis reports usually include a brief discussion of the features of the data distribution based on the graphical display. For categorical data, this discussion might be a few sentences on the relative proportion for each category, possibly pointing out categories that were either common or rare compared to other categories. For numerical data sets, the discussion of the graphical display usually summarizes the information that the display provides on three characteristics of the data distribution: center or location, spread, and shape. For bivariate numerical data, the discussion of the scatterplot would typically focus on the nature of the relationship between the two variables used to construct the plot. For data collected over time, any trends or patterns in the time-series plot would be described.
Interpreting the Results of Statistical Analyses When someone uses a web search engine, do they rely on the ranking of the search results returned or do they first scan the results looking for the most relevant? The authors of the paper “Learning User Interaction Models for Predicting Web Search
Result Preferences” (Proceedings of the 29th Annual ACM Conference on Research and Development in Information Retrieval, 2006) attempted to answer this question by observing user behavior when they varied the position of the most relevant result in the list of resources returned in response to a web search. They concluded that people clicked more often on results near the top of the list, even when they were not relevant. They supported this conclusion with the comparative bar graph in Figure 3.37.
Relative click frequency
1.0 PTR = 1 PTR = 2 PTR = 3 PTR = 5 PTR = 10 Background
0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1
FIGURE 3.37 Comparative bar graph for click frequency data.
0 1
2
3
5
10
Result position
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
144
Chapter 3 Graphical Methods for Describing Data
Although this comparative bar chart is a bit complicated, we can learn a great deal from this graphical display. Let’s start by looking at the first group of bars. The different bars correspond to where in the list of search results the result that was considered to be most relevant was located. For example, in the legend PTR ⫽ 1 means that the most relevant result was in position 1 in the list returned. PTR ⫽ 2 means that the most relevant result was in the second position in the list returned, and so on. PTR ⫽ Background means that the most relevant result was not in the first 10 results returned. The first group of bars shows the proportion of times users clicked on the first result returned. Notice that all users clicked on the first result when it was the most relevant, but nearly half clicked on the first result when the most relevant result was in the second position and more than half clicked on the first result when the most relevant result was even farther down the list. The second group of bars represents the proportion of users who clicked on the second result. Notice that the proportion who clicked on the second result was highest when the most relevant result was in that position. Stepping back to look at the entire graphical display, we see that users tended to click on the most relevant result if it was in one of the first three positions, but if it appeared after that, very few selected it. Also, if the most relevant result was in the third or a later position, users were more likely to click on the first result returned, and the likelihood of a click on the most relevant result decreased the farther down the list it appeared. To fully understand why the researchers’ conclusions are justified, we need to be able to extract this kind of information from graphical displays. The use of graphical data displays is quite common in newspapers, magazines, and journals, so it is important to be able to extract information from such displays. For example, data on test scores for a standardized math test given to eighth graders in 37 states, 2 territories (Guam and the Virgin Islands), and the District of Columbia were used to construct the stem-and-leaf display and histogram shown in Figure 3.38. Careful examination of these displays reveals the following: 1. Most of the participating states had average eighth-grade math scores between 240 and 280. We would describe the shape of this display as negatively skewed, because of the longer tail on the low end of the distribution. 2. Three of the average scores differed substantially from the others. These turn out to be 218 (Virgin Islands), 229 (District of Columbia), and 230 (Guam). These Frequency
8
FIGURE 3.38 Stem-and-leaf display and histogram for math test scores.
21H 22L 22H 23L 23H 24L 24H 25L 25H 26L 26H 27L 27H 28L
8
6
9 0 4
79 014 6667779999 0003344 55778 12233 Stem: Tens 667 Leaf: Ones 01
2
0 220
230
240 250 260 Average test score
270
280
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
3.5 Interpreting and Communicating the Results of Statistical Analyses
145
three scores could be described as outliers. It is interesting to note that the three unusual values are from the areas that are not states. 3. There do not appear to be any outliers on the high side. 4. A “typical” average math score for the 37 states would be somewhere around 260. 5. There is quite a bit of variability in average score from state to state. How would the displays have been different if the two territories and the District of Columbia had not participated in the testing? The resulting histogram is shown in Figure 3.39. Note that the display is now more symmetric, with no noticeable outliers. The display still reveals quite a bit of state-to-state variability in average score, and 260 still looks reasonable as a “typical” average score. Now suppose that the two highest values among the 37 states (Montana and North Dakota) had been even higher. The stem-and-leaf display might then look like the one given in Figure 3.40. In this stem-and-leaf display, two values stand out from the main part of the display. This would catch our attention and might cause us to look carefully at these two states to determine what factors may be related to high math scores. Frequency
8
6
4
2
0 245
255 265 275 Average test score
24H 25L 25H 26L 26H 27L 27H 28L 28H 29L 29H
79 014 6667779999 0003344 55778 12233 667
68
Stem: Tens Leaf: Ones
FIGURE 3.39
FIGURE 3.40
Histogram frequency for the modified math score data.
Stem-and-leaf display for modified math score data.
What to Look for in Published Data Here are some questions you might ask yourself when attempting to extract information from a graphical data display: • Is the chosen display appropriate for the type of data collected? • For graphical displays of univariate numerical data, how would you describe the
shape of the distribution, and what does this say about the variable being summarized? • Are there any outliers (noticeably unusual values) in the data set? Is there any
plausible explanation for why these values differ from the rest of the data? (The presence of outliers often leads to further avenues of investigation.) • Where do most of the data values fall? What is a typical value for the data set? What does this say about the variable being summarized? • Is there much variability in the data values? What does this say about the variable being summarized? Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
146
Chapter 3 Graphical Methods for Describing Data
Of course, you should always think carefully about how the data were collected. If the data were not gathered in a reasonable manner (based on sound sampling methods or experimental design principles), you should be cautious in formulating any conclusions based on the data. Consider the histogram in Figure 3.41, which is based on data published by the National Center for Health Statistics. The data set summarized by this histogram consisted of infant mortality rates (deaths per 1000 live births) for the 50 states in the United States. A histogram is an appropriate way of summarizing these data (although with only 50 observations, a stem-and-leaf display would also have been reasonable). The histogram itself is slightly positively skewed, with most mortality rates between 7.5 and 12. There is quite a bit of variability in infant mortality rate from state to state— perhaps more than we might have expected. This variability might be explained by differences in economic conditions or in access to health care. We may want to look further into these issues. Although there are no obvious outliers, the upper tail is a little longer than the lower tail. The three largest values in the data set are 12.1 (Alabama), 12.3 (Georgia), and 12.8 (South Carolina)—all Southern states. Again, this may suggest some interesting questions that deserve further investigation. A typical infant mortality rate would be about 9.5 deaths per 1000 live births. This represents an improvement, because researchers at the National Center for Health Statistics stated that the overall rate for 1988 was 10 deaths per 1000 live births. However, they also point out that the United States still ranked 22 out of 24 industrialized nations surveyed, with only New Zealand and Israel having higher infant mortality rates.
A Word to the Wise: Cautions and Limitations When constructing and interpreting graphical displays, you need to keep in mind these things: 1. Areas should be proportional to frequency, relative frequency, or magnitude of the number being represented. The eye is naturally drawn to large areas in graphical displays, and it is natural for the observer to make informal comparisons based
Frequency
10
8
6
4
2
FIGURE 3.41 Histogram of infant mortality rates.
0 7.0
8.0
9.0 10.0 11.0 Mortality rate
12.0
13.0
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
3.5 Interpreting and Communicating the Results of Statistical Analyses
147
USA TODAY. October 03, 2002. Reprinted with permission.
on area. Correctly constructed graphical displays, such as pie charts, bar charts, and histograms, are designed so that the areas of the pie slices or the bars are proportional to frequency or relative frequency. Sometimes, in an effort to make graphical displays more interesting, designers lose sight of this important principle, and the resulting graphs are misleading. For example, consider the following graph (USA Today, October 3, 2002):
In trying to make the graph more visually interesting by replacing the bars of a bar chart with milk buckets, areas are distorted. For example, the two buckets for 1980 represent 32 cows, whereas the one bucket for 1970 represents 19 cows. This is misleading because 32 is not twice as big as 19. Other areas are distorted as well. Another common distortion occurs when a third dimension is added to bar charts or pie charts. For example, the pie chart at the bottom left of the page appeared in USA Today (September 17, 2009). Adding the third dimension distorts the areas and makes it much more difficult to interpret correctly. A correctly drawn pie chart is shown below.
Category 3–5 times a week Never 1–3 times a week
3–5 times a week
Image not available due to copyright restrictions Never
1–3 times a week
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
148
Chapter 3 Graphical Methods for Describing Data
Image not available due to copyright restrictions
2. Be cautious of graphs with broken axes. Although it is common to see scatterplots with broken axes, be extremely cautious of time-series plots, bar charts, or histograms with broken axes. The use of broken axes in a scatterplot does not distort information about the nature of the relationship in the bivariate data set used to construct the display. On the other hand, in time-series plots, broken axes can sometimes exaggerate the magnitude of change over time. Although it is not always inadvisable to break the vertical axis in a time-series plot, it is something you should watch for, and if you see a time-series plot with a broken axis, as in the accompanying time-series plot of mortgage rates (USA Today, October 25, 2002), you should pay particular attention to the scale on the vertical axis and take extra care in interpreting the graph. In bar charts and histograms, the vertical axis (which represents frequency, relative frequency, or density) should never be broken. If the vertical axis is broken in this type of graph, the resulting display will violate the “proportional area” principle and the display will be misleading. For example, the accompanying bar chart is similar to one appearing in an advertisement for a software product designed to help teachers raise student test scores. By starting the vertical axis at 50, the gain for students using the software is exaggerated. Areas of the bars are not proportional to the magnitude of the numbers represented—the area for the rectangle representing 68 is more than three times the area of the rectangle representing 55! Percentile score
Pretest Post-test
70 65 60 55 50 Traditional instruction Using software Group
3. Watch out for unequal time spacing in time-series plots. If observations over time are not made at regular time intervals, special care must be taken in constructing the timeseries plot. Consider the accompanying time-series plot, which is similar to one appearing in the San Luis Obispo Tribune (September 22, 2002) in an article on online banking: Number using online banking (in millions) 20
10
0 Jan 94
May 95
May 96
Dec. 97
Dec. 98
Feb. 00
Sept. 01
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
3.5 Interpreting and Communicating the Results of Statistical Analyses
149
Notice that the intervals between observations are irregular, yet the points in the plot are equally spaced along the time axis. This makes it difficult to make a coherent assessment of the rate of change over time. This could have been remedied by spacing the observations differently along the time axis, as shown in the following plot: Number using online banking (in millions) 20
10
0 Jan 94
May May 95 96
Dec. Dec. 97 98
Feb. 00
Sept. 01
Time
USA TODAY. June 25, 2002. Used with permission.
4. Be careful how you interpret patterns in scatterplots. A strong pattern in a scatterplot means that the two variables tend to vary together in a predictable way, but it does not mean that there is a cause-and-effect relationship between the two variables. We will consider this point further in Chapter 5, but in the meantime, when describing patterns in scatterplots, be careful not to use wording that implies that changes in one variable cause changes in the other. 5. Make sure that a graphical display creates the right first impression. For example, consider the graph below from USA Today (June 25, 2002). Although this graph does not violate the proportional area principle, the way the “bar” for the “none” category is displayed makes this graph difficult to read, and a quick glance at this graph would leave the reader with an incorrect impression.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
150
Chapter 3
Graphical Methods for Describing Data
EX E RC I S E S 3 . 4 6 - 3 . 5 1 3.46 The accompanying comparative bar chart is from
3.47 Figure EX-3.47 is from the Fall 2008 Census
the report “More and More Teens on Cell Phones” (Pew Research Center, www.pewresearch.org, August 19, 2009).
Enrollment Report at Cal Poly, San Luis Obispo. It uses both a pie chart and a segmented bar graph to summarize data on ethnicity for students enrolled at the university in Fall 2008. a. Use the information in the graphical display to construct a single segmented bar graph for the ethnicity data. b. Do you think that the original graphical display or the one you created in Part (a) is more informative? Explain your choice. c. Why do you think that the original graphical display format (combination of pie chart and segmented bar graph) was chosen over a single pie chart with 7 slices?
Image not available due to copyright restrictions
3.48 The accompanying graph appeared in USA Today (August 5, 2008). This graph is a modified comparative
Suppose that you plan to include this graph in an article that you are writing for your school newspaper. Write a few paragraphs that could accompany the graph. Be sure to address what the graph reveals about how teen cell phone ownership is related to age and how it has changed over time.
Nonresident alien 1.2%
Native American 0.8%
Unknown/other 9.6%
Fall 2008 total enrollment
Hispanic/ Latino 11.3% White 65.0%
Nonwhite 24.2%
bar graph. Most likely, the modifications (incorporating hands and the earth) were made to try to make a display that readers would find more interesting. a. Use the information in the USA Today graph to construct a traditional comparative bar graph. b. Explain why the modifications made in the USA Today graph may make interpretation more difficult than with the traditional comparative bar graph.
Image not available due to copyright restrictions
African American 1.1%
Asian American 11.0%
FIGURE EX-3.47 Bold exercises answered in back
Data set available online
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
3.5 Interpreting and Communicating the Results of Statistical Analyses
151
3.49 The two graphical displays below appeared in
3.50 The following graphical display is meant to be a
USA Today (June 8, 2009 and July 28, 2009). One is an appropriate representation and the other is not. For each of the two, explain why it is or is not drawn appropriately.
comparative bar graph (USA Today, August 3, 2009). Do you think that this graphical display is an effective summary of the data? If so, explain why. If not, explain why not and construct a display that makes it easier to compare the ice cream preferences of men and women.
Images not available due to copyright restrictions
Bold exercises answered in back
Data set available online
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
152
Chapter 3
Graphical Methods for Describing Data
AC TI V I TY 3 . 1
Locating States
Background: A newspaper article bemoaning the state of students’ knowledge of geography claimed that more students could identify the island where the 2002 season of the TV show Survivor was filmed than could locate Vermont on a map of the United States. In this activity, you will collect data that will allow you to estimate the proportion of students who can correctly locate the states of Vermont and Nebraska. 1. Working as a class, decide how you will select a sample that you think will be representative of the students from your school. 2. Use the sampling method from Step 1 to obtain the subjects for this study. Subjects should be shown the accompanying map of the United States and asked to point out the state of Vermont. After the subject has given his or her answer, ask the subject to point out the state of Nebraska. For each subject, record whether or not Vermont was correctly identified and whether or not Nebraska was correctly identified.
AC TI V I TY 3 . 2
3. When the data collection process is complete, summarize the resulting data in a table like the one shown here: Response
Frequency
Correctly identified both states Correctly identified Vermont but not Nebraska Correctly identified Nebraska but not Vermont Did not correctly identify either state
4. Construct a pie chart that summarizes the data in the table from Step 3. 5. What proportion of sampled students were able to correctly identify Vermont on the map? 6. What proportion of sampled students were able to correctly identify Nebraska on the map? 7. Construct a comparative bar chart that shows the proportion correct and the proportion incorrect for each of the two states considered. 8. Which state, Vermont or Nebraska, is closer to the state in which your school is located? Based on the pie chart, do you think that the students at your school were better able to identify the state that was closer than the one that was farther away? Justify your answer. 9. Write a paragraph commenting on the level of knowledge of U.S. geography demonstrated by the students participating in this study. 10. Would you be comfortable generalizing your conclusions in Step 8 to the population of students at your school? Explain why or why not.
Bean Counters!
Materials needed: A large bowl of dried beans (or marbles, plastic beads, or any other small, fairly regular objects) and a coin. In this activity, you will investigate whether people can hold more in the right hand or in the left hand. 1. Flip a coin to determine which hand you will measure first. If the coin lands heads side up, start with the right hand. If the coin lands tails side up, start with the left hand. With the designated hand, reach into the bowl and grab as many beans as possible. Raise the hand over the bowl and count to 4.
If no beans drop during the count to 4, drop the beans onto a piece of paper and record the number of beans grabbed. If any beans drop during the count, restart the count. That is, you must hold the beans for a count of 4 without any beans falling before you can determine the number grabbed. Repeat the process with the other hand, and then record the following information: (1) right-hand number, (2) left-hand number, and (3) dominant hand (left or right, depending on whether you are left- or right-handed).
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
Summary of Key Concepts and Formulas
2. Create a class data set by recording the values of the three variables listed in Step 1 for each student in your class. 3. Using the class data set, construct a comparative stem-and-leaf display with the right-hand counts displayed on the right and the left-hand counts displayed on the left of the stem-and-leaf display. Comment on the interesting features of the display and include a comparison of the right-hand count and left-hand count distributions. 4. Now construct a comparative stem-and-leaf display that allows you to compare dominant-hand count to nondominant-hand count. Does the display support
153
the theory that dominant-hand count tends to be higher than nondominant-hand count? 5. For each observation in the data set, compute the difference dominant-hand count 2 nondominant-hand count Construct a stem-and-leaf display of the differences. Comment on the interesting features of this display. 6. Explain why looking at the distribution of the differences (Step 5) provides more information than the comparative stem-and-leaf display (Step 4). What information is lost in the comparative display that is retained in the display of the differences?
Summary of Key Concepts and Formulas TERM OR FORMULA
COMMENT
Frequency distribution
A table that displays frequencies, and sometimes relative and cumulative relative frequencies, for categories (categorical data), possible values (discrete numerical data), or class intervals (continuous data).
Comparative bar chart
Two or more bar charts that use the same set of horizontal and vertical axes.
Pie chart
A graph of a frequency distribution for a categorical data set. Each category is represented by a slice of the pie, and the area of the slice is proportional to the corresponding frequency or relative frequency.
Segmented bar graph
A graph of a frequency distribution for a categorical data set. Each category is represented by a segment of the bar, and the area of the segment is proportional to the corresponding frequency or relative frequency.
Stem-and-leaf display
A method of organizing numerical data in which the stem values (leading digit(s) of the observations) are listed in a column, and the leaf (trailing digit(s)) for each observation is then listed beside the corresponding stem. Sometimes stems are repeated to stretch the display.
Histogram
A picture of the information in a frequency distribution for a numerical data set. A rectangle is drawn above each possible value (discrete data) or class interval. The rectangle’s area is proportional to the corresponding frequency or relative frequency.
Histogram shapes
A (smoothed) histogram may be unimodal (a single peak), bimodal (two peaks), or multimodal. A unimodal histogram may be symmetric, positively skewed (a long right or upper tail), or negatively skewed. A frequently occurring shape is one that is approximately normal.
Cumulative relative frequency plot
A graph of a cumulative relative frequency distribution.
Scatterplot
A picture of bivariate numerical data in which each observation (x, y) is represented as a point with respect to a horizontal x-axis and a vertical y-axis.
Time-series plot
A graphical display of numerical data collected over time.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
154
Chapter 3
Graphical Methods for Describing Data
Chapter Review Exercises 3.52 - 3.71 3.52 The article “Most Smokers Wish They Could Quit” (Gallup Poll Analyses, November 21, 2002) noted that smokers and nonsmokers perceive the risks of smoking differently. The accompanying relative frequency table summarizes responses regarding the perceived harm of smoking for each of three groups: a sample of 241 smokers, a sample of 261 former smokers, and a sample of 502 nonsmokers. Construct a comparative bar chart for these data. Do not forget to use relative frequencies in constructing the bar chart because the three sample sizes are different. Comment on how smokers, former smokers, and nonsmokers differ with respect to perceived risk of smoking. Frequency Perceived Risk of Smoking Very harmful Somewhat harmful Not too harmful Not at all harmful
Smokers
Former Smokers
Nonsmokers
145 72 17 7
204 42 10 5
432 50 15 5
3.53 Each year the College Board publishes a profile of students taking the SAT. In the report “2005 College Bound Seniors: Total Group Profile Report,” the average SAT scores were reported for three groups defined by first language learned. Use the data in the accompanying table to construct a bar chart of the average verbal SAT score for the three groups. First Language Learned English English and another language A language other than English
Average Verbal SAT 519 486 462
3.54 The report referenced in Exercise 3.53 also gave average math SAT scores for the three language groups, as shown in the following table. First Language Learned English English and another language A language other than English
Bold exercises answered in back
Average Math SAT 521 513 521
Data set available online
Construct a comparative bar chart for the average verbal and math scores for the three language groups. Write a few sentences describing the differences and similarities between the three language groups as shown in the bar chart.
3.55 The Connecticut Agricultural Experiment Station conducted a study of the calorie content of different types of beer. The calorie content (calories per 100 ml) for 26 brands of light beer are (from the web site brewery.org): 29 28 33 31 30 33 30 28 27 41 39 31 29 23 32 31 32 19 40 22 34 31 42 35 29 43
Construct a stem-and-leaf display using stems 1, 2, 3, and 4. Write a sentence or two describing the calorie content of light beers.
3.56 The stem-and-leaf display of Exercise 3.16 uses only four stems. Construct a stem-and-leaf display for these data using repeated stems 1H, 2L, 2H, . . . , 4L. For example, the first observation, 29, would have a stem of 2 and a leaf of 9. It would be entered into the display for the stem 2H, because it is a “high” 2—that is, it has a leaf that is on the high end (5, 6, 7, 8, 9). The article “A Nation Ablaze with Change” (USA Today, July 3, 2001) gave the accompanying data
3.57
on percentage increase in population between 1990 and 2000 for the 50 U.S. states. Also provided in the table is a column that indicates for each state whether the state is in the eastern or western part of the United States (the states are listed in order of population size): State
Percentage Change
California Texas New York Florida Illinois Pennsylvania Ohio Michigan New Jersey Georgia North Carolina
13.8 22.8 5.5 23.5 8.6 3.4 4.7 6.9 8.9 26.4 21.4
East/West W W E E E E E E E E E (continued)
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
155
Chapter Review Exercises
State
Percentage Change
East/West
14.4 5.5 9.7 21.1 16.7 9.3 9.6 10.8 40.0 12.4 5.9 10.1 30.6 9.7 15.1 9.7 20.4 3.6 5.4 10.5 8.5 13.7 29.6 66.3 20.1 0.8 8.4 28.5 3.9 11.4 9.3 4.5 12.9 17.6 8.5 0.5 14.0 8.2 8.9
E E E W E E E E W E E E W E E W W E E E W E W W W E W W E E W E W E W W W E W
Virginia Massachusetts Indiana Washington Tennessee Missouri Wisconsin Maryland Arizona Minnesota Louisiana Alabama Colorado Kentucky South Carolina Oklahoma Oregon Connecticut Iowa Mississippi Kansas Arkansas Utah Nevada New Mexico West Virginia Nebraska Idaho Maine New Hampshire Hawaii Rhode Island Montana Delaware South Dakota North Dakota Alaska Vermont Wyoming
a. Construct a stem-and-leaf display for percentage growth for the data set consisting of all 50 states. Hints: Regard the observations as having two digits to the left of the decimal place. That is, think of an observation such as 8.5 as 08.5. It will also be easier to truncate leaves to a single digit; for example, a leaf of 8.5 could be truncated to 8 for purposes of constructing the display. b. Comment on any interesting features of the data set. Do any of the observations appear to be outliers? Bold exercises answered in back
Data set available online
c. Now construct a comparative stem-and-leaf display for the eastern and western states. Write a few sentences comparing the percentage growth distributions for eastern and western states. People suffering from Alzheimer’s disease often have difficulty performing basic activities of daily living (ADLs). In one study (“Functional Status and Clinical
3.58
Findings in Patients with Alzheimer’s Disease,” Journal of Gerontology [1992]: 177–182), investigators focused on six such activities: dressing, bathing, transferring, toileting, walking, and eating. Here are data on the number of ADL impairments for each of 240 patients: Number of impairments 0 Frequency 100
1 43
2 36
3 17
4 24
5 9
6 11
a. Determine the relative frequencies that correspond to the given frequencies. b. What proportion of these patients had at most two impairments? c. Use the result of Part (b) to determine what proportion of patients had more than two impairments. d. What proportion of the patients had at least four impairments? Does the size of a transplanted organ matter? A study that attempted to answer this question (“Minimum
3.59
Graft Size for Successful Living Donor Liver Transplantation,” Transplantation [1999]:1112–1116) presented a scatterplot much like the following (“graft weight ratio” is the weight of the transplanted liver relative to the ideal size liver for the recipient): a. Discuss interesting features of this scatterplot. b. Why do you think the overall relationship is negative? Graft weight ratio (%) 200
100
x
0 0
50 Recipient body weight (kg)
100
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
156
Chapter 3 Graphical Methods for Describing Data
The National Telecommunications and Information Administration published a report titled “Falling Through the Net: Toward Digital Inclusion” (U.S. Department of Commerce, October 2000) that included
3.60
the following information on access to computers in the home:
Year
Percentage of Households with a Computer
1985 1990 1994 1995 1998 1999 2000
8.2 15.0 22.8 24.1 36.6 42.1 51.0
tobacco use produced by Walt Disney, Inc., were as follows: 223 176 548 37 158 51 299 37 11 165 74 92 6 23 206 9 Data for 11 G-rated animated films showing tobacco use that were produced by MGM/United Artists, Warner Brothers, Universal, and Twentieth Century Fox were also given. The tobacco exposure times (in seconds) for these films was as follows: 205 162 6 1 117 5 91 155 24 55 17 Construct a comparative stem-and-leaf display for these data. Comment on the interesting features of this display. The accompanying data on household expenditures on transportation for the United Kingdom appeared in “Transport Statistics for Great Britain: 2002
3.64 a. Construct a time-series plot for these data. Be careful—the observations are not equally spaced in time. The points in the plot should not be equally spaced along the x-axis. b. Comment on any trend over time.
3.61 According to the National Association of Home Builders, the average size of a home in 1950 was 983 ft 2. The average size increased to 1500 ft 2 in 1970, 2080 ft 2 in 1990; and 2330 ft 2 in 2003 (San Luis Obispo Tribune, October 16, 2005). a. Construct a time-series plot that shows how the average size of a home has changed over time. b. If the trend of the time-series plot were to continue, what would you predict the average home size to be in 2010?
3.62 The paper “Community Colleges Start to Ask, Where Are the Men?” (Chronicle of Higher Education, June 28, 2002) gave data on gender for community college students. It was reported that 42% of students enrolled at community colleges nationwide were male and 58% were female. Construct a segmented bar graph for these data. The article “Tobacco and Alcohol Use in G-Rated Children’s Animated Films” ( Journal of the American Medical Association [1999]: 1131–1136) re-
3.63
ported exposure to tobacco and alcohol use in all G-rated animated films released between 1937 and 1997 by five major film studios. The researchers found that tobacco use was shown in 56% of the reviewed films. Data on the total tobacco exposure time (in seconds) for films with Bold exercises answered in back
Data set available online
Edition” (in Family Spending: A Report on the Family Expenditure Survey [The Stationary Office, 2002]). Expenditures (in pounds per week) included costs of purchasing and maintaining any vehicles owned by members of the household and any costs associated with public transportation and leisure travel.
Year
Average Transportation
Percentage of Household Expenditures for Transportation
1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000
247.20 259.00 271.80 276.70 283.60 289.90 309.10 328.80 352.20 359.40 385.70
16.2 15.3 15.8 15.6 15.1 14.9 15.7 16.7 17.0 17.2 16.7
a. Construct time-series plots of the transportation expense data and the percent of household expense data. b. Do the time-series plots of Part (a) support the statement that follows? Explain why or why not. Statement: Although actual expenditures have been increasing, the percentage of the total household expenditures that go toward transportation has remained relatively stable. Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
Chapter Review Exercises
The article “The Healthy Kids Survey: A Look at
Year
Cost (in billions of dollars)
the Findings” (San Luis Obispo Tribune, October 25, 2002) gave the accompanying information for a sample
1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005
7.5 2.5 4.0 22.5 5.0 18.0 9.0 8.0 2.6 10.0 9.0 3.0 27.0 5.0 12.0 28.5 56.8
3.65
of fifth graders in San Luis Obispo County. Responses are to the question: “After school, are you home alone without adult supervision?” Response Never Some of the time Most of the time All of the time
Percentage 8 15 16 61
a. Summarize these data using a pie chart. b. Construct a segmented bar graph for these data. c. Which graphing method—the pie chart or the segmented bar graph—do you think does a better job of conveying information about response? Explain.
3.66 “If you were taking a new job and had your choice of a boss, would you prefer to work for a man or a woman?” That was the question posed to individuals in a sample of 576 employed adults (Gallup at a Glance, October 16, 2002). Responses are summarized in the following table: Response Prefer to work for a man Prefer to work for a woman No difference No opinion
Frequency 190 92 282 12
a. Construct a pie chart to summarize this data set, and write a sentence or two summarizing how people responded to this question. b. Summarize the given data using a segmented bar graph. 2005 was a record year for hurricane devastation in the United States (San Luis Obispo Tribune, November 30, 2005). Of the 26 tropical storms and hurricanes in the season, four hurricanes hit the mainland: Dennis, Katrina, Rita, and Wilma. The United States insured catastrophic losses since 1989 (approximate values read from a graph that appeared in the San Luis Obispo Tribune, November 30, 2005) are as follows:
3.67
Bold exercises answered in back
Data set available online
157
Construct a time-series plot that shows the insured catastrophic loss over time. What do you think causes the peaks in the graph?
3.68 An article in the San Luis Obispo Tribune (November 20, 2002) stated that 39% of those with critical housing needs (those who pay more than half their income for housing) lived in urban areas, 42% lived in suburban areas, and the rest lived in rural areas. Construct a pie chart that shows the distribution of type of residential area (urban, suburban, or rural) for those with critical housing needs. Living-donor kidney transplants are becoming more common. Often a living donor has chosen to donate a kidney to a relative with kidney disease. The following data appeared in a USA Today article on organ transplants (“Kindness Motivates Newest Kidney Donors,” June 19, 2002):
3.69
Number of Kidney Transplants Year
Living-Donor to Relative
Living-Donor to Unrelated Person
1994 1995 1996 1997 1998 1999 2000 2001
2390 2906 2916 3144 3324 3359 3679 3879
202 400 526 607 814 930 1325 1399
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
158
Chapter 3
Graphical Methods for Describing Data
a. Construct a time-series plot for the number of living-donor kidney transplants where the donor is a relative of the recipient. Describe the trend in this plot. b. Use the data from 1994 and 2001 to construct a comparative bar chart for the type of donation (relative or unrelated). Write a few sentences commenting on your display. Many nutritional experts have expressed concern about the high levels of sodium in prepared foods. The following data on sodium content (in milligrams) per frozen meal appeared in the article “Comparison of ‘Light’ Frozen Meals” (Boston Globe, April 24, 1991):
3.70
720 530 800 690 880 1050 340 810 760 300 400 680 780 390 950 520 500 630 480 940 450 990 910 420 850 390 600
Two histograms for these data are shown: a. Do the two histograms give different impressions about the distribution of values? b. Use each histogram to determine approximately the proportion of observations that are less than 800, and compare to the actual proportion. Frequency 10
5
Frequency 6 5 4 3 2 1 0 375
525
675
825
975
1125
Sodium
Americium 241 (241Am) is a radioactive material used in the manufacture of smoke detectors. The article “Retention and Dosimetry of Injected 241Am in Beagles” (Radiation Research [1984]: 564–575) described a study in which 55 beagles were injected with a dose of 241Am (proportional to each animal’s weight). Skeletal retention of 241Am (in microcuries per kilogram) was recorded for each beagle, resulting in the following data:
3.71
0.196 0.300 0.287 0.315 0.335 0.333 0.306 0.353
0.451 0.346 0.243 0.447 0.332 0.408 0.367 0.357
0.498 0.448 0.334 0.585 0.292 0.399 0.345 0.320
0.411 0.188 0.299 0.291 0.375 0.303 0.428 0.354
0.324 0.399 0.292 0.186 0.349 0.318 0.345 0.361
0.190 0.305 0.419 0.393 0.324 0.468 0.412 0.329
0.489 0.304 0.236 0.419 0.301 0.441 0.337
a. Construct a frequency distribution for these data, and draw the corresponding histogram. b. Write a short description of the important features of the shape of the histogram.
0 300
550
800
1050
Sodium
Bold exercises answered in back
Data set available online
Video Solution available
Cumulative Review Exercises CR3.1 - CR3.16 CR3.1 Does eating broccoli reduce the risk of prostate cancer? According to an observational study from the Fred Hutchinson Cancer Research Center (see the CNN.com web site article titled “Broccoli, Not Pizza Bold exercises answered in back
Data set available online
Sauce, Cuts Cancer Risk, Study Finds,” January 5, 2000), men who ate more cruciferous vegetables (broccoli, cauliflower, brussels sprouts, and cabbage) had a lower risk of prostate cancer. This study made separate Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
Cumulative Review Exercises
comparisons for men who ate different levels of vegetables. According to one of the investigators, “at any given level of total vegetable consumption, as the percent of cruciferous vegetables increased, the prostate cancer risk decreased.” Based on this study, is it reasonable to conclude that eating cruciferous vegetables causes a reduction in prostate cancer risk? Explain. An article that appeared in USA Today (August 11, 1998) described a study on prayer and blood pressure.
CR3.2
In this study, 2391 people 65 years or older, were followed for 6 years. The article stated that people who attended a religious service once a week and prayed or studied the Bible at least once a day were less likely to have high blood pressure. The researcher then concluded that “attending religious services lowers blood pressure”. The headline for this article was “Prayer Can Lower Blood Pressure.” Write a few sentences commenting on the appropriateness of the researcher’s conclusion and on the article headline. Sometimes samples are composed entirely of volunteer responders. Give a brief description of the dangers of using voluntary response samples.
CR3.3
CR3.4 A newspaper headline stated that at a recent budget workshop, nearly three dozen people supported a sales tax increase to help deal with the city’s financial deficit (San Luis Obispo Tribune, January 22, 2005). This conclusion was based on data from a survey acknowledged to be unscientific, in which 34 out of the 43 people who chose to attend the budget workshop recommended raising the sales tax. Briefly discuss why the survey was described as “unscientific” and how this might limit the conclusions that can be drawn from the survey data. CR3.5 “More than half of California’s doctors say they are so frustrated with managed care they will quit, retire early, or leave the state within three years.” This conclusion from an article titled “Doctors Feeling Pessimistic, Study Finds” (San Luis Obispo Tribune, July 15, 2001) was based on a mail survey conducted by the California Medical Association. Surveys were mailed to 19,000 California doctors, and 2000 completed surveys were returned. Describe any concerns you have regarding the conclusion drawn.
CR3.6 Based on observing more than 400 drivers in the Atlanta area, two investigators at Georgia State University concluded that people exiting parking spaces did so more Bold exercises answered in back
Data set available online
159
slowly when a driver in another car was waiting for the space than when no one was waiting (“Territorial Defense
in Parking Lots: Retaliation Against Waiting Drivers,” Journal of Applied Social Psychology [1997]: 821-834). Describe how you might design an experiment to determine whether this phenomenon is true for your city. What is the response variable? What are some extraneous variables and how does your design control for them?
CR3.7 An article from the Associated Press (May 14, 2002) led with the headline “Academic Success Lowers Pregnancy Risk.” The article described an evaluation of a program that involved about 350 students at 18 Seattle schools in high crime areas. Some students took part in a program beginning in elementary school in which teachers showed children how to control their impulses, recognize the feelings of others, and get what they want without aggressive behavior. Others did not participate in the program. The study concluded that the program was effective because by the time young women in the program reached age 21, the pregnancy rate among them was 38%, compared to 56% for the women in the experiment who did not take part in the program. Explain why this conclusion is valid only if the women in the experiment were randomly assigned to one of the two experimental groups.
CR3.8 Researchers at the University of Pennsylvania suggest that a nasal spray derived from pheromones (chemicals emitted by animals when they are trying to attract a mate) may be beneficial in relieving symptoms of premenstrual syndrome (PMS) (Los Angeles Times, January 17, 2003). a. Describe how you might design an experiment using 100 female volunteers who suffer from PMS to determine whether the nasal spray reduces PMS symptoms. b. Does your design from Part (a) include a placebo treatment? Why or why not? c. Does your design from Part (a) involve blinding? Is it single-blind or double-blind? Explain.
CR3.9 Students in California are required to pass an exit exam in order to graduate from high school. The pass rate for San Luis Obispo High School has been rising, as have the rates for San Luis Obispo County and the state of California (San Luis Obispo Tribune, August 17, 2004). The percentage of students who passed the test was as follows: Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
160
Chapter 3
Graphical Methods for Describing Data
Year
District
Pass Rate
2002 2003 2004 2002 2003 2004 2002 2003 2004
San Luis Obispo High School
66% 72% 93% 62% 57% 85% 32% 43% 74%
San Luis Obispo County
State of California
a. Construct a comparative bar chart that allows the change in the pass rate for each group to be compared. b. Is the change the same for each group? Comment on any difference observed.
CR3.10 A poll conducted by the Associated Press–Ipsos on public attitudes found that most Americans are convinced that political corruption is a major problem (San Luis Obispo Tribune, December 9, 2005). In the poll, 1002 adults were surveyed. Two of the questions and the summarized responses to these questions follow: How widespread do you think corruption is in public service in America? Hardly anyone A small number A moderate number A lot of people Almost everyone Not sure
1% 20% 39% 28% 10% 2%
In general, which elected officials would you say are more ethical? Democrats Republicans Both equally Neither Not sure
36% 33% 10% 15% 6%
a. For each question, construct a pie chart summarizing the data. b. For each question, construct a segmented bar chart displaying the data. c. Which type of graph (pie chart or segmented bar graph) does a better job of presenting the data? Explain. The article “Determination of Most Representative Subdivision” (Journal of Energy Engineering [1993]: 43–55) gave data on various characteristics of
CR3.11
Bold exercises answered in back
Data set available online
subdivisions that could be used in deciding whether to provide electrical power using overhead lines or underground lines. Data on the variable x ⫽ total length of streets within a subdivision are as follows: 1280 360 3350 450 1850 3150
5320 3330 540 2250 2460 1890
4390 3380 3870 2320 5850 510
2100 340 1250 2400 2700 240
1240 1000 2400 3150 2730 396
3060 960 960 5700 1670 1419
4770 1320 1120 5220 100 2109
1050 530 2120 500 5770 5770
a. Construct a stem-and-leaf display for these data using the thousands digit as the stem. Comment on the various features of the display. b. Construct a histogram using class boundaries of 0 to ,1000, 1000 to ,2000, and so on. How would you describe the shape of the histogram? c. What proportion of subdivisions has total length less than 2000? between 2000 and 4000? The paper “Lessons from Pacemaker Implantations” ( Journal of the American Medical Association [1965]: 231–232) gave the results of a study that
CR3.12
followed 89 heart patients who had received electronic pacemakers. The time (in months) to the first electrical malfunction of the pacemaker was recorded: 24 8 24 34 24 26 14 30
20 16 26 18 10 20 18 22
16 12 28 20 14 6 24 24
32 24 18 22 16 14 22 22
14 22 14 20 14 10 24 26 22 24 16 18 28 24 328
2 18 12 18 22 24 30
12 14 24 2 20 18 34
24 16 6 18 24 16 26
6 18 12 12 28 6 24
10 20 18 12 20 16 22
20 22 16 8 22 10 28
a. Summarize these data in the form of a frequency distribution, using class intervals of 0 to ⬍6, 6 to ⬍12, and so on. b. Compute the relative frequencies and cumulative relative frequencies for each class interval of the frequency distribution of Part (a). c. Show how the relative frequency for the class interval 12 to ⬍18 could be obtained from the cumulative relative frequencies. d. Use the cumulative relative frequencies to give approximate answers to the following: i. What proportion of those who participated in the study had pacemakers that did not malfunction within the first year? Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
161
Cumulative Review Exercises
ii. If the pacemaker must be replaced as soon as the first electrical malfunction occurs, approximately what proportion required replacement between 1 and 2 years after implantation? e. Construct a cumulative relative frequency plot, and use it to answer the following questions. i. What is the approximate time at which about 50% of the pacemakers had failed? ii. What is the approximate time at which only about 10% of the pacemakers initially implanted were still functioning?
CR3.13 How does the speed of a runner vary over the course of a marathon (a distance of 42.195 km)? Consider determining both the time (in seconds) to run the first 5 km and the time (in seconds) to run between the 35 km and 40 km points, and then subtracting the 5-km time from the 35–40-km time. A positive value of this difference corresponds to a runner slowing down toward the end of the race. The histogram below is based on times of runners who participated in several different Japanese marathons (“Factors Affecting Runners’ Marathon Performance,” Chance [Fall 1993]: 24–30). What are some interesting features of this histogram? What is a typical difference value? Roughly what proportion of the runners ran the late distance more quickly than the early distance? Frequency 200 150
Dropout rate
15
10
5 10
15 Poverty rate
20
Write a few sentences commenting on this scatterplot. Would you describe the relationship between poverty rate and dropout rate as positive (y tends to increase as x increases), negative (y tends to decrease as x increases), or as having no discernible relationship between x and y? One factor in the development of tennis elbow, a malady that strikes fear into the hearts of all serious players of that sport, is the impact-induced vibration of the racket-and-arm system at ball contact. It is well known that the likelihood of getting tennis elbow depends on various properties of the racket used. Consider the accompanying scatterplot of x ⫽ racket resonance frequency (in hertz) and y ⫽ sum of peak-to-peak accelerations (a characteristic of arm vibration, in meters per second per second) for n ⫽ 23 different rackets
CR3.15
(“Transfer of Tennis Racket Vibrations into the Human Forearm,” Medicine and Science in Sports and Exercise [1992]: 1134–1140). Discuss interesting features
100
of the data and of the scatterplot. 50
−100
100 200 300 400 500 600 700
800
y
Time difference
Data on x ⫽ poverty rate (%) and y ⫽ high school dropout rate (%) for the 50 U.S. states and the District of Columbia were used to construct the following scatterplot (Chronicle of Higher Education, August 31, 2001):
CR3.14
38 36 34 32 30 28 26 24 22 100 110 120 130 140 150 160 170 180
Bold exercises answered in back
Data set available online
x 190
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
162
Chapter 3 Graphical Methods for Describing Data
CR3.16 An article that appeared in USA Today (September 3, 2003) included a graph similar to the one shown here summarizing responses from polls conducted in 1978, 1991, and 2003 in which a sample of American adults were asked whether or not it was a good time or a bad time to buy a house. a. Construct a time-series plot that shows how the percentage that thought it was a good time to buy a house has changed over time. b. Add a new line to the plot from Part (a) showing the percentage that thought it was a bad time to buy a house over time. Be sure to label the lines clearly. c. Which graph, the given bar chart or the time-series plot, best shows the trend over time? Bold exercises answered in back
Data set available online
Mar-78 90% 78% 78%
80%
Jun-03
67%
70% 60%
Mar-91
Sep-03
53%
50% 40% 29%
30%
25% 19% 20%
20% 10% 0% Good time
Bad time
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
CHAPTER
4
Numerical Methods for Describing Data In 2006, Medicare introduced a new prescription drug program. The article “Those Most in Need May Miss Drug Benefit Sign-Up” (USA Today, May 9, 2006) notes that only 24% of those eligible for low-income subsidies under this program had signed up just 2 weeks before the enrollment deadline. The article also gave the percentage of those eligible who had signed up in each of 49 states and the District of Columbia (information was not available for Vermont):
Hideji Watanabe/Sebun Photo/ amana images/Getty Images
24 16 27 25 14
27 21 22 19 18
12 28 19 17
38 20 22 21
21 21 22 27
26 41 22 19
23 22 30 27
33 16 20 34
19 29 21 20
19 26 34 30
26 22 26 20
28 16 20 21
Make the most of your study time by accessing everything you need to succeed online with CourseMate. Visit http://www.cengagebrain.com where you will find: • An interactive eBook, which allows you to take notes, highlight, bookmark, search • • • • • •
the text, and use in-context glossary definitions Step-by-step instructions for Minitab, Excel, TI-83/84, SPSS, and JMP Video solutions to selected exercises Data sets available for selected examples and exercises Online quizzes Flashcards Videos
163 Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
164
Chapter 4
Numerical Methods for Describing Data
What is a typical value for this data set? Is the nationwide figure of 24% representative of the individual state percentages? The enrollment percentages differ widely from state to state, ranging from a low of 12% (Arizona) to a high of 41% (Kentucky). How might we summarize this variability numerically? In this chapter, we show how to calculate numerical summary measures that describe more precisely both the center and the extent of spread in a data set. In Section 4.1, we introduce the mean and the median, the two most widely used measures of the center of a distribution. The variance and the standard deviation are presented in Section 4.2 as measures of variability. In later sections, we will see some additional ways that measures of center and spread can be used to describe data distributions.
4.1
Describing the Center of a Data Set When describing numerical data, it is common to report a value that is representative of the observations. Such a number describes roughly where the data are located or “centered” along the number line, and is called a measure of center. The two most widely used measures of center are the mean and the median.
The Mean The mean of a numerical data set is just the familiar arithmetic average: the sum of the observations divided by the number of observations. It is helpful to have concise notation for the variable on which observations were made, for the number of observations in the data set, and for the individual observations: x 5 the variable for which we have sample data n 5 the number of observations in the data set (the sample size) x1 5 the first observation in the data set x2 5 the second observation in the data set ( xn 5 the nth (last) observation in the data set For example, we might have a sample consisting of n 4 observations on x battery lifetime (in hours): x1 5.9 x2 7.3 x3 6.6 x4 5.7 Notice that the value of the subscript on x has no relationship to the magnitude of the observation. In this example, x1 is just the first observation in the data set and not necessarily the smallest observation, and xn is the last observation but not necessarily the largest. The sum of x1, x2, . . . , xn can be denoted by x1 x2 c xn, but this is cumbersome. The Greek letter g is traditionally used in mathematics to denote summation. In particular, g x denotes the sum of all the x values in the data set under consideration.*
DEFINITION The sample mean of a sample consisting of numerical observations x1, x2, . . . , xn, denoted by x , is x5
sum of all observations in the sample x1 1 x2 1 c1 xn gx 5 5 n n number of observations in the sample n
*It is also common to see g x written as g xi or even as a xi, but for simplicity we will usually omit the summation i51 indices.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
4.1
EXAMPLE 4.1
Describing the Center of a Data Set
165
Improving Knee Extension
Increasing joint extension is one goal of athletic trainers. In a study to investigate the effect of a therapy that uses ultrasound and stretching (Trae Tashiro, Masters Thesis, University of Virginia, 2004) passive knee extension was measured after treatment. Passive knee extension (in degrees) is given for each of 10 participants in the study: x1 5 59 x2 5 46 x3 5 64 x4 5 49 x5 5 56 x6 5 70 x7 5 45 x8 5 52 x9 5 63 x10 5 52
The sum of these sample values is 59 46 64 c 52 556, and the sample mean passive knee extension is x5
556 gx 5 5 55.6 n 10
We would report 55.6 degrees as a representative value of passive knee extension for this sample (even though there is no person in the sample that actually had a passive knee extension of 55.6 degrees).
The data values in Example 4.1 were all integers, yet the mean was given as 55.6. It is common to use more digits of decimal accuracy for the mean. This allows the value of the mean to fall between possible observable values (for example, the average number of children per family could be 1.8, whereas no single family will have 1.8 children). The sample mean x is computed from sample observations, so it is a characteristic of the particular sample in hand. It is customary to use Roman letters to denote sample characteristics, as we have done with x. Characteristics of the population are usually denoted by Greek letters. One of the most important of such characteristics is the population mean.
DEFINITION The population mean, denoted by m, is the average of all x values in the entire population.
For example, the average fuel efficiency for all 600,000 cars of a certain type under specified conditions might be m 5 27.5 mpg. A sample of n 5 5 cars might yield efficiencies of 27.3, 26.2, 28.4, 27.9, 26.5, from which we obtain x 5 27.26 for this particular sample (somewhat smaller than m). However, a second sample might give x 5 28.52, a third x 5 26.85, and so on. The value of x varies from sample to sample, whereas there is just one value for m. In later chapters, we will see how the value of x from a particular sample can be used to draw various conclusions about the value of m. Example 4.2 illustrates how the value of x from a particular sample can differ from the value of m and how the value of x differs from sample to sample. Data set available online Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
166
Chapter 4 Numerical Methods for Describing Data
E X A M P L E 4 . 2 County Population Sizes The 50 states plus the District of Columbia contain 3137 counties. Let x denote the number of residents of a county. Then there are 3137 values of the variable x in the population. The sum of these 3137 values is 293,655,404 (2004 Census Bureau estimate), so the population average value of x is m5
293,655,404 5 93,610.27 residents per county 3137
We used the Census Bureau web site to select three different samples at random from this population of counties, with each sample consisting of five counties. The results appear in Table 4.1, along with the sample mean for each sample. Not only are the three x values different from one another—because they are based on three different samples and the value of x depends on the x values in the sample—but also none of the three values comes close to the value of the population mean, m. If we did not know the value of m but had only Sample 1 available, we might use x as an estimate of m, but our estimate would be far off the mark.
TABLE 4 . 1 Three Samples from the Population of All U.S. Counties (x number of residents) SAMPLE 1
County Fayette, TX Monroe, IN Greene, NC Shoshone, ID Jasper, IN
SAMPLE 2
x Value
County
22,513 121,013 20,219 12,827 31,624 gx 5 208,196 x 5 41,639.2
Stoddard, MO Johnston, OK Sumter, AL Milwaukee, WI Albany, WY
SAMPLE 3
x Value 29,773 10,440 14,141 928,018 31,473 gx 5 1,013,845 x 5 202,769.0
County Chattahoochee, GA Petroleum, MT Armstrong, PA Smith, MI Benton, MO
x Value 13,506 492 71,395 14,306 18,519 gx 5 118,218 x 5 23,643.6
Alternatively, we could combine the three samples into a single sample with n 15 observations: x1 5 22,513, . . . , x5 5 31,624, . . . , x15 5 18,519 g x 5 1,340,259 x5
1,340,259 5 89,350.6 15
This value is closer to the value of m but is still somewhat unsatisfactory as an estimate. The problem here is that the population of x values exhibits a lot of variability (the largest value is x 9,937,739 for Los Angeles County, California, and the smallest value is x 52 for Loving County, Texas, which evidently few people love). Therefore, it is difficult for a sample of 15 observations, let alone just 5, to be reasonably representative of the population. In Chapter 9, you will see how to take variability into account when deciding on a sample size.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
4.1
167
Describing the Center of a Data Set
One potential drawback to the mean as a measure of center for a data set is that its value can be greatly affected by the presence of even a single outlier (an unusually large or small observation) in the data set.
EXAMPLE 4.3
Number of Visits to a Class Web Site
Forty students were enrolled in a section of a general education course in statistical reasoning during one fall quarter at Cal Poly, San Luis Obispo. The instructor made course materials, grades, and lecture notes available to students on a class web site, and course management software kept track of how often each student accessed any of the web pages on the class site. One month after the course began, the instructor requested a report that indicated how many times each student had accessed a web page on the class site. The 40 observations were: 20 0 4 13
37 22 0 12
4 3 5 8
20 13 23 42
0 14 19
84 36 7
14 4 12
36 0 8
5 18 13
331 8 16
19 0 21
0 26 7
The sample mean for this data set is x 5 23.10. Figure 4.1 is a Minitab dotplot of the data. Many would argue that 23.10 is not a very representative value for this sample, because 23.10 is larger than most of the observations in the data set— only 7 of 40 observations, or 17.5%, are larger than 23.10. The two outlying values of 84 and 331 (no, that was not a typo!) have a substantial impact on the value of x.
FIGURE 4.1 A Minitab dotplot of the data in Example 4.3.
100
200 Number of accesses
300
We now turn our attention to a measure of center that is not as sensitive to outliers—the median.
The Median The median strip of a highway divides the highway in half, and the median of a numerical data set does the same thing for a data set. Once the data values have been listed in order from smallest to largest, the median is the middle value in the list, and it divides the list into two equal parts. Depending on whether the sample size n is even or odd, the process of determining the median is slightly different. When n is an odd number (say, 5), the sample median is the single middle value. But when n is even (say, 6), there are two middle values in the ordered list, and we average these two middle values to obtain the sample median. Step-by-Step technology instructions available online Data set available online
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
168
Chapter 4
Numerical Methods for Describing Data
DEFINITION The sample median is obtained by first ordering the n observations from smallest to largest (with any repeated values included, so that every sample observation appears in the ordered list). Then sample median 5 e
EXAMPLE 4.4
the single middle value if n is odd the average of the middle two values if n is even
Web Site Data Revised
The sample size for the web site access data of Example 4.3 was n 40, an even number. The median is the average of the 20th and 21st values (the middle two) in the ordered list of the data. Arranging the data in order from smallest to largest produces the following ordered list (with the two middle values highlighted): 0 7 16 37
0 7 18 42
0 8 19 84
0 8 19 331
0 8 20
0 12 20
3 12 21
4 13 22
4 13 23
4 13 26
5 14 36
5 14 36
The median can now be determined: median 5
13 1 13 5 13 2
Looking at the dotplot (Figure 4.1), we see that this value appears to be a more typical value for the data set than the sample mean x 5 23.10 is.
The sample mean can be sensitive to even a single value that lies far above or below the rest of the data. The value of the mean is pulled out toward such an outlying value or values. The median, on the other hand, is quite insensitive to outliers. For example, the largest sample observation (331) in Example 4.4 can be increased by any amount without changing the value of the median. Similarly, an increase in the second or third largest observations does not affect the median, nor would a decrease in several of the smallest observations. This stability of the median is what sometimes justifies its use as a measure of center in some situations. For example, the article “Educating Undergraduates on Using Credit Cards” (Nellie Mae, 2005) reported that the mean credit card debt for undergraduate students in 2001 was $2327, whereas the median credit card debt was only $1770. In this case, the small percentage of students with unusually high credit card debt may be resulting in a mean that is not representative of a typical student’s credit card debt.
Comparing the Mean and the Median Figure 4.2 shows several smoothed histograms that might represent either a distribution of sample values or a population distribution. Pictorially, the median is the value on the measurement axis that separates the smoothed histogram into two parts, with .5 (50%) of the area under each part of the curve. The mean is a bit harder to visualize. If the Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
4.1
Describing the Center of a Data Set
169
histogram were balanced on a triangle (a fulcrum), it would tilt unless the triangle was positioned at the mean. The mean is the balance point for the distribution. Equal areas
Fulcrum
FIGURE 4.2 The mean and the median.
Median
Mean
Mean
When the histogram is symmetric, the point of symmetry is both the dividing point for equal areas and the balance point, and the mean and the median are equal. However, when the histogram is unimodal (single-peaked) with a longer upper tail (positively skewed), the outlying values in the upper tail pull the mean up, so it generally lies above the median. For example, an unusually high exam score raises the mean but does not affect the median. Similarly, when a unimodal histogram is negatively skewed, the mean is generally smaller than the median (see Figure 4.3).
FIGURE 4.3 Relationship between the mean and the median.
Mean = Median
Median
Mean
Mean
Median
Trimmed Means The extreme sensitivity of the mean to even a single outlier and the extreme insensitivity of the median to a substantial proportion of outliers can sometimes make both of them suspect as a measure of center. A trimmed mean is a compromise between these two extremes.
DEFINITION A trimmed mean is computed by first ordering the data values from smallest to largest, deleting a selected number of values from each end of the ordered list, and finally averaging the remaining values. The trimming percentage is the percentage of values deleted from each end of the ordered list. Sometimes the number of observations to be deleted from each end of the data set is specified. Then the corresponding trimming percentage is calculated as trimming percentage 5 a
number deleted from each end # b 100 n
In other cases, the trimming percentage is specified and then used to determine how many observations to delete from each end, with number deleted from each end 5 a
trimming percentage # b n 100
If the number of observations to be deleted from each end resulting from this calculation is not an integer, it can be rounded to the nearest integer (which changes the trimming percentage a bit). Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
170
Chapter 4 Numerical Methods for Describing Data
EXAMPLE 4.5
NBA Salaries
The web site HoopsHype (hoopshype.com/salaries) publishes salaries of NBA players. Salaries for the players of the Chicago Bulls in 2009 were
Player
2009 Salary
Brad Miller Luol Deng Kirk Hinrich Jerome James Tim Thomas John Salmons Derrick Rose Tyrus Thomas Joakim Noah Jannero Pargo James Johnson Lindsey Hunter Taj Gibson Aaron Gray
$12,250,000 $10,370,425 $9,500,000 $6,600,000 $6,466,600 $5,456,000 $5,184,480 $4,743,598 $2,455,680 $2,000,000 $1,594,080 $1,306,455 $1,039,800 $1,000,497
A Minitab dotplot of these data is shown in Figure 4.4(a). Because the data distribution is not symmetric and there are outliers, a trimmed mean is a reasonable choice for describing the center of this data set. There are 14 observations in this data set. Deleting the two largest and the two smallest observations from the data set and then averaging the remaining values 2 would result in a a b 11002 5 14% trimmed mean. Based on the Bulls’ salary data, 14 the two largest salaries are $12,250,000 and $10,370,425, and the two smallest are $1,039,800 and $1,000,497. The average of the remaining 10 observations is 9,500,000 1 c1 1,306,445 45,306,893 14% trimmed mean 5 5 5 4,530,689 10 10
Data set available online
The mean ($4,997,687) is larger than the trimmed mean because of the few unusually large values in the data set. For the L.A. Lakers, the difference between the mean ($7,035,947) and the 14% trimmed mean ($5,552,607) is even more dramatic because in 2009 one
5,000,000
(a)
FIGURE 4.4 Minitab dotplots for NBA salary data (a) Bulls (b) Lakers.
0 (b)
10,000,000
15,000,000
20,000,000
25,000,000
20,000,000
25,000,000
2009 Salary (Bulls)
5,000,000
10,000,000
15,000,000
2009 Salary (Lakers)
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
4.1
Describing the Center of a Data Set
171
player on the Lakers earned over $23 million and two players earned well over $10 million (see Figure 4.4(b)).
Categorical Data The natural numerical summary quantities for a categorical data set are the relative frequencies for the various categories. Each relative frequency is the proportion (fraction) of responses that is in the corresponding category. Often there are only two possible responses (a dichotomy)—for example, male or female, does or does not have a driver’s license, did or did not vote in the last election. It is convenient in such situations to label one of the two possible responses S (for success) and the other F (for failure). As long as further analysis is consistent with the labeling, it does not matter which category is assigned the S label. When the data set is a sample, the fraction of S’s in the sample is called the sample proportion of successes.
DEFINITION The sample proportion of successes, denoted by p^ , is number of S’s in the sample p^ 5 sample proportion of successes 5 n where S is the label used for the response designated as success.
E X A M P L E 4 . 6 Can You Hear Me Now?
Getty Images
It is not uncommon for a cell phone user to complain about the quality of his or her service provider. Suppose that each person in a sample of n 15 cell phone users is asked if he or she is satisfied with the cell phone service. Each response is classified as S (satisfied) or F (not satisfied). The resulting data are S S
F S
S S
S F
S F
F
F
S
S
F
This sample contains nine S’s, so p^ 5
9 5 .60 15
That is, 60% of the sample responses are S’s. Of those surveyed, 60% are satisfied with their cell phone service.
The letter p is used to denote the population proportion of S’s.* We will see later how the value of p^ from a particular sample can be used to make inferences about p. *Note that this is one situation in which we will not use a Greek letter to denote a population characteristic. Some statistics books use the symbol p for the population proportion and p for the sample proportion. We will not use p in this context so there is no confusion with the mathematical constant p 3.14. . . .
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
172
Chapter 4 Numerical Methods for Describing Data
EX E RC I S E S 4 . 1 - 4 . 1 6 4.1 The Insurance Institute for Highway Safety (www.iihs.org, June 11, 2009) published data on repair
4.3
costs for cars involved in different types of accidents. In one study, seven different 2009 models of mini- and micro-cars were driven at 6 mph straight into a fixed barrier. The following table gives the cost of repairing damage to the bumper for each of the seven models.
feine concentration (mg/cup) for 12 brands of coffee:
Model Smart Fortwo Chevrolet Aveo Mini Cooper Toyota Yaris Honda Fit Hyundai Accent Kia Rio
Repair Cost $1,480 $1,071 $2,291 $1,688 $1,124 $3,476 $3,701
Compute the values of the mean and median. Why are these values so different? Which of the two—mean or median—appears to be better as a description of a typical value for this data set?
4.2
The article “Caffeinated Energy Drinks—A
Growing Problem” (Drug and Alcohol Dependence [2009]: 1–10) gave the following data on caffeine concentration (mg/ounce) for eight top-selling energy drinks:
Energy Drink Red Bull Monster Rockstar Full Throttle No Fear Amp SoBe Adrenaline Rush Tab Energy
Caffeine Concentration (mg/oz) 9.6 10.0 10.0 9.0 10.9 8.9 9.5 9.1
a. What is the value of the mean caffeine concentration for this set of top-selling energy drinks? x 5 9.625 b. Coca-Cola has 2.9 mg/ounce of caffeine and Pepsi Cola has 3.2 mg/ounce of caffeine. Write a sentence explaining how the caffeine concentration of topselling energy drinks compares to that of these colas.
Bold exercises answered in back
Data set available online
Consumer Reports Health (www.consumer reports.org/health) reported the accompanying cafCaffeine Concentration (mg/cup)
Coffee Brand Eight O’Clock Caribou Kickapoo Starbucks Bucks Country Coffee Co. Archer Farms Gloria Jean’s Coffees Chock Full o’Nuts Peet’s Coffee Maxwell House Folgers Millstone
140 195 155 115 195 180 110 110 130 55 60 60
Use at least one measure of center to compare caffeine concentration for coffee with that of the energy drinks of the previous exercise. (Note: 1 cup 5 8 ounces)
4.4
Consumer Reports Health (www.consumer reports.org/health) reported the sodium content (mg)
per 2 tablespoon serving for each of 11 different peanut butters: 120 170
50 250
140 110
120
150
150
150
65
a. Display these data using a dotplot. Comment on any unusual features of the plot. b. Compute the mean and median sodium content for the peanut butters in this sample. c. The values of the mean and the median for this data set are similar. What aspect of the distribution of sodium content—as pictured in the dotplot from Part (a)—provides an explanation for why the values of the mean and median are similar?
4.5 In August 2009, Harris Interactive released the results of the “Great Schools” survey. In this survey, 1086 parents of children attending a public or private school were asked approximately how much time they spent volunteering at school per month over the last school year. For this sample, the mean number of hours per month was 5.6 hours and the median number of hours was 1.0. What does the large difference between the mean and median tell you about this data set? Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
4.1
The accompanying data on number of minutes used for cell phone calls in one month was generated to be consistent with summary statistics published in a report of a marketing study of San Diego residents (TeleTruth, March 2009):
4.6
189 0 189 177 106 201 0 212 0 306 0 0 59 224 0 189 142 83 71 165 236 0 142 236 130 a. Would you recommend the mean or the median as a measure of center for this data set? Give a brief explanation of your choice. (Hint: It may help to look at a graphical display of the data.) b. Compute a trimmed mean by deleting the three smallest observations and the three largest observations in the data set and then averaging the remaining 19 observations. What is the trimming percentage for this trimmed mean? c. What trimming percentage would you need to use in order to delete all of the 0 minute values from the data set? Would you recommend a trimmed mean with this trimming percentage? Explain why or why not.
4.7 USA Today (May 9, 2006) published the accompanying average weekday circulation for the 6-month period ending March 31, 2006, for the top 20 newspapers in the country: 2,272,815 2,049,786 1,142,464 851,832 724,242 708,477 673,379 579,079 513,387 438,722 427,771 398,329 398,246 397,288 365,011 362,964 350,457 345,861 343,163 323,031 a. Do you think the mean or the median will be larger for this data set? Explain. b. Compute the values of the mean and the median of this data set. c. Of the mean and median, which does the best job of describing a typical value for this data set? d. Explain why it would not be reasonable to generalize from this sample of 20 newspapers to the population of all daily newspapers in the United States. The chapter introduction gave the accompanying data on the percentage of those eligible for a lowincome subsidy who had signed up for a Medicare drug plan in each of 49 states (information was not available for Vermont) and the District of Columbia (USA Today, May 9, 2006).
4.8
Bold exercises answered in back
Data set available online
24 19 21 27 21 27 14
27 19 41 22 34 19 18
173
Describing the Center of a Data Set
12 26 22 19 26 27
38 28 16 22 20 34
21 16 29 22 25 20
26 21 26 22 19 30
23 28 22 30 17 20
33 20 16 20 21 21
a. Compute the mean for this data set. b. The article stated that nationwide, 24% of those eligible had signed up. Explain why the mean of this data set from Part (a) is not equal to 24. (No information was available for Vermont, but that is not the reason that the mean differs—the 24% was calculated excluding Vermont.)
4.9 The U.S. Department of Transportation reported the number of speeding-related crash fatalities for the 20 days of the year that had the highest number of these fatalities between 1994 and 2003 (Traffic Safety Facts, July 2005). Date
Speeding-Related Fatalities
Date
Speeding-Related Fatalities
Jan 1 Jul 4 Aug 12 Nov 23 Jul 3 Dec 26 Aug 4 Aug 31 May 25 Dec 23
521 519 466 461 458 455 455 446 446 446
Aug 17 Dec 24 Aug 25 Sep 2 Aug 6 Aug 10 Sept 21 Jul 27 Sep 14 May 27
446 436 433 433 431 426 424 422 422 420
a. Compute the mean number of speeding-related fatalities for these 20 days. b. Compute the median number of speeding-related fatalities for these 20 days. c. Explain why it is not reasonable to generalize from this sample of 20 days to the other 345 days of the year.
4.10 The ministry of Health and Long-Term Care in Ontario, Canada, publishes information on its web site (www.health.gov.on.ca) on the time that patients must wait for various medical procedures. For two cardiac procedures completed in fall of 2005, the following information was provided:
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
174
Chapter 4 Numerical Methods for Describing Data
Angioplasty Bypass surgery
Number of Completed Procedures
Median Wait Time (days)
Mean Wait Time (days)
90% Completed Within (days)
847 539
14 13
18 19
39 42
a. The median wait time for angioplasty is greater than the median wait time for bypass surgery but the mean wait time is shorter for angioplasty than for bypass surgery. What does this suggest about the distribution of wait times for these two procedures? b. Is it possible that another medical procedure might have a median wait time that is greater than the time reported for “90% completed within”? Explain.
4.11 Houses in California are expensive, especially on the Central Coast where the air is clear, the ocean is blue, and the scenery is stunning. The median home price in San Luis Obispo County reached a new high in July 2004, soaring to $452,272 from $387,120 in March 2004. (San Luis Obispo Tribune, April 28, 2004). The article included two quotes from people attempting to explain why the median price had increased. Richard Watkins, chairman of the Central Coast Regional Multiple Listing Services was quoted as saying, “There have been some fairly expensive houses selling, which pulls the median up.” Robert Kleinhenz, deputy chief economist for the California Association of Realtors explained the volatility of house prices by stating: “Fewer sales means a relatively small number of very high or very low home prices can more easily skew medians.” Are either of these statements correct? For each statement that is incorrect, explain why it is incorrect and propose a new wording that would correct any errors in the statement.
4.12 Consider the following statement: More than 65% of the residents of Los Angeles earn less than the average wage for that city. Could this statement be correct? If so, how? If not, why not?
Suppose that one more piece is selected and denote its weight by x5. Find a value of x5 such that x sample median.
4.14 Suppose that 10 patients with meningitis received treatment with large doses of penicillin. Three days later, temperatures were recorded, and the treatment was considered successful if there had been a reduction in a patient’s temperature. Denoting success by S and failure by F, the 10 observations are S
S
F
S
S
S
F
F
S
S
a. What is the value of the sample proportion of successes? b. Replace each S with a 1 and each F with a 0. Then calculate x for this numerically coded sample. How does x compare to p^ ? c. Suppose that it is decided to include 15 more patients in the study. How many of these would have to be S’s to give p^ .80 for the entire sample of 25 patients?
4.15 An experiment to study the lifetime (in hours) for a certain brand of light bulb involved putting 10 light bulbs into operation and observing them for 1000 hours. Eight of the light bulbs failed during that period, and those lifetimes were recorded. The lifetimes of the two light bulbs still functioning after 1000 hours are recorded as 10001. The resulting sample observations were 480 170
790 290
1000
350
920
860
570
1000
Which of the measures of center discussed in this section can be calculated, and what are the values of those measures?
4.16 An instructor has graded 19 exam papers submitted by students in a class of 20 students, and the average so far is 70. (The maximum possible score is 100.) How high would the score on the last paper have to be to raise the class average by 1 point? By 2 points?
A sample consisting of four pieces of luggage was selected from among those checked at an airline counter, yielding the following data on x 5 weight (in pounds):
4.13
x1 33.5, x2 27.3, x3 36.7, x4 30.5
Bold exercises answered in back
Data set available online
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
4.2
4.2
Describing Variability in a Data Set
175
Describing Variability in a Data Set Reporting a measure of center gives only partial information about a data set. It is also important to describe how much the observations differ from one another. The three different samples displayed in Figure 4.5 all have mean 5 median 5 45. There is a lot of variability in the first sample compared to the third sample. The second sample shows less variability than the first and more variability than the third; most of the variability in the second sample is due to the two extreme values being so far from the center. Sample 1.
20, 40, 50, 30, 60, 70
2.
47, 43, 44, 46, 20, 70
3.
44, 43, 40, 50, 47, 46 20
FIGURE 4.5
30
Three samples with the same center and different amounts of variability.
40
50
60
70
Mean = Median
The simplest numerical measure of variability is the range.
DEFINITION The range of a data set is defined as range 5 largest observation 2 smallest observation
In general, more variability will be reflected in a larger range. However, variability is a characteristic of the entire data set, and each observation contributes to variability. The first two samples plotted in Figure 4.5 both have a range of 70 2 20 5 50, but there is less variability in the second sample.
Deviations from the Mean The most widely used measures of variability describe the extent to which the sample observations deviate from the sample mean x. Subtracting x from each observation gives a set of deviations from the mean.
DEFINITION The n deviations from the sample mean are the differences 1 x1 2 x 2 , 1 x2 2 x 2 , p , 1 x n 2 x 2
A particular deviation is positive if the corresponding x value is greater than x and negative if the x value is less than x. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
176
Chapter 4 Numerical Methods for Describing Data
EXAMPLE 4.7
The Big Mac Index
McDonald’s fast-food restaurants are now found in many countries around the world. But the cost of a Big Mac varies from country to country. Table 4.2 shows data on the cost of a Big Mac (converted to U.S. dollars based on the July 2009 exchange rates) taken from the article “Cheesed Off” (The Economist, July 18, 2009).
T AB LE 4.2 Big Mac Prices for 7 Countries Country
Big Mac Price in U.S. Dollars
Argentina Brazil Chile Colombia Costa Rica Peru Uruguay
3.02 4.67 3.28 3.51 3.42 2.76 2.87
Notice that there is quite a bit of variability in the Big Mac prices. For this data set, g x 5 23.53 and x 5 $3.36. Table 4.3 displays the data along with the corresponding deviations, formed by subtracting x 5 3.36 from each observation. Three of the deviations are positive because three of the observations are larger than x. The negative deviations correspond to observations that are smaller than x. Some of the deviations are quite large in magnitude (1.31 and 0.60, for example), indicating observations that are far from the sample mean.
T A B L E 4 .3 Deviations from the Mean for the Big Mac Data Country Argentina Brazil Chile Colombia Costa Rica Peru Uruguay
Data set available online
Big Mac Price in U.S. Dollars
Deviations from Mean
3.02 4.67 3.28 3.51 3.42 2.76 2.87
0.34 1.31
0.08 0.15 0.06 0.60 0.49
In general, the greater the amount of variability in the sample, the larger the magnitudes (ignoring the signs) of the deviations. We now consider how to combine the deviations into a single numerical measure of variability. A first thought might be to calculate the average deviation, by adding the deviations together 1this sum can be denoted compactly by g 1x 2 x 2 2 and then dividing by n. This does not work, though, because negative and positive deviations counteract one another in the summation. As a result of rounding, the value of the sum of the seven deviations in Example 4.7 is g 1x 2 x 2 5 0.01. If we used even more decimal accuracy in computing x the sum would be even closer to zero.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
4.2
Describing Variability in a Data Set
177
Except for the effects of rounding in computing the deviations, it is always true that g 1 x 2 x2 5 0 Since this sum is zero, the average deviation is always zero and so it cannot be used as a measure of variability.
The Variance and Standard Deviation The customary way to prevent negative and positive deviations from counteracting one another is to square them before combining. Then deviations with opposite signs but with the same magnitude, such as 12 and 22, make identical contributions to variability. The squared deviations are 1x1 2 x 2 2, 1x2 2 x 2 2, . . . , 1xn 2 x 2 2 and their sum is 1x1 2 x 2 2 1 1x2 2 x 2 2 1 c1 1xn 2 x 2 2 5 g 1x 2 x 2 2 Common notation for g 1x 2 x2 2 is Sxx. Dividing this sum by the sample size n gives the average squared deviation. Although this seems to be a reasonable measure of variability, we use a divisor slightly smaller than n. (The reason for this will be explained later in this section and in Chapter 9.)
DEFINITION The sample variance, denoted by s 2, is the sum of squared deviations from the mean divided by n 2 1. That is, g 1x 2 x 2 2 S 5 xx n21 n21 The sample standard deviation is the positive square root of the sample variance and is denoted by s. s2 5
A large amount of variability in the sample is indicated by a relatively large value of s 2 or s, whereas a value of s 2 or s close to zero indicates a small amount of variability. Notice that whatever unit is used for x (such as pounds or seconds), the squared deviations and therefore s 2 are in squared units. Taking the square root gives a measure expressed in the same units as x. Thus, for a sample of heights, the standard deviation might be s 5 3.2 inches, and for a sample of textbook prices, it might be s 5 $12.43.
E X A M P L E 4 . 8 Big Mac Revisited Let’s continue using the Big Mac data and the computed deviations from the mean given in Example 4.7 to calculate the sample variance and standard deviation. Table 4.4 shows the observations, deviations from the mean, and squared deviations. Combining the squared deviations to compute the values of s 2 and s gives g 1x 2 x2 5 Sxx 5 2.4643 and s2 5 Step-by-Step technology instructions available online
2.4643 2.4643 g 1x 2 x 2 2 5 5 5 0.4107 n21 721 6
s 5 "0.4107 5 0.641
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
178
Chapter 4
Numerical Methods for Describing Data
T A B L E 4 .4 Deviations and Squared Deviations for the Big Mac Data Big Mac Price in U.S. Dollars
Deviations from Mean
Squared Deviations
3.02 4.67 3.28 3.51 3.42 2.76 2.87
0.34 1.31 0.08 0.15 0.06 0.60 0.49
0.1156 1.7161 0.0064 0.0225 0.0036 0.3600 0.2401 g 1x 2 x 2 2 5 2.4643
The computation of s 2 can be a bit tedious, especially if the sample size is large. Fortunately, many calculators and computer software packages compute the variance and standard deviation upon request. One commonly used statistical computer package is Minitab. The output resulting from using the Minitab Describe command with the Big Mac data follows. Minitab gives a variety of numerical descriptive measures, including the mean, the median, and the standard deviation. Descriptive Statistics: Big Mac Price in U.S. Dollars Variable Big Mac Price
N 7
Mean 3.361
Variable Big Mac Price
Q3 3.510
Maximum 4.670
SE Mean 0.242
StDev 0.641
Minimum 2.760
Q1 2.870
Median 3.280
The standard deviation can be informally interpreted as the size of a “typical” or “representative” deviation from the mean. Thus, in Example 4.8, a typical deviation from x is about 0.641; some observations are closer to x than 0.641 and others are farther away. We computed s 5 0.641 in Example 4.8 without saying whether this value indicated a large or a small amount of variability. At this point, it is better to use s for comparative purposes than for an absolute assessment of variability. If Big Mac prices for a different group of countries resulted in a standard deviation of s 5 1.25 (this is the standard deviation for all 45 countries for which Big Mac data was available) then we would conclude that our original sample has much less variability than the data set consisting of all 45 countries. There are measures of variability for the entire population that are analogous to s 2 and s for a sample. These measures are called the population variance and the population standard deviation and are denoted by 2 and , respectively. (We again use a lowercase Greek letter for a population characteristic.)
Notation s2 2 s
sample variance population variance sample standard deviation population standard deviation
In many statistical procedures, we would like to use the value of , but unfortunately it is not usually known. Therefore, in its place we must use a value computed Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
4.2
Describing Variability in a Data Set
179
from the sample that we hope is close to (i.e., a good estimate of ). We use the divisor 1n 2 12 in s 2 rather than n because, on average, the resulting value tends to be a bit closer to 2. We will say more about this in Chapter 9. An alternative rationale for using 1n 2 12 is based on the property g 1x 2 x 2 5 0. Suppose that n 5 5 and that four of the deviations are x1 2 x 5 24 x2 2 x 5 6 x3 2 x 5 1 x5 2 x 5 28 Then, because the sum of these four deviations is 5, the remaining deviation must be x4 2 x 5 5 (so that the sum of all five is zero). Although there are five deviations, only four of them contain independent information about variability. More generally, once any 1n 2 12 of the deviations are available, the value of the remaining deviation is determined. The n deviations actually contain only 1n 2 12 independent pieces of information about variability. Statisticians express this by saying that s 2 and s are based on 1n 2 12 degrees of freedom (df ).
The Interquartile Range As with x, the value of s can be greatly affected by the presence of even a single unusually small or large observation. The interquartile range is a measure of variability that is resistant to the effects of outliers. It is based on quantities called quartiles. The lower quartile separates the bottom 25% of the data set from the upper 75%, and the upper quartile separates the top 25% from the bottom 75%. The middle quartile is the median, and it separates the bottom 50% from the top 50%. Figure 4.6 illustrates the locations of these quartiles for a smoothed histogram.
25%
25%
25%
25%
FIGURE 4.6 The quartiles for a smoothed histogram.
Lower quartile
Median
Upper quartile
The quartiles for sample data are obtained by dividing the n ordered observations into a lower half and an upper half; if n is odd, the median is excluded from both halves. The two extreme quartiles are then the medians of the two halves. (Note: The median is only temporarily excluded for the purpose of computing quartiles. It is not excluded from the data set.)
DEFINITION* lower quartile median of the lower half of the sample upper quartile median of the upper half of the sample (If n is odd, the median of the entire sample is excluded from both halves when computing quartiles.) The interquartile range (iqr), a measure of variability that is not as sensitive to the presence of outliers as the standard deviation, is given by iqr upper quartile 2 lower quartile *There are several other sensible ways to define quartiles. Some calculators and software packages use an alternative definition.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
180
Chapter 4
Numerical Methods for Describing Data
The resistant nature of the interquartile range follows from the fact that up to 25% of the smallest sample observations and up to 25% of the largest sample observations can be made more extreme without affecting the value of the interquartile range.
E X A M P L E 4 . 9 Higher Education The Chronicle of Higher Education (Almanac Issue, 2009–2010) published the accompanying data on the percentage of the population with a bachelor’s or higher degree in 2007 for each of the 50 U.S. states and the District of Columbia. The 51 data values are 21 24 19 22 17 N = 51 Leaf Unit = 1.0 1 1 2 2 2 2 2 3 3 3 3 3 4 4 4 4
7 99 001 222333 444455555 66666677777 8999 00001 23 444555
27 29 24 28 25
30 29 23 23
35 20 34 25
35 20 34 22
26 27 25 25
47 35 32 29
26 38 26 33
27 25 26 34
30 31 24 30
Ordered Data Lower Half: 23 26 Median:
7
Upper Half: 30 35
Stem-and-leaf display: Percent with bachelor’s or higher degree
19 24 27 30
Figure 4.7 gives a stem-and-leaf display (using repeated stems) of the data. The smallest value in the data set is 17% (West Virginia), and two values stand out on the high end—38% (Massachusetts) and 47% (District of Columbia). To compute the quartiles and the interquartile range, we first order the data and use the median to divide the data into a lower half and an upper half. Because there is an odd number of observations 1n 5 512 , the median is excluded from both the upper and lower halves when computing the quartiles.
8
FIGURE 4.7
26 22 27 26 23
17 23 26
19 24 26
19 24 26
20 24
20 24
21 25
22 25
22 25
22 25
23 25
27 30 47
27 31
27 32
27 33
28 34
29 34
29 34
29 35
26 26 30 35
27 30 38
Each half of the sample contains 25 observations. The lower quartile is just the median of the lower half of the sample (24 for this data set), and the upper quartile is the median of the upper half (30 for this data set). This gives lower quartile 5 24 upper quartile 5 30 iqr 5 30 2 24 5 6 The sample mean and standard deviation for this data set are 27.18 and 5.53, respectively. If we were to change the two largest values from 38 and 47 to 58 and 67 (so that they still remain the two largest values), the median and interquartile range would not be affected, whereas the mean and the standard deviation would change to 27.96 and 8.40, respectively. The value of the interquartile range is not affected by a few extreme values in the data set.
Data set available online Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
4.2
181
Describing Variability in a Data Set
The population interquartile range is the difference between the upper and lower population quartiles. If a histogram of the data set under consideration (whether a population or a sample) can be reasonably well approximated by a normal curve, then the relationship between the standard deviation (sd) and the interquartile range is roughly sd 5 iqr/1.35. A value of the standard deviation much larger than iqr/1.35 suggests a distribution with heavier (or longer) tails than a normal curve. For the degree data of Example 4.9, we had s 5 5.53, whereas iqr/1.35 5 6/1.35 5 4.44. This suggests that the distribution of data values in Example 4.9 is indeed heavytailed compared to a normal curve. This can be seen in the stem-and-leaf display of Figure 4.7.
E X E RC I S E S 4 . 1 7 - 4 . 3 1 The following data are cost (in cents) per ounce for nine different brands of sliced Swiss cheese (www .consumerreports.org):
4.17
29
62
37
41
70
82
47
52
49
a. Compute the variance and standard deviation for this data set. s 2 5 279.111; s 5 16.707 b. If a very expensive cheese with a cost per slice of 150 cents was added to the data set, how would the values of the mean and standard deviation change? Cost per serving (in cents) for six high-fiber cereals rated very good and for nine high-fiber cereals rated good by Consumer Reports are shown below. Write a few sentences describing how these two data sets differ with respect to center and variability. Use summary statistics to support your statements.
4.18
Cereals Rated Very Good 46 49 62 41 19
77
Cereals Rated Good 71 30 53 53
43
67
48
28
54
Combining the cost-per-serving data for highfiber cereals rated very good and those rated good from the previous exercise gives the following data set:
4.19
46 49 62 41 19 77 71 30 53 53 67 43 48 28 54 a. Compute the quartiles and the interquartile range for this combined data set. b. Compute the interquartile range for just the cereals rated good. Is this value greater than, less than, or about equal to the interquartile range computed in Part (a)? Bold exercises answered in back
Data set available online
4.20
The paper “Caffeinated Energy Drinks—A
Growing Problem” (Drug and Alcohol Dependence [2009]: 1–10) gave the accompanying data on caffeine per ounce for eight top-selling energy drinks and for 11 high-caffeine energy drinks: Top-Selling Energy Drinks 9.6 10.0 10.0 9.0 10.9
8.9
High-Caffeine Energy Drinks 21.0 25.0 15.0 21.5 33.3 11.9 16.3 31.3
9.5
35.7 30.0
9.1 15.0
The mean caffeine per ounce is clearly higher for the highcaffeine energy drinks, but which of the two groups of energy drinks (top-selling or high-caffeine) is the most variable with respect to caffeine per ounce? Justify your choice.
4.21
The Insurance Institute for Highway Safety (www.iihs.org, June 11, 2009) published data on repair costs for cars involved in different types of accidents. In one study, seven different 2009 models of mini- and micro-cars were driven at 6 mph straight into a fixed barrier. The following table gives the cost of repairing damage to the bumper for each of the seven models: Model Smart Fortwo Chevrolet Aveo Mini Cooper Toyota Yaris Honda Fit Hyundai Accent Kia Rio
Repair Cost $1,480 $1,071 $2,291 $1,688 $1,124 $3,476 $3,701
a. Compute the values of the variance and standard deviation. The standard deviation is fairly large. What does this tell you about the repair costs? Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
182
Chapter 4 Numerical Methods for Describing Data
b. The Insurance Institute for Highway Safety (referenced in the previous exercise) also gave bumper repair costs in a study of six models of minivans (December 30, 2007). Write a few sentences describing how mini- and micro-cars and minivans differ with respect to typical bumper repair cost and bumper repair cost variability.
4.24 Give two sets of five numbers that have the same mean but different standard deviations, and give two sets of five numbers that have the same standard deviation but different means.
4.25 Going back to school can be an expensive time
4.22
for parents—second only to the Christmas holiday season in terms of spending (San Luis Obispo Tribune, August 18, 2005). Parents spend an average of $444 on their children at the beginning of the school year stocking up on clothes, notebooks, and even iPods. Of course, not every parent spends the same amount of money and there is some variation. Do you think a data set consisting of the amount spent at the beginning of the school year for each student at a particular elementary school would have a large or a small standard deviation? Explain.
concentration (mg/cup) for 12 brands of coffee:
4.26 The article “Rethink Diversification to Raise
Model Honda Odyssey Dodge Grand Caravan Toyota Sienna Chevrolet Uplander Kia Sedona Nissan Quest
Repair Cost $1,538 $1,347 $840 $1,631 $1,176 $1,603
Consumer Reports Health (www.consumer reports.org/health) reported the accompanying caffeine
Coffee Brand Eight O’Clock Caribou Kickapoo Starbucks Bucks Country Coffee Co. Archer Farms Gloria Jean’s Coffees Chock Full o’Nuts Peet’s Coffee Maxwell House Folgers Millstone
Caffeine concentration (mg/cup) 140 195 155 115 195 180 110 110 130 55 60 60
Compute the values of the quartiles and the interquartile range for this data set.
4.23 The accompanying data on number of minutes used for cell phone calls in 1 month was generated to be consistent with summary statistics published in a report of a marketing study of San Diego residents (TeleTruth, March 2009): 189 0 189 177 106 201 0 212 0 306 0 0 59 224 0 189 142 83 71 165 236 0 142 236 130 a. Compute the values of the quartiles and the interquartile range for this data set. b. Explain why the lower quartile is equal to the minimum value for this data set. Will this be the case for every data set? Explain. Bold exercises answered in back
Data set available online
Returns, Cut Risk” (San Luis Obispo Tribune, January 21, 2006) included the following paragraph: In their research, Mulvey and Reilly compared the results of two hypothetical portfolios and used actual data from 1994 to 2004 to see what returns they would achieve. The first portfolio invested in Treasury bonds, domestic stocks, international stocks, and cash. Its 10-year average annual return was 9.85% and its volatility— measured as the standard deviation of annual returns—was 9.26%. When Mulvey and Reilly shifted some assets in the portfolio to include funds that invest in real estate, commodities, and options, the 10-year return rose to 10.55% while the standard deviation fell to 7.97%. In short, the more diversified portfolio had a slightly better return and much less risk. Explain why the standard deviation is a reasonable measure of volatility and why it is reasonable to interpret a smaller standard deviation as meaning less risk. The U.S. Department of Transportation reported the accompanying data (see next page) on the number of speeding-related crash fatalities during holiday periods for the years from 1994 to 2003 (Traffic Safety Facts, July 20, 2005). a. Compute the standard deviation for the New Year’s Day data. b. Without computing the standard deviation of the Memorial Day data, explain whether the standard deviation for the Memorial Day data would be larger
4.27
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
4.2
183
Describing Variability in a Data Set
Data for Exercise 4.27
Speeding-Related Fatalities Holiday Period
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
New Year’s Day Memorial Day July 4th Labor Day Thanksgiving Christmas
141 193 178 183 212 152
142 178 219 188 198 129
178 185 202 166 218 66
72 197 179 179 210 183
219 138 169 162 205 134
138 183 176 171 168 193
171 156 219 180 187 155
134 190 64 138 217 210
210 188 234 202 210 60
70 181 184 189 202 198
or smaller than the standard deviation of the New Year’s Day data. c. Memorial Day and Labor Day are holidays that always occur on Monday and Thanksgiving always occurs on a Thursday, whereas New Year’s Day, July 4th and Christmas do not always fall on the same day of the week every year. Based on the given data, is there more or less variability in the speeding-related crash fatality numbers from year to year for same day of the week holiday periods than for holidays that can occur on different days of the week? Support your answer with appropriate measures of variability.
4.28 The Ministry of Health and Long-Term Care in Ontario, Canada, publishes information on the time that patients must wait for various medical procedures on its web site (www.health.gov.on.ca). For two cardiac procedures completed in fall of 2005, the following information was provided:
Procedure Angioplasty Bypass surgery
Number of Completed Procedures
Median Wait Time (days)
Mean Wait Time (days)
90% Completed Within (days)
847 539
14 13
18 19
39 42
a. Which of the following must be true for the lower quartile of the data set consisting of the 847 wait times for angioplasty? i. The lower quartile is less than 14. ii. The lower quartile is between 14 and 18. iii. The lower quartile is between 14 and 39. iv. The lower quartile is greater than 39. b. Which of the following must be true for the upper quartile of the data set consisting of the 539 wait times for bypass surgery? i. The upper quartile is less than 13. ii. The upper quartile is between 13 and 19. Bold exercises answered in back
Data set available online
iii. The upper quartile is between 13 and 42. iv. The upper quartile is greater than 42. c. Which of the following must be true for the number of days for which only 5% of the bypass surgery wait times would be longer? i. It is less than 13. ii. It is between 13 and 19. iii. It is between 13 and 42. iv. It is greater than 42. The accompanying table shows the low price, the high price, and the average price of homes sold in 15 communities in San Luis Obispo County between January 1, 2004, and August 1, 2004 (San Luis Obispo Tribune, September 5, 2004):
4.29
Community Cayucos Pismo Beach Cambria Avila Beach Morro Bay Arroyo Grande Templeton San Luis Obispo Nipomo Los Osos Santa Margarita Atascadero Grover Beach Paso Robles Oceano
Average Number Price Sold
Low
High $2,450,000 $2,500,000 $2,000,000 $1,375,000 $2,650,000 $1,526,000
$937,366 $804,212 $728,312 $654,918 $606,456 $595,577
31 71 85 16 114 214
$380,000 $439,000 $340,000 $475,000 $257,000 $178,000
$578,249 $557,628
89 277
$265,000 $2,350,000 $258,000 $2,400,000
$528,572 $511,866 $430,354
138 123 22
$263,000 $1,295,000 $140,000 $3,500,000 $290,000 $583,000
$420,603 $416,405 $412,584 $390,354
270 97 439 59
$140,000 $1,600,000 $242,000 $720,000 $170,000 $1,575,000 $177,000 $1,350,000
a. Explain why the average price for the combined areas of Los Osos and Morro Bay is not just the average of $511,866 and $606,456. Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
184
Chapter 4
Numerical Methods for Describing Data
b. Houses sold in Grover Beach and Paso Robles have very similar average prices. Based on the other information given, which is likely to have the higher standard deviation for price? c. Consider houses sold in Grover Beach and Paso Robles. Based on the other information given, which is likely to have the higher median price? In 1997, a woman sued a computer keyboard manufacturer, charging that her repetitive stress injuries were caused by the keyboard (Genessey v. Digital Equipment Corporation). The jury awarded about $3.5 million for pain and suffering, but the court then set aside that award as being unreasonable compensation. In making this determination, the court identified a “normative” group of 27 similar cases and specified a reasonable award as one within 2 standard deviations of the mean of the awards in the 27 cases. The 27 award amounts were (in thousands of dollars)
4.30
37 60 75 115 135 140 149 150 238 290 340 410 600 750 750 750 1050 1100 1139 1150 1200 1200 1250 1576 1700 1825 2000 What is the maximum possible amount that could be awarded under the “2-standard deviations rule?”
Bold exercises answered in back
4.3
Data set available online
The standard deviation alone does not measure relative variation. For example, a standard deviation of $1 would be considered large if it is describing the variability from store to store in the price of an ice cube tray. On the other hand, a standard deviation of $1 would be considered small if it is describing store-to-store variability in the price of a particular brand of freezer. A quantity designed to give a relative measure of variability is the coefficient of variation. Denoted by CV, the coefficient of variation expresses the standard deviation as a percentage s of the mean. It is defined by the formula CV 5 100a b. x Consider two samples. Sample 1 gives the actual weight (in ounces) of the contents of cans of pet food labeled as having a net weight of 8 ounces. Sample 2 gives the actual weight (in pounds) of the contents of bags of dry pet food labeled as having a net weight of 50 pounds. The weights for the two samples are
4.31
Sample 1 Sample 2
8.3 8.3 52.3 47.0
7.1 8.2 50.6 50.4
7.6 7.7 52.1 50.3
8.1 7.7 48.4 48.7
7.6 7.5 48.8 48.2
a. For each of the given samples, calculate the mean and the standard deviation. b. Compute the coefficient of variation for each sample. Do the results surprise you? Why or why not? Video Solution available
Summarizing a Data Set: Boxplots In Sections 4.1 and 4.2, we looked at ways of describing the center and variability of a data set using numerical measures. It would be nice to have a method of summarizing data that gives more detail than just a measure of center and spread and yet less detail than a stem-and-leaf display or histogram. A boxplot is one way to do this. A boxplot is compact, yet it provides information about the center, spread, and symmetry or skewness of the data. We will consider two types of boxplots: the skeletal boxplot and the modified boxplot.
Construction of a Skeletal Boxplot 1. Draw a horizontal (or vertical) measurement scale. 2. Construct a rectangular box with a left (or lower) edge at the lower quartile and a right (or upper) edge at the upper quartile. The box width is then equal to the iqr. 3. Draw a vertical (or horizontal) line segment inside the box at the location of the median. 4. Extend horizontal (or vertical) line segments, called whiskers, from each end of the box to the smallest and largest observations in the data set.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
4.3
185
Summarizing a Data Set: Boxplots
E X A M P L E 4 . 1 0 Revisiting the Degree Data Let’s reconsider the data on percentage of the population with a bachelor’s or higher degree for the 50 U.S. states and the District of Columbia (Example 4.9). The ordered observations are
Ordered Data Lower Half: 23 26
17 23 26
Median: Upper Half: 30 35
19 24 26
19 24 26
20 24
20 24
21 25
22 25
22 25
22 25
23 25
27 30 47
27 31
27 32
27 33
28 34
29 34
29 34
29 35
26 26 30 35
27 30 38
To construct a boxplot of these data, we need the following information: the smallest observation, the lower quartile, the median, the upper quartile, and the largest observation. This collection of summary measures is often referred to as a five-number summary. For this data set we have smallest observation 17 lower quartile median of the lower half 24 median 26th observation in the ordered list 26 upper quartile median of the upper half 30 largest observation 47 Figure 4.8 shows the corresponding boxplot. The median line is somewhat closer to the lower edge of the box than to the upper edge, suggesting a concentration of values in the lower part of the middle half. The upper whisker is longer than the lower whisker. These observations are consistent with the stem-and-leaf display of Figure 4.7.
FIGURE 4.8 Skeletal boxplot for the degree data of Example 4.10.
20
25
30 35 40 45 Percent of population with bachelor’s or higher degree
50
The sequence of steps used to construct a skeletal boxplot is easily modified to give information about outliers.
DEFINITION An observation is an outlier if it is more than 1.5(iqr) away from the nearest quartile (the nearest end of the box). An outlier is extreme if it is more than 3(iqr) from the nearest quartile and it is mild otherwise.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
186
Chapter 4 Numerical Methods for Describing Data
A modified boxplot represents mild outliers by solid circles and extreme outliers by open circles, and the whiskers extend on each end to the most extreme observations that are not outliers.
Construction of a Modified Boxplot 1. Draw a horizontal (or vertical) measurement scale. 2. Construct a rectangular box with a left (or lower) edge at the lower quartile and right (or upper) edge at the upper quartile. The box width is then equal to the iqr. 3. Draw a vertical (or horizontal) line segment inside the box at the location of the median. 4. Determine if there are any mild or extreme outliers in the data set. 5. Draw whiskers that extend from each end of the box to the most extreme observation that is not an outlier. 6. Draw a solid circle to mark the location of any mild outliers in the data set. 7. Draw an open circle to mark the location of any extreme outliers in the data set.
E X A M P L E 4 . 1 1 Golden Rectangles The accompanying data came from an anthropological study of rectangular shapes
(Lowie’s Selected Papers in Anthropology, Cora Dubios, ed. [Berkeley, CA: University of California Press, 1960]: 137–142). Observations were made on the variable x width/length for a sample of n 20 beaded rectangles used in Shoshoni Indian leather handicrafts: .553 .654
.570 .662
.576 .668
.601 .670
.606 .672
.606 .690
.609 .693
.611 .749
.615 .844
.628 .933
The quantities needed for constructing the modified boxplot follow: median .641 lower quartile .606 upper quartile .681
iqr .681 .606 .075 1.5(iqr) .1125 3(iqr) .225
Thus, (upper quartile) 1.5(iqr) .681 .1125 .7935 (lower quartile) 1.5(iqr) .606 .1125 .4935 So 0.844 and 0.933 are both outliers on the upper end (because they are larger than 0.7935), and there are no outliers on the lower end (because no observations are smaller than 0.4935). Because (upper quartile) 3(iqr) 0.681 0.225 0.906
Step-by-Step technology instructions available online Data set available online
0.933 is an extreme outlier and 0.844 is only a mild outlier. The upper whisker extends to the largest observation that is not an outlier, 0.749, and the lower whisker extends to 0.553. The boxplot is shown in Figure 4.9. The median line is not at the center of the box, so there is a slight asymmetry in the middle half of the data. However, the most striking feature is the presence of the two outliers. These two x values considerably exceed the “golden ratio” of 0.618, used since antiquity as an aesthetic standard for rectangles.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
4.3
Summarizing a Data Set: Boxplots
Largest observation that isn’t an outlier
0.4
0.5
0.6
0.7
0.8
Mild outlier
187
Extreme outlier
0.9
Median
FIGURE 4.9
Mild outlier cutoffs
Boxplot for the rectangle data in Example 4.11.
Extreme outlier cutoffs
E X A M P L E 4 . 1 2 Another Look at Big Mac Prices Big Mac prices in U.S. dollars for 45 different countries were given in the article “Cheesed Off” first introduced in Example 4.7. The 45 Big Mac prices were: 3.57 5.89 3.06 3.03 2.87
3.01 3.04 1.99 2.37
3.97 2.36 2.48 2.91
4.67 4.92 3.54 1.83
3.80 1.72 7.03 5.57
3.64 3.89 2.28 6.39
3.28 5.20 2.76 2.31
1.83 2.21 2.09 1.93
3.51 3.98 2.66 3.80
3.42 3.54 2.31 2.72
3.92 3.24 2.93 1.70
Figure 4.10 shows a Minitab boxplot for the Big Mac price data. Note that the upper whisker is longer than the lower whisker and that there are two outliers on the high end (Norway with a Big Mac price of $7.04 and Switzerland with a price of $6.29).
FIGURE 4.10 Minitab boxplot of the Big Mac price data of Example 4.12.
1
2
3 4 5 Price of a Big Mac in U.S. dollars
6
7
Note that Minitab does not distinguish between mild outliers and extreme outliers in the boxplot. For the Big Mac price data, lower quartile 2.335 upper quartile 3.845 iqr 3.845 2.335 1.510 Then 1.5(iqr) 2.265 3(iqr) 4.530 We can compute outlier boundaries as follows: upper quartile 1.5(iqr) 3.845 2.265 6.110 upper quartile 3(iqr) 3.845 4.530 8.375 The observation for Switzerland (6.39) is a mild outlier because it is greater than 6.110 (the upper quartile 1.5(iqr)) but less than 8.375 (the upper quartile 3(iqr)). The observation for Norway is also a mild outlier. There are no extreme outliers in this data set.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
188
Chapter 4 Numerical Methods for Describing Data
With two or more data sets consisting of observations on the same variable (for example, fuel efficiencies for four types of car or weight gains for a control group and a treatment group), comparative boxplots (more than one boxplot drawn using the same scale) can tell us a lot about similarities and differences between the data sets.
E X A M P L E 4 . 1 3 NBA Salaries Revisited The 2009–2010 salaries of NBA players published on the web site hoopshype.com were used to construct the comparative boxplot of the salary data for five teams shown in Figure 4.11. Bulls
Lakers
Knicks
Grizzlies
Nuggets
FIGURE 4.11
Comparative boxplot for salaries for five NBA teams.
5,000,000
10,000,000
15,000,000
20,000,000
25,000,000
Data
The comparative boxplot reveals some interesting similarities and differences in the salary distributions of the five teams. The minimum salary is lower for the Grizzlies, but is about the same for the other four teams. The median salary was lowest for the Nuggets—in fact the median for the Nuggets is about the same as the lower quartile for the Knicks and the Lakers, indicating that half of the players on the Nuggets have salaries less than about $2.5 million, whereas only about 25% of the Knicks and the Lakers have salaries less than about $2.5 million. The Lakers had the player with by far the highest salary. The Grizzlies and the Lakers were the only teams that had any salary outliers. With the exception of one highly paid player, salaries for players on the Grizzlies team were noticeably lower than for the other four teams.
EX E RC I S E S 4 . 3 2 - 4 . 3 7 4.32 Based on a large national sample of working adults, the U.S. Census Bureau reports the following information on travel time to work for those who do not work at home: Bold exercises answered in back
Data set available online
lower quartile 7 minutes median 18 minutes upper quartile 31 minutes
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
4.3
Also given was the mean travel time, which was reported as 22.4 minutes. a. Is the travel time distribution more likely to be approximately symmetric, positively skewed, or negatively skewed? Explain your reasoning based on the given summary quantities. b. Suppose that the minimum travel time was 1 minute and that the maximum travel time in the sample was 205 minutes. Construct a skeletal boxplot for the travel time data. c. Were there any mild or extreme outliers in the data set? How can you tell? The report “Who Moves? Who Stays Put? Where’s Home?” (Pew Social and Demographic Trends, December 17, 2008) gave the accompanying
4.33
data for the 50 U.S. states on the percentage of the population that was born in the state and is still living there. The data values have been arranged in order from largest to smallest. 75.8 65.1 59.2 54.5 48.6
71.4 64.4 59.0 54.0 47.1
69.6 64.3 58.7 54.0 43.4
69.0 63.8 57.3 53.9 40.4
68.6 63.7 57.1 53.5 35.7
67.5 62.8 55.6 52.8 28.2
66.7 62.6 55.6 52.5
66.3 61.9 55.5 50.2
66.1 61.9 55.3 50.2
66.0 61.5 54.9 48.9
The National Climate Data Center gave the accompanying annual rainfall (in inches) for Medford, Oregon, from 1950 to 2008 (www.ncdc.noaa.gov/oa/ climate/research/cag3/city.html): 20.15 20.85 20.68 15.95 14.95 31.57 23.97
18.88 19.86 23.43 20.46 13.86 18.13 21.99
25.72 23.34 19.55 16.05 15.30 28.87 17.25
16.42 19.08 20.82 22.08 13.71 16.69 14.07
20.18 29.23 19.04 19.44 14.68 18.81
28.96 18.32 18.77 30.38 15.16 15.15
20.72 21.27 19.63 18.79 16.77 18.16
23.58 18.93 12.39 10.89 12.33 19.99
a. Compute the quartiles and the interquartile range. b. Are there outliers in this data set? If so, which observations are mild outliers? Which are extreme outliers? c. Draw a boxplot for this data set that shows outliers.
Bold exercises answered in back
30.3 27.2 37.0 28.3 37.5
Data set available online
39.0 52.9 34.4 39.1 31.5
33.9 45.8 35.5 55.0 32.0
38.6 63.3 62.2 35.0 35.5
44.6 36.0 30.3 28.8 37.5
31.4 64.0 40.0 25.7 41.0
26.7 31.4 36.0 62.7 37.5
51.9 42.2 39.4 32.4 48.6
31.9 41.1 34.4 31.9 28.1
Fiber content (in grams per serving) and sugar content (in grams per serving) for 18 high fiber cereals (www.consumerreports.com) are shown below.
4.36
Fiber Content 7 13
10 10
10 8
7 12
8 7
7 14
12 7
12 8
8 8
Sugar Content
a.
4.34
28.84 10.62 15.47 22.39 17.25 21.93 19.00
The accompanying data on annual maximum wind speed (in meters per second) in Hong Kong for each year in a 45-year period were given in an article that appeared in the journal Renewable Energy (March 2007). Use the annual maximum wind speed data to construct a boxplot. Is the boxplot approximately symmetric?
4.35
66.0 61.1 54.7 48.7
a. Find the values of the median, the lower quartile, and the upper quartile. b. The two smallest values in the data set are 28.2 (Alaska) and 35.7 (Wyoming). Are these two states outliers? c. Construct a boxplot for this data set and comment on the interesting features of the plot.
189
Summarizing a Data Set: Boxplots
b. c. d.
e.
11 6 14 13 0 18 9 10 19 6 10 17 10 10 0 9 5 11 Find the median, quartiles, and interquartile range for the fiber content data set. Find the median, quartiles, and interquartile range for the sugar content data set. Are there any outliers in the sugar content data set? Explain why the minimum value for the fiber content data set and the lower quartile for the fiber content data set are equal. Construct a comparative boxplot and use it to comment on the differences and similarities in the fiber and sugar distributions.
Shown here are the number of auto accidents per year for every 1000 people in each of 40 occupations (Knight Ridder Tribune, June 19, 2004):
4.37
Occupation Student Physician Lawyer Architect Real estate broker Enlisted military
Accidents per 1000 Occupation 152 109 106 105 102 99
Social worker Manual laborer Analyst Engineer Consultant Sales
Accidents per 1000 98 96 95 94 94 93 (continued)
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
190
Chapter 4
Numerical Methods for Describing Data
Accidents per 1000 Occupation
Occupation Military officer Nurse School administrator Skilled laborer Librarian Creative arts Executive Insurance agent Banking, finance Customer service Manager Medical support Computer-related Dentist
91 90 90 90 90 90 89 89 89 88 88 87 87 86
Bold exercises answered in back
4.4
a. Would you recommend using the standard deviation or the iqr as a measure of variability for this data set? b. Are there outliers in this data set? If so, which observations are mild outliers? Which are extreme outliers? c. Draw a modified boxplot for this data set. d. If you were asked by an insurance company to decide which, if any, occupations should be offered a professional discount on auto insurance, which occupations would you recommend? Explain.
Accidents per 1000
Pharmacist Proprietor Teacher, professor Accountant Law enforcement Physical therapist Veterinarian Clerical, secretary Clergy Homemaker Politician Pilot Firefighter Farmer
85 84 84 84 79 78 78 77 76 76 76 75 67 43
Data set available online
Video Solution available
Interpreting Center and Variability: Chebyshev’s Rule, the Empirical Rule, and z Scores The mean and standard deviation can be combined to make informative statements about how the values in a data set are distributed and about the relative position of a particular value in a data set. To do this, it is useful to be able to describe how far away a particular observation is from the mean in terms of the standard deviation. For example, we might say that an observation is 2 standard deviations above the mean or that an observation is 1.3 standard deviations below the mean.
E X A M P L E 4 . 1 4 Standardized Test Scores Consider a data set of scores on a standardized test with a mean and standard deviation of 100 and 15, respectively. We can make the following statements: 1. Because 100 15 85, we say that a score of 85 is “1 standard deviation below the mean.” Similarly, 100 15 115 is “1 standard deviation above the mean” (see Figure 4.12). 1 sd
70
85
1 sd
100
115
130
FIGURE 4.12 Values within 1 standard deviation of the mean (Example 4.14).
Mean
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
4.4 Interpreting Center and Variability: Chebyshev’s Rule, the Empirical Rule, and z Scores
191
2. Because 2 times the standard deviation is 2(15) 30, and 100 30 130 and 100 30 70, scores between 70 and 130 are those within 2 standard deviations of the mean (see Figure 4.13). 3. Because 100 (3)(15) 145, scores above 145 are greater than the mean by more than 3 standard deviations. Within 2 sd’s of the mean 2 sd’s 70
2 sd’s
85
100
115
130
FIGURE 4.13 Values within 2 standard deviations of the mean (Example 4.14).
Mean
Sometimes in published articles, the mean and standard deviation are reported, but a graphical display of the data is not given. However, using a result called Chebyshev’s Rule, it is possible to get a sense of the distribution of data values based on our knowledge of only the mean and standard deviation.
Chebyshev’s Rule Consider any number k, where k $ 1. Then the percentage of observations that 1 are within k standard deviations of the mean is at least 100a1 2 2 b%. Subk stituting selected values of k gives the following results. Number of Standard Deviations, k 2 3 4 4.472 5 10
12
1 k2
1 5 .75 4 1 1 2 5 .89 9 1 12 5 .94 16 1 12 5 .95 20 1 12 5 .96 25 1 12 5 .99 100 12
Percentage Within k Standard Deviations of the Mean at least 75% at least 89% at least 94% at least 95% at least 96% at least 99%
E X A M P L E 4 . 1 5 Child Care for Preschool Kids The article “Piecing Together Child Care with Multiple Arrangements: Crazy Quilt or Preferred Pattern for Employed Parents of Preschool Children?” ( Journal of Marriage and the Family [1994]: 669–680) examined various modes of care for Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
192
Chapter 4 Numerical Methods for Describing Data
preschool children. For a sample of families with one preschool child, it was reported that the mean and standard deviation of child care time per week were approximately 36 hours and 12 hours, respectively. Figure 4.14 displays values that are 1, 2, and 3 standard deviations from the mean. FIGURE 4.14
0 – x – 3s
Ariel Skelley/Blend Images/Jupiter Images
Measurement scale for child care time (Example 4.15).
12 –x – 2s
24 – x–s
36 x–
48 – x+s
60 – x + 2s
72 – x + 3s
Chebyshev’s Rule allows us to assert the following: 1. At least 75% of the sample observations must be between 12 and 60 hours (within 2 standard deviations of the mean). 2. Because at least 89% of the observations must be between 0 and 72, at most 11% are outside this interval. Time cannot be negative, so we conclude that at most 11% of the observations exceed 72. 3. The values 18 and 54 are 1.5 standard deviations to either side of x, so using k 1.5 in Chebyshev’s Rule implies that at least 55.6% of the observations must be between these two values. Thus, at most 44.4% of the observations are less than 18—not at most 22.2%, because the distribution of values may not be symmetric.
Because Chebyshev’s Rule is applicable to any data set (distribution), whether symmetric or skewed, we must be careful when making statements about the proportion above a particular value, below a particular value, or inside or outside an interval that is not centered at the mean. The rule must be used in a conservative fashion. There is another aspect of this conservatism. The rule states that at least 75% of the observations are within 2 standard deviations of the mean, but for many data sets substantially more than 75% of the values satisfy this condition. The same sort of understatement is frequently encountered for other values of k (numbers of standard deviations).
E X A M P L E 4 . 1 6 IQ Scores Figure 4.15 gives a stem-and-leaf display of IQ scores of 112 children in one of the early studies that used the Stanford revision of the Binet–Simon intelligence scale (The Intelligence of School Children, L. M. Terman [Boston: Houghton-Mifflin, 1919]). Summary quantities include x 5 104.5 s 5 16.3 2s 5 32.6 3s 5 48.9
FIGURE 4.15 Stem-and-leaf display of IQ scores used in Example 4.16.
6 7 8 9 10 11 12 13 14 15
1 25679 0000124555668 0000112333446666778889 0001122222333566677778899999 00001122333344444477899 01111123445669 006 26 Stem: Tens 2 Leaf: Ones
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
4.4 Interpreting Center and Variability: Chebyshev’s Rule, the Empirical Rule, and z Scores
193
In Figure 4.15, all observations that are within two standard deviations of the mean are shown in blue. Table 4.5 shows how Chebyshev’s Rule can sometimes considerably understate actual percentages.
T A B L E 4.5 Summarizing the Distribution of IQ Scores k ⫽ Number of sd’s
x 6 ks
Chebyshev
Actual
2.0 2.5 3.0
71.9 to 137.1 63.7 to 145.3 55.6 to 153.4
at least 75% at least 84% at least 89%
96% (108) 97% (109) 100% (112)
the blue leaves in Figure 4.15
Empirical Rule The fact that statements based on Chebyshev’s Rule are frequently conservative suggests that we should look for rules that are less conservative and more precise. One useful rule is the Empirical Rule, which can be applied whenever the distribution of data values can be reasonably well described by a normal curve (distributions that are “mound” shaped).
The Empirical Rule If the histogram of values in a data set can be reasonably well approximated by a normal curve, then Approximately 68% of the observations are within 1 standard deviation of the mean. Approximately 95% of the observations are within 2 standard deviations of the mean. Approximately 99.7% of the observations are within 3 standard deviations of the mean.
The Empirical Rule makes “approximately” instead of “at least” statements, and the percentages for k 5 1, 2, and 3 standard deviations are much higher than those of Chebyshev’s Rule. Figure 4.16 illustrates the percentages given by the Empirical Rule. In contrast to Chebyshev’s Rule, dividing the percentages in half is permissible, because a normal curve is symmetric.
34%
34%
2.35%
2.35% 13.5%
FIGURE 4.16 Approximate percentages implied by the Empirical Rule.
13.5%
1 sd
1 sd Mean
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
194
Chapter 4 Numerical Methods for Describing Data
The Image Bank/Paul Thomas/ Getty Images
E X A M P L E 4 . 1 7 Heights of Mothers and the Empirical Rule One of the earliest articles to argue for the wide applicability of the normal distribution was “On the Laws of Inheritance in Man. I. Inheritance of Physical Characters” (Biometrika [1903]: 375–462). Among the data sets discussed in the article was one consisting of 1052 measurements of the heights of mothers. The mean and standard deviation were x 5 62.484 in. s 5 2.390 in. The data distribution was described as approximately normal. Table 4.6 contrasts actual percentages with those obtained from Chebyshev’s Rule and the Empirical Rule.
T A B L E 4 .6 Summarizing the Distribution of Mothers’ Heights Number of sd’s
Interval
Actual
Empirical Rule
Chebyshev Rule
1 2 3
60.094 to 64.874 57.704 to 67.264 55.314 to 69.654
72.1% 96.2% 99.2%
Approximately 68% Approximately 95% Approximately 99.7%
At least 0%0 At least 75% At least 89%
Clearly, the Empirical Rule is much more successful and informative in this case than Chebyshev’s Rule.
Our detailed study of the normal distribution and areas under normal curves in Chapter 7 will enable us to make statements analogous to those of the Empirical Rule for values other than k 5 1, 2, or 3 standard deviations. For now, note that it is unusual to see an observation from a normally distributed population that is farther than 2 standard deviations from the mean (only 5%), and it is very surprising to see one that is more than 3 standard deviations away. If you encountered a mother whose height was 72 inches, you might reasonably conclude that she was not part of the population described by the data set in Example 4.17.
Measures of Relative Standing When you obtain your score after taking a test, you probably want to know how it compares to the scores of others who have taken the test. Is your score above or below the mean, and by how much? Does your score place you among the top 5% of those who took the test or only among the top 25%? Questions of this sort are answered by finding ways to measure the position of a particular value in a data set relative to all values in the set. One measure of relative standing is a z score.
DEFINITION The z score corresponding to a particular value is value 2 mean z score 5 standard deviation The z score tells us how many standard deviations the value is from the mean. It is positive or negative according to whether the value lies above or below the mean.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
4.4 Interpreting Center and Variability: Chebyshev’s Rule, the Empirical Rule, and z Scores
195
The process of subtracting the mean and then dividing by the standard deviation is sometimes referred to as standardization, and a z score is one example of what is called a standardized score.
E X A M P L E 4 . 1 8 Relatively Speaking, Which Is the Better Offer? Suppose that two graduating seniors, one a marketing major and one an accounting major, are comparing job offers. The accounting major has an offer for $45,000 per year, and the marketing student has an offer for $43,000 per year. Summary information about the distribution of offers follows: Accounting: mean 46,000 Marketing: mean 42,500
standard deviation 1500 standard deviation 1000
Then, accounting z score 5
45,000 2 46,000 5 2.67 1500
(so $45,000 is .67 standard deviation below the mean), whereas marketing z score 5
43,000 2 42,500 5 .5 1000
Relative to the appropriate data sets, the marketing offer is actually more attractive than the accounting offer (although this may not offer much solace to the marketing major).
The z score is particularly useful when the distribution of observations is approximately normal. In this case, from the Empirical Rule, a z score outside the interval from 22 to 12 occurs in about 5% of all cases, whereas a z score outside the interval from 23 to 13 occurs only about 0.3% of the time.
Percentiles A particular observation can be located even more precisely by giving the percentage of the data that fall at or below that observation. If, for example, 95% of all test scores are at or below 650, whereas only 5% are above 650, then 650 is called the 95th percentile of the data set (or of the distribution of scores). Similarly, if 10% of all scores are at or below 400 and 90% are above 400, then the value 400 is the 10th percentile.
DEFINITION For any particular number r between 0 and 100, the rth percentile is a value such that r percent of the observations in the data set fall at or below that value.
Figure 4.17 illustrates the 90th percentile. We have already met several percentiles in disguise. The median is the 50th percentile, and the lower and upper quartiles are the 25th and 75th percentiles, respectively. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
196
Chapter 4
Numerical Methods for Describing Data
Shaded area = 90% of total area
FIGURE 4.17 Ninetieth percentile for a smoothed histogram.
90th percentile
E X A M P L E 4 . 1 9 Head Circumference at Birth In addition to weight and length, head circumference is another measure of health in newborn babies. The National Center for Health Statistics reports the following summary values for head circumference (in cm) at birth for boys (approximate values read from graphs on the Center for Disease Control web site):
Percentile
5
Head Circumference (cm)
32.2
10
25
50
75
90
95
33.2
34.5
35.8
37.0
38.2
38.6
Interpreting these percentiles, we know that half of newborn boys have head circumferences of less than 35.8 cm, because 35.8 is the 50th percentile (the median). The middle 50% of newborn boys have head circumferences between 34.5 cm and 37.0 cm, with about 25% of the head circumferences less than 34.5 cm and about 25% greater than 37.0 cm. We can tell that the head circumference distribution for newborn boys is not symmetric, because the 5th percentile is 3.6 cm below the median, whereas the 95th percentile is only 2.8 cm above the median. This suggests that the bottom part of the distribution stretches out more than the top part of the distribution. This would be consistent with a distribution that is negatively skewed, as shown in Figure 4.18.
Shaded area = .05
FIGURE 4.18 Negatively skewed distribution.
Shaded area .05 32.2
35.8
38.6
5th percentile
Median
95th percentile
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
4.4 Interpreting Center and Variability: Chebyshev’s Rule, the Empirical Rule, and z Scores
197
E X E RC I S E S 4 . 3 8 - 4 . 5 2 4.38 The average playing time of compact discs in a large collection is 35 minutes, and the standard deviation is 5 minutes. a. What value is 1 standard deviation above the mean? 1 standard deviation below the mean? What values are 2 standard deviations away from the mean? b. Without assuming anything about the distribution of times, at least what percentage of the times is between 25 and 45 minutes? c. Without assuming anything about the distribution of times, what can be said about the percentage of times that are either less than 20 minutes or greater than 50 minutes? d. Assuming that the distribution of times is approximately normal, about what percentage of times are between 25 and 45 minutes? less than 20 minutes or greater than 50 minutes? less than 20 minutes? In a study investigating the effect of car speed on accident severity, 5000 reports of fatal automobile accidents were examined, and the vehicle speed at impact was recorded for each one. For these 5000 accidents, the average speed was 42 mph and the standard deviation was 15 mph. A histogram revealed that the vehicle speed at impact distribution was approximately normal. a. Roughly what proportion of vehicle speeds were between 27 and 57 mph? b. Roughly what proportion of vehicle speeds exceeded 57 mph?
4.39
4.40 The U.S. Census Bureau (2000 census) reported the following relative frequency distribution for travel time to work for a large sample of adults who did not work at home: Travel Time (minutes)
Relative Frequency
0 to 5 5 to 10 10 to 15 15 to 20 20 to 25 25 to 30 30 to 35 35 to 40 40 to 45 45 to 60 60 to 90 90 or more
.04 .13 .16 .17 .14 .05 .12 .03 .03 .06 .05 .02
Bold exercises answered in back
Data set available online
a. Draw the histogram for the travel time distribution. In constructing the histogram, assume that the last interval in the relative frequency distribution (90 or more) ends at 200; so the last interval is 90 to 200. Be sure to use the density scale to determine the heights of the bars in the histogram because not all the intervals have the same width. b. Describe the interesting features of the histogram from Part (a), including center, shape, and spread. c. Based on the histogram from Part (a), would it be appropriate to use the Empirical Rule to make statements about the travel time distribution? Explain why or why not. d. The approximate mean and standard deviation for the travel time distribution are 27 minutes and 24 minutes, respectively. Based on this mean and standard deviation and the fact that travel time cannot be negative, explain why the travel time distribution could not be well approximated by a normal curve. e. Use the mean and standard deviation given in Part (d) and Chebyshev’s Rule to make a statement about i. the percentage of travel times that were between 0 and 75 minutes ii. the percentage of travel times that were between 0 and 47 minutes f. How well do the statements in Part (e) based on Chebyshev’s Rule agree with the actual percentages for the travel time distribution? (Hint: You can estimate the actual percentages from the given relative frequency distribution.)
4.41 Mobile homes are tightly constructed for energy conservation. This can lead to a buildup of indoor pollutants. The paper “A Survey of Nitrogen Dioxide Lev-
els Inside Mobile Homes” (Journal of the Air Pollution Control Association [1988]: 647–651) discussed various aspects of NO2 concentration in these structures. a. In one sample of mobile homes in the Los Angeles area, the mean NO2 concentration in kitchens during the summer was 36.92 ppb, and the standard deviation was 11.34. Making no assumptions about the shape of the NO2 distribution, what can be said about the percentage of observations between 14.24 and 59.60? b. Inside what interval is it guaranteed that at least 89% of the concentration observations will lie? c. In a sample of non–Los Angeles mobile homes, the average kitchen NO2 concentration during the winVideo Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
198
Chapter 4 Numerical Methods for Describing Data
ter was 24.76 ppb, and the standard deviation was 17.20. Do these values suggest that the histogram of sample observations did not closely resemble a normal curve? (Hint: What is x 2 2s?)
4.42 The article “Taxable Wealth and Alcoholic Beverage Consumption in the United States” (Psychological Reports [1994]: 813–814) reported that the mean annual adult consumption of wine was 3.15 gallons and that the standard deviation was 6.09 gallons. Would you use the Empirical Rule to approximate the proportion of adults who consume more than 9.24 gallons (i.e., the proportion of adults whose consumption value exceeds the mean by more than 1 standard deviation)? Explain your reasoning.
4.43 A student took two national aptitude tests. The national average and standard deviation were 475 and 100, respectively, for the first test and 30 and 8, respectively, for the second test. The student scored 625 on the first test and 45 on the second test. Use z scores to determine on which exam the student performed better relative to the other test takers. 4.44 Suppose that your younger sister is applying for entrance to college and has taken the SATs. She scored at the 83rd percentile on the verbal section of the test and at the 94th percentile on the math section of the test. Because you have been studying statistics, she asks you for an interpretation of these values. What would you tell her?
4.45 A sample of concrete specimens of a certain type is selected, and the compressive strength of each specimen is determined. The mean and standard deviation are calculated as x 5 3000 and s 5 500, and the sample histogram is found to be well approximated by a normal curve. a. Approximately what percentage of the sample observations are between 2500 and 3500? b. Approximately what percentage of sample observations are outside the interval from 2000 to 4000? c. What can be said about the approximate percentage of observations between 2000 and 2500? d. Why would you not use Chebyshev’s Rule to answer the questions posed in Parts (a)–(c)? 4.46 The paper “Modeling and Measurements of Bus Service Reliability” (Transportation Research [1978]: 253–256) studied various aspects of bus service and presented data on travel times (in minutes) from several different routes. The accompanying frequency distribution Bold exercises answered in back
Data set available online
is for bus travel times from origin to destination on one particular route in Chicago during peak morning traffic periods: Travel Time
Frequency
Relative Frequency
15 to 16 16 to 17 17 to 18 18 to 19 19 to 20 20 to 21 21 to 22 22 to 23 23 to 24 24 to 25 25 to 26
4 0 26 99 36 8 12 0 0 0 16
.02 .00 .13 .49 .18 .04 .06 .00 .00 .00 .08
a. Construct the corresponding histogram. b. Compute (approximately) the following percentiles: i. 86th iv. 95th ii. 15th v. 10th iii. 90th
4.47 An advertisement for the “30 inch Wonder” that appeared in the September 1983 issue of the journal Packaging claimed that the 30 inch Wonder weighs cases and bags up to 110 pounds and provides accuracy to within 0.25 ounce. Suppose that a 50 ounce weight was repeatedly weighed on this scale and the weight readings recorded. The mean value was 49.5 ounces, and the standard deviation was 0.1. What can be said about the proportion of the time that the scale actually showed a weight that was within 0.25 ounce of the true value of 50 ounces? (Hint: Use Chebyshev’s Rule.)
4.48 Suppose that your statistics professor returned your first midterm exam with only a z score written on it. She also told you that a histogram of the scores was approximately normal. How would you interpret each of the following z scores? a. 2.2 d. 1.0 b. 0.4 e. 0 c. 1.8
4.49 The paper “Answer Changing on MultipleChoice Tests” (Journal of Experimental Education [1980]: 18–21) reported that for a group of 162 college students, the average number of responses changed from the correct answer to an incorrect answer on a test containing 80 multiple-choice items was 1.4. The corresponding standard deviation was reported to be 1.5. Based on this mean and standard deviation, what can Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
4.5 Interpreting and Communicating the Results of Statistical Analyses
you tell about the shape of the distribution of the variable number of answers changed from right to wrong? What can you say about the number of students who changed at least six answers from correct to incorrect?
4.50 The average reading speed of students completing a speed-reading course is 450 words per minute (wpm). If the standard deviation is 70 wpm, find the z score associated with each of the following reading speeds. a. 320 wpm c. 420 wpm b. 475 wpm d. 610 wpm
a. Summarize this data set with a frequency distribution. Construct the corresponding histogram. b. Use the histogram in Part (a) to find approximate values of the following percentiles: i. 50th iv. 90th ii. 70th v. 40th iii. 10th
4.52 The accompanying table gives the mean and standard deviation of reaction times (in seconds) for each of two different stimuli:
4.51 The following data values are 2009 per capita expenditures on public libraries for each of the 50 U.S. states (from www.statemaster.com): 16.84 16.17 11.74 11.11 8.65 7.03 6.20 6.20 5.95 5.72 5.43 5.33 4.84 4.63 4.59 3.81 3.75 3.74 3.67 3.40 3.18 3.16 2.91 2.78 2.61 2.30 2.19 2.06 1.78 1.54 1.20 1.19 1.09 0.70 0.66 0.30 0.01 Bold exercises answered in back
4.5
7.69 5.61 4.58 3.35 2.58 1.31 0.54
7.48 5.47 3.92 3.29 2.45 1.26 0.49
Data set available online
199
Mean Standard deviation
Stimulus 1
Stimulus 2
6.0 1.2
3.6 0.8
If your reaction time is 4.2 seconds for the first stimulus and 1.8 seconds for the second stimulus, to which stimulus are you reacting (compared to other individuals) relatively more quickly?
Video Solution available
Interpreting and Communicating the Results of Statistical Analyses As was the case with the graphical displays of Chapter 3, the primary function of the descriptive tools introduced in this chapter is to help us better understand the variables under study. If we have collected data on the amount of money students spend on textbooks at a particular university, most likely we did so because we wanted to learn about the distribution of this variable (amount spent on textbooks) for the population of interest (in this case, students at the university). Numerical measures of center and spread and boxplots help to inform us, and they also allow us to communicate to others what we have learned from the data.
Communicating the Results of Statistical Analyses When reporting the results of a data analysis, it is common to start with descriptive information about the variables of interest. It is always a good idea to start with a graphical display of the data, and, as we saw in Chapter 3, graphical displays of numerical data are usually described in terms of center, variability, and shape. The numerical measures of this chapter can help you to be more specific in describing the center and spread of a data set. When describing center and spread, you must first decide which measures to use. Common choices are to use either the sample mean and standard deviation or the sample median and interquartile range (and maybe even a boxplot) to describe center and spread. Because the mean and standard deviation can be sensitive to extreme Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
200
Chapter 4 Numerical Methods for Describing Data
values in the data set, they are best used when the distribution shape is approximately symmetric and when there are few outliers. If the data set is noticeably skewed or if there are outliers, then the observations are more spread out in one part of the distribution than in the others. In this situation, a five-number summary or a boxplot conveys more information than the mean and standard deviation do.
Interpreting the Results of Statistical Analyses It is relatively rare to find raw data in published sources. Typically, only a few numerical summary quantities are reported. We must be able to interpret these values and understand what they tell us about the underlying data set. For example, a university conducted an investigation of the amount of time required to enter the information contained in an application for admission into the university computer system. One of the individuals who performs this task was asked to note starting time and completion time for 50 randomly selected application forms. The resulting entry times (in minutes) were summarized using the mean, median, and standard deviation: x 5 7.854 median 5 7.423 s 5 2.129 What do these summary values tell us about entry times? The average time required to enter admissions data was 7.854 minutes, but the relatively large standard deviation suggests that many observations differ substantially from this mean. The median tells us that half of the applications required less than 7.423 minutes to enter. The fact that the mean exceeds the median suggests that some unusually large values in the data set affected the value of the mean. This last conjecture is confirmed by the stem-and-leaf display of the data given in Figure 4.19.
FIGURE 4.19 Stem-and-leaf display of data entry times.
4 5 6 7 8 9 10 11 12 13 14
8 02345679 00001234566779 223556688 23334 002 011168 134 2 Stem: Ones Leaf: Tenths 3
The administrators conducting the data-entry study looked at the outlier 14.3 minutes and at the other relatively large values in the data set; they found that the five largest values came from applications that were entered before lunch. After talking with the individual who entered the data, the administrators speculated that morning entry times might differ from afternoon entry times because there tended to be more distractions and interruptions (phone calls, etc.) during the morning hours, when the admissions office generally was busier. When morning and afternoon entry times were separated, the following summary statistics resulted: Morning (based on n 20 applications): Afternoon (based on n 30 applications):
x 9.093 x 7.027
median 8.743 median 6.737
s 2.329 s 1.529
Clearly, the average entry time is higher for applications entered in the morning; also, the individual entry times differ more from one another in the mornings than in the Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
4.5 Interpreting and Communicating the Results of Statistical Analyses
201
afternoons (because the standard deviation for morning entry times, 2.329, is about 1.5 times as large as 1.529, the standard deviation for afternoon entry times).
What to Look for in Published Data Here are a few questions to ask yourself when you interpret numerical summary measures. • Is the chosen summary measure appropriate for the type of data collected? In
particular, watch for inappropriate use of the mean and standard deviation with categorical data that has simply been coded numerically. • If both the mean and the median are reported, how do the two values compare? What does this suggest about the distribution of values in the data set? If only the mean or the median was used, was the appropriate measure selected? • Is the standard deviation large or small? Is the value consistent with your expectations regarding variability? What does the value of the standard deviation tell you about the variable being summarized? • Can anything of interest be said about the values in the data set by applying Chebyshev’s Rule or the Empirical Rule? For example, consider a study that investigated whether people tend to spend more money when they are paying with a credit card than when they are paying with cash. The authors of the paper “Monopoly Money: The Effect of Payment Coupling
and Form on Spending Behavior” ( Journal of Experimental Psychology: Applied [2008]: 213–225) randomly assigned each of 114 volunteers to one of two experimental groups. Participants were given a menu for a new restaurant that showed nine menu items. They were then asked to estimate the amount they would be willing to pay for each item. A price index was computed for each participant by averaging the nine prices assigned. The difference between the two experimental groups was that the menu viewed by one group showed a credit card logo at the bottom of the menu while there was no credit card logo on the menu that those in the other group viewed. The following passage appeared in the results section of the paper: On average, participants were willing to pay more when the credit card logo was present (M $4.53, SD 1.15) than when it was absent (M $4.11, SD 1.06). Thus, even though consumers were not explicitly informed which payment mode they would be using, the mere presence of a credit card logo increased the price that they were willing to pay. The price index data was also described as mound shaped with no outliers for each of the two groups. Because price index (the average of the prices that a participant assigned to the nine menu items) is a numerical variable, the mean and standard deviation are reasonable measures for summarizing center and spread in the data set. Although the mean for the credit-card-logo group is higher than the mean for the no-logo group, the two standard deviations are similar, indicating similar variability in price index from person to person for the two groups. Because the distribution of price index values was mound shaped for each of the two groups, we can use the Empirical Rule to tell us a bit more about the distribution. For example, for those in the group who viewed the menu with a credit card logo, approximately 95% of the price index values would have been between 4.53 2 2(1.15) 5 4.53 – 2.3 5 2.23 and 4.53 1 2(1.15) 5 4.53 1 2.30 5 6.83. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
202
Chapter 4 Numerical Methods for Describing Data
A Word to the Wise: Cautions and Limitations When computing or interpreting numerical descriptive measures, you need to keep in mind the following: 1. Measures of center don’t tell all. Although measures of center, such as the mean and the median, do give us a sense of what might be considered a typical value for a variable, this is only one characteristic of a data set. Without additional information about variability and distribution shape, we don’t really know much about the behavior of the variable. 2. Data distributions with different shapes can have the same mean and standard deviation. For example, consider the following two histograms: Frequency
Frequency
10
20
5
10
0 5.5
6.5
7.5 8.5
5.5
9.5 10.5 11.5 12.5 13.5 14.5
6.5 7.5 8.5 9.5 10.5 11.5 12.5 13.5 14.5
Both histograms summarize data sets that have a mean of 10 and a standard deviation of 2, yet they have different shapes. 3. Both the mean and the standard deviation are sensitive to extreme values in a data set, especially if the sample size is small. If a data distribution is skewed or if the data set has outliers, the median and the interquartile range may be a better choice for describing center and spread. 4. Measures of center and variability describe the values of the variable studied, not the frequencies in a frequency distribution or the heights of the bars in a histogram. For example, consider the following two frequency distributions and histograms: FREQUENCY DISTRIBUTION A
FREQUENCY DISTRIBUTION B
Value
Frequency
Value
Frequency
1 2 3 4 5
10 10 10 10 10
1 2 3 4 5
5 10 20 10 5
Frequency
Frequency 10
20
5
10
0 1
2 3 4 Histogram A
5
1
2 3 4 Histogram B
5
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
4.5 Interpreting and Communicating the Results of Statistical Analyses
203
There is more variability in the data summarized by Frequency Distribution and Histogram A than in the data summarized by Frequency Distribution and Histogram B. This is because the values of the variable described by Histogram and Frequency Distribution B are more concentrated near the mean than are the values for the variable described by Histogram and Frequency Distribution A. Don’t be misled by the fact that there is no variability in the frequencies in Frequency Distribution A or the heights of the bars in Histogram A. 5. Be careful with boxplots based on small sample sizes. Boxplots convey information about center, variability, and shape, but when the sample size is small, you should be hesitant to overinterpret shape information. It is really not possible to decide whether a data distribution is symmetric or skewed if only a small sample of observations from the distribution is available. 6. Not all distributions are normal (or even approximately normal). Be cautious in applying the Empirical Rule in situations in which you are not convinced that the data distribution is at least approximately normal. Using the Empirical Rule in such situations can lead to incorrect statements. 7. Watch out for outliers! Unusual observations in a data set often provide important information about the variable under study, so it is important to consider outliers in addition to describing what is typical. Outliers can also be problematic— both because the values of some descriptive measures are influenced by outliers and because some of the methods for drawing conclusions from data may not be appropriate if the data set has outliers.
E X E RC I S E S 4 . 5 3 - 4 . 5 4 4.53 The authors of the paper “Delayed Time to Defibrillation after In-Hospital Cardiac Arrest” (New England Journal of Medicine [2008]: 9–16) described a study of how survival is related to the length of time it takes from the time of a heart attack to the administration of defibrillation therapy. The following is a statement from the paper: We identified 6789 patients from 369 hospitals who had in-hospital cardiac arrest due to ventricular fibrillation (69.7%) or pulseless ventricular trachycardia (30.3%). Overall, the median time to defibrillation was 1 minute (interquartile range [was] 3 minutes).
Data from the paper on time to defibrillation (in minutes) for these 6789 patients was used to produce the following Minitab output and boxplot. a. Why is there no lower whisker in the given boxplot? b. How is it possible for the median, the lower quartile, and the minimum value in the data set to all be equal? (Note—this is why you do not see a median line in the box part of the boxplot.) c. The authors of the paper considered a time to defibrillation of greater than 2 minutes as unacceptable. Based on the given boxplot and summary statistics, is it possible that the percentage of patients having an unacceptable time to defibrillation is greater than 50%? Greater than 25%? Less than 25%? Explain.
Descriptive Statistics: Time to Defibrillation Variable Time
1
N 6789
2
Mean 2.3737
StDev 2.0713
3 4 5 Time to defibrillation (minutes)
Bold exercises answered in back
Minimum 1.0000
6
Data set available online
Q1 1.0000
Median 1.0000
Q3 3.0000
Maximum 7.0000
7
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
204
Chapter 4 Numerical Methods for Describing Data
d. Is the outlier shown at 7 a mild outlier or an extreme outlier?
4.54 The paper “Portable Social Groups: Willingness to Communicate, Interpersonal Communication Gratifications, and Cell Phone Use among Young Adults” (International Journal of Mobile Communications [2007]: 139–156) describes a study of young adult cell phone use patterns. a. Comment on the following quote from the paper. Do you agree with the authors? Seven sections of an Introduction to Mass Communication course at a large southern university were surveyed in the spring and fall of 2003. The Bold exercises answered in back
AC TI V I TY 4 . 1
Data set available online
b. Below is another quote from the paper. In this quote, the author reports the mean number of minutes of cell phone use per week for those who participated in the survey. What additional information would have been provided about cell phone use behavior if the author had also reported the standard deviation? Based on respondent estimates, users spent an average of 629 minutes (about 10.5 hours) per week using their cell phone on or off line for any reason. Video Solution available
Collecting and Summarizing Numerical Data
In this activity, you will work in groups to collect data that will provide information about how many hours per week, on average, students at your school spend engaged in a particular activity. You will use the sampling plan designed in Activity 2.1 to collect the data. 1. With your group, pick one of the following activities to be the focus of your study: i. Surfing the web ii. Studying or doing homework iii. Watching TV iv. Exercising v. Sleeping
AC TI V I TY 4 . 2
sample was chosen because it offered an excellent representation of the population under study— young adults.
or you may choose a different activity, subject to the approval of your instructor. 2. Use the plan developed in Activity 2.1 to collect data on the variable you have chosen for your study. 3. Summarize the resulting data using both numerical and graphical summaries. Be sure to address both center and variability. 4. Write a short article for your school paper summarizing your findings regarding student behavior. Your article should include both numerical and graphical summaries.
Airline Passenger Weights
The article “Airlines Should Weigh Passengers, Bags, NTSB Says” (USA Today, February 27, 2004) states that the National Transportation Safety Board recommended that airlines weigh passengers and their bags to prevent overloaded planes from attempting to take off. This recommendation was the result of an investigation into the crash of a small commuter plane in 2003, which determined that too much weight contributed to the crash. Rather than weighing passengers, airlines currently use estimates of average passenger and luggage weights. After the 2003 accident, this estimate was increased by 10 pounds for passengers and 5 pounds for luggage. Although an airplane can fly if it is somewhat overweight if
all systems are working properly, if one of the plane’s engines fails an overweight plane becomes difficult for the pilot to control. Assuming that the new estimate of the average passenger weight is accurate, discuss the following questions with a partner and then write a paragraph that answers these questions. 1. What role does variability in passenger weights play in creating a potentially dangerous situation for an airline? 2. Would an airline have a lower risk of a potentially dangerous situation if the variability in passenger weight is large or if it is small?
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
205
Summary of Key Concepts and Formulas
A C TI V I T Y 4 . 3
Boxplot Shapes
In this activity, you will investigate the relationship between boxplot shapes and the corresponding five-number summary. The accompanying figure shows four boxplots, labeled A–D. Also given are 4 five-number summaries, labeled I–IV. Match each five-number summary to the appropriate boxplot. Note that scales are not included on the boxplots, so you will have to think about what the five-number summary implies about characteristics of the boxplot.
Five-Number Summaries I
II
III
IV
Minimum Lower quartile Median Upper quartile Maximum
4 8 16 25 30
0.0 0.1 0.9 2.2 5.1
10 34 44 82 132
40 45 71 88 106
A
B
C
D
Summary of Key Concepts and Formulas TERM OR FORMULA
COMMENT
x1, x2, p , xn
Notation for sample data consisting of observations on a variable x, where n is the sample size.
Sample mean, x
The most frequently used measure of center of a sample. It can be very sensitive to the presence of even a single outlier (unusually large or small observation).
Population mean, m
The average x value in the entire population.
Sample median
The middle value in the ordered list of sample observations. (For n even, the median is the average of the two middle values.) It is very insensitive to outliers.
Trimmed mean
A measure of center in which the observations are first ordered from smallest to largest, one or more observations are deleted from each end, and the remaining ones are averaged. In terms of sensitivity to outliers, it is a compromise between the mean and the median.
Deviations from the mean: x1 2 x, x2 2 x, p , xn 2 x
Quantities used to assess variability in a sample. Except for rounding effects, g 1x 2 x 2 5 0.
g 1x 2 x2 2 and n21 standard deviation s 5 "s2
The most frequently used measures of variability for sample data.
The sample variance s2 5
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
206
Chapter 4 Numerical Methods for Describing Data
TERM OR FORMULA
COMMENT
The population variance s2 and standard deviation s
Measures of variability for the entire population.
Quartiles and the interquartile range
The lower quartile separates the smallest 25% of the data from the remaining 75%, and the upper quartile separates the largest 25% from the smallest 75%. The interquartile range (iqr), a measure of variability less sensitive to outliers than s, is the difference between the upper and lower quartiles.
Chebyshev’s Rule
This rule states that for any number k 1, at least 1 100a1 2 2 b% of the observations in any data set are k within k standard deviations of the mean. It is typically conservative in that the actual percentages often considerably exceed the stated lower bound.
Empirical Rule
This rule gives the approximate percentage of observations within 1 standard deviation (68%), 2 standard deviations (95%), and 3 standard deviations (99.7%) of the mean when the histogram is well approximated by a normal curve.
z score
This quantity gives the distance between an observation and the mean expressed as a certain number of standard deviations. It is positive (negative) if the observation lies above (below) the mean.
r th percentile
The value such that r % of the observations in the data set fall at or below that value.
Five-number summary
A summary of a data set that includes the minimum, lower quartile, median, upper quartile, and maximum.
Boxplot
A picture that conveys information about the most important features of a numerical data set: center, spread, extent of skewness, and presence of outliers.
Chapter Review Exercises 4.55 - 4.73 4.55 Research by the Food and Drug Administration (FDA) shows that acrylamide (a possible cancer-causing substance) forms in high-carbohydrate foods cooked at high temperatures and that acrylamide levels can vary widely even within the same brand of food (Associated Press, December 6, 2002). FDA scientists analyzed McDonald’s French fries purchased at seven different locations and found the following acrylamide levels: 497
193
328
Bold exercises answered in back
155
326
245
270
Data set available online
a. Compute the mean acrylamide level and the seven deviations from the mean. b. Verify that, except for the effect of rounding, the sum of the deviations from mean is equal to 0 for this data set. (If you rounded the sample mean or the deviations, your sum may not be exactly zero, but it should be close to zero if you have computed the deviations correctly.) c. Calculate the variance and standard deviation for this data set. Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
Chapter Review Exercises
The technical report “Ozone Season Emissions by State” (U.S. Environmental Protection Agency, 2002) gave the following nitrous oxide emissions (in
4.56
thousands of tons) for the 48 states in the continental U.S. states: 76 22 40 7 30 5 6 136 72 33 0 89 136 39 92 40 13 27 1 63 33 60 27 16 63 32 20 2 15 36 19 39 130 40 4 85 38 7 68 151 32 34 0 6 43 89 34 0 Use these data to construct a boxplot that shows outliers. Write a few sentences describing the important characteristics of the boxplot.
4.57 The San Luis Obispo Telegram-Tribune (October 1, 1994) reported the following monthly salaries for supervisors from six different counties: $5354 (Kern), $5166 (Monterey), $4443 (Santa Cruz), $4129 (Santa Barbara), $2500 (Placer), and $2220 (Merced). San Luis Obispo County supervisors are supposed to be paid the average of the two counties among these six in the middle of the salary range. Which measure of center determines this salary, and what is its value? Why is the other measure of center featured in this section not as favorable to these supervisors (although it might appeal to taxpayers)?
4.58 A sample of 26 offshore oil workers took part in a simulated escape exercise, resulting in the accompanying data on time (in seconds) to complete the escape (“Oxygen Consumption and Ventilation During Escape from an Offshore Platform,” Ergonomics [1997]: 281–292): 389 356 359 363 375 424 325 394 402 373 373 370 364 366 364 325 339 393 392 369 374 359 356 403 334 397 a. Construct a stem-and-leaf display of the data. Will the sample mean or the sample median be larger for this data set? b. Calculate the values of the sample mean and median. x 5 370.692
c. By how much could the largest time be increased without affecting the value of the sample median? By how much could this value be decreased without affecting the sample median?
Bold exercises answered in back
Data set available online
207
4.59 Because some homes have selling prices that are much higher than most, the median price is usually used to describe a “typical” home price for a given location. The three accompanying quotes are all from the San Luis Obispo Tribune, but each gives a different interpretation of the median price of a home in San Luis Obispo County. Comment on each of these statements. (Look carefully. At least one of the statements is incorrect.) a. “So we have gone from 23% to 27% of county residents who can afford the median priced home at $278,380 in SLO County. That means that half of the homes in this county cost less than $278,380 and half cost more.” (October 11, 2001) b. “The county’s median price rose to $285,170 in the fourth quarter, a 9.6% increase from the same period a year ago, the report said. (The median represents the midpoint of a range.)” (February 13, 2002) c. “ ‘Your median is going to creep up above $300,000 if there is nothing available below $300,000,’ Walker said.” (February 26, 2002) Although bats are not known for their eyesight, they are able to locate prey (mainly insects) by emitting high-pitched sounds and listening for echoes. A paper appearing in Animal Behaviour (“The Echolocation of Flying Insects by Bats” [1960]: 141–154) gave the following distances (in centimeters) at which a bat first detected a nearby insect:
4.60
62 23 27 56 52 34 42 40 68 45 83 a. Compute the sample mean distance at which the bat first detects an insect. b. Compute the sample variance and standard deviation for this data set. Interpret these values.
4.61 For the data in Exercise 4.60, subtract 10 from each sample observation. For the new set of values, compute the mean and the deviations from the mean. How do these deviations compare to the deviations from the mean for the original sample? How does s 2 for the new values compare to s 2 for the old values? In general, what effect does subtracting (or adding) the same number to each observation have on s 2 and s? Explain.
4.62 For the data of Exercise 4.60, multiply each data value by 10. How does s for the new values compare to s for the original values? More generally, what happens to s if each observation is multiplied by the same positive constant c?
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
208
Chapter 4 Numerical Methods for Describing Data
4.63
The percentage of juice lost after thawing for 19 different strawberry varieties appeared in the article
4.66 The paper cited in Exercise 4.65 also reported
“Evaluation of Strawberry Cultivars with Different Degrees of Resistance to Red Scale” (Fruit Varieties Journal [1991]: 12–17):
values of single-leg power for a low workload. The sample mean for n 13 observations was x 119.8 (actually 119.7692), and the 14th observation, somewhat of an outlier, was 159. What is the value of x for the entire sample?
46 51 44 50 33 46 60 41 55 46 53 53 42 44 50 54 46 41 48 a. Are there any observations that are mild outliers? Extreme outliers? b. Construct a boxplot, and comment on the important features of the plot.
4.64 The risk of developing iron deficiency is especially high during pregnancy. Detecting such a deficiency is complicated by the fact that some methods for determining iron status can be affected by the state of pregnancy itself. Consider the following data on transferrin receptor concentration for a sample of women with laboratory evidence of overt iron-deficiency anemia (“Serum Transferrin Receptor for the Detection of Iron Deficiency in Pregnancy,” American Journal of Clinical Nutrition [1991]: 1077–1081): 15.2 20.4
9.3 9.4
7.6 11.5
11.9 16.2
10.4 9.4
9.7 8.3
The amount of aluminum contamination (in parts per million) in plastic was determined for a sample of 26 plastic specimens, resulting in the following data
4.67
(“The Log Normal Distribution for Modeling Quality Data When the Mean Is Near Zero,” Journal of Quality Technology [1990]: 105–110): 30 30 60 63 70 79 87 90 101 102 115 118 119 119 120 125 140 145 172 182 183 191 222 244 291 511 Construct a boxplot that shows outliers, and comment on the interesting features of this plot.
4.68 The article “Can We Really Walk Straight?” (American Journal of Physical Anthropology [1992]: 19–27) reported on an experiment in which each of 20 healthy men was asked to walk as straight as possible to a target 60 m away at normal speed. Consider the following data on cadence (number of strides per second):
Compute the values of the sample mean and median. Why are these values different here? Which one do you regard as more representative of the sample, and why?
0.95 0.85 0.92 0.95 0.93 0.86 1.00 0.92 0.85 0.81 0.78 0.93 0.93 1.05 0.93 1.06 1.06 0.96 0.81 0.96
The paper “The Pedaling Technique of Elite Endurance Cyclists” (International Journal of Sport Biomechanics [1991]: 29–53) reported the following data
Use the methods developed in this chapter to summarize the data; include an interpretation or discussion whenever appropriate. (Note: The author of the paper used a rather sophisticated statistical analysis to conclude that people cannot walk in a straight line and suggested several explanations for this.)
4.65
on single-leg power at a high workload: 244 191 160 187 180 176 174 205 211 183 211 180 194 200 a. Calculate and interpret the sample mean and median. x 5 192.571 b. Suppose that the first observation had been 204, not 244. How would the mean and median change? c. Calculate a trimmed mean by eliminating the smallest and the largest sample observations. What is the corresponding trimming percentage? d. Suppose that the largest observation had been 204 rather than 244. How would the trimmed mean in Part (c) change? What if the largest value had been 284?
Bold exercises answered in back
Data set available online
4.69 The article “Comparing the Costs of Major Hotel Franchises” (Real Estate Review [1992]: 46–51) gave the following data on franchise cost as a percentage of total room revenue for chains of three different types: Budget
2.7 5.9 7.7 Midrange 1.5 7.4 First-class 1.8 7.2 9.6
2.8 6.7 7.9 4.0 7.8 5.8 7.5
3.8 7.0 7.9 6.6 8.0 6.0 7.6
3.8 7.2 8.1 6.7 8.1 6.6 7.6
4.0 7.2 8.2 7.0 8.3 6.6 7.8
4.1 7.5 8.5 7.2 8.6 6.6 7.8
5.5 7.5 7.2 9.0 7.1 8.2
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
Chapter Review Exercises
Construct a boxplot for each type of hotel, and comment on interesting features, similarities, and differences. The accompanying data on milk volume (in grams per day) were taken from the paper “Smoking
4.70
During Pregnancy and Lactation and Its Effects on Breast Milk Volume” (American Journal of Clinical Nutrition [1991]: 1011–1016): Smoking 621 793 593 545 mothers 895 767 714 598 Nonsmoking 947 945 1086 1202 mothers 930 745 903 899 Compare and contrast the two samples
753 693 973 961
655 981
4.71 The Los Angeles Times (July 17, 1995) reported that in a sample of 364 lawsuits in which punitive damages were awarded, the sample median damage award was $50,000, and the sample mean was $775,000. What does this suggest about the distribution of values in the sample? Age at diagnosis for each of 20 patients under treatment for meningitis was given in the paper “Penicil-
4.72
209
a. Calculate the values of the sample mean and the standard deviation. b. Calculate the 10% trimmed mean. How does the value of the trimmed mean compare to that of the sample mean? Which would you recommend as a measure of center? Explain. c. Compute the upper quartile, the lower quartile, and the interquartile range. d. Are there any mild or extreme outliers present in this data set? e. Construct the boxplot for this data set.
4.73 Suppose that the distribution of scores on an exam is closely described by a normal curve with mean 100. The 16th percentile of this distribution is 80. a. What is the 84th percentile? b. What is the approximate value of the standard deviation of exam scores? c. What z score is associated with an exam score of 90? d. What percentile corresponds to an exam score of 140? e. Do you think there were many scores below 40? Explain.
lin in the Treatment of Meningitis” ( Journal of the American Medical Association [1984]: 1870–1874). The ages (in years) were as follows: 18 18 25 19 23 20 69 18 21 18 20 18 18 20 18 19 28 17 18 18 Bold exercises answered in back
Data set available online
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
CHAPTER
5
Summarizing Bivariate Data
© Spencer Platt/Getty Images
Unusually large brain size at age 2 to 5 years is one indicator that a child may be at risk for autism. The authors of a paper that appeared in the Journal of the American Medical Association (July 2003) investigated whether head circumference at age 6 to 14 months could serve as a predictor of cerebral grey matter at age 2 to 5 years. Data on head circumference (measured at age 6 to 14 months) and cerebral grey matter (measured at age 2 to 5 years) for 18 male children with autism were used to explore the relationship between these two variables. Questions of interest are: Is there a relationship between head circumference at age 6 to 14 months and the cerebral grey matter measurement at age 2 to 5 years? If so, can a head circumference measurement taken at an early age be used to predict what the grey matter measurement will be, potentially allowing doctors to detect autism at a younger age? How accurate are such predictions of grey matter? In this chapter, we introduce methods for describing relationships between two numerical variables and for assessing the strength of a relationship. These methods allow us to answer questions such as the ones just posed regarding the relationship between head circumference at age 6 to 14 months and the grey matter measurement at age 2 to 5 years. Make the most of your study time by accessing everything you need to succeed online with CourseMate. Visit http://www.cengagebrain.com where you will find: • An interactive eBook, which allows you to take notes, highlight, bookmark, search • • • • • •
the text, and use in-context glossary definitions Step-by-step instructions for Minitab, Excel, TI-83/84, SPSS, and JMP Video solutions to selected exercises Data sets available for selected examples and exercises Online quizzes Flashcards Videos
211 Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
212
Chapter 5
Summarizing Bivariate Data
In Chapter 13, methods for drawing conclusions from this type of data are developed. The techniques introduced in this chapter are also important stepping stones for analyzing data consisting of observations on three or more variables, the topic of Chapter 14.
5.1
Correlation An investigator is often interested in how two or more variables are related to one another. For example, an environmental researcher might wish to know how the lead content of soil varies with distance from a major highway. Researchers in early childhood education might investigate how vocabulary size is related to age. College admissions officers, who must try to predict whether an applicant will succeed in college, might use a model relating college grade point average to high school grades, and ACT or SAT scores. Recall that a scatterplot of bivariate numerical data gives a visual impression of how strongly x values and y values are related. However, to make precise statements and draw conclusions from data, we must go beyond pictures. A correlation coefficient (from co- and relation) is a numerical assessment of the strength of relationship between the x and y values in a bivariate data set consisting of (x, y) pairs. In this section, we introduce the most commonly used correlation coefficient. Figure 5.1 displays several scatterplots that show different relationships between the x and y values. The plot in Figure 5.1(a) suggests a strong positive relationship between x and y; for every pair of points in the plot, the one with the larger x value also has the larger y value. That is, an increase in x is paired with an increase in y. The plot in Figure 5.1(b) shows a strong tendency for y to increase as x does, but there are a few exceptions. For example, the x and y values of the two points with the largest x values (shown in a different color) go in opposite directions (for this pair of points, x increases but y decreases in value). Nevertheless, a plot like this still indicates a fairly strong positive relationship. Figure 5.1(c) suggests that x and y are negatively related—as x increases, y tends to decrease. The y
y
y
x
x (a)
x
(b)
y
(c) y
FIGURE 5.1 Scatterplots illustrating various types of relationships: (a) positive linear relationship; (b) another positive linear relationship; (c) negative linear relationship; (d) no relationship; (e) curved relationship.
x
x (d)
(e)
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.1
Correlation
213
negative relationship in this plot is not as strong as the positive relationship in Figure 5.1(b), although both plots show a well-defined linear pattern. The plot of Figure 5.1(d) indicates that there is not a strong relationship between x and y; there is no tendency for y either to increase or to decrease as x increases. Finally, as illustrated in Figure 5.1(e), a scatterplot can show evidence of a strong relationship that is curved rather than linear.
Pearson’s Sample Correlation Coefficient Pearson’s sample correlation coefficient measures the strength of any linear relationship between two numerical variables. It does this by using z scores in a clever way. Consider replacing each x value by the corresponding z score, zx (by subtracting x and then dividing by sx) and similarly replacing each y value by its z score. Note that x values that are larger than x will have positive z scores and those smaller than x will have negative z scores. Also y values larger than y will have positive z scores and those smaller will have negative z scores. Pearson’s sample correlation coefficient is based on the sum of the products of zx and zy for each observation in the bivariate data set. In algebraic notation, this is g zxzy. To see how this works, let’s look at some scatterplots. The scatterplot in Figure 5.2(a) indicates a strong positive relationship. A vertical line through x and a horizontal line through y divide the plot into four regions. In Region I, both x and y exceed their mean values, so the z score for x and the z score for y are both positive numbers. It follows that zx zy is positive. The product of the z scores is also positive for any point in Region III, because both z scores are negative in Region III and multiplying two negative numbers gives a positive number. In each of the other two regions, one z score is positive and the other is negative, so zx zy is negative. But because the points generally fall in Regions I and III, the products of z scores tend to be positive. Thus, the sum of the products will be a relatively large positive number. Similar reasoning for the data displayed in Figure 5.2(b), which exhibits a strong negative relationship, implies that g zxzy will be a relatively large (in magnitude) negative number. When there is no strong relationship, as in Figure 5.2(c), positive and negative products tend to counteract one another, producing a value of g zxzy that is close to zero. In summary, g zxzy seems to be a reasonable measure of the degree of association between x and y; it can be a large positive number, a large negative number, or a number close to 0, depending on whether there is a strong positive, a strong negative, or no strong linear relationship. Pearson’s sample correlation coefficient, denoted r, is obtained by dividing g zxzy by (n 2 1).
DEFINITION Pearson’s sample correlation coefficient r is given by r5
g zx zy
n21 Although there are several different correlation coefficients, Pearson’s correlation coefficient is by far the most commonly used, and so the name “Pearson’s” is often omitted and it is referred to as simply the correlation coefficient. Hand calculation of the correlation coefficient is quite tedious. Fortunately, all statistical software packages and most scientific calculators can compute r once the x and y values have been input. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
214
Chapter 5 Summarizing Bivariate Data
y
y zx is negative zy is positive zxzy is negative
II
zx is positive zy is positive zxzy is positive
I
–y
II
I
III
IV
–y
zx is negative zy is negative zxzy is positive
III
zx is positive zy is negative zxzy is negative
IV
x
–x
x
–x
(a)
(b)
y
II
I
III
IV
–y
FIGURE 5.2 Viewing a scatterplot according to the signs of zx and zy: (a) a positive relation; (b) a negative relation; (c) no strong relation.
–x
x
(c)
EXAMPLE 5.1
Graduation Rates and Student-Related Expenditures
The web site www.collegeresults.org (The Education Trust) publishes data on U.S. colleges and universities. For the seven primarily undergraduate public universities in California with enrollments between 10,000 and 20,000, six-year graduation rates and student-related expenditures per full-time student for 2007 were reported as follows:
Data set available online
Observation
Graduation Rate (percent)
Student-Related Expenditure (dollars)
1 2 3 4 5 6 7
66.1 52.4 48.9 48.1 42.0 38.3 31.3
8,810 7,780 8,112 8,149 8,477 7,342 7,984
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.1
Correlation
215
70.00
Graduation rate
60.00
50.00
40.00
30.00
FIGURE 5.3
7500.00
SPSS scatterplot for the data of Example 5.1.
8000.00 8500.00 Student-related expenditure
9000.00
Figure 5.3 is a scatterplot of these data generated using SPSS, a widely used statistics package. Let x denote the student-related expenditure per full-time student and y denote the six-year graduation rate. It is easy to verify that x 5 8093.43 sx 5 472.39 y 5 46.73 sy 5 11.15 To illustrate the calculation of the correlation coefficient, we begin by computing z scores for each (x, y) pair in the data set. For example, the first observation is (8810, 66.1). The corresponding z scores are zx 5
8810 2 8093.43 66.1 2 46.73 5 1.52 zy 5 5 1.74 472.39 11.15
The following table shows the z scores and the product zx zy for each observation: y
x
zx
zy
zxzy
66.1 52.4 48.9 48.1 42.0 38.3 31.3
8810 7780 8112 8149 8477 7342 7984
1.52 0.66 0.04 0.12 0.81 1.59 0.23
1.74 0.51 0.19 0.12 0.42 0.76 1.38
2.65 0.34 0.01 0.01 0.34 1.21 0.32
a zx zy 5 3.52 Then, with n 7 g zx zy 3.52 5 5 .587 r5 n21 6 SPSS was used to compute the correlation coefficient, producing the following computer output. Correlations Gradrate
Pearson Correlation
.583
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
216
Chapter 5
Summarizing Bivariate Data
The difference between the correlation coefficient reported by SPSS and what we obtained is the result of rounding in the z scores when carrying out the calculations by hand. Based on the scatterplot and the properties of the correlation coefficient presented in the discussion that follows this example, we conclude that there is a moderate positive linear relationship between student-related expenditure and graduation rate for these seven universities.
Properties of r 1. The value of r does not depend on the unit of measurement for either variable. For example, if x is height, the corresponding z score is the same whether height is expressed in inches, meters, or miles, and thus the value of the correlation coefficient is not affected. The correlation coefficient measures the inherent strength of the linear relationship between two numerical variables. 2. The value of r does not depend on which of the two variables is considered x. Thus, if we had let x 5 graduation rate and y 5 student-related expenditure in Example 5.1, the same value, r 5 0.587, would have resulted. 3. The value of r is between 21 and 11. A value near the upper limit, 11, indicates a strong positive relationship, whereas an r close to the lower limit, 21, suggests a strong negative relationship. Figure 5.4 shows a useful way to describe the strength of relationship based on r. It may seem surprising that a value of r as extreme as .5 or .5 should be in the weak category; an explanation for this is given later in the chapter. Even a weak correlation can indicate a meaningful relationship. Strong Moderate
Weak
Moderate Strong
FIGURE 5.4 Describing the strength of a linear relationship.
−1
−0.8
−0.5
0.5
0.8
1
4. A correlation coefficient of r 5 1 occurs only when all the points in a scatterplot of the data lie exactly on a straight line that slopes upward. Similarly, r 5 21 only when all the points lie exactly on a downward-sloping line. Only when there is a perfect linear relationship between x and y in the sample does r take on one of its two possible extreme values. 5. The value of r is a measure of the extent to which x and y are linearly related—that is, the extent to which the points in the scatterplot fall close to a straight line. A value of r close to 0 does not rule out any strong relationship between x and y; there could still be a strong relationship that is not linear.
E X A M P L E 5 . 2 Tannin Concentration in Wine Astringency is the characteristic of a wine that makes the wine drinker’s mouth feel dry and puckery. The paper “Analysis of Tannins in Red Wine Using Multiple Data set available online
Methods: Correlation with Perceived Astringency” (American Journal of Enology and Viticulture [2006]: 481–485) describes a study to determine if there is a relation-
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.1
Correlation
217
ship between astringency and the concentration of tannins (chemical compounds found in the bark and fruit of some plants) in the wine. The accompanying data on x 5 tannin concentration and y 5 perceived astringency as determined by a panel of tasters for 32 red wines was provided by the authors.
x
y
x
y
x
y
0.72 0.81 0.92 1.00 0.67 0.53 0.51 0.56 0.77 0.47 0.73
0.43 0.48 0.49 0.99 0.32 0.30 0.22 0.20 0.33 0.34 0.77
0.76 0.67 0.56 0.38 0.78 0.67 0.85 0.41 0.93 0.31 0.32
0.19 0.07 0.22 0.90 0.84 0.13 0.30 0.58 0.78 0.71 0.61
0.52 0.69 0.91 0.64 0.23 0.78 0.33 0.43 0.32 0.24
0.65 0.15 1.01 0.09 1.13 0.54 1.10 0.58 0.86 0.55
Minitab was used to construct a scatterplot of the data (Figure 5.5) and to compute the value of the correlation coefficient for these data with the following result: Correlations: x, y Pearson correlation of x and y 0.916
The correlation coefficient of r 5 .916 indicates a strong positive relationship between tannin concentration and astringency rating. This indicates that higher astringency ratings are associated with higher tannin concentrations. We will return to this data set again in Section 5.2 to see how the relationship between tannin concentration and astringency can be described in a way that will allow us to predict what the astringency rating will be for a given tannin concentration.
Astringency rating
1.0
0.5
0.0
–0.5
–1.0
FIGURE 5.5 Minitab scatterplot for the wine data of Example 5.2.
0.2
0.3
0.4
0.5 0.6 0.7 0.8 Tannin concentration
0.9
1.0
1.1
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
218
Chapter 5
Summarizing Bivariate Data
EXAMPLE 5.3
Does It Pay to Pay More for a Bike Helmet?
Are more expensive bike helmets safer than less expensive ones? The accompanying data on x 5 price and y 5 quality rating for 12 different brands of bike helmets appeared on the Consumer Reports web site (www.consumerreports.org/health). Quality rating was a number from 0 (the worst possible rating) to 100, and was determined based on factors that included how well the helmet absorbed the force of an impact, the strength of the helmet, ventilation, and ease of use. Figure 5.6 shows a scatterplot of the data. Price
Quality Rating
35 20 30 40 50 23 30 18 40 28 20 25
65 61 60 55 54 47 47 43 42 41 40 32
65
Quality rating
60 55 50 45 40 35 30
FIGURE 5.6 Minitab scatterplot for the bike helmet data of Example 5.3.
20
25
30
35 Price
40
45
50
From the scatterplot, it appears that there is only a weak positive relationship between price and quality rating. The correlation coefficient, obtained using Minitab, is Correlations: Price, Quality Rating Pearson correlation of Price and Quality Rating 0.303
A correlation coefficient of r 5 .303 confirms that there is a tendency for higher quality ratings to be associated with higher priced helmets, but that the relationship is not very strong. In fact, the highest quality rating was for a helmet priced near the middle of the price values. Data set available online Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.1
EXAMPLE 5.4
Correlation
219
Age and Marathon Times
The article “Master’s Performance in the New York City Marathon” (British Journal of Sports Medicine [2004]: 408–412) gave the following data on the average finishing time by age group for female participants in the New York City marathon.
Age Group
Representative Age
Average Finish Time
10–19 20–29 30–39 40–49 50–59 60–69
15 25 35 45 55 65
302.38 193.63 185.46 198.49 224.30 288.71
The scatterplot of average finish time versus representative age is shown in Figure 5.7.
Average finish time
300 275 250 225 200
FIGURE 5.7
Scatterplot of y average finish time and x age for the data of Example 5.4.
10
20
30
40 Age
50
60
70
Using Minitab to compute Pearson’s correlation coefficient between age and average finish time results in the following: Correlations: Age, Average Finish Time Pearson correlation of Age and Average Finish Time 0.038
This example shows the importance of interpreting r as a measure of the strength of a linear association. Here, r is not large, but there is a strong nonlinear relationship between age and average finish time. This is an important point—we should not conclude that there is no relationship whatsoever simply because the value of r is small in absolute value. Be sure to look at the scatterplot of the data before concluding that there is no relationship between two variables based on a correlation coefficient with a value near 0. Data set available online
The Population Correlation Coefficient The sample correlation coefficient r measures how strongly the x and y values in a sample of pairs are linearly related to one another. There is an analogous measure of how strongly x and y are related in the entire population of pairs from which the Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
220
Chapter 5 Summarizing Bivariate Data
sample was obtained. It is called the population correlation coefficient and is denoted r. (Notice again the use of a Greek letter for a population characteristic and a Roman letter for a sample characteristic.) We will never have to calculate r from an entire population of pairs, but it is important to know that r satisfies properties paralleling those of r : 1. r is a number between 1 and 1 that does not depend on the unit of measurement for either x or y, or on which variable is labeled x and which is labeled y. 2. r 1 or 1 if and only if all (x, y) pairs in the population lie exactly on a straight line, so r measures the extent to which there is a linear relationship in the population. In Chapter 13, we show how the sample correlation coefficient r can be used to draw conclusions about the value of the population correlation coefficient r.
Correlation and Causation A value of r close to 1 indicates that the larger values of one variable tend to be associated with the larger values of the other variable. This is far from saying that a large value of one variable causes the value of the other variable to be large. Correlation measures the extent of association, but association does not imply causation. It frequently happens that two variables are highly correlated not because one is causally related to the other but because they are both strongly related to a third variable. Among all elementary school children, the relationship between the number of cavities in a child’s teeth and the size of his or her vocabulary is strong and positive. Yet no one advocates eating foods that result in more cavities to increase vocabulary size (or working to decrease vocabulary size to protect against cavities). Number of cavities and vocabulary size are both strongly related to age, so older children tend to have higher values of both variables than do younger ones. In the ABCNews.com series “Who’s Counting?” (February 1, 2001), John Paulos reminded readers that correlation does not imply causation and gave the following example: Consumption of hot chocolate is negatively correlated with crime rate (high values of hot chocolate consumption tend to be paired with lower crime rates), but both are responses to cold weather. Scientific experiments can frequently make a strong case for causality by carefully controlling the values of all variables that might be related to the ones under study. Then, if y is observed to change in a “smooth” way as the experimenter changes the value of x, a plausible explanation would be that there is a causal relationship between x and y. In the absence of such control and ability to manipulate values of one variable, we must admit the possibility that an unidentified underlying third variable is influencing both the variables under investigation. A high correlation in many uncontrolled studies carried out in different settings can also marshal support for causality—as in the case of cigarette smoking and cancer—but proving causality is an elusive task.
EX E RC I S E S 5. 1 - 5 . 1 3 5.1 For each of the following pairs of variables, indicate whether you would expect a positive correlation, a negative correlation, or a correlation close to 0. Explain your choice. a. Maximum daily temperature and cooling costs Bold exercises answered in back
Data set available online
b. Interest rate and number of loan applications c. Incomes of husbands and wives when both have fulltime jobs d. Height and IQ e. Height and shoe size Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.1
f. Score on the math section of the SAT exam and score on the verbal section of the same test g. Time spent on homework and time spent watching television during the same day by elementary school children h. Amount of fertilizer used per acre and crop yield (Hint: As the amount of fertilizer is increased, yield tends to increase for a while but then tends to start decreasing.)
5.2 Is the following statement correct? Explain why or why not. A correlation coefficient of 0 implies that no relationship exists between the two variables under study.
5.3 Draw two scatterplots, one for which r 5 1 and a second for which r 5 21.
5.4 The article “That’s Rich: More You Drink, More You Earn” (Calgary Herald, April 16, 2002) reported that there was a positive correlation between alcohol consumption and income. Is it reasonable to conclude that increasing alcohol consumption will increase income? Give at least two reasons or examples to support your answer. The accompanying data are x 5 cost (cents per serving) and y 5 fiber content (grams per serving) for 18 high-fiber cereals rated by Consumer Reports (www .consumerreports.org/health).
5.5
Correlation
221
a. Compute and interpret the correlation coefficient for this data set. b. The serving size differed for the different cereals, with serving sizes varying from 1⁄2 cup to 11⁄4 cups. Converting price and fiber content to “per cup” rather than “per serving” results in the accompanying data. Is the correlation coefficient for the per cup data greater than or less than the correlation coefficient for the per serving data? Cost per Cup
Fiber per Cup
9.3 10 10 7 6.4 7 12 9.6 8 13 10 8 12 7 28 7 16 10.7
44 46 49 62 32.8 19 77 56.8 30 53 53 67 43 48 56 54 54 77.3
5.6 The authors of the paper “Flat-footedness is Not a Cost per serving
Fiber per serving
33 46 49 62 41 19 77 71 30 53 53 67 43 48 28 54 27 58
7 10 10 7 8 7 12 12 8 13 10 8 12 7 14 7 8 8
Bold exercises answered in back
Data set available online
Disadvantage for Athletic Performance in Children Aged 11 to 15 Years” (Pediatrics [2009]: e386–e392) studied the relationship between y 5 arch height and scores on a number of different motor ability tests for 218 children. They reported the following correlation coefficients:
Motor Ability Test Height of counter movement jump Hopping: average height Hopping: average power Balance, closed eyes, one leg Toe flexion
Correlation between Test Score and Arch Height 0.02 0.10 0.09 0.04 0.05
a. Interpret the value of the correlation coefficient between average hopping height and arch height. What does the fact that the correlation coefficient is Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
222
Chapter 5 Summarizing Bivariate Data
negative say about the relationship? Do higher arch heights tend to be paired with higher or lower average hopping heights? b. The title of the paper suggests that having a small value for arch height (flat-footedness) is not a disadvantage when it comes to motor skills. Do the given correlation coefficients support this conclusion? Explain.
Data from the U.S. Federal Reserve Board (Household Debt Service Burden, 2002) on the per-
5.9
centage of disposable personal income required to meet consumer loan payments and mortgage payments for selected years are shown in the following table: Consumer Household Debt Debt 7.88 7.91 7.65 7.61 7.48 7.49 7.37 6.57
5.7 In a study of 200 Division I athletes, variables related to academic performance were examined. The paper “Noncognitive Predictors of Student Athletes’
Academic Performance” (Journal of College Reading and Learning [2000]: e167) reported that the correlation coefficient for college GPA and a measure of academic self-worth was r 5 0.48. Also reported were the correlation coefficient for college GPA and high school GPA (r 5 0.46) and the correlation coefficient for college GPA and a measure of tendency to procrastinate (r 5 20.36). Higher scores on the measure of self-worth indicate higher self-worth, and higher scores on the measure of procrastination indicate a higher tendency to procrastinate. Write a few sentences summarizing what these correlation coefficients tell you about the academic performance of the 200 athletes in the sample.
5.8 The following time-series plot is based on data from the article “Bubble Talk Expands: Corporate Debt Is Latest Concern Turning Heads” (San Luis Obispo Tribune, September 13, 2002) and shows how house-
8 7 Household debt
6 5 4
Corporate debt
3 2 2
4
6 Year
8
10
5.73 5.95 6.09 6.28 6.08 5.79 5.81
The accompanying data were read from graphs that appeared in the article “Bush Timber Pro-
5.10
posal Runs Counter to the Record” (San Luis Obispo Tribune, September 22, 2002). The variables shown are the number of acres burned in forest fires in the western United States and timber sales.
Year
Number of Acres Burned (thousands)
Timber Sales (billions of board feet)
1945 1950 1955 1960 1965 1970 1975 1980 1985 1990 1995
200 250 260 380 80 450 180 240 440 400 180
2.0 3.7 4.4 6.8 9.7 11.0 11.0 10.2 10.0 11.0 3.8
12
Based on the time-series plot, would the correlation coefficient between household debt and corporate debt be positive or negative? Weak or strong? What aspect of the time-series plot supports your answer? Bold exercises answered in back
6.24 6.09 6.32 6.97 7.38 7.52 7.84
a. What is the value of the correlation coefficient for this data set? b. Is it reasonable to conclude in this case that there is no strong relationship between the variables (linear or otherwise)? Use a graphical display to support your answer.
hold debt and corporate debt have changed over the time period from 1991 (year 1 in the graph) to 2002: Debt (trillions of dollars)
6.22 6.14 5.95 5.83 5.83 5.85 5.81 5.79
Consumer Household Debt Debt
Data set available online
a. Is there a correlation between timber sales and acres burned in forest fires? Compute and interpret the value of the correlation coefficient. b. The article concludes that “heavier logging led to large forest fires.” Do you think this conclusion is justified based on the given data? Explain. Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.2
Linear Regression: Fitting a Line to Bivariate Data
5.11 It may seem odd, but one of the ways biologists can tell how old a lobster is involves measuring the concentration of a pigment called neurolipofuscin in the eyestalk of a lobster. (We are not making this up!) The authors of the paper “Neurolipofuscin is a Measure of
Age in Panulirus argus, the Caribbean Spiny Lobster, in Florida” (Biological Bulletin [2007]: 55–66) wondered if it was sufficient to measure the pigment in just one eye stalk, which would be the case if there is a strong relationship between the concentration in the right and left eyestalks. Pigment concentration (as a percentage of tissue sample) was measured in both eyestalks for 39 lobsters, resulting is the following summary quantities (based on data read from a graph that appeared in the paper): g x 5 88.8 g y 5 86.1 n 5 39 g x 2 5 288.0 g y 2 5 286.6 g xy 5 281.1 An alternative formula for computing the correlation coefficient that is based on raw data and is algebraically equivalent to the one given in the text is Bold exercises answered in back
5.2
Data set available online
g xy 2 r5
223
1 g x2 1 g y2 n
1 g y2 2 1 g x2 2 2 gx 2 gy 2 n Å n Å 2
Use this formula to compute the value of the correlation coefficient, and interpret this value.
5.12 An auction house released a list of 25 recently sold paintings. Eight artists were represented in these sales. The sale price of each painting also appears on the list. Would the correlation coefficient be an appropriate way to summarize the relationship between artist (x) and sale price ( y)? Why or why not?
5.13 A sample of automobiles traversing a certain stretch of highway is selected. Each one travels at roughly a constant rate of speed, although speed does vary from auto to auto. Let x 5 speed and y 5 time needed to traverse this segment of highway. Would the sample correlation coefficient be closest to .9, .3, 2.3, or 2.9? Explain. Video Solution available
Linear Regression: Fitting a Line to Bivariate Data The objective of regression analysis is to use information about one variable, x, to draw some sort of conclusion concerning a second variable, y. For example, we might want to predict y 5 product sales during a given period when the amount spent on advertising is x 5 $10,000. The two variables in a regression analysis play different roles: y is called the dependent or response variable, and x is referred to as the independent, predictor, or explanatory variable. Scatterplots frequently exhibit a linear pattern. When this is the case, it makes sense to summarize the relationship between the variables by finding a line that is as close as possible to the points in the plot. Before seeing how this is done, let’s review some elementary facts about lines and linear relationships. The equation of a line is y 5 a 1 bx. A particular line is specified by choosing values of a and b. For example, one line is y 5 10 1 2x; another is y 5 100 2 5x. If we choose some x values and compute y 5 a 1 bx for each value, the points in the plot of the resulting (x, y) pairs will fall exactly on a straight line.
DEFINITION The equation of a line is Intercept
y 5 a 1 bx Slope
The value of b, called the slope of the line, is the amount by which y increases when x increases by 1 unit. The value of a, called the intercept (or sometimes the y-intercept or vertical intercept) of the line, is the height of the line above the value x 5 0.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
224
Chapter 5 Summarizing Bivariate Data
The line y 5 10 1 2x has slope b 5 2, so each 1-unit increase in x is paired with an increase of 2 in y. When x 5 0, y 5 10, so the height at which the line crosses the vertical axis (where x 5 0) is 10. This is illustrated in Figure 5.8(a). The slope of the line y 5 100 2 5x is 25, so y increases by 25 (or equivalently, decreases by 5) when x increases by 1. The height of the line above x 5 0 is a 5 100. The resulting line is pictured in Figure 5.8(b). y
100
y
y changes by b = −5
y = 10 + 2x
30 25
75
y increases by b = 2
20 15
a = 100
10
y = 100 − 5x
25 a = 10
5
FIGURE 5.8
Graphs of two lines: (a) slope b 2, intercept a 10; (b) slope b 5, intercept a 100.
x increases by 1
50
x increases by 1
5
10
15
20
x 0
5
10
(a)
15
x
(b)
It is easy to draw the line corresponding to any particular linear equation. Choose any two x values and substitute them into the equation to obtain the corresponding y values. Then plot the resulting two (x, y) pairs as two points. The desired line is the one passing through these points. For the equation y 5 10 1 2x, substituting x 5 5 yields y 5 20, whereas using x 5 10 gives y 5 30. The resulting two points are then (5, 20) and (10, 30). The line in Figure 5.8(a) passes through these points.
Fitting a Straight Line: The Principle of Least Squares Figure 5.9 shows a scatterplot with two lines superimposed on the plot. Line II is a better fit to the data than Line I is. In order to measure the extent to which a particular line provides a good fit to data, we focus on the vertical deviations from the line. For example, Line II in Figure 5.9 has equation y 5 10 1 2x, and the third and fourth points from the left in the scatterplot are (15, 44) and (20, 45). For these two points, the vertical deviations from this line are 3rd deviation 5 y3 2 height of the line above x3 5 44 2 3 10 1 2 1152 4 54 and 4th deviation 45 [10 2(20)] 5 A positive vertical deviation results from a point that lies above the chosen line, and a negative deviation results from a point that lies below this line. A particular line is said to be a good fit to the data if the deviations from the line are small in magnitude. Line I in Figure 5.9 fits poorly, because all deviations from that line are larger in magnitude (some are much larger) than the corresponding deviations from Line II. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.2
225
Linear Regression: Fitting a Line to Bivariate Data
y 70 60
Slope = 2 Vertical intercept = 10
Line II
50 (15, 44)
40
(20, 45) Line I
30 20
FIGURE 5.9 Line I gives a poor fit and Line II gives a good fit to the data.
10 5
10
15
20
25
30
x
To assess the overall fit of a line, we need a way to combine the n deviations into a single measure of fit. The standard approach is to square the deviations (to obtain nonnegative numbers) and then to sum these squared deviations.
DEFINITION The most widely used measure of the goodness of fit of a line y 5 a 1 bx to bivariate data 1x1, y1 2 , p , 1xn, yn 2 is the sum of the squared deviations about the line g 3 y 2 1a 1 bx2 4 2 5 3 y1 2 1a 1 bx12 4 2 1 3 y2 2 1a 1 bx22 4 2 1 c1 3 yn 2 1a 1 bxn2 4 2
The least-squares line, also called the sample regression line, is the line that minimizes this sum of squared deviations. Fortunately, the equation of the least-squares line can be obtained without having to calculate deviations from any particular line. The accompanying box gives relatively simple formulas for the slope and intercept of the least-squares line.
The slope of the least-squares line is b5
g 1x 2 x 2 1 y 2 y 2 g 1x 2 x 2 2
and the y intercept is a 5 y 2 bx We write the equation of the least-squares line as
y^
y^ 5 a 1 bx where the ^ above y indicates that y^ (read as y-hat) is the prediction of y that results from substituting a particular x value into the equation. Statistical software packages and many calculators can compute the slope and intercept of the least-squares line. If the slope and intercept are to be computed by hand, the following computational formula can be used to reduce the amount of time required to perform the calculations. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
226
Chapter 5 Summarizing Bivariate Data
Calculating Formula for the Slope of the Least-Squares Line 1 g x2 1 g y2 n 1 g x2 2 g x2 2 n
g xy 2 b5
EXAMPLE 5.5
Pomegranate Juice and Tumor Growth
Pomegranate, a fruit native to Persia, has been used in the folk medicines of many cultures to treat various ailments. Researchers are now studying pomegranate’s antioxidant properties to see if it might have any beneficial effects in the treatment of cancer. One such study, described in the paper “Pomegranate Fruit Juice for Che-
moprevention and Chemotherapy of Prostate Cancer” (Proceedings of the National Academy of Sciences [October 11, 2005]: 14813–14818), investigated whether pomegranate fruit extract (PFE) was effective in slowing the growth of prostate cancer tumors. In this study, 24 mice were injected with cancer cells. The mice were then randomly assigned to one of three treatment groups. One group of eight mice received normal drinking water, the second group of eight mice received drinking water supplemented with .1% PFE, and the third group received drinking water supplemented with .2% PFE. The average tumor volume for the mice in each group was recorded at several points in time. The accompanying data on y 5 average tumor volume (in mm3) and x 5 number of days after injection of cancer cells for the mice that received plain drinking water was approximated from a graph that appeared in the paper: x y
Data set available online
11 150
15 270
19 450
23 580
27 740
A scatterplot of these data (Figure 5.10) shows that the relationship between number of days after injection of cancer cells and average tumor volume could reasonably be summarized by a straight line.
Control average tumor size
800 700 600 500 400 300 200
FIGURE 5.10 Minitab scatterplot for the data of Example 5.5.
100 10
12
14
16 18 20 22 Days after injection
24
26
28
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.2
Linear Regression: Fitting a Line to Bivariate Data
227
The summary quantities necessary to compute the equation of the least-squares line are g x 5 95 g y 5 2190
g x 2 5 1965 g y 2 5 1,181,900
g xy 5 47,570
From these quantities, we compute y 5 438
x 5 19
1 g x2 1 g y2 1952 121902 47,570 2 n 5 5960 5 5 5 37.25 2 2 1 g x2 1952 160 2 gx 2 1965 2 n 5
g xy 2 b5
and a 5 y 2 bx 5 438 2 137.252 1192 5 2269.75 The least-squares line is then y^ 5 2269.75 1 37.25x This line is also shown on the scatterplot of Figure 5.10. If we wanted to predict average tumor volume 20 days after injection of cancer cells, we could use the y value of the point on the least-squares line above x 5 20: y^ 5 2269.75 1 37.25 1202 5 475.25 Predicted average tumor volume for other numbers of days after injection of cancer cells could be computed in a similar way. But, be careful in making predictions—the least-squares line should not be used to predict average tumor volume for times much outside the range 11 to 27 days (the range of x values in the data set) because we do not know whether the linear pattern observed in the scatterplot continues outside this range. This is sometimes referred to as the danger of extrapolation. In this example, we can see that using the least-squares line to predict average tumor volume for fewer than 10 days after injection of cancer cells can lead to nonsensical predictions. For example, if the number of days after injection is five the predicted average tumor volume is negative: y^ 5 2269.75 1 37.25 152 5 283.5 Because it is impossible for average tumor volume to be negative, this is a clear indication that the pattern observed for x values in the 11 to 27 range does not continue outside this range. Nonetheless, the least-squares line can be a useful tool for making predictions for x values within the 11- to 27-day range. Figure 5.11 shows a scatterplot for average tumor volume versus number of days after injection of cancer cells for both the group of mice that drank only water and the group that drank water supplemented by .2% PFE. Notice that the tumor growth seems to be much slower for the mice that drank water supplemented with PFE. For the .2% PFE group, the relationship between average tumor volume and number of days after injection of cancer cells appears to be curved rather than linear. We will see in Section 5.4 how a curve (rather than a straight line) can be used to summarize this relationship. Calculations involving the least-squares line can obviously be tedious. This is when the computer or a graphing calculator comes to our rescue. All the standard statistical packages can fit a straight line to bivariate data. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
228
Chapter 5 Summarizing Bivariate Data
Variable Water average tumor size .2% PFE average tumor size
800
Average tumor size
700 600 500 400 300 200 100
FIGURE 5.11
Scatterplot of average tumor volume versus number of days after injection of cancer cells for the water group and the .2% PFE group.
10
15
20 25 30 35 Days after injection
40
USE CAUTION—The Danger of Extrapolation The least-squares line should not be used to make predictions outside the range of the x values in the data set because we have no evidence that the linear relationship continues outside this range.
E X A M P L E 5 . 6 Revisiting the Tannin Concentration Data Data on x 5 tannin concentration and y 5 perceived astringency for n 5 32 red wines was given in Example 5.2. In that example, we saw that the correlation coefficient was 0.916, indicating a strong positive linear relationship. This linear relationship can be summarized using the least-squares line, as shown in Figure 5.12.
Perceived astringency
1.0
0.5
0.0
–0.5
–1.0
FIGURE 5.12 Scatterplot and least-squares line for the data of Example 5.6.
0.2
0.3
0.4
0.5 0.6 0.7 0.8 Tannin concentration
0.9
1.0
1.1
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.2
Linear Regression: Fitting a Line to Bivariate Data
229
Minitab was used to fit the least-squares line, and Figure 5.13 shows part of the resulting output. Instead of x and y, the variable labels “Perceived Astringency” and “Tannin Concentration” are used. The equation at the top is that of the leastsquares line. In the rectangular table just below the equation, the first row gives information about the intercept, a, and the second row gives information concerning the slope, b. In particular, the coefficient column labeled “Coef” contains the values of a and b using more digits than in the rounded values that appear in the equation. The regression equation is Perceived Astringency = – 1.59 + 2.59 Tannin concentration Predictor Constant Tannin concentration
FIGURE 5.13 Partial Minitab output for Example 5.6.
e of
Valu
Coef –1.5908 2.5946
a
SE Coef 0.1339 O.2079
e of
Valu
T –11.88 12.48
Equation ˆy = a + bx
P 0.000 0.000
b
The least-squares line should not be used to predict the perceived astringency for wines with tannin concentrations such as x 5 0.10 or x 5 0.15. These x values are well outside the range of the data, and we do not know if the linear relationship continues outside the observed range.
Regression The least-squares line is often called the sample regression line. This terminology comes from the relationship between the least-squares line and Pearson’s correlation coefficient. To understand this relationship, we first need alternative expressions for the slope b and the equation of the line itself. With sx and sy denoting the sample standard deviations of the x’s and y’s, respectively, a bit of algebraic manipulation gives sy b 5 ra b sx sy y^ 5 y 1 r a b 1x 2 x 2 sx You do not need to use these formulas in any computations, but several of their implications are important for appreciating what the least-squares line does. 1. When x 5 x is substituted in the equation of the line, y^ 5 y results. That is, the least-squares line passes through the point of averages 1x, y 2 . 2. Suppose for the moment that r 5 1, so that all points lie exactly on the line whose equation is sy y^ 5 y 1 1x 2 x 2 sx Now substitute x 5 x 1 sx, which is 1 standard deviation above x: sy y^ 5 y 1 1x 1 sx 2 x 2 5 y 1 sy sx Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
230
Chapter 5
Summarizing Bivariate Data
That is, with r 5 1, when x is 1 standard deviation above its mean, we predict that the associated y value will be 1 standard deviation above its mean. Similarly, if x 5 x 2 2sx (2 standard deviations below its mean), then sy y^ 5 y 1 1x 2 2sx 2 x2 5 y 2 2sy sx which is also 2 standard deviations below the mean. If r 5 21, then x 5 x 1 sx results in y^ 5 y 2 sy, so the predicted y is also 1 standard deviation from its mean but on the opposite side of y from where x is relative to x. In general, if x and y are perfectly correlated, the predicted y value associated with a given x value will be the same number of standard deviations (of y) from its mean y as x is from its mean x. 3. Now suppose that x and y are not perfectly correlated. For example, suppose r 5 .5, so the least-squares line has the equation sy y^ 5 y 1 .5a b 1x 2 x 2 sx Then substituting x 5 x 1 sx gives sy y^ 5 y 1 .5a b 1 x 1 sx 2 x 2 5 y 1 .5sy sx That is, for r 5 .5, when x lies 1 standard deviation above its mean, we predict that y will be only 0.5 standard deviation above its mean. Similarly, we can predict y when r is negative. If r 5 2.5, then the predicted y value will be only half the number of standard deviations from y that x is from x but x and the predicted y will now be on opposite sides of their respective means.
Consider using the least-squares line to predict the value of y associated with an x value some specified number of standard deviations away from x. Then the predicted y value will be only r times this number of standard deviations from y. In terms of standard deviations, except when r 1 or 1, the predicted y will always be closer to y than x is to x.
Using the least-squares line for prediction results in a predicted y that is pulled back in, or regressed, toward the mean of y compared to where x is relative to the mean of x. This regression effect was first noticed by Sir Francis Galton (1822–1911), a famous biologist, when he was studying the relationship between the heights of fathers and their sons. He found that predicted heights of sons whose fathers were above average in height were also above average (because r is positive here) but not by as much as the father’s height; he found a similar relationship for fathers whose heights were below average. This regression effect has led to the term regression analysis for the collection of methods involving the fitting of lines, curves, and more complicated functions to bivariate and multivariate data. The alternative form of the regression (least-squares) line emphasizes that predicting y from knowledge of x is not the same problem as predicting x from knowledge of y. The slope of the least-squares line for predicting x is r(sx /sy) rather than r(sy /sx) and the intercepts of the lines are almost always different. For purposes of prediction, it makes a difference whether y is regressed on x, as we have done, or x is regressed on y. The regression line of y on x should not be used to predict x, because it is not the line that minimizes the sum of squared deviations in the x direction. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.2
231
Linear Regression: Fitting a Line to Bivariate Data
E X E RC I S E S 5 . 1 4 - 5 . 2 8 The article “Air Pollution and Medical Care
5.14
Mean Temperature (x)
Use by Older Americans” (Health Affairs [2002]: 207–214) gave data on a measure of pollution (in micro-
6.17 8.06 8.62 10.56 12.45 11.99 12.50 17.98 18.29 19.89 20.25 19.07 17.73 19.62
grams of particulate matter per cubic meter of air) and the cost of medical care per person over age 65 for six geographical regions of the United States: Region North Upper South Deep South West South Big Sky West
Pollution
Cost of Medical Care
30.0 31.8 32.1 26.8 30.4 40.0
915 891 968 972 952 899
a. Construct a scatterplot of the data. Describe any interesting features of the scatterplot. b. Find the equation of the least-squares line describing the relationship between y 5 medical cost and x 5 pollution. y^ 5 1082.24 2 4.691x c. Is the slope of the least-squares line positive or negative? Is this consistent with your description of the relationship in Part (a)? d. Do the scatterplot and the equation of the leastsquares line support the researchers’ conclusion that elderly people who live in more polluted areas have higher medical costs? Explain.
5.15
The authors of the paper “Evaluating Existing
Movement Hypotheses in Linear Systems Using Larval Stream Salamanders” (Canadian Journal of Zoology [2009]: 292–298) investigated whether water temperature was related to how far a salamander would swim and whether it would swim upstream or downstream. Data for 14 streams with different mean water temperatures where salamander larvae were released are given (approximated from a graph that appeared in the paper). The two variables of interest are x 5 mean water temperature (°C) and y 5 net directionality, which was defined as the difference in the relative frequency of the released salamander larvae moving upstream and the relative frequency of released salamander larvae moving downstream. A positive value of net directionality means a higher proportion were moving upstream than downstream. A negative value of net directionality means a higher proportion were moving downstream than upstream. Bold exercises answered in back
Data set available online
Net Directionality (y) 0.08 0.25 0.14 0.00 0.08 0.03 0.07 0.29 0.23 0.24 0.19 0.14 0.05 0.07
a. Construct a scatterplot of the data. How would you describe the relationship between x and y? b. Find the equation of the least-squares line describing the relationship between y 5 net directionality and x 5 mean water temperature. c. What value of net directionality would you predict for a stream that had mean water temperature of 15 °C? d. The authors state that “when temperatures were warmer, more larvae were captured moving upstream, but when temperatures were cooler, more larvae were captured moving downstream.” Do the scatterplot and least-squares line support this statement? e. Approximately what mean temperature would result in a prediction of the same number of salamander larvae moving upstream and downstream?
5.16
The article “California State Parks Closure List
Due Soon” (The Sacramento Bee, August 30, 2009) gave the following data on x 5 number of visitors in fiscal year 2007–2008 and y 5 percentage of operating costs covered by park revenues for the 20 state park districts in California:
Number of Visitors
Percentage of Operating Costs Covered by Park Revenues
2,755,849 1,124,102 1,802,972
37 19 32 (continued)
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
232
Chapter 5 Summarizing Bivariate Data
Number of Visitors
Percentage of Operating Costs Covered by Park Revenues
1,757,386 1,424,375 1,524,503 1,943,208 819,819 1,292,942 3,170,290 3,984,129 1,575,668 1,383,898 14,519,240 3,983,963 14,598,446 4,551,144 10,842,868 1,351,210 603,938
80 17 34 36 32 38 40 53 31 35 108 34 97 62 36 36 34
predictor variable. For each of the following potential dependent variables, indicate whether you expect the slope of the least-squares line to be positive or negative and give a brief explanation for your choice. a. y a measure of nurse’s job satisfaction (higher values indicate higher satisfaction) b. y a measure of patient satisfaction with hospital care (higher values indicate higher satisfaction) c. y a measure of patient quality of care. The accompanying data on x 5 head circumference z score (a comparison score with peers of the same age—a positive score suggests a larger size than for peers) at age 6 to 14 months and y 5 volume of cerebral grey matter (in ml) at age 2 to 5 years were read from a graph in the article described in the chapter introduction (Journal of the American Medical Association [2003]).
5.19
Cerebral Grey Matter (ml) 2–5 yr 680 690 700 720 740 740 750 750 760 780 790 810 815 820 825 835 840 845
a. Use a statistical software package or a graphing calculator to construct a scatterplot of the data. Describe any interesting features of the scatterplot. b. Find the equation of the least-squares regression line (use software or a graphing calculator). c. Is the slope of the least-squares line positive or negative? Is this consistent with your description in Part (a)? d. Based on the scatterplot, do you think that the correlation coefficient for this data set would be less than 0.5 or greater than 0.5? Explain.
5.17 A sample of 548 ethnically diverse students from Massachusetts were followed over a 19-month period from 1995 and 1997 in a study of the relationship between TV viewing and eating habits (Pediatrics [2003]: 1321–1326). For each additional hour of television viewed per day, the number of fruit and vegetable servings per day was found to decrease on average by 0.14 serving. a. For this study, what is the dependent variable? What is the predictor variable? b. Would the least-squares line for predicting number of servings of fruits and vegetables using number of hours spent watching TV as a predictor have a positive or negative slope? Explain.
5.18 The relationship between hospital patient-tonurse ratio and various characteristics of job satisfaction and patient care has been the focus of a number of research studies. Suppose x 5 patient-to-nurse ratio is the Bold exercises answered in back
Data set available online
Head Circumference z Scores at 6–14 Months .75 1.2 .3 .25 .3 1.5 1.1 2.0 1.1 1.1 2.0 2.1 2.8 2.2 .9 2.35 2.3 2.2
a. b. c. d.
Construct a scatterplot for these data. What is the value of the correlation coefficient? Find the equation of the least-squares line. Predict the volume of cerebral grey matter at age 2 to 5 years for a child whose head circumference z score at age 12 months was 1.8. e. Explain why it would not be a good idea to use the least-squares line to predict the volume of grey matter for a child whose head circumference z score was 3.0.
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.2
233
Linear Regression: Fitting a Line to Bivariate Data
5.20 Studies have shown that people who suffer sud-
5.23
den cardiac arrest have a better chance of survival if a defibrillator shock is administered very soon after cardiac arrest. How is survival rate related to the time between when cardiac arrest occurs and when the defibrillator shock is delivered? This question is addressed in the paper “Improving Survival from Sudden Cardiac Arrest:
ysis in Real Estate Appraisal” (Appraisal Journal [2002]: 424–430):
The Role of Home Defibrillators” (by J. K. Stross, University of Michigan, February 2002; available at www.heartstarthome.com). The accompanying data give y 5 survival rate (percent) and x 5 mean call-toshock time (minutes) for a cardiac rehabilitation center (in which cardiac arrests occurred while victims were hospitalized and so the call-to-shock time tended to be short) and for four communities of different sizes: Mean call-to-shock time, x Survival rate, y
2 90
6 45
7 30
9 5
12 2
a. Construct a scatterplot for these data. How would you describe the relationship between mean call-toshock time and survival rate? b. Find the equation of the least-squares line. c. Use the least-squares line to predict survival rate for a community with a mean call-to-shock time of 10 minutes.
5.21 The data given in the previous exercise on x 5 call-to-shock time (in minutes) and y 5 survival rate (percent) were used to compute the equation of the leastsquares line, which was y^ 5 101.33 2 9.30x The newspaper article “FDA OKs Use of Home
Defibrillators” (San Luis Obispo Tribune, November 13, 2002) reported that “every minute spent waiting for paramedics to arrive with a defibrillator lowers the chance of survival by 10 percent.” Is this statement consistent with the given least-squares line? Explain.
5.22 An article on the cost of housing in California that appeared in the San Luis Obispo Tribune (March 30, 2001) included the following statement: “In Northern California, people from the San Francisco Bay area pushed into the Central Valley, benefiting from home prices that dropped on average $4000 for every mile traveled east of the Bay area.” If this statement is correct, what is the slope of the least-squares regression line, y^ 5 a 1 bx, where y 5 house price (in dollars) and x 5 distance east of the Bay (in miles)? Explain.
Bold exercises answered in back
Data set available online
The following data on sale price, size, and land-to-building ratio for 10 large industrial properties appeared in the paper “Using Multiple Regression Anal-
Property
Sale Price (millions of dollars)
Size (thousands of sq. ft.)
Land-toBuilding Ratio
1 2 3 4 5 6 7 8 9 10
10.6 2.6 30.5 1.8 20.0 8.0 10.0 6.7 5.8 4.5
2166 751 2422 224 3917 2866 1698 1046 1108 405
2.0 3.5 3.6 4.7 1.7 2.3 3.1 4.8 7.6 17.2
a. Calculate and interpret the value of the correlation coefficient between sale price and size. b. Calculate and interpret the value of the correlation coefficient between sale price and land-to-building ratio. c. If you wanted to predict sale price and you could use either size or land-to-building ratio as the basis for making predictions, which would you use? Explain. d. Based on your choice in Part (c), find the equation of the least-squares regression line you would use for predicting y sale price. y^ 5 1.333 1 0.00525x
5.24 Representative data read from a plot that appeared in the paper “Effect of Cattle Treading on Erosion from Hill Pasture: Modeling Concepts and Analysis of Rainfall Simulator Data” (Australian Journal of Soil Research [2002]: 963–977) on runoff sediment concentration for plots with varying amounts of grazing damage, measured by the percentage of bare ground in the plot, are given for gradually sloped plots and for steeply sloped plots. Gradually Sloped Plots Bare ground (%) 5 Concentration 50
10 200
Bare ground (%) Concentration
40 500
30 600
15 250
25 500
(continued)
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
234
Chapter 5
Summarizing Bivariate Data
Steeply Sloped Plots Bare ground (%) Concentration
5 100
5 250
10 15 300 600
Bare ground (%) Concentration
20 500
25 500
20 30 900 800
Bare ground (%) Concentration
35 40 35 1100 1200 1000
a. Using the data for steeply sloped plots, find the equation of the least-squares line for predicting y 5 runoff sediment concentration using x 5 percentage of bare ground. y^ 5 59.9 1 27.46x b. What would you predict runoff sediment concentration to be for a steeply sloped plot with 18% bare ground? c. Would you recommend using the least-squares equation from Part (a) to predict runoff sediment concentration for gradually sloped plots? If so, explain why it would be appropriate to do so. If not, provide an alternative way to make such predictions.
5.25 Explain why it can be dangerous to use the leastsquares line to obtain predictions for x values that are substantially larger or smaller than those contained in the sample.
5.26 The sales manager of a large company selected a random sample of n 5 10 salespeople and determined for each one the values of x 5 years of sales experience and y 5 annual sales (in thousands of dollars). A scatterplot of the resulting (x, y) pairs showed a linear pattern. a. Suppose that the sample correlation coefficient is r 5 .75 and that the average annual sales is y 5 100. If a particular salesperson is 2 standard deviations above the mean in terms of experience, what would you predict for that person’s annual sales? Bold exercises answered in back
5.3
Data set available online
b. If a particular person whose sales experience is 1.5 standard deviations below the average experience is predicted to have an annual sales value that is 1 standard deviation below the average annual sales, what is the value of r?
5.27 Explain why the slope b of the least-squares line always has the same sign (positive or negative) as does the sample correlation coefficient r. The accompanying data resulted from an experiment in which weld diameter x and shear strength y (in pounds) were determined for five different spot welds on steel. A scatterplot shows a strong linear pattern. With g 1x 2 x 2 2 5 1000 and g 1x 2 x 2 1 y 2 y 2 5 8577, the least-squares line is y^ 5 2936.22 1 8.577x.
5.28
x y
200.1 813.7
210.1 785.3
220.1 960.4
230.1 1118.0
240.0 1076.2
a. Because 1 lb 5 0.4536 kg, strength observations can be re-expressed in kilograms through multiplication by this conversion factor: new y 5 0.4536(old y). What is the equation of the least-squares line when y is expressed in kilograms? y^ 5 2424.7 1 3.891x b. More generally, suppose that each y value in a data set consisting of n (x, y) pairs is multiplied by a conversion factor c (which changes the units of measurement for y). What effect does this have on the slope b (i.e., how does the new value of b compare to the value before conversion), on the intercept a, and on the equation of the least-squares line? Verify your conjectures by using the given formulas for b and a. (Hint: Replace y with cy, and see what happens— and remember, this conversion will affect y.)
Video Solution available
Assessing the Fit of a Line Once the least-squares regression line has been obtained, the next step is to examine how effectively the line summarizes the relationship between x and y. Important questions to consider are 1. Is a line an appropriate way to summarize the relationship between the two variables? 2. Are there any unusual aspects of the data set that we need to consider before proceeding to use the regression line to make predictions? 3. If we decide that it is reasonable to use the regression line as a basis for prediction, how accurate can we expect predictions based on the regression line to be? In this section, we look at graphical and numerical methods that will allow us to answer these questions. Most of these methods are based on the vertical deviations of the data
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.3
Assessing the Fit of a Line
235
points from the regression line. These vertical deviations are called residuals, and each represents the difference between an actual y value and the corresponding predicted value, y^ , that would result from using the regression line to make a prediction.
Predicted Values and Residuals The predicted value corresponding to the first observation in a data set is obtained by substituting that value, x1, into the regression equation to obtain y^ 1, where y^ 1 5 a 1 bx1 The difference between the actual y value for the first observation, y1, and the corresponding predicted value is y1 2 y^ 1 This difference, called a residual, is the vertical deviation of a point in the scatterplot from the regression line. An observation falling above the line results in a positive residual, whereas a point falling below the line results in a negative residual. This is shown in Figure 5.14. y (x1, y1) y1 is greater than ˆy1 so y1 – yˆ 1 is positive
(x2, yˆ 2) y2 is less than ˆy2 so y2 – yˆ 2 is negative (x1, ˆy1)
(x2, y2)
FIGURE 5.14 Positive and negative deviations from the least-squares line (residuals).
x1
x2
x
DEFINITION The predicted or fitted values result from substituting each sample x value in turn into the equation for the least-squares line. This gives y^ 1 5 first predicted value 5 a 1 bx1 y^ 2 5 second predicted value 5 a 1 bx2 ( y^ n 5 nth predicted value 5 a 1 bxn The residuals from the least-squares line are the n quantities y1 2 y^ 1 5 first residual y2 2 y^ 2 5 second residual ( yn 2 y^ n 5 nth residual Each residual is the difference between an observed y value and the corresponding predicted y value.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
236
Chapter 5 Summarizing Bivariate Data
EXAMPLE 5.7
It May Be a Pile of Debris to You, but It Is Home to a Mouse
The accompanying data is a subset of data read from a scatterplot that appeared in the paper “Small Mammal Responses to fine Woody Debris and Forest Fuel
Reduction in Southwest Oregon” (Journal of Wildlife Management [2005]: 625–632). The authors of the paper were interested in how the distance a deer mouse will travel for food is related to the distance from the food to the nearest pile of fine woody debris. Distances were measured in meters. The data are given in Table 5.1.
T A B L E 5 .1 Predicted Values and Residuals for the Data of Example 5.7 Distance From Debris (x)
Distance Traveled (y)
Predicted Distance Traveled (y^ )
Residual 1 y 2 y^ 2
6.94 5.23 5.21 7.10 8.16 5.50 9.19 9.05 9.36
0.00 6.13 11.29 14.35 12.03 22.72 20.11 26.16 30.65
14.76 9.23 9.16 15.28 18.70 10.10 22.04 21.58 22.59
14.76 3.10 2.13 0.93 6.67 12.62 1.93 4.58 8.06
Minitab was used to fit the least-squares regression line. Partial computer output follows: Regression Analysis: Distance Traveled versus Distance to Debris The regression equation is Distance Traveled 7.7 3.23 Distance to Debris Predictor Coef SE Coef Constant 7.69 13.33 Distance to Debris 3.234 1.782 S 8.67071 R-Sq 32.0% R-Sq(adj) 22.3%
Data set available online
T 0.58 1.82
P 0.582 0.112
The resulting least-squares line is y^ 5 27.69 1 3.234x. A plot of the data that also includes the regression line is shown in Figure 5.15. The residuals for this data set are the signed vertical distances from the points to the line.
30
Distance traveled
25 20 15 10 5 0
FIGURE 5.15 Scatterplot for the data of Example 5.7.
5
6
7 Distance to debris
8
9
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.3
Assessing the Fit of a Line
237
For the mouse with the smallest x value (the third observation with x3 5 5.21 and y3 5 11.29), the corresponding predicted value and residual are predicted value 5 y^ 3 5 27.69 3.234(x3) 5 27.69 1 3.234(5.21) 5 9.16 residual 5 y3 2 y^ 3 5 11.29 2 9.16 5 2.13 The other predicted values and residuals are computed in a similar manner and are included in Table 5.1. Computing the predicted values and residuals by hand can be tedious, but Minitab and other statistical software packages, as well as many graphing calculators, include them as part of the output, as shown in Figure 5.16. The predicted values and residuals can be found in the table at the bottom of the Minitab output in the columns labeled “Fit” and “Residual,” respectively. The regression equation is Distance Traveled = – 7.7 + 3.23 Distance to Debris Predictor Constant Distance to Debris S = 8.67071
Coef –7.69 3.234
R-Sq = 32.0%
SE Coef 13.33 1.782
T –0.58 1.82
P 0.582 0.112
R-Sq(adj) = 22.3%
Analysis of Variance Source Regression Residual Error Total Obs
FIGURE 5.16 Minitab output for the data of Example 5.7.
1 2 3 4 5 6 7 8 9
Distance to Debris 6.94 5.23 5.21 7.10 8.16 5.50 9.19 9.05 9.36
DF 1 7 8 Distance Traveled 0.00 6.13 11.29 14.35 12.03 22.72 20.11 26.16 30.65
SS 247.68 526.27 773.95
MS 247.68 75.18
Fit 14.76 9.23 9.16 15.28 18.70 10.10 22.04 21.58 22.59
SE Fit 2.96 4.69 4.72 2.91 3.27 4.32 4.43 4.25 4.67
F 3.29
Residual –14.76 –3.10 2.13 –0.93 –6.67 12.62 –1.93 4.58 8.06
P 0.112
St Resid –1.81 –0.42 0.29 –0.11 –0.83 1.68 –0.26 0.61 1.10
Plotting the Residuals A careful look at residuals can reveal many potential problems. A residual plot is a good place to start when assessing the appropriateness of the regression line.
DEFINITION A residual plot is a scatterplot of the (x, residual) pairs. Isolated points or a pattern of points in the residual plot indicate potential problems. A desirable residual plot is one that exhibits no particular pattern, such as curvature. Curvature in the residual plot is an indication that the relationship between x and y is not linear and that a curve would be a better choice than a line for describing the relationship between x and y. This is sometimes easier to see in a residual plot than in a scatterplot of y versus x, as illustrated in Example 5.8. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
238
Chapter 5
Summarizing Bivariate Data
E X A M P L E 5 . 8 Heights and Weights of American Women
Data set available online
Consider the accompanying data on x 5 height (in inches) and y 5 average weight (in pounds) for American females, age 30–39 (from The World Almanac and Book of Facts). The scatterplot displayed in Figure 5.17(a) appears rather straight. However, when the residuals from the least-squares line 1 y^ 5 98.23 1 3.59x2 are plotted (Figure 5.17(b)), substantial curvature is apparent (even though r ⬇ .99). It is not accurate to say that weight increases in direct proportion to height (linearly with height). Instead, average weight increases somewhat more rapidly for relatively large heights than it does for relatively small heights. x y
58 113
59 115
60 118
61 121
62 124
63 128
64 131
x y
66 137
67 141
68 145
69 150
70 153
71 159
72 164
y
65 134
Residual 3
170 2 160 1 150
62
70 x
0 140
58
66
−1 130 −2
Strong curved pattern
120
FIGURE 5.17 Plots for the data of Example 5.8: (a) scatterplot; (b) residual plot.
58
62
66 (a)
70
74
x
−3
(b)
There is another common type of residual plot—one that plots the residuals versus the corresponding y^ values rather than versus the x values. Because y^ 5 a 1 bx is simply a linear function of x, the only real difference between the two types of residual plots is the scale on the horizontal axis. The pattern of points in the residual plots will be the same, and it is this pattern of points that is important, not the scale. Thus the two plots give equivalent information, as can be seen in Figure 5.18, which gives both plots for the data of Example 5.7. It is also important to look for unusual values in the scatterplot or in the residual plot. A point falling far above or below the horizontal line at height 0 corresponds to a large residual, which may indicate some type of unusual behavior, such as a recording error, a nonstandard experimental condition, or an atypical experimental subject. A point whose x value differs greatly from others in the data set may have exerted excessive influence in determining the fitted line. One method for assessing the impact of such an isolated point on the fit is to delete it from the data set, recompute the best-fit line, and evaluate the extent to which the equation of the line has changed. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.3
Assessing the Fit of a Line
239
10
Residuals
5 0 –5 –10 –15 5
6
7 Distance to debris
8
9
(a)
10
Residuals
5 0 –5 –10 –15
FIGURE 5.18 Plots for the data of Example 5.7. (a) Plot of residuals versus x; (b) plot of residuals versus y^ .
10
12
14
16 18 Predicted y
20
22
24
(b)
E X A M P L E 5 . 9 Older Than Your Average Bear
Data set available online
The accompanying data on x 5 age (in years) and y 5 weight (in kg) for 12 black bears appeared in the paper “Habitat Selection by Black Bears in an Intensively Logged Boreal Forest” (Canadian Journal of Zoology [2008]: 1307–1316). A scatterplot and residual plot are shown in Figures 5.19(a) and 5.19(b), respectively. One bear in the sample was much older than the other bears (bear 3 with an age of x 5 28.5 years and a weight of y 5 62.00 kg). This results in a point in the scatterplot that is far to the right of the other points in the scatterplot. Because the least-squares line minimizes the sum of squared residuals, the line is pulled toward this observation. This single observation plays a big role in determining the slope of the least-squares line, and it is therefore called an influential observation. Notice that this influential observation is not necessarily one with a large residual, because the least-squares line actually passes near this point. Figure 5.20 shows what happens when the influential observation is removed from the data set. Both the slope and intercept of the least-squares line are quite different from the slope and intercept of the line with this influential observation included.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
240
Chapter 5
Summarizing Bivariate Data
Bear
Age
Weight
1 2 3 4 5 6 7 8 9 10 11 12
10.5 6.5 28.5 10.5 6.5 7.5 6.5 5.5 7.5 11.5 9.5 5.5
54 40 62 51 55 56 62 42 40 59 51 50
Fitted Line Plot Weight = 45.90 + 0.6141 Age
65 Observation with large residual 60 Weight
Influential observation 55 50 45 40 5
10
15
20
25
30
Age (a) Residuals vs Age 15 Observation with large residual
Residuals
10 5 0 –5
Influential observation
–10
FIGURE 5.19 Minitab plots for the bear data of Example 5.9: (a) scatterplot; (b) residual plot.
5
10
15
20
25
30
Age (b)
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.3
Assessing the Fit of a Line
241
Fitted Line Plot—Omit Bear 3 Weight = 41.13 + 1.230 Age 65
Weight
60 55 50 45 40
FIGURE 5.20 Scatterplot and least-squares line with bear 3 removed from data set.
5
6
7
8
9
10
11
12
Age
Some points in the scatterplot may fall far from the least-squares line in the y direction, resulting in a large residual. These points are sometimes referred to as outliers. In this example, the observation with the largest residual is bear 7 with an age of x 5 6.5 years and a weight of y 5 62.00 kg. This observation is labeled in Figure 5.19. Even though this observation has a large residual, this observation is not influential. The equation of the least-squares line for the data set consisting of all 12 observations is y^ 5 45.90 1 0.6141x, which is not much different from the equation that results from deleting bear 7 from the data set 1 y^ 5 43.81 1 0.7131x2 .
Unusual points in a bivariate data set are those that fall away from most of the other points in the scatterplot in either the x direction or the y direction. An observation is potentially an influential observation if it has an x value that is far away from the rest of the data (separated from the rest of the data in the x direction). To determine if the observation is in fact influential, we assess whether removal of this observation has a large impact on the value of the slope or intercept of the least-squares line. An observation is an outlier if it has a large residual. Outlier observations fall far away from the least-squares line in the y direction.
Careful examination of a scatterplot and a residual plot can help us determine the appropriateness of a line for summarizing a relationship. If we decide that a line is appropriate, the next step is to think about assessing the accuracy of predictions based on the least-squares line and whether these predictions (based on the value of x) are better in general than those made without knowledge of the value of x. Two numerical measures that are helpful in this assessment are the coefficient of determination and the standard deviation about the regression line.
Coefficient of Determination Suppose that we would like to predict the price of homes in a particular city. A random sample of 20 homes that are for sale is selected, and y 5 price and x 5 size (in square feet) are recorded for each house in the sample. There will be variability in house price (the houses will differ with respect to price), and it is this variability that Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
242
Chapter 5
Summarizing Bivariate Data
makes accurate prediction of price a challenge. How much of the variability in house price can be explained by the fact that price is related to house size and that houses differ in size? If differences in size account for a large proportion of the variability in price, a price prediction that takes house size into account is a big improvement over a prediction that is not based on size. The coefficient of determination is a measure of the proportion of variability in the y variable that can be “explained” by a linear relationship between x and y.
DEFINITION The coefficient of determination, denoted by r 2, gives the proportion of variation in y that can be attributed to an approximate linear relationship between x and y. The value of r 2 is often converted to a percentage (by multiplying by 100) and interpreted as the percentage of variation in y that can be explained by an approximate linear relationship between x and y. To understand how r 2 is computed, we first consider variation in the y values. Variation in y can effectively be explained by an approximate straight-line relationship when the points in the scatterplot fall close to the least-squares line—that is, when the residuals are small in magnitude. A natural measure of variation about the leastsquares line is the sum of the squared residuals. (Squaring before combining prevents negative and positive residuals from counteracting one another.) A second sum of squares assesses the total amount of variation in observed y values by considering how spread out the y values are from the mean y value.
DEFINITION The total sum of squares, denoted by SSTo, is defined as SSTo 5 1 y1 2 y 2 2 1 1 y2 2 y 2 2 1 c1 1 yn 2 y 2 2 5 g 1 y 2 y 2 2 The residual sum of squares (sometimes referred to as the error sum of squares), denoted by SSResid, is defined as SSResid 5 1 y1 2 y^ 1 2 2 1 1 y2 2 y^ 2 2 2 1 c1 1 yn 2 y^ n 2 2 5 g 1 y 2 y^ 2 2 These sums of squares can be found as part of the regression output from most standard statistical packages or can be obtained using the following computational formulas: 1 g y2 2 2 SSTo 5 g y 2 n SSResid 5 g y2 2 ag y 2 b g xy
E X A M P L E 5 . 1 0 Revisiting the Deer Mice Data Figure 5.21 displays part of the Minitab output that results from fitting the leastsquares line to the data on y 5 distance traveled for food and x 5 distance to nearest woody debris pile from Example 5.7. From the output, SSTo 773.95 and SSResid 526.27 Notice that SSResid is fairly large relative to SSTo. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.3
243
Assessing the Fit of a Line
Regression Analysis: Distance Traveled versus Distance to Debris The regression equation is Distance Traveled = – 7.7 + 3.23 Distance to Debris Predictor Constant Distance to Debris S = 8.67071
Coef –7.69 3.234
R-Sq = 32.0%
SE Coef 13.33 1.782
T –0.58 1.82
P 0.582 0.112
R-Sq(adj) = 22.3%
Analysis of Variance
FIGURE 5.21
Source Regression Residual Error Total
DF 1 7 8
SS 247.68 526.27 773.95
o
Minitab output for the data of Example 5.10.
SST
MS 247.68 75.18
F 3.29
P 0.112
esid
SSR
The residual sum of squares is the sum of squared vertical deviations from the least-squares line. As Figure 5.22 illustrates, SSTo is also a sum of squared vertical deviations from a line—the horizontal line at height y. The least-squares line is, by definition, the one having the smallest sum of squared deviations. It follows that SSResid SSTo. The two sums of squares are equal only when the least-squares line is the horizontal line. y
y
Least-squares line y– Horizontal line at height –y
FIGURE 5.22 Interpreting sums of squares: (a) SSResid 5 sum of squared vertical deviations from the least-squares line; (b) SSTo 5 sum of squared vertical deviations from the horizontal line at height y.
x
x (a)
(b)
SSResid is often referred to as a measure of unexplained variation—the amount of variation in y that cannot be attributed to the linear relationship between x and y. The more the points in the scatterplot deviate from the least-squares line, the larger the value of SSResid and the greater the amount of y variation that cannot be explained by the approximate linear relationship. Similarly, SSTo is interpreted as a measure of total variation. The larger the value of SSTo, the greater the amount of variability in y1, y2, . . . , yn. The ratio SSResid/SSTo is the fraction or proportion of total variation that is unexplained by a straight-line relation. Subtracting this ratio from 1 gives the proportion of total variation that is explained:
The coefficient of determination is computed as r2 5 1 2
SSResid SSTo
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
244
Chapter 5
Summarizing Bivariate Data
Multiplying r 2 by 100 gives the percentage of y variation attributable to the approximate linear relationship. The closer this percentage is to 100%, the more successful is the relationship in explaining variation in y.
E X A M P L E 5 . 1 1 r 2 for the Deer Mice Data For the data on distance traveled for food and distance to nearest debris pile from Example 5.10, we found SSTo 773.95 and SSResid 526.27. Thus r2 5 1 2
SSResid 526.27 512 5 .32 SSTo 773.95
This means that only 32% of the observed variability in distance traveled for food can be explained by an approximate linear relationship between distance traveled for food and distance to nearest debris pile. Note that the r 2 value can be found in the Minitab output of Figure 5.21, labeled “R-Sq.”
The symbol r was used in Section 5.1 to denote Pearson’s sample correlation coefficient. It is not coincidental that r 2 is used to represent the coefficient of determination. The notation suggests how these two quantities are related: (correlation coefficient) 2 coefficient of determination Thus, if r .8 or r 2.8, then r 2 .64, so 64% of the observed variation in the dependent variable can be explained by the linear relationship. Because the value of r does not depend on which variable is labeled x, the same is true of r 2. The coefficient of determination is one of the few quantities computed in a regression analysis whose value remains the same when the roles of dependent and independent variables are interchanged. When r .5, we get r 2 .25, so only 25% of the observed variation is explained by a linear relation. This is why a value of r between 2.5 and .5 is not considered evidence of a strong linear relationship.
E X A M P L E 5 . 1 2 Lead Exposure and Brain Volume The authors of the paper “Decreased Brain Volume in Adults with Childhood Lead Exposure” (Public Library of Science Medicine [May 27, 2008]: e112) studied the relationship between childhood environmental lead exposure and a measure of brain volume change in a particular region of the brain. Data on x 5 mean childhood blood lead level (mg/dL) and y 5 brain volume change (percent) read from a graph that appeared in the paper was used to produce the scatterplot in Figure 5.23. The least-squares line is also shown on the scatterplot. Figure 5.24 displays part of the Minitab output that results from fitting the leastsquares line to the data. Notice that although there is a slight tendency for smaller y values (corresponding to a brain volume decrease) to be paired with higher values of mean blood lead levels, the relationship is weak. The points in the plot are widely scattered around the least-squares line. From the computer output, we see that 100r 2 5 13.6%, so r 2 5 .136. This means that differences in childhood mean blood lead level explain only 13.6% of the variability in adult brain volume change. Because the coefficient of determination is the square of the correlation coefficient, we can compute the value of the correlation Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.3
Assessing the Fit of a Line
245
Brain volume change
0.05
FIGURE 5.23
0.0
–0.05
–0.10 0
Scatterplot and least-squares line for the data of Example 5.12.
10
20 Mean blood lead
30
40
Regression Analysis: Brain Volume Change versus Mean Blood Lead The regression equation is Brain Volume Change = 0.01559 – 0.001993 Mean Blood Lead S = 0.0310931
R-Sq = 13.6%
R-Sq(adj) = 12.9%
Analysis of Variance
FIGURE 5.24 Minitab output for the data of Example 5.12.
Source Regression Error Total
DF 1 111 112
SS 0.016941 0.107313 0.124254
MS 0.0169410 0.0009668
F 17.52
P 0.000
coefficient by taking the square root of r 2. In this case, we know that the correlation coefficient will be negative (because there is a negative relationship between x and y), so we want the negative square root: r 5 2".136 5 2.369 Based on the values of the correlation coefficient and the coefficient of determination, we would conclude that there is a weak negative linear relationship and that childhood mean blood lead level explains only about 13.6% of adult change in brain volume.
Standard Deviation About the Least-Squares Line The coefficient of determination measures the extent of variation about the best-fit line relative to overall variation in y. A high value of r 2 does not by itself promise that the deviations from the line are small in an absolute sense. A typical observation could deviate from the line by quite a bit, yet these deviations might still be small relative to overall y variation. Recall that in Chapter 4 the sample standard deviation s5
g 1x 2 x 2 2 Å n21
was used as a measure of variability in a single sample; roughly speaking, s is the typical amount by which a sample observation deviates from the mean. There is an analogous measure of variability when a least-squares line is fit. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
246
Chapter 5
Summarizing Bivariate Data
DEFINITION The standard deviation about the least-squares line is given by se 5
SSResid Än22
Roughly speaking, se is the typical amount by which an observation deviates from the least-squares line. Justification for division by (n 2 2) and the use of the subscript e is given in Chapter 13.
E X A M P L E 5 . 1 3 Predicting Graduation Rates Consider the accompanying data from 2007 on six-year graduation rate (%), student-related expenditure per full-time student, and median SAT score for the 38 primarily undergraduate public universities and colleges in the United States with enrollments between 10,000 and 20,000 (Source: College Results Online, The Education Trust). Graduation Rate
Expenditure
Median SAT
81.2 66.8 66.4 66.1 64.9 63.7 62.6 62.5 61.2 59.8 56.6 54.8 52.7 52.4 52.4 50.5 49.9 48.9 48.1 46.5 45.3 45.2 43.7 43.5 42.9 42.1 42.0 38.9 38.8 38.3 35.9
7462 7310 6959 8810 7657 8063 8352 7789 8106 7776 8515 7037 8715 7780 7198 7429 7551 8112 8149 6744 8842 7743 5587 7166 5749 6268 8477 7076 8153 7342 8444
1160 1115 1070 1205 1135 1060 1130 1200 1015 1100 990 1085 1040 1040 1105 975 1030 1030 950 1010 1223 990 1010 1010 950 955 985 990 990 910 1075
Data set available online Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.3
Assessing the Fit of a Line
Graduation Rate
Expenditure
Median SAT
32.8 32.6 32.3 31.8 31.3 31.0 26.0
7245 6408 4981 7333 7984 5811 7410
885 1060 990 970 905 1010 1005
247
Figure 5.25 displays scatterplots of graduation rate versus student-related expenditure and graduation rate versus median SAT score. The least-squares lines and the values of r 2 and se are also shown. Graduation Rate = 12.41 + 0.004834 Student-related Expenditure S 12.2846 R-Sq 11.9% R-Sq(adj) 9.4%
80
Graduation rate
70 60 50 40 30 20 5000
6000 7000 8000 Student-related expenditure
9000
(a) Graduation Rate = –57.43 + 0.1023 Median SAT S 9.95214 R-Sq 42.1% R-Sq(adj) 40.5%
80
Graduation rate
70 60 50 40 30
FIGURE 5.25 Scatterplots for the data of Example 5.13: (a) graduation rate versus student-related expenditure; (b) graduation rate versus median SAT.
20 900
950
1000
1050 1100 Median SAT
1150
1200
1250
(b)
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
248
Chapter 5
Summarizing Bivariate Data
Notice that while there is a positive linear relationship between student-related expenditure and graduation rate, the relationship is weak. The value of r 2 is only .119 (11.9%), indicating that only about 11.9% of the variability in graduation rate from university to university can be explained by student-related expenditures. The standard deviation about the regression line is se 5 12.2846, which is larger than se for the predictor median SAT, a reflection of the fact that the points in the scatterplot of graduation rate versus student-related expenditure tend to fall farther from the regression line than is the case for the line that describes graduation rate versus median SAT. The value of r 2 for graduation rate versus median SAT is .421 (42.1%) and se 5 9.95214, indicating that the predictor median SAT does a better job of explaining variability in graduation rates and the corresponding least-squares line would be expected to produce more accurate estimates of graduation rates than would be the case for the predictor student-related expenditure. Based on the values of r 2 and se, median SAT would be a better choice for predicting graduation rates than student-related expenditures. It is also possible to develop a prediction equation that would incorporate both potential predictors— techniques for doing this are introduced in Chapter 14.
EX E RC I S E S 5. 2 9 - 5 . 4 3 The data in the accompanying table is from the paper “Six-Minute Walk Test in Children and Adolescents” (The Journal of Pediatrics [2007]: 395–399). Two hundred and eighty boys completed a test that measures the distance that the subject can walk on a flat, hard surface in 6 minutes. For each age group shown in the table, the median distance walked by the boys in that age group is also given.
5.29
Age Group
Representative Age (Midpoint of Age Group)
Median Six-minute Walk Distance (meters)
3–5 6–8 9–11 12–15 16–18
4 7 10 13.5 17
544.3 584.0 667.3 701.1 727.6
a. With x 5 representative age and y 5 median distance walked in 6 minutes, construct a scatterplot. Does the pattern in the scatterplot look linear? b. Find the equation of the least-squares regression line that describes the relationship between median distance walked in 6 minutes and representative age. c. Compute the five residuals and construct a residual plot. Are there any unusual features in the plot?
Bold exercises answered in back
Data set available online
The paper referenced in the previous exercise also gave the 6-minute walk distances for 248 girls age 3 to 18 years. The median 6-minute walk times for girls for the five age groups were
5.30
492.4
578.3
655.8
657.6
660.9
a. With x 5 representative age and y 5 median distance walked in 6 minutes, construct a scatterplot. How does the pattern in the scatterplot for girls differ from the pattern in the scatterplot for boys from Exercise 5.29? b. Find the equation of the least-squares regression line that describes the relationship between median distance walked in 6 minutes and representative age for girls. y^ 5 479.997 1 12.525x c. Compute the five residuals and construct a residual plot. The authors of the paper decided to use a curve rather than a straight line to describe the relationship between median distance walked in 6 minutes and age for girls. What aspect of the residual plot supports this decision?
5.31 Data on pollution and cost of medical care for elderly people were given in Exercise 5.14 and are also shown here. The accompanying data are a measure of pollution (micrograms of particulate matter per cubic meter of air) and the cost of medical care per person over age 65 for six geographic regions of the United States. Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.3
Region North Upper South Deep South West South Big Sky West
Pollution
Cost of Medical Care
30.0 31.8 32.1 26.8 30.4 40.0
915 891 968 972 952 899
The equation of the least-squares regression line for this data set is y^ 5 1082.2 2 4.691x, where y medical cost and x pollution. a. Compute the six residuals. b. What is the value of the correlation coefficient for this data set? Does the value of r indicate that the linear relationship between pollution and medical cost is strong, moderate, or weak? Explain. c. Construct a residual plot. Are there any unusual features of the plot? d. The observation for the West, (40.0, 899), has an x value that is far removed from the other x values in the sample. Is this observation influential in determining the values of the slope and /or intercept of the least-squares line? Justify your answer. Northern flying squirrels eat lichen and fungi, which makes for a relatively low quality diet. The authors of the paper “Nutritional Value and Diet Prefer-
5.32
ence of Arboreal Lichens and Hypogeous Fungi for Small Mammals in the Rocky Mountain” (Canadian Journal of Zoology [2008]: 851–862) measured nitrogen intake and nitrogen retention in six flying squirrels that were fed the fungus Rhizopogon. Data read from a graph that appeared in the paper are given in the table below. (The negative value for nitrogen retention for the first squirrel represents a net loss in nitrogen.) Nitrogen Intake, x (grams)
Nitrogen Retention, y (grams)
0.03 0.10 0.07 0.06 0.07 0.25
0.04 0.00 0.01 0.01 0.04 0.11
a. Construct a scatterplot of these data. b. Find the equation of the least-squares regression line. Based on this line, what would you predict nitrogen retention to be for a flying squirrel whose Bold exercises answered in back
Data set available online
Assessing the Fit of a Line
249
nitrogen intake is 0.06 grams? What is the residual associated with the observation (0.06, 0.01)? c. Look again at the scatterplot from Part (a). Which observation is potentially influential? Explain the reason for your choice. d. When the potentially influential observation is deleted from the data set, the equation of the leastsquares regression line fit to the remaining five observations is y^ 5 20.037 1 0.627x. Use this equation to predict nitrogen retention for a flying squirrel whose nitrogen intake is 0.06. Is this prediction much different than the prediction made in Part (b)? y^ 5 0.00062, small difference The relationship between x 5 total number of salmon in a creek and y 5 percentage of salmon killed by bears that were transported away from the stream prior to the bear eating the salmon was examined in the paper “Transportation of Pacific Salmon Carcasses
5.33
from Streams to Riparian Forests by Bears” (Canadian Journal of Zoology [2009]: 195–203). Data for the 10 years from 1999 to 2008 is given in the accompanying table. Total Number
Percentage Transported
19,504 3,460 1,976 8,439 11,142 3,467 3,928 20,440 7,850 4,134
77.8 28.7 28.9 27.9 55.3 20.4 46.8 76.3 40.3 24.1
a. Construct a scatterplot of the data. Does there appear to be a relationship between the total number of salmon in the stream and the percentage of salmon killed by bears that are transported away from the stream? b. Find the equation of the least-squares regression line. Draw the regression line for the scatterplot from Part (a). y^ 5 18.483 1 0.00287x c. The residuals from the least-squares line are shown in the accompanying table. The observation (3928, 46.8) has a large residual. Is this data point also an influential observation?
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
250
Chapter 5 Summarizing Bivariate Data
Total Number
Percent Transported
19,504
77.8
3.43
3,460 1,976 8,439 11,142 3,467 3,928 20,440 7,850 4,134
28.7 28.9 27.9 55.3 20.4 46.8 76.3 40.3 24.1
0.30 4.76 14.76 4.89 8.02 17.06 0.75 0.68 6.23
d. The two points with unusually large x values (19,504 and 20,440) were not thought to be influential observations even though they are far removed in the x direction from the rest of the points in the scatterplot. Explain why these two points are not influential. e. Partial Minitab output resulting from fitting the least-squares line is shown here. What is the value of se? Write a sentence interpreting this value. Regression Analysis: Percent Transported versus Total Number The regression equation is
Average Swim Distance (meters)
50–59 60–69 70–79 80–89
55 65 75 85
3361.9 3000.1 2649.0 2118.4
a. Find the equation of the least-squares line with x 5 representative age and y 5 average swim distance. b. Compute the seven residuals and use them to construct a residual plot. What does the residual plot suggest about the appropriateness of using a line to describe the relationship between representative age and swim distance? c. Would it be reasonable to use the least-squares line from Part (a) to predict the average swim distance for women age 40 to 49 by substituting the representative age of 45 into the equation of the least-squares line? Explain.
5.35 Data on x 5 representative age and y 5 6-minute walk time for boys were given in Exercise 5.29. Compute the values of se and r 2 for these data. What do these values tell you about the fit of the least-squares line? Cost-to-charge ratio (the percentage of the amount billed that represents the actual cost) for inpatient and outpatient services at 11 Oregon hospitals is shown in the following table (Oregon Department of Health Services, 2002). A scatterplot of the data is also shown.
5.36
Percent Transported 18.5 0.00287 Total Number Predictor Constant Total Number S 9.16217
Age Group
Representative Age (Midpoint of Age Group)
Residual
Coef SE Coef T P 18.483 4.813 3.84 0.005 0.0028655 0.0004557 6.29 0.000 R-Sq 83.2% R-Sq(adj) 81.1%
f. What is the value of r 2 for this data set (see Minitab output in Part (e))? Is the value of r 2 large or small? Write a sentence interpreting the value of r 2.
5.34 The paper “Effects of Age and Gender on Physical Performance” (Age [2007]: 77–85) describes a study of the relationship between age and 1-hour swimming performance. Data on age and swim distance for over 10,000 men participating in a national long-distance 1-hour swimming competition are summarized in the accompanying table.
Age Group
Representative Age (Midpoint of Age Group)
Average Swim Distance (meters)
20–29 30–39 40–49
25 35 45
3913.5 3728.8 3579.4
Cost-to-Charge Ratio Hospital
Outpatient Care
Inpatient Care
1 2 3 4 5 6 7 8 9 10 11
62 66 63 51 75 65 56 45 48 71 54
80 76 75 62 100 88 64 50 54 83 100
(continued)
Bold exercises answered in back
Data set available online
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.3
Inpatient
251
Assessing the Fit of a Line
Perceived Stress
Telomere Length
Perceived Stress
Telomere Length
5 6 6 7 10 11 12 13 14 14 15 15 15 17 17 17 18 18 19
1.25 1.32 1.5 1.35 1.3 1 1.18 1.1 1.08 1.3 0.92 1.22 1.24 1.12 1.32 1.4 1.12 1.46 0.84
20 20 20 21 21 21 22 22 22 23 24 24 25 26 27 27 28 29 33
1.22 1.30 1.32 1.24 1.26 1.30 1.18 1.22 1.24 1.18 1.12 1.50 0.94 0.84 1.02 1.12 1.22 1.30 0.94
100 90 80 70 60 50 45
55
65
75
Outpatient
The least-squares regression line with y inpatient costto-charge ratio and x outpatient cost-to-charge ratio is y^ 5 21.1 1 1.29x. a. Is the observation for Hospital 11 an influential observation? Justify your answer. b. Is the observation for Hospital 11 an outlier? Explain. c. Is the observation for Hospital 5 an influential observation? Justify your answer. d. Is the observation for Hospital 5 an outlier? Explain.
5.37 The article “Examined Life: What Stanley H. Kaplan Taught Us About the SAT” (The New Yorker [December 17, 2001]: 86–92) included a summary of findings regarding the use of SAT I scores, SAT II scores, and high school grade point average (GPA) to predict firstyear college GPA. The article states that “among these, SAT II scores are the best predictor, explaining 16 percent of the variance in first-year college grades. GPA was second at 15.4 percent, and SAT I was last at 13.3 percent.” a. If the data from this study were used to fit a leastsquares line with y 5 first-year college GPA and x 5 high school GPA, what would be the value of r 2? b. The article stated that SAT II was the best predictor of first-year college grades. Do you think that predictions based on a least-squares line with y 5 first-year college GPA and x 5 SAT II score would be very accurate? Explain why or why not.
5.38
The paper “Accelerated Telomere Shortening
in Response to Life Stress” (Proceedings of the National Academy of Sciences [2004]: 17312–17315) described a study that examined whether stress accelerates aging at a cellular level. The accompanying data on a measure of perceived stress (x) and telomere length ( y) were read from a scatterplot that appeared in the paper. Telomere length is a measure of cell longevity. Bold exercises answered in back
Data set available online
a. Compute the equation of the least-squares line. b. What is the value of r 2? c. Does the linear relationship between perceived stress and telomere length account for a large or small proportion of the variability in telomere length? Justify your answer. The article “California State Parks Closure List Due Soon” (The Sacramento Bee, August 30, 2009)
5.39
gave the following data on y 5 number of employees in fiscal year 2007–2008 and x 5 total size of parks (in acres) for the 20 state park districts in California: Number of Employees, y
Total Park Size, x
95 95 102 69 67 77 81 116 51 36 96 71 76 112 43
39,334 324 17,315 8,244 620,231 43,501 8,625 31,572 14,276 21,094 103,289 130,023 16,068 3,286 24,089 Continued
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
252
Chapter 5 Summarizing Bivariate Data
Number of Employees, y
Total Park Size, x
87 131 138 80 52
6,309 14,502 62,595 23,666 35,833
a. Construct a scatterplot of the data. b. Find the equation of the least-squares line. Do you think the least-squares line gives accurate predictions? Explain. c. Delete the observation with the largest x value from the data set and recalculate the equation of the leastsquares line. Does this observation greatly affect the equation of the line? The article referenced in the previous exercise also gave data on the percentage of operating costs covered by park revenues for the 2007–2008 fiscal year.
5.40
Number of Employees, x
Percent of Operating Cost Covered by Park Revenues, y
95 95 102 69 67 77 81 116 51 36
37 19 32 80 17 34 36 32 38 40
Number of Employees, x
Percent of Operating Cost Covered by Park Revenues, y
96 71 76 112 43 87 131 138 80 52
53 31 35 108 34 97 62 36 36 34
a. Find the equation of the least-squares line relating y 5 percent of operating costs covered by park revenues and x 5 number of employees. b. Based on the values of r 2 and se, do you think that the least-squares regression line does a good job of describing the relationship between y 5 percent of operating costs covered by park revenues and x 5 number of employees? Explain. c. The graph in Figure EX5.40 is a scatterplot of y 5 percent of operating costs covered by park revenues and x 5 number of employees. The least-squares line is also shown. Which observations are outliers? Do the observations with the largest residuals correspond to the park districts with the largest number of employees?
5.41 A study was carried out to investigate the rela-
(continued)
tionship between the hardness of molded plastic ( y, in Brinell units) and the amount of time elapsed since termination of the molding process (x, in hours). Summary quantities include n 5 15, SSResid 5 1235.470, and
Fitted Line Plot Percentage of operating costs = 27.71 + 0.2011 number of employees
Percentage of operating costs
110
S 24.1475 R-Sq 5.3% R-Sq(adj) 0.1%
100 90 80 70 60 50 40 30 20 50
75 100 125 Number of employees
150
FIGURE EX5.40 Bold exercises answered in back
Data set available online
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.4
253
a. For what value of r is se as large as sy? What is the least-squares line in this case? r 5 0, y^ 5 y b. For what values of r will se be much smaller than sy? c. A study by the Berkeley Institute of Human Development (see the book Statistics by Freedman et al., listed in the back of the book) reported the following summary data for a sample of n 5 66 California boys:
SSTo 5 25,321.368. Calculate and interpret the coefficient of determination.
5.42 Both r 2 and se are used to assess the fit of a line. a. Is it possible that both r 2 and se could be large for a bivariate data set? Explain. (A picture might be helpful.) b. Is it possible that a bivariate data set could yield values of r 2 and se that are both small? Explain. (Again, a picture might be helpful.) c. Explain why it is desirable to have r 2 large and se small if the relationship between two variables x and y is to be described using a straight line.
r ⬇ .80 At age 6, average height ⬇ 46 inches, standard deviation ⬇ 1.7 inches. At age 18, average height ⬇ 70 inches, standard deviation ⬇ 2.5 inches.
5.43 With a bit of algebra, we can show that
What would se be for the least-squares line used to predict 18-year-old height from 6-year-old height? d. Referring to Part (c), suppose that you wanted to predict the past value of 6-year-old height from knowledge of 18-year-old height. Find the equation for the appropriate least-squares line. What is the corresponding value of se? y^ 5 7.92 1 .544x, se 5 1.02
SSResid 5 11 2 r 22 g 1 y 2 y 2 2 from which it follows that se 5
Nonlinear Relationships and Transformations
n21 "1 2 r 2 sy Ån 2 2
Unless n is quite small, (n 1)/(n 2) ⬇ 1, so se < "1 2 r 2 sy Bold exercises answered in back
5.4
Data set available online
Video Solution available
Nonlinear Relationships and Transformations As we have seen in previous sections, when the points in a scatterplot exhibit a linear pattern and the residual plot does not reveal any problems with the linear fit, the least-squares line is a sensible way to summarize the relationship between x and y. A linear relationship is easy to interpret, departures from the line are easily detected, and using the line to predict y from our knowledge of x is straightforward. Often, though, a scatterplot or residual plot exhibits a curved pattern, indicating a more complicated relationship between x and y. In this case, finding a curve that fits the observed data well is a more complex task. In this section, we consider two common approaches to fitting nonlinear relationships: polynomial regression and transformations.
Polynomial Regression Let’s reconsider the data first introduced in Example 5.4 on x 5 age and y 5 average marathon finish time: Age Group
x ⫽ Representative Age
y ⫽ Average Finish Time
10–19 20–29 30–39 40–49 50–59 60–69
15 25 35 45 55 65
302.38 193.63 185.46 198.49 224.30 288.71
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
254
Chapter 5 Summarizing Bivariate Data
The scatterplot of these data is reproduced here as Figure 5.26. Because this plot shows a marked curved pattern, it is clear that no straight line can do a reasonable job of describing the relationship between x and y. However, the relationship can be described by a curve, and in this case the curved pattern in the scatterplot looks like a parabola (the graph of a quadratic function). This suggests trying to find a quadratic function of the form y^ 5 a 1 b1x 1 b2x 2 that would reasonably describe the relationship. That is, the values of the coefficients a, b1, and b2 in this function must be selected to obtain a good fit to the data.
Average finish time
300 275 250 225 200
10
FIGURE 5.26
20
30
Scatterplot for the marathon data.
40 Age
50
60
70
What are the best choices for the values of a, b1, and b2? In fitting a line to data, we used the principle of least squares to guide our choice of slope and intercept. Least squares can be used to fit a quadratic function as well. The deviations, y 2 y^ , are still represented by vertical distances in the scatterplot, but now they are vertical distances from the points to a parabola (the graph of a quadratic function) rather than to a line, as shown in Figure 5.27. We then choose values for the coefficients in the quadratic function so that the sum of squared deviations is as small as possible. y
Deviation
FIGURE 5.27 Deviation for a quadratic function.
x
x1
For a quadratic regression, the least squares estimates of a, b1, and b2 are those values that minimize the sum of squared deviationsg 1 y 2 y^ 2 2 where y^ 5 a 1 b1 x 1 b2 x 2. For quadratic regression, a measure that is useful for assessing fit is R2 5 1 2
SSResid SSTo
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.4
255
Nonlinear Relationships and Transformations
where SSResid 5 g 1y 2 y^ 2 2. The measure R 2 is defined in a way similar to r 2 for simple linear regression and is interpreted in a similar fashion. The notation r 2 is used only with linear regression to emphasize the relationship between r 2 and the correlation coefficient, r, in the linear case. The general expressions for computing the least-squares estimates are somewhat complicated, so we rely on a statistical software package or graphing calculator to do the computations for us.
E X A M P L E 5 . 1 4 Marathon Data Revisited: Fitting a Quadratic Model
300
75
275
50 25
Residual
Average finish time
For the marathon data, the scatterplot (see Figure 5.26) showed a marked curved pattern. If the least-squares line is fit to these data, it is no surprise that the line does not do a good job of describing the relationship (r 2 5 .001 or .1% and se 5 56.9439), and the residual plot shows a distinct curved pattern as well (Figure 5.28).
250 225
0 −25
200 −50 10
20
30
FIGURE 5.28 Plots for the marathon data of Example 5.14: (a) least-square regression line; (b) residual plot.
40 Age (a)
50
60
70
20
10
30
40 x (b)
50
60
70
Part of the Minitab output from fitting a quadratic regression function to these data is as follows: The regression equation is y = 462 – 14.2 x + 0.179 x-squared Predictor Constant x x-squared S = 18.4813
Coef 462.00 –14.205 0.17888
SE Coef 43.99 2.460 0.03025
R-Sq = 92.1%
T 10.50 –5.78 5.91
P 0.002 0.010 0.010
R-Sq(adj) = 86.9%
Analysis of Variance Source Regression Residual Error Total
DF 2 3 5
SS 11965.0 1024.7 12989.7
MS 5982.5 341.6
F 17.52
P 0.022
The least-squares coefficients are a 462.00
b1 14.205
b2 0.17888
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
256
Chapter 5 Summarizing Bivariate Data
and the least-squares quadratic is y^ 5 462.00 2 14.205x 1 0.17888x 2 A plot showing the curve and the corresponding residual plot for the quadratic regression are given in Figure 5.29. Notice that there is no strong pattern in the residual plot for the quadratic case, as there was in the linear case. For the quadratic regression, R 2 5 .921 (as opposed to .001 for the least-squares line), which means that 92.1% of the variability in average marathon finish time can be explained by an approximate quadratic relationship between average finish time and age. 300
10 0
260 Residual
Average finish time
280
240 220
−10 −20
200
−30
180 10
20
30
40 Age (a)
50
60
10
70
20
30
40 x (b)
50
60
70
FIGURE 5.29 Quadratic regression of Example 5.13: (a) scatterplot; (b) residual plot.
Linear and quadratic regression are special cases of polynomial regression. A polynomial regression curve is described by a function of the form y^ 5 a 1 b1x 1 b2 x 2 1 b3 x 3 1 c1 bk x k which is called a kth-degree polynomial. The case of k 5 1 results in linear regression 1 y^ 5 a 1 b1x2 and k 5 2 yields a quadratic regression 1 y^ 5 a 1 b1x 1 b2 x 22 . A quadratic curve has only one bend (see Figure 5.30(a) and (b)). A less frequently encountered special case is for k 5 3, where y^ 5 a 1 b1x 1 b2 x 2 1 b3 x 3, which is called a cubic regression curve. Cubic curves have two bends, as shown in Figure 5.30(c). y
y
y
FIGURE 5.30 Polynomial regression curves: (a) quadratic curve with b2 0; (b) quadratic curve with b2 0; (c) cubic curve.
x (a)
x
x (b)
(c)
E X A M P L E 5 . 1 5 Fish Food
Data set available online
Sea bream are one type of fish that are often raised in large fish farming enterprises. These fish are usually fed a diet consisting primarily of fish meal. The authors of the paper “Growth and Economic Profit of Gilthead Sea Bream (Sparus aurata, L.) Fed Sunflower Meal (Aquaculture [2007]: 528–534) describe a study to investigate
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.4
Nonlinear Relationships and Transformations
257
whether it would be more profitable to substitute plant protein in the form of sunflower meal for some of the fish meal in the sea bream’s diet. The accompanying data are consistent with summary quantities given in the paper for x 5 percent of sunflower meal in the diet and y 5 average weight (in grams) of fish after 248 days. Sunflower Meal (%)
Average Fish Weight
0 6 12 18 24 30 36
432 450 455 445 427 422 421
Figure 5.31 shows a scatterplot of these data. The relationship between x and y does not appear to be linear, so we might try using a quadratic regression to describe the relationship between sunflower meal content and average fish weight.
Averagee fish weight
460
FIGURE 5.31 Scatterplot of average fish weight versus sunflower meal content for the data of Example 5.15.
450
440
430
420 0
10
20 Sunflower meal (%)
30
40
Minitab was used to fit a quadratic regression function and to compute the corresponding residuals. The least-squares quadratic regression is y^ 439 1.22x 0.053x 2 A plot of the quadratic regression curve and the corresponding residual plot are shown in Figure 5.32. Notice that the residual plot in Figure 5.32(b) shows a curved pattern (cubic)—not something we like to see in a residual plot. This suggests that we may want to consider something other than a quadratic curve to describe the relationship between x and y. Looking again at the scatterplot of Figure 5.31, we see that a cubic function might be a better choice because there appear to be two “bends” in the curved relationship—one at around x 5 12 and another at the far right hand side of the scatterplot. Using the given data, Minitab was used to fit a cubic regression, resulting in the curve shown in Figure 5.33(a). The cubic regression is then y^ 431.5 5.39x 0.37x 2 0.006x 3 The corresponding residual plot, shown in Figure 5.33(b), does not reveal any troublesome patterns that would suggest a choice other than the cubic regression. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
258
Chapter 5 Summarizing Bivariate Data
Fitted Line Plot Average fish weight = 439.0 + 1.220 sunflower meal (%) −0.05324 Sunflower Meal (%)**2 S 9.46736 R-Sq 69.0% R-Sq(adj) 53.5%
450
10
5
440 Residuals
Averagee fish weight
460
430
420
−5
410
−10 0
10
20 Sunflower meal (%)
30
40
10
(a)
20 Sunflower meal (%)
30
40
(b)
FIGURE 5.32 Quadratic regression plots for the fish food data of Example 5.15: (a) least-squares quadratic regression; (b) residual plot for quadratic regression.
Fitted Line Plot Average fish weight = 431.5 + 5.387 sunflower meal (%) −0.3657 sunflower meal (%)**2 + 0.005787 sunflower meal (%)**3 S 2.64725 R-Sq 98.2% R-Sq(adj) 96.4%
450
2 1 Residual
Averagee fish weight
460
440
0 −1
430 −2 −3
420 0
10
20 Sunflower meal (%)
30
40
(a)
10
20 Sunflower meal (%)
30
40
(b)
FIGURE 5.33 Cubic regression plots for the fish food data of Example 5.15: (a) least-squares cubic regression; (b) residual plot for cubic regression.
Based on analysis of these data, we might recommend using sunflower meal for about 12% of the diet. Sunflower meal is less costly than fish meal, but using more than about 12% sunflower meal is associated with a decrease in the average fish weight. It is not clear what happens to average fish weight when sunflower meal is used for more than 36% of the diet, the largest x value in the data set.
Transformations An alternative to finding a curve to fit the data is to find a way to transform the x values and/or y values so that a scatterplot of the transformed data has a linear appearance. A transformation (sometimes called a reexpression) involves using a simple funcCopyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.4
Nonlinear Relationships and Transformations
259
tion of a variable in place of the variable itself. For example, instead of trying to describe the relationship between x and y, it might be easier to describe the relationship between !x and y or between x and log( y). And, if we can describe the relationship between, say, !x and y, we will still be able to predict the value of y for a given x value. Common transformations involve taking square roots, logarithms, or reciprocals.
E X A M P L E 5 . 1 6 River Water Velocity and Distance from Shore As fans of white-water rafting know, a river flows more slowly close to its banks (because of friction between the river bank and the water). To study the nature of the relationship between water velocity and the distance from the shore, data were gathered on velocity (in centimeters per second) of a river at different distances (in meters) from the bank. Suppose that the resulting data were as follows: Distance .5 1.5 2.5 3.5 4.5 5.5 6.5 7.5 8.5 9.5 Velocity 22.00 23.18 25.48 25.25 27.15 27.83 28.49 28.18 28.50 28.63 A graph of the data exhibits a curved pattern, as seen in both the scatterplot and the residual plot from a linear fit (see Figures 5.34(a) and 5.34(b)). Velocity
FIGURE 5.34 Plots for the data of Example 5.16: (a) scatterplot of the river data; (b) residual plot from linear fit.
Residual 1
30 29 28 27 26 25 24 23 22
–1 0 1 2 3 4 5 6 7 8 9 10 Distance (a)
1
2
3
4 5 6 Distance (b)
7
8
9 10
Let’s try transforming the x values by replacing each x value by its square root. We define xr 5 !x The resulting transformed data are given in Table 5.2.
T AB LE 5 .2 Original and Transformed Data of Example 5.16 Original Data
Data set available online
Transformed Data
x
y
x⬘
y
0.5 1.5 2.5 3.5 4.5 5.5 6.5 7.5 8.5 9.5
22.00 23.18 25.48 25.25 27.15 27.83 28.49 28.18 28.50 28.63
0.7071 1.2247 1.5811 1.8708 2.1213 2.3452 2.5495 2.7386 2.9155 3.0822
22.00 23.18 25.48 25.25 27.15 27.83 28.49 28.18 28.50 28.63
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
260
Chapter 5
Summarizing Bivariate Data
Figure 5.35(a) shows a scatterplot of y versus x (or equivalently y versus !x). The pattern of points in this plot looks linear, and so we can fit a least-squares line using the transformed data. The Minitab output from this regression appears below. Regression Analysis The regression equation is Velocity = 20.1 + 3.01 sqrt distance Predictor Constant Sqrt dis
Coef 20.1102 3.0085
S = 0.6292
StDev 0.6097 0.2726
R-Sq = 93.8%
T 32.99 11.03
P 0.000 0.000
R-Sq(adj) = 93.1%
Analysis of Variance Source Regression Residual Error Total
DF 1 8 9
SS 48.209 3.168 51.376
MS 48.209 0.396
F 121.76
P 0.000
The residual plot in Figure 5.35(b) shows no indication of a pattern. The resulting regression equation is y^ 5 20.1 1 3.01xr An equivalent equation is y^ 5 20.1 1 3.01!x The values of r 2 and se (see the Minitab output) indicate that a line is a reasonable way to describe the relationship between y and x. To predict velocity of the river at a distance of 9 meters from shore, we first compute xr 5 !x 5 !9 5 3 and then use the sample regression line to obtain a prediction of y: y^ 5 20.1 1 3.01xr 5 20.1 1 13.012 132 5 29.13 Velocity
Residual
29 28 27 26 25 24
FIGURE 5.35 Plots for the transformed data of Example 5.16: (a) scatterplot of y versus x; (b) residual plot resulting from a linear fit to the transformed data.
23 22 1
2 Distance (a)
3
.8 .6 .4 .2 0 −.2 −.4 −.6 −.8
1
2 Distance (b)
3
In Example 5.16, transforming the x values using the square root function worked well. In general, how can we choose a transformation that will result in a linear pattern? Table 5.3 gives some guidance and summarizes some of the properties of the most commonly used transformations. Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.4
Nonlinear Relationships and Transformations
261
TA BLE 5. 3 Commonly Used Transformations Transformation
Mathematical Description
No transformation
y^ 5 a 1 bx
Square root of x
y^ 5 a 1 b!x
Log of x *
y^ 5 a 1 b log10 1x2 or y^ 5 a 1 b ln 1x2 1 y^ 5 a 1 ba b x
Reciprocal of x
Log of y * (Exponential growth or decay)
Try This Transformation When
log10 1 y^ 2 5 a 1 bx or ln 1 ^y2 5 a 1 bx
The change in y is constant as x changes. A 1-unit increase in x is associated with, on average, an increase of b in the value of y. The change in y is not constant. A 1-unit increase in x is associated with smaller increases or decreases in y for larger x values. The change in y is not constant. A 1-unit increase in x is associated with smaller increases or decreases in the value of y for larger x values. The change in y is not constant. A 1-unit increase in x is associated with smaller increases or decreases in the value of y for larger x values. In addition, y has a limiting value of a as x increases. The change in y is not constant. A 1-unit increase in x is associated with larger increases or decreases in the value of y for larger x values.
* The values of a and b in the regression equation will depend on whether log10 or ln is used, but the y^ ’s and r 2 values will be identical.
E X A M P L E 5 . 1 7 Loons on Acidic Lakes A study of factors that affect the survival of loon chicks is described in the paper
“Does Prey Biomass or Mercury Exposure Affect Loon Chick Survival in Wisconsin?” (The Journal of Wildlife Management [2005]: 57–67). In this study, a relationship between the pH of lake water and blood mercury level in loon chicks was observed. The researchers thought that this might be because the pH of the lake water might be related to the type of fish that the loons ate. The accompanying data (read from a graph in the paper and shown in Table 5.4) is x 5 lake pH and y 5 blood
T A B LE 5 .4 Data and Transformed Data from Example 5.17
Data set available online
Lake pH (x)
Blood Mercury Level (y)
Log(y)
5.28 5.69 5.56 5.51 4.90 5.02 5.02 5.04 5.30 5.33 5.64 5.83 5.83 6.17 6.22 6.15
1.10 0.76 0.74 0.60 0.48 0.43 0.29 0.09 0.10 0.20 0.28 0.17 0.18 0.55 0.43 0.40
0.0414 0.1192 0.1308 0.2218 0.3188 0.3665 0.5376 1.0458 1.0000 0.6990 0.5528 0.7696 0.7447 0.2596 0.3665 0.3979 (continued)
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
262
Chapter 5 Summarizing Bivariate Data
T A B L E 5 .4 Data and Transformed Data from Example 5.17—cont'd Lake pH (x)
Blood Mercury Level (y)
Log(y)
6.05 6.04 6.24 6.30 6.80 6.58 6.65 7.06 6.99 6.97 7.03 7.20 7.89 7.93 7.99 7.99 8.30 8.42 8.42 8.95 9.49
0.33 0.26 0.18 0.16 0.45 0.30 0.28 0.22 0.21 0.13 0.12 0.15 0.11 0.11 0.09 0.06 0.09 0.09 0.04 0.12 0.14
0.4815 0.5850 0.7447 0.7959 0.3468 0.5229 0.5528 0.6576 0.6778 0.8861 0.9208 0.8239 0.9586 0.9586 1.0458 1.2218 1.0458 1.0458 1.3979 0.9208 0.8539
mercury level (mg/g) for 37 loon chicks from different lakes in Wisconsin. A scatterplot is shown in Figure 5.36(a). The pattern in this scatterplot is typical of exponential decay, with the change in y as x increases much smaller for large x values than for small x values. You can see that a change of 1 in pH is associated with a much larger change in blood mercury level in the part of the plot where the x values are small than in the part of the plot where the x values are large. Table 5.3 suggests transforming the y values (blood mercury level in this example) by taking their logarithms. Two standard logarithmic functions are commonly used for such transformations— the common logarithm (log base 10, denoted by log or log10) and the natural logarithm (log base e, denoted ln). Either the common or the natural logarithm can be used; the only difference in the resulting scatterplots is the scale of the transformed y variable. This can be seen in Figures 5.36(b) and 5.36(c). These two scatterplots show the same pattern, and it looks like a line would be appropriate to describe this relationship. Table 5.4 displays the original data along with the transformed y values using y 5 log(y). The following Minitab output shows the result of fitting the least-squares line to the transformed data: Regression Analysis: Log(y) versus Lake pH The regression equation is Log(y) 0.458 0.172 Lake pH Predictor Coef SE Coef T P Constant 0.4582 0.2404 1.91 0.065 Lake Ph 0.17183 0.03589 4.79 0.000 S 0.263032 R-Sq 39.6% R-Sq(adj) 37.8%
(continued)
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.4
Nonlinear Relationships and Transformations
263
Analysis of Variance Source Regression Residual Error Total
DF 1 35 36
SS 1.5856 2.4215 4.0071
MS 1.5856 0.0692
F 22.92
P 0.000
1.2
Blood mercury level (y)
1.0 0.8 0.6 0.4 0.2 0.0 5
6
7 8 Lake pH (x)
9
10
(a) 0.0 −0.2
Log (y)
−0.4 −0.6 −0.8 −1.0 −1.2 −1.4 5
6
7 8 Lake pH (x)
9
10
9
10
(b) 0.0 −0.5
ln y
−1.0 −1.5 −2.0 −2.5 −3.0
FIGURE 5.36 Plots for the data of Example 5.17: (a) scatterplot of the loon data; (b) scatterplot of the transformed data with y log(y); (c) scatterplot of transformed data with y ln(y).
−3.5 5
6
7 8 Lake pH (x) (c)
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
264
Chapter 5
Summarizing Bivariate Data
The resulting regression equation is y 0.458 0.172x or, equivalently log(y) 0.458 0.172x
Fitting a Curve Using Transformations The objective of a regression analysis is usually to describe the approximate relationship between x and y with an equation of the form y some function of x. If we have transformed only x, fitting a least-squares line to the transformed data results in an equation of the desired form, for example, y^ 5 5 1 3xr 5 5 1 3 !x where xr 5 !x or 1 1 y^ 5 4 1 .2xr 5 4 1 .2 where xr 5 x x These functions specify lines when graphed using y and x, and they specify curves when graphed using y and x, as illustrated in Figure 5.37 for the square root transformation. 160 140 120
y
100 80 60 40 20 0 0
10
20 30 Transformed x = square root x
40
50
(a)
160 140 120
y
100 80 60 40 20 0
FIGURE 5.37
(a) A plot of y^ 5 5 1 3xr where xr 5 !x; (b) a plot of y^ 5 5 1 3!x.
500
1000
1500
2000
2500
x (b)
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.4
Nonlinear Relationships and Transformations
265
If the y values have been transformed, after obtaining the least-squares line the transformation can be undone to yield an expression of the form y 5 some function of x (as opposed to y 5 some function of x). For example, to reverse a logarithmic transformation 1 yr 5 log 1 y2 2 , we can take the antilogarithm of each side of the equation. To reverse a square root transformation 1 yr 5 !y 2 , we can square both sides of the equation, and to reverse a reciprocal transformation 1 yr 5 1 /y2 , we can take the reciprocal of each side of the equation. This is illustrated in Example 5.18.
E X A M P L E 5 . 1 8 Revisiting the Loon Data For the loon data of Example 5.17, yr 5 log 1 y2 and the least-squares line relating y and x was y 0.458 0.172x or, equivalently log 1 y2 5 0.458 2 0.172x To reverse this transformation, we take the antilog log of both sides of the equation: 10log1 y2 5 100.45820.172x Using properties of logs and exponents we know that 10log1 y2 5 y and 100.45820.172x 5 1100.4582 11020.172x2 Finally, we get y^ 5 1100.4582 11020.172x2 5 2.8708 11020.172x2 This equation can now be used to predict the y value (blood mercury level) for a given x (lake pH). For example, the predicted blood mercury level when lake pH is 6 is y^ 5 2.8708 11020.172x2 5 2.8708 11020.1721622 5 12.87082 10.09292 5 0.2667
It should be noted that the process of transforming data, fitting a line to the transformed data, and then undoing the transformation to get an equation for a curved relationship between x and y usually results in a curve that provides a reasonable fit to the sample data, but it is not the least-squares curve for the data. For example, in Example 5.18, a transformation was used to fit the curve y 5 1100.4582 11020.172x2 . However, there may be another equation of the form y^ 5 a 110 bx2 that has a smaller sum of squared residuals for the original data than the one we obtained using transformations. Finding the least-squares estimates for a and b in an equation of this form is complicated. Fortunately, the curves found using transformations usually provide reasonable predictions of y.
Power Transformations Frequently, an appropriate transformation is suggested by the data. One type of transformation that statisticians have found useful for straightening a plot is a power transformation. A power (exponent) is first selected, and each original value is raised to that power to obtain the corresponding transformed Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
266
Chapter 5 Summarizing Bivariate Data
value. Table 5.5 displays a “ladder” of the most frequently used power transformations. The power 1 corresponds to no transformation at all. Using the power 0 would transform every value to 1, which is certainly not informative, so statisticians use the logarithmic transformation in its place in the ladder of transformations. Other powers intermediate to or more extreme than those listed can be used, of course, but they are not used as frequently as those on the ladder. Notice that all the transformations previously presented are included in this ladder. Figure 5.38 is designed to suggest where on the ladder we should go to find an appropriate transformation. The four curved segments, labeled 1, 2, 3, and 4, represent shapes of curved scatterplots that are commonly encountered. Suppose that a scatterplot looks like the curve labeled 1. Then, to straighten the plot, we should use a power of x that is up the ladder from the no-transformation row (x 2 or x 3) and/or a power on y that is also up the ladder from the power 1. Thus, we might be led to squaring each x value, cubing each y, and plotting the transformed pairs. If the curvature looks like curved segment 2, a power up the ladder from no transformation for x and/or a power down the ladder for y (e.g., !y or log( y)) should be used.
T A B LE 5 .5 Power Transformation Ladder Power 3 2 1 1 2 1 3
0 1
Transformed Value
Name
(Original value)3 (Original value)2 (Original value) !Original value
Cube Square No transformation Square root
3 ! Original value Log(Original value) 1 Original value
Cube root Logarithm Reciprocal
The scatterplot for the loon data (Figure 5.36(a)) has the pattern of segment 3 in Figure 5.38. This suggests going down the ladder of transformations for x and/or for y. We found that transforming the y values in this data set by taking logarithms worked well, and this is consistent with the suggestion of going down the ladder of transformations for y. y up
y up
4
1 y
x down
x up
x
x down
FIGURE 5.38 Scatterplot shapes and where to go on the transformation ladder to straighten the plot.
x up
3
2
y down
y down
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.4
Nonlinear Relationships and Transformations
267
E X A M P L E 5 . 1 9 How Old Is That Lobster? Can you tell how old a lobster is by its size? This question was investigated by the authors of a paper that appeared in the Biological Bulletin (August 2007). Researchers measured carapace (the exterior shell) length (in mm) of 27 laboratory-raised lobsters of known age. The data on x 5 carapace length and y 5 age (in years) in Table 5.6 were read from a graph that appeared in the paper.
T AB LE 5 .6 Original and Transformed Data for Example 5.19
Data set available online
y
x
!y
x2
1.00 1.00 1.00 1.00 1.42 1.42 1.42 1.82 1.82 1.82 2.18 2.18 2.17 2.17 2.33 2.50 2.51 2.50 2.93 2.92 2.92 2.92 3.17 3.41 3.42 3.75 4.08
63.32 67.50 69.58 73.41 79.32 82.80 85.59 105.07 107.16 117.25 109.24 110.64 118.99 122.81 138.47 133.95 125.25 123.51 146.82 139.17 136.73 122.81 142.30 152.73 145.78 148.21 152.04
1.00 1.00 1.00 1.00 1.19 1.19 1.19 1.35 1.35 1.35 1.48 1.48 1.47 1.47 1.53 1.58 1.58 1.58 1.71 1.71 1.71 1.71 1.78 1.85 1.85 1.94 2.02
4,009.4 4,556.3 4,841.4 5,389.0 6,291.7 6,855.8 7,325.7 11,039.7 11,483.3 13,747.6 11,933.4 12,241.2 14,158.6 15,082.3 19,173.9 17,942.6 15,687.6 15,254.7 21,556.1 19,368.3 18,695.1 15,082.3 20,249.3 23,326.5 21,251.8 21,966.2 23,116.2
The scatterplot of the data in Figure 5.39 shows a clear curved pattern, which resembles the curved segment 2 in Figure 5.38. This suggests that we should consider transforming x using a power up the ladder (such as 1x 2 or x 32 ) or transforming y using a power down the ladder (such as !y or log( y)).
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
268
Chapter 5
Summarizing Bivariate Data
4.5 4.0
Age
3.5 3.0 2.5 2.0 1.5 1.0
FIGURE 5.39
50
Scatterplot of age versus carapace length.
75
100 Carapace length
125
150
Figure 5.40 shows scatterplots of !y versus x (Figure 5.40(a)), of y versus x 2 (Figure 5.40(b)), and of !y versus x 2 (Figure 5.40(c)). The relationship in the scatterplot of Figure 5.40(c) is more nearly linear than in the other two plots, so we can fit a line to the transformed data with yr 5 !y and xr 5 x 2.
4.5 4.0
1.8
3.5 1.6 Age
Square root (age)
2.0
1.4
3.0 2.5 2.0
1.2
1.5
1.0
1.0 50
75
100 Carapace length (a)
125
150
5000
10,000
15,000 20,000 Length squared (b)
25,000
Square root (age)
2.0 1.8 1.6 1.4 1.2 1.0
FIGURE 5.40 Scatterplots using transformed data from example 5.19: (a) !y versus x; (b) y versus x 2; (c) !y versus x 2.
5000
10,000
15,000 20,000 Length squared (c)
25,000
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.4
Nonlinear Relationships and Transformations
269
Using Minitab to fit a least-squares line to the transformed data results in the following output: Regression Analysis: Sqrt(Age) versus Length Squared The regression equation is Sqrt(Age) 0.829 0.000046 Length Squared Predictor Constant Length Squared S 0.0828744
Coef 0.82867 0.00004637 R-Sq 92.6%
SE Coef T 0.04022 20.60 0.00000261 17.75 R-Sq(adj) 92.4%
P 0.000 0.000
The least-squares line is y 0.829 0.000046x or equivalently !y 0.829 0.000046x 2 This transformation can be reversed by squaring both sides to obtain an equation of the form y 5 some function of x: 1 yr2 2 5 10.829 1 0.000046xr2 2 Since 1 yr2 2 5 y and xr 5 x 2, we get y^ 5 10.829 1 0.000046x 22 2 To predict the age of a lobster with a carapace length of 100 mm, we substitute 100 into the equation to obtain a predicted value y^ : y^ 5 10.829 1 0.000046x 22 2 5 10.829 1 0.000046 11002 22 2 5 11.2892 2 5 1.66 years
E X E RC I S E S 5 . 4 4 - 5 . 5 5 Example 5.5 described a study of the effectiveness of pomegranate fruit extract (PFE) in slowing the growth of prostate cancer tumors (Proceedings of the
5.44
National Academy of Sciences [October 11, 2005]: 14813–14818]. Figure 5.11 from that example is repro-
duced here. Based on this figure, we noted that for the .2% PFE group, the relationship between average tumor volume and number of days after injection of cancer cells appears curved rather than linear. Variable Water average tumor size .2% PFE average tumor size
800
Average tumor size
700 600 500 400 300 200 100 0 10
Bold exercises answered in back
15
20 25 30 35 Days after injection
Data set available online
40
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
270
Chapter 5 Summarizing Bivariate Data
a. One transformation that might result in a relationship that is more nearly linear is log( y). The values of x 5 number of days since injection of cancer cells, y 5 average tumor volume (in mm3), and y 5 log(y) for the 0.2% PFE group are given in the accompanying table. Construct a scatterplot of y versus x. Does this scatterplot look more nearly linear than the plot of the original data? Days After Injection x
0.2% PFE Average Tumor Size y
Log(y)
11 15 19 23 27 31 35 39
40 75 90 210 230 330 450 600
1.60 1.88 1.95 2.32 2.36 2.52 2.65 2.78
5.45 Example 5.15 described a study that involved substituting sunflower meal for a portion of the usual diet of farm-raised sea breams (Aquaculture [2007]: 528–534). This paper also gave data on y 5 feed intake (in grams per 100 grams of fish per day) and x 5 percentage sunflower meal in the diet (read from a graph in the paper). x y
b. Based on the accompanying Minitab output, does the least-squares line effectively summarize the relationship between y and x? The regression equation is log(average tumor size) 1.23 0.0413 Days After Injection Predictor Coef Constant 1.22625 Days After Injection 0.041250 S 0.0777817 R-Sq 96.9%
SE Coef T 0.07989 15.35 0.003000 13.75 R-Sq(adj) 96.4%
d. Residual plots of the residuals from the least-squares line using the untransformed data and using y and x are shown at the bottom of the page in Figure EX5.44. What feature of the residual plot for the original data suggests that the least-squares line is not the best way to describe the relationship between y and x? In what way does the residual plot for the transformed data suggest that a line is a reasonable description of the relationship for the transformed data (y and x)?
P 0.000 0.000
0 0.86
12 0.82
18 0.86
24 0.87
30 1.00
36 1.09
A scatterplot of these data is curved and the pattern in the plot resembles a quadratic curve. a. Using a statistical software package or a graphing calculator, find the equation of the least-squares quadratic curve that can be used to describe the relationship between percentage sunflower meal and feed intake. y^ 5 0.866 2 0.008x 1 0.0004x 2 b. Use the least-squares equation from Part (a) to predict feed intake for fish fed a diet that included 20% sunflower meal. The paper “Commercially Available Plant
5.46
c. Use the Minitab output to predict average tumor size 30 days after injection of cancer cells for a mouse that received water supplemented with 0.2% PFE.
6 0.84
Growth Regulators and Promoters Modify Bulk Tissue Abscisic Acid Concentrations in Spring Barley, but not Root Growth and Yield Response to Drought” (ApResidual Plot Using Transformed Data
Residual Plot for Original Data 0.15 75 0.10 Residuals
Residuals
50 25 0 −25
0.05 0.00 −0.05
−50
−0.10 10
15
20 25 30 Days after injection
35
40
10
15
20 25 30 Days after injection
35
40
FIGURE EX5.44 Bold exercises answered in back
Data set available online
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.4
plied Biology [2006]: 291–304) describes a study of the drought response of barley. The accompanying data on x days after sowing and y soil moisture deficit (in mm) was read from a graph that appeared in the paper. Days After Sowing
Soil Moisture Deficit
37 63 68 75 82 98 104 111 132
0.00 69.36 79.15 85.11 93.19 104.26 108.94 112.34 115.74
a. Construct a scatterplot of y soil moisture deficit versus x days after sowing. Does the relationship between these two variables appear to be linear or nonlinear? b. Fit a least-squares line to the given data and construct a residual plot. Does the residual plot support your conclusion in Part (a)? Explain. c. Consider transforming the data by leaving y un1 changed and using either xr 5 !x or xs 5 . x Which of these transformations would you recommend? Justify your choice using appropriate graphical displays. d. Using the transformation you recommend in Part (c), find the equation of the least-squares line that describes the relationship between y and the transformed x. y^ 5 166.490 2 6109.479 11x 2 e. What would you predict for soil moisture deficit 50 days after sowing? For 100 days after sowing? f. Explain why it would not be reasonable to predict soil moisture deficit 200 days after sowing. Is electromagnetic radiation from phone antennae associated with declining bird populations? This is one of the questions addressed by the authors of the paper
5.47
“The Urban Decline of the House Sparrow (Passer domesticus): A Possible Link with Electromagnetic Radiation” (Electromagnetic Biology and Medicine [2007]: 141–151). The accompanying data on x electromagnetic field strength (V/m) and y sparrow density (birds/hectare) was read from a graph that appeared in the paper.
Bold exercises answered in back
Data set available online
Nonlinear Relationships and Transformations
Field Strength
Sparrow Density
0.11 0.20 0.29 0.40 0.50 0.61 1.01 1.10 0.70 0.80 0.90 1.20 1.30 1.41 1.50 1.80 1.90 3.01 3.10 3.41
41.71 33.60 24.74 19.50 19.42 18.74 24.23 22.04 16.29 14.69 16.29 16.97 12.83 13.17 4.64 2.11 0.00 0.00 14.69 0.00
271
a. Construct a scatterplot of y sparrow density versus x field strength. Does the relationship between these two variables appear to be linear or nonlinear? b. Consider transforming the data by leaving y unchanged and using either xr 5 !x or xs 5 log 1x2 . The log(x) values given in the same order as the x values in the table above are 0.96 0.70 0.54 0.40 0.30 0.21 0.00 0.04 0.15 0.10 0.05 0.08 0.11 0.15 0.18 0.26 0.28 0.48 0.49 0.53 Which of these transformations would you recommend? Justify your choice using appropriate graphical displays. c. Using the transformation you recommend in Part (b), find the equation of the least-squares line that describes the relationship between y and the transformed x. y^ 5 14.805 2 24.280 # log 1x2 d. What would you predict for sparrow density if the field strength is 0.5? What would you predict for sparrow density if the field strength is 2.5? The paper “Effects of Age and Gender on Physical Performance” (Age [2007]: 77–85) describes a
5.48
study of the relationship between age and 1-hour swimming performance. Data on age and swim distance for over 10,000 men participating in a national long-distance 1-hour swimming competition are summarized in Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
272
Chapter 5 Summarizing Bivariate Data
The paper “Developmental and Individual Differences in Pure Numerical Estimation” (Developmental Psychology [2006]: 189–201) describes a study
the accompanying table. In Exercise 5.34 from Section 5.3, a plot of the residuals from the least-squares line showed a curved pattern that suggested that a quadratic curve would do a better job of summarizing the relationship between x representative age and y average swim distance. Find the equation of the least-squares quadratic curve and use it to predict average swim distance at 40 years of age. y^ 5 3843.027 1 10.619x 2 0.360x 2
5.50
of how young children develop the ability to estimate lengths. Children were shown a piece of paper with two lines. One line was a short line labeled as having length 1 zip. The second line was a much longer line labeled as having length 1000 zips. The child was then asked to draw a line that had a length of a specified number of zips, such as 438 zips. The data in the accompanying table gives the length requested and the average of the actual lengths of the lines drawn by 30 second graders.
y^ 5 3691.293
Age Group
Representative Age (Midpoint of Age Group)
Average Swim Distance (meters)
20–29 30–39 40–49 50–59 60–69 70–79 80–89
25 35 45 55 65 75 85
3913.5 3728.8 3579.4 3361.9 3000.1 2649.0 2118.4
5.49 The paper “Feeding Ecology of the Great Basin Rattlesnake” (Canadian Journal of Zoology [2008]: 723–734) investigated whether there was a relationship between the size of a rattlesnake and the size of the unsuspecting rodent the snake would hunt. The authors collected data on x snake body mass (in grams) and prey body mass (in grams) for 22 snakes and their prey. Because the relationship between snake mass and prey mass was not linear, the authors transformed both x and y by taking natural logarithms. The transformed data was used to construct the following scatterplot, which shows a linear pattern. If a scatterplot had been constructed using the untransformed data, which of the four curved segments in Figure 5.38 do you think it would most resemble? Explain your choice.
Second Grade Average Length Drawn
3 7 19 52 103 158 240 297 346 391 438 475 502 586 613 690 721 760 835 874 907 962
37.15 92.88 207.43 272.45 458.20 442.72 371.52 467.49 487.62 530.96 482.97 544.89 515.48 595.98 575.85 605.26 637.77 674.92 701.24 662.54 758.51 749.23
a. Construct a scatterplot of y second grade average length drawn versus x requested length. b. Based on the scatterplot in Part (a), would you suggest using a line, a quadratic curve, or a cubic curve to describe the relationship between x and y? Explain your choice. c. Using a statistical software package or a graphing calculator, fit a cubic curve to this data and use it to predict average length drawn for a requested length of 500 zips.
4.5 4.0 3.5 ln(prey mass)
Requested Length
3.0 2.5 2.0 1.5
y^ 5 138.471 1 19276x 2 0.0032x 2 1 0.000002x y^ 5 555.219
1.0 3
4
Bold exercises answered in back
5 ln(snake mass)
6
7
Data set available online
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.4
Researchers have examined a number of climatic variables in an attempt to understand the mechanisms that govern rainfall runoff. The paper “The Applicability of
Nonlinear Relationships and Transformations
5.51
Age
Fatality Rate
Age
Fatality Rate
Morton’s and Penman’s Evapotranspiration Estimates in Rainfall-Runoff Modeling” (Water Resources Bulletin [1991]: 611–620) reported on a study that examined the
40 45 50 55 60 65
0.75 0.75 0.95 1.05 1.15 1.20
70 75 80 85 90
1.30 1.65 2.20 3.00 3.20
relationship between x cloud cover index and y sunshine index. The cloud cover index can have values between 0 and 1. The accompanying data are consistent with summary quantities in the article. The authors of the article used a cubic regression to describe the relationship between cloud cover and sunshine. Cloud Cover Index (x)
Sunshine Index (y)
0.2 0.5 0.3 0.1 0.2 0.4 0.0 0.4 0.3
10.98 10.94 10.91 10.94 10.97 10.89 10.88 10.92 10.86
a. Construct a scatterplot of the data. What characteristics of the plot suggest that a cubic regression would be more appropriate for summarizing the relationship between sunshine index and cloud cover index than a linear or quadratic regression? b. Find the equation of the least-squares cubic function. y^ 5 10.8768 1 1.4604x 2 7.259x 2 1 9.2342x 3 c. Construct a residual plot by plotting the residuals from the cubic regression model versus x. Are there any troubling patterns in the residual plot that suggest that a cubic regression is not an appropriate way to summarize the relationship? d. Use the cubic regression to predict sunshine index when the cloud cover index is 0.25. e. Use the cubic regression to predict sunshine index when the cloud cover index is 0.45. f. Explain why it would not be a good idea to use the cubic regression equation to predict sunshine index for a cloud cover index of 0.75.
5.52 The report “Older Driver Involvement in Injury Crashes in Texas” (Texas Transportation Institute, 2004) included a scatterplot of y fatality rate
(percentage of drivers killed in injury crashes) versus x driver age. The accompanying data are approximate values read from the scatterplot.
Bold exercises answered in back
Data set available online
273
a. Construct a scatterplot of these data. b. Using Table 5.5 and the ladder of transformations in Figure 5.38, suggest a transformation that might result in variables for which the scatterplot would exhibit a pattern that was more nearly linear. c. Reexpress x and/or y using the transformation you recommended in Part (b). Construct a scatterplot of the transformed data. d. Does the scatterplot in Part (c) suggest that the transformation was successful in straightening the plot? e. Using the transformed variables, fit the least-squares line and use it to predict the fatality rate for 78-yearold drivers. 1.893
5.53 The article “Organ Transplant Demand Rises Five Times as Fast as Existing Supply” (San Luis Obispo Tribune, February 23, 2001) included a graph that showed the number of people waiting for organ transplants each year from 1990 to 1999. The following data are approximate values and were read from the graph in the article: Year
Number Waiting for Transplant (in thousands)
1 (1990) 2 3 4 5 6 7 8 9 10 (1999)
22 25 29 33 38 44 50 57 64 72
a. Construct a scatterplot of the data with y number waiting for transplant and x year. Describe how the number of people waiting for transplants has changed over time from 1990 to 1999.
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
274
Chapter 5 Summarizing Bivariate Data
b. The scatterplot in Part (a) is shaped like segment 2 in Figure 5.38. Find a transformation of x and/or y that straightens the plot. Construct a scatterplot for your transformed variables. c. Using the transformed variables from Part (b), fit a least-squares line and use it to predict the number waiting for an organ transplant in 2000 (Year 11). d. The prediction made in Part (c) involves prediction for an x value that is outside the range of the x values in the sample. What assumption must you be willing to make for this to be reasonable? Do you think this assumption is reasonable in this case? Would your answer be the same if the prediction had been for the year 2010 rather than 2000? Explain.
5.54 Penicillin was administered orally to five different horses, and the concentration of penicillin in the blood was determined after five different lengths of time. The following data on x 5 elapsed time (in hours) and y 5 penicillin concentration (in mg/ml) appeared in the paper “Absorption and Distribution Patterns of Oral Phenoxymethyl Penicillin in the Horse” (Cornell Veterinarian [1983]: 314–323): x y
1 1.8
2 1.0
3 0.5
6 0.1
8 0.1
Construct scatterplots using the following variables. Which transformation, if any, would you recommend? a. x and y c. x and !y b. !x and y d. !x and !y Bold exercises answered in back
5.5
Data set available online
e. x and log(y) (values of log( y) are 0.26, 0, 0.30, 1, and 1). Determining the age of an animal can sometimes be a difficult task. One method of estimating the age of harp seals is based on the width of the pulp canal in the seal’s canine teeth. To investigate the relationship between age and the width of the pulp canal, researchers measured age and canal width in seals of known age. The following data on x age (in years) and y canal length (in millimeters) are a portion of a larger data set that appeared in the paper “Validation of Age Estimation in
5.55
the Harp Seal Using Dentinal Annuli” (Canadian Journal of Fisheries and Aquatic Science [1983]: 1430–1441): x y
0.25 0.25 0.50 0.50 0.50 0.75 0.75 1.00 700 675 525 500 400 350 300 300
x y
1.00 1.00 1.00 1.00 1.25 1.25 1.50 1.50 250 230 150 100 200 100 100 125
x y
2.00 2.00 2.50 2.75 3.00 4.00 4.00 5.00 60 140 60 50 10 10 10 10
x y
5.00 5.00 5.00 6.00 6.00 15 10 10 15 10
Construct a scatterplot for this data set. Would you describe the relationship between age and canal length as linear? If not, suggest a transformation that might straighten the plot. Video Solution available
Logistic Regression (Optional) The correlation and regression techniques we have seen up to this point require that both variables of interest be numerical. But what if the dependent variable in a study is not numerical? This situation requires a different approach. For a dependent variable that is categorical with just two possible values (a binary variable), logistic regression can be used to describe the way in which such a dependent variable is related to a numerical predictor variable.
E X A M P L E 5 . 2 0 Look Out for Those Wolf Spiders The paper “Sexual Cannibalism and Mate Choice Decisions in Wolf Spiders: Influence of Male Size and Secondary Sexual Characteristics” (Animal Behaviour [2005]: 83–94) described a study in which researchers were interested in what variData set available online
ables might be related to a female wolf spider’s decision to kill and consume her part-
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.5
Logistic Regression (Optional)
275
ner during courtship or mating. The accompanying data (approximate values read from a graph in the paper) are values of x 5 difference in body width (female – male) and y 5 cannibalism, coded as 0 for no cannibalism and 1 for cannibalism for 52 pairs of courting wolf spiders.
Size Difference (mm)
Cannibalism
1 1 0.8 0.8 0.6 0.6 0.4 0.4 0.4 0.4 0.2 0.2 0.2 0.2 0.0 0.0 0.0 0.0 0.0 0.0 0.2 0.2 0.2 0.2 0.2 0.2
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
Size Difference (mm) 0.4 0.4 0.4 0.4 0.4 0.6 0.6 0.6 0.6 0.6 0.6 0.6 0.8 0.8 0.8 0.8 0.8 1.0 1.0 1.0 1.0 1.2 1.4 1.6 1.8 2.0
Cannibalism 0 0 0 0 1 0 0 0 0 0 1 1 0 0 1 1 1 0 0 1 1 0 0 1 1 1
A Minitab scatterplot of the data is shown in Figure 5.41. Note that the plot was constructed so that if two points fell in exactly the same position, one was offset a bit so that all observations would be visible. (This is called jittering.)
1.0
Cannibalism
0.8 0.6 0.4 0.2 0.0
FIGURE 5.41 Scatterplot of the wolf spider data.
−1.0 −0.5
0.0
0.5 1.0 Size difference
1.5
2.0
2.5
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
276
Chapter 5 Summarizing Bivariate Data
The scatterplot doesn’t look like others we have seen before—its odd appearance is due to the fact that all y values are either 0 or 1. But, we can see from the plot that there are more occurrences of cannibalism for large x values (where the female is bigger than the male) than for smaller x values. In this situation, it makes sense to consider the probability of cannibalism (or equivalently, the proportion of the time cannibalism would occur) as being related to size difference. For example, we might focus on a single x value, say x 5 0 where the female and male are the same size. Based on the data at hand, what can we say about the cannibalism proportion for pairs where the size difference is 0? We will return to this question after introducing the logistic regression equation.
A logistic regression equation is used to describe how the probability of “success” (for example, cannibalism in the wolf spider example) changes as a numerical predictor variable, x, changes.
With p denoting the probability of success, the logistic regression function is p5
ea1bx 1 1 ea1bx
where a and b are constants.
The logistic regression equation looks complicated, but it has some very convenient properties. For any x value, the value of e a1bx / 11 1 e a1bx 2 is between 0 and 1. As x changes, the graph of this equation has an “S” shape. Consider the two S-shaped curves of Figure 5.42. The blue curve starts near 0 and increases to 1 as x increases. This is the type of behavior exhibited by p 5 e a1bx / 11 1 e a1bx 2 when b . 0. The red curve starts near 1 for small x values and then decreases as x increases. This happens when b , 0 in the logistic regression equation. The steepness of the curve—how quickly it rises or falls—also depends on the value of b. The farther b is from 0, the steeper the curve. Most statistics packages, such as Minitab and SPSS, have the capability of using sample data to compute values for a and b in the logistic regression equation to produce an equation relating the probability of success to the predictor x. An explanation of an alternate method for computing reasonable values of a and b is given later in this section.
1.0 0.8 b<0
0.6 b>0 0.4 0.2
FIGURE 5.42 Two logistic regression curves.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.5
Logistic Regression (Optional)
277
E X A M P L E 5 . 2 1 Cannibal Spiders II Minitab was used to fit a logistic regression equation to the wolf spider data of Example 5.20. The resulting Minitab output is given in Figure 5.43, and Figure 5.44 shows a scatterplot of the original data with the logistic regression curve superimposed. Response Information Variable Cannibalism
Value 1 0 Total
Count 11 41 52
(Event)
Logistic Regression Table
FIGURE 5.43 Minitab output for the data of Example 5.21.
Predictor Constant Size difference
Coef −3.08904 3.06928
SE Coef 0.828780 1.00407
Z −3.73 3.06
P 0.000 0.002
Odds Ratio 21.53
95% CI Lower Upper 3.01
154.05
With a 5 23.08904 and b 5 3.06928, the equation of the logistic regression function is p5
e23.0890413.06928x 1 1 e23.0890413.06928x
To predict or estimate the probability of cannibalism when the size difference between the female and male 0, we substitute 0 into the logistic regression equation to obtain p5
e23.0890413.06928102 e23.08904 5 5 .044 1 1 e23.08904 1 1 e23.0890413.06928102
The probabilities of cannibalism for other values of x 5 size difference can be computed in a similar manner.
1.0
Variable Cannibalism Logistic model
Cannibalism
0.8 0.6 0.4 0.2
FIGURE 5.44 Scatterplot and logistic regression curve for data of Example 5.21.
0.0 −1.0
−0.5
0.0 0.5 1.0 Size difference
1.5
2.0
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
278
Chapter 5 Summarizing Bivariate Data
Relating Logistic Regression to Data Transformation Consider an important question in drug development—what strength of dose of a drug is needed to elicit a response? For example, suppose that we are marketing a poison, RatRiddance, to be used to eradicate rats. We want to use enough of the toxic agent to dispose of the little critters, but for safety and ecological reasons we don’t want to use more poison than necessary. Imagine that an experiment is conducted to assess the toxicity of RatRiddance, where the amount of the active ingredient is varied. Eleven different concentrations are tested, with about 500 rats in each treatment. The results of the experiment are given in Table 5.7. A plot of the data is shown in Figure 5.45.
TABLE 5 . 7 Mortality Data for RatRiddance Concentration Number Exposed Mortality Rate
20
40
60
80
100
120
140
160
180
200
240
440
462
500
467
515
561
469
550
542
479
497
.225
.236
.398
.628
.678
.795
.853
.860
.921
.940
.968
The original data consisted of about 5000 observations; for each individual rat there was a (dose, response) pair, where the response was categorical—survived or did not survive. The data were then summarized in Table 5.7 by computing the proportion that did not survive (the mortality rate) for each dose. It is these proportions that were plotted in the scatterplot of Figure 5.45 and that exhibit the typical “S” shape of the logistic regression equation. Let’s use the logistic regression equation to describe the relationship between the proportion of rats who did not survive (mortality rate) and dose. The model is then p5
e a1bx 1 1 e a1bx
Mortality rate 1.0
0.8
0.6
0.4
0.2
FIGURE 5.45 Scatterplot of mortality versus dose.
20
40
60
80
100 120 140 160 180 200 220 240 260 Concentration
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.5
Logistic Regression (Optional)
279
For data that has been converted into proportions, some tedious but straightforward algebra demonstrates how we can use a transformation to calculate values for a and b in the logistic regression equation: p5
e a1bx 1 1 e a1bx
p 11 1 e a1bx 2 5 e a1bx
complete the multiplication on the left hand side of the equation
p 1 pe a1bx 5 e a1bx
subtract pea1bx from each side
p 5 e a1bx 2 pe a1bx
factor out ea1bx in the right hand side of the equation
p 5 e a1bx 11 2 p2
divide both sides of the equation by 11 2 p2
p 5 e a1bx 12p lna
multiply both sides by 1 1 ea1bx
take the natural log of both sides
p b 5 a 1 bx 12p
This means that if the logistic regression curve is a reasonable way to describe p the relationship between p and x, the relationship between lna b and x is linear. 12p A consequence of this is that if we transform p using yr 5 lna
p b 12p
we can use least squares to fit a line to the (x, y) data. For the RatRiddance example, the transformed data are
x
p
p 12p
20 40 60 80 100 120 140 160 180 200
0.225 0.236 0.398 0.628 0.678 0.795 0.853 0.860 0.921 0.940
0.290 0.309 0.661 1.688 2.106 3.878 5.803 6.143 11.658 15.667
yr 5 lna
p b 12p
1.237 1.175 0.414 0.524 0.745 1.355 1.758 1.815 2.456 2.752
The resulting best fit line is yr 5 a 1 bx 5 21.6033 1 0.221x We can check the transformed linear model fit in the customary way, checking the scatterplot and the residual plot, as shown in Figure 5.46(a) and (b). Although there
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
280
Chapter 5 Summarizing Bivariate Data
seems to be an ever-so-slight hint of curvature in the data, the linear model appears to fit quite well. In fact, the linear fit accounts for about 97% of the variation in yr. 4
0.5
3 0.25 Residuals
ln(p/(1−p))
2 1
0 −0.25 −1 −2
50
FIGURE 5.46
100 150 Concentration
200
250
−0.5
50
100 150 Concentration
(a)
200
250
(b)
Scatterplot (a) and residual plot (b) for the transformed mortality data.
Alan and Sandy Carey/Photodisc/ Getty Images
EXAMPLE 5.22
The Call of the Wild Amazonian . . . Frog
The Amazonian tree frog uses vocal communication to call for a mate. In a study of the relationship between calling behavior and the amount of rainfall (“How,
When, and Where to Perform Visual Displays: The Case of the Amazonian Frog Hyla parviceps,” Herpetologica [2004]: 420–429), the daily rainfall (in mm) was recorded as well as observations of calling behavior by male Amazonian frogs. Calling behavior was used to compute the call rate, which is the proportion of frogs exhibiting calling behavior. Data consistent with the article are given in Table 5.8.
T A B L E 5 .8 Proportion Calling versus Daily Rainfall (mm) Rainfall Call rate Rainfall Call rate Rainfall Call rate Rainfall Call rate
0.2 .17
0.3 .19
0.4 .20
0.5 .21
0.7 .27
0.8 .28
0.9 .29
1.1 .34
1.2 .39
1.3 .41
1.5 .46
1.6 .49
1.7 .53
2.0 .60
2.2 .67
2.4 .71
2.6 .75
2.8 .82
2.9 .84
3.2 .88
3.5 .90
4.2 .97
4.9 .98
5.0 .98
Inspection of the scatterplot in Figure 5.47(a) reveals a pattern that is consistent with a logistic relationship between the daily rainfall and the proportion of frogs exhibiting calling behavior. The transformed data in Figure 5.47(b) show a clearly linear pattern. For these data the least-squares line is given by the equation yr 5 21.871 1 1.177(Rainfall). To predict calling proportion for a location with daily rainfall of 4.0 mm, we use the computed values of a and b in the logistic regression equation: e 21.87111.177x e21.87111.17714.02 p5 5 5 .945 1 1 e 21.87111.177x 1 1 e21.87111.17714.02 Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
1.0
4.0
0.8
3.0 ln(p/(1−p))
Calling rate (p)
5.5
0.6 0.4
281
Logistic Regression (Optional)
2.0 1.0 0.0
0.2
−1.0 −2.0
0.0 1.0
2.0
3.0 4.0 Daily rain
5.0
6.0
1.0
2.0
3.0 4.0 Daily rain
5.0
6.0
(a) Linear Fit ln(p/(1−p)) = −1.871 + 1.177 Daily Rain Summary of Fit 0.996 0.996 0.103
RSquare RSquare Adj s
FIGURE 5.47 Scatterplot of original and transformed data of Example 5.22.
(b)
E X E RC I S E S 5 . 5 6 - 5 . 6 2 Anabolic steroid abuse has been increasing despite increased press reports of adverse medical and psychiatric consequences. In a recent study, medical researchers studied the potential for addiction to testosterone in hamsters (Neuroscience [2004]: 971–981). Hamsters were allowed to self-administer testosterone over a period of days, resulting in the death of some of the animals. The data below show the proportion of hamsters surviving versus the peak self-administration of testosterone (mg). Fit a logistic regression equation and use the equation to predict the probability of survival for a hamster with a peak intake of 40mg.
5.56
lna
Survival Peak Intake Proportion (micrograms) (p) 10 30 50 70 90
0.980 0.900 0.880 0.500 0.170
Bold exercises answered in back
p 12p
b 5 4.589 2 0.0659x;
p 12p 49.0000 9.0000 7.3333 1.0000 0.2048
yr 5 lna
p b 12p
3.8918 2.1972 1.9924 0.0000 1.5856
Data set available online
Does high school GPA predict success in firstyear college English? The proportion with a grade of C or better in freshman English for students with various high school GPAs for freshmen at Cal Poly, San Luis Obispo, in fall of 2007 is summarized in the accompanying table. Fit a logistic regression equation that would allow you to predict the probability of passing freshman English based on high school GPA. Use the resulting equation to predict the probability of passing freshman English for students with a high school GPA of 2.2.
5.57
High School GPA
Proportion C or Better
p 12p
3.36 2.94 2.68 2.49 2.33 2.19 2.06
0.95 0.90 0.85 0.80 0.75 0.70 0.65
19.00 9.00 5.67 4.00 3.00 2.33 1.86
y9 5 lna
p b 12p
2.94 2.20 1.73 1.39 1.10 0.85 0.62 (continued)
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
282
Chapter 5 Summarizing Bivariate Data
High School GPA
Proportion C or Better
p 12p
p y9 5 lna b 12p
1.94 1.83 1.72 1.61 1.49 1.38 1.25 1.11 0.95 0.75 0.05 0.08
0.60 0.55 0.50 0.45 0.40 0.34 0.30 0.25 0.20 0.15 0.10 0.05
1.50 1.22 1.00 0.82 0.67 0.52 0.43 0.33 0.25 0.18 0.11 0.05
0.41 0.20 0.00 0.20 0.41 0.66 0.85 1.10 1.39 1.73 2.20 2.94
Persistently Aphid-Borne Viruses in Lupin Stands” (Annals of Applied Biology [2005]: 337–350) was used to fit a least-squares regression line to describe the relationship p between x and yr 5 lna b. Minitab output re12p sulting from fitting the least-squares line is given below. The regression equation is ln(p/(1-p)) 20.917 2 0.107 Distance to Crop Edge P 0.000 0.000
a. What is the logistic regression function relating x and the proportion of plants with virus symptoms? b. What would you predict for the proportion of plants with virus symptoms at a distance of 15 meters from the edge of the field? (Note: the x values in the data set ranged from 0 to 20.) The paper “The Shelf Life of Bird Eggs: Test-
ing Egg Viability Using a Tropical Climate Gradient” (Ecology [2005]: 2164–2175) investigated the effect of altitude and length of exposure on the hatch rate of thrasher eggs. Data consistent with the estimated probabilities of hatching after a number of days of exposure given in the paper are shown here.
Bold exercises answered in back
1
2
3
4
5
Proportion 0.75 0.67 0.36 0.31 0.14 (cloud forest)
to spread from the edges of a field inward. The data on x 5 distance from the edge of the field (in meters) and y 5 proportion of plants with virus symptoms that appeared in the paper “Patterns of Spread of Two Non-
5.59
Exposure (days)
6
7
8
Proportion 0.81 0.83 0.68 0.42 0.13 0.07 0.04 0.02 (lowland) Proportion 0.73 0.49 0.24 0.14 0.037 0.040 0.024 0.030 (mid-elevation)
5.58 Some plant viruses are spread by insects and tend
Predictor Coef SE Coef T Constant 0.9171 0.1249 7.34 Distance to Crop Edge 0.10716 0.01062 10.09 S 0.387646 R-Sq 72.8% R-Sq(adj) 72.1%
Probability of Hatching
Data set available online
0.09
0.06
0.07
a. Plot the data for the low- and mid-elevation experimental treatments versus exposure. Are the plots generally the shape you would expect from “logistic” plots? b. Using the techniques introduced in this section, p calculate yr 5 lna b for each of the exposure 12p times in the cloud forest and fit the line yr 5 a 1 b(Days). What is the significance of a negative slope to this line? c. Using your best-fit line from Part (b), what would you estimate the proportion of eggs that would, on average, hatch if they were exposed to cloud forest conditions for 3 days? 5 days? d. At what point in time does the estimated proportion of hatching for cloud forest conditions seem to cross from greater than 0.5 to less than 0.5?
5.60 As part of a study of the effects of timber management strategies (Ecological Applications [2003]: 1110– 1123) investigators used satellite imagery to study abundance of the lichen Lobaria oregano at different elevations. Abundance of a species was classified as “common” if there were more than 10 individuals in a plot of land. In the table below, approximate proportions of plots in which Lobaria oregano were common are given. Proportions of Plots Where Lobaria oregano Are Common Elevation (m) Prop. of plots with lichen common
400 600 800 1000 1200 1400 1600 0.99 0.96 0.75 0.29 0.077 0.035 0.01
a. As elevation increases, does the proportion of plots for which lichen is common become larger or smaller? What aspect(s) of the table support your answer? Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.6 Interpreting and Communicating the Results of Statistical Analyses
b. Using the techniques introduced in this section, calp culate yr 5 lna b for each of the elevations 12p and fit the line yr 5 a 1 b(Elevation). What is the equation of the best-fit line? c. Using the best-fit line from Part (b), estimate the proportion of plots of land on which Lobaria oregano are classified as “common” at an elevation of 900 m.
5.61 The hypothetical data below are from a toxicity study designed to measure the effectiveness of different doses of a pesticide on mosquitoes. The table below summarizes the concentration of the pesticide, the sample sizes, and the number of critters dispatched. Concentration (g/cc) Number of mosquitoes Number killed
0.10
0.15
0.20
0.30
0.50
0.70
0.95
48
52
56
51
47
53
51
10
13
25
31
39
51
49
a. Make a scatterplot of the proportions of mosquitoes killed versus the pesticide concentration. b. Using the techniques introduced in this section, calcup late yr 5 lna b for each of the concentrations 12p and fit the line yr 5 a 1 b(Concentration). What is the significance of a positive slope for this line? c. The point at which the dose kills 50% of the pests is sometimes called LD50, for “Lethal dose 50%.” What would you estimate to be LD50 for this pesticide when used on mosquitoes? Bold exercises answered in back
5.6
Data set available online
283
In the study of textiles and fabrics, the strength of a fabric is a very important consideration. Suppose that a significant number of swatches of a certain fabric are subjected to different “loads” or forces applied to the fabric. The data from such an experiment might look as follows:
5.62
Hypothetical Data on Fabric Strength Load (lb/sq in.) Proportion failing
5
15
35
50
70
80
90
0.02
0.04
0.20
0.23
0.32
0.34
0.43
a. Make a scatterplot of the proportion failing versus the load on the fabric. b. Using the techniques introduced in this section, calcup late yr 5 lna b for each of the loads and fit 12p the line yr 5 a 1 b(Load). What is the significance of a positive slope for this line? c. What proportion of the time would you estimate this fabric would fail if a load of 60 lb/sq in. were applied? d. In order to avoid a “wardrobe malfunction,” one would like to use fabric that has less than a 5% chance of failing. Suppose that this fabric is our choice for a new shirt. To have less than a 5% chance of failing, what would you estimate to be the maximum “safe” load in lb/sq in.?
Video Solution available
Interpreting and Communicating the Results of Statistical Analyses Using either a least-squares line to summarize a linear relationship or a correlation coefficient to describe the strength of a linear relationship is common in investigations that focus on more than a single variable. In fact, the methods described in this chapter are among the most widely used of all statistical tools. When numerical bivariate data are analyzed in journal articles and other published sources, it is common to find a scatterplot of the data and a least-squares line or a correlation coefficient.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
284
Chapter 5 Summarizing Bivariate Data
Communicating the Results of Statistical Analyses When reporting the results of a data analysis involving bivariate numerical data, it is important to include graphical displays as well as numerical summaries. Including a scatterplot and providing a description of what the plot reveals about the form of the relationship between the two variables under study establish the context in which numerical summary measures, such as the correlation coefficient or the equation of the least-squares line, can be interpreted. In general, the goal of an analysis of bivariate data is to give a quantitative description of the relationship, if any, between the two variables. If there is a relationship, you can describe how strong or weak the relationship is or model the relationship in a way that allows various conclusions to be drawn. If the goal of the study is to describe the strength of the relationship and the scatterplot shows a linear pattern, you can report the value of the correlation coefficient or the coefficient of determination as a measure of the strength of the linear relationship. When you interpret the value of the correlation coefficient, it is a good idea to relate the interpretation to the pattern observed in the scatterplot. This is especially important before making a statement of no relationship between two variables, because a correlation coefficient near 0 does not necessarily imply that there is no relationship of any form. Similarly, a correlation coefficient near 1 or 1, by itself, does not guarantee that the relationship is linear. A curved pattern, such as the one we saw in Figure 5.17, can produce a correlation coefficient that is near 1. If the goal of a study is prediction, then, when you report the results of the study, you should not only give a scatterplot and the equation of the least-squares line but also address how well the linear prediction model fits the data. At a minimum, you should include both the values of se (the standard deviation about the regression line) and r 2 (the coefficient of determination). Including a residual plot can also provide support for the appropriateness of the linear model for describing the relationship between the two variables.
What to Look for in Published Data Here are a few things to consider when you read an article that includes an analysis of bivariate data: • What two variables are being studied? Are they both numerical? Is a distinction
made between a dependent variable and an independent variable? • Does the article include a scatterplot of the data? If so, does there appear to be a
relationship between the two variables? Can the relationship be described as linear, or is some type of nonlinear relationship a more appropriate description? • Does the relationship between the two variables appear to be weak or strong? Is the value of a correlation coefficient reported? • If the least-squares line is used to summarize the relationship between the dependent and independent variables, is any measure of goodness of fit reported, such as r 2 or se? How are these values interpreted, and what do they imply about the usefulness of the least-squares line? • If a correlation coefficient is reported, is it interpreted properly? Be cautious of interpretations that claim a causal relationship. The authors of the paper “Recycling and Ambivalence: Quantitative and Qualitative Analyses of Household Recycling Among Young Adults” (Environment and Behavior [2008]: 777–797) describe a study of recycling among young adults in Sweden. They considered y 5 a numerical measure of recycling behavior that was based
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.6 Interpreting and Communicating the Results of Statistical Analyses
285
on how often six types of household waste (newspapers, glass, hard plastic, soft plastic, metal, and paper) were recycled and x 5 distance to the nearest recycling facility. They reported that “a check of a plot between the variables recycling behavior and distance to the nearest recycling facility showed an approximately linear association between the two variables.” Based on this observation, a least-squares regression line is an appropriate way to summarize the relationship between recycling behavior and distance to nearest recycling facility. A related article, “Rubbish Regression and the Census Undercount” (Chance [1992]: 33), describes work done by the Garbage Project at the University of Arizona. Project researchers had analyzed different categories of garbage for a number of households. They were asked by the Census Bureau to see whether any of the garbage data variables were related to household size. They reported that “the weight data for different categories of garbage were plotted on graphs against data on household size, dwelling by dwelling, and the resulting scatterplots were analyzed to see in which categories the weight showed a steady, monotonic rise relative to household size.” The researchers determined that the strongest linear relationship appeared to be that between the amount of plastic discarded and household size. The line used to summarize this relationship was stated to be y^ 5 0.2815x, where y 5 household size and x 5 weight (in pounds) of plastic during a 5-week collection. Note that this line has an intercept of 0. Scientists at the Census Bureau believed that this relationship would extend to entire neighborhoods, and so the amount of plastic discarded by a neighborhood could be measured (rather than having to measure house by house) and then used to approximate the neighborhood size. An example of the use of the correlation coefficient is found in a paper describing a study of nightingales (a species of bird known for its song). For male songbirds, both physical characteristics and the quality of the song play a role in a female’s choice of a mate. The authors of the article “Song Repertoire Is Correlated with Body
Measures and Arrival Date in Common Nightingales” (Animal Behaviour [2005]: 211–217) used data from n 5 20 nightingales to reach the conclusion that there was a positive correlation between the number of different songs in a nightingale’s repertoire and both body weight (r 5 .53) and wing length (r 5 .47), and a negative correlation between repertoire size and arrival date (r 5 2.47). This means that heavier birds tend to know more songs, as do birds with longer wings. The authors of the paper indicated that the observed correlation between repertoire size and body characteristics was unlikely to be due solely to the age of the bird, since all nightingales in the study were more than 3 years old and prior research indicates that repertoire size does not continue to increase with age after the third year. The negative correlation between repertoire size and arrival date was interpreted as meaning that male nightingales who knew more songs tended to arrive at their breeding habitats earlier than those who knew fewer songs. A nonlinear regression was used by the authors of the paper “Maternal Blood Manganese Levels and Infant Birth Weight” (Epidemiology [2009]: 367–373) to describe the relationship between y 5 birth weight and x 5 maternal blood-manganese level at delivery for 470 mother-infant pairs. The paper states: In this cross-sectional study, there was an inverted U-shaped association between maternal blood-manganese levels at delivery and birth weight in full-term infants. This suggests that both lower and higher manganese exposures are associated with lower birth weight, although the association of higher manganese with lower weight was rather weak and imprecise. This is the first epidemiologic study to provide clear evidence of a nonlinear association between maternal manganese exposure and birth weight.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
286
Chapter 5 Summarizing Bivariate Data
The paper goes on to suggest why the relationship may be best described by a quadratic rather than a linear equation: One possible explanation for this effect would be oxidative stress caused by high manganese levels, leading to impairment of cellular function and growth. Manganese, like iron, is a transitional metal and can catalyze oxidative cellular reactions. Exposure to high levels of iron, a metal with overlapping chemical properties to manganese, has been associated with low birth weight.
A Word to the Wise: Cautions and Limitations There are a number of ways to get into trouble when analyzing bivariate numerical data! Here are some of the things you need to keep in mind when conducting your own analyses or when reading reports of such analyses: 1. Correlation does not imply causation. A common media blunder is to infer a cause-and-effect relationship between two variables simply because there is a strong correlation between them. Don’t fall into this trap! A strong correlation implies only that the two variables tend to vary together in a predictable way, but there are many possible explanations for why this is occurring besides one variable causing changes in the other. For example, the article “Ban Cell Phones? You May as Well Ban Talking Instead” (USA Today, April 27, 2000) gave data that showed a strong negative correlation between the number of cell phone subscribers and traffic fatality rates. During the years from 1985 to 1998, the number of cell phone subscribers increased from 200,000 to 60,800,000, and the number of traffic deaths per 100 million miles traveled decreased from 2.5 to 1.6 over the same period. However, based on this correlation alone, the conclusion that cell phone use improves road safety is not reasonable! Similarly, the Calgary Herald (April 16, 2002) reported that heavy and moderate drinkers earn more than light drinkers or those who do not drink. Based on the correlation between number of drinks consumed and income, the author of the study concluded that moderate drinking “causes” higher earnings. This is obviously a misleading statement, but at least the article goes on to state that “there are many possible reasons for the correlation. It could be because better-off men simply choose to buy more alcohol. Or it might have something to do with stress: Maybe those with stressful jobs tend to earn more after controlling for age, occupation, etc., and maybe they also drink more in order to deal with the stress.” 2. A correlation coefficient near 0 does not necessarily imply that there is no relationship between two variables. Before such an interpretation can be given, it is important to examine a scatterplot of the data carefully. Although it may be true that the variables are unrelated, there may in fact be a strong but nonlinear relationship. 3. The least-squares line for predicting y from x is not the same line as the leastsquares line for predicting x from y. The least-squares line is, by definition, the line that has the smallest possible sum of squared deviations of points from the line in the y direction (it minimizes g 1 y 2 y^ 2 2 ). The line that minimizes the sum of squared deviations in the y direction is not generally the same as the line that minimizes the sum of the squared deviations in the x direction. So, for example, it is not appropriate to fit a line to data using y 5 house price and x 5 house size and then use the resulting least-squares line Price 5 a 1 b(Size) to predict the size of a house by substituting in a price and then solving for size. Make sure that Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.6 Interpreting and Communicating the Results of Statistical Analyses
4.
5.
6.
7.
8.
9.
287
the dependent and independent variables are clearly identified and that the appropriate line is fit. Beware of extrapolation. It is dangerous to assume that a linear model fit to data is valid over a wider range of x values. Using the least-squares line to make predictions outside the range of x values in the data set often leads to poor predictions. Be careful in interpreting the value of the slope and intercept in the least-squares line. In particular, in many instances interpreting the intercept as the value of y that would be predicted when x 5 0 is equivalent to extrapolating way beyond the range of the x values in the data set, and this should be avoided unless x 5 0 is within the range of the data. Remember that the least-squares line may be the “best” line (in that it has a smaller sum of squared deviations than any other line), but that doesn’t necessarily mean that the line will produce good predictions. Be cautious of predictions based on a least-squares line without any information about the adequacy of the linear model, such as se and r 2. It is not enough to look at just r 2 or just se when assessing a linear model. These two measures address different aspects of the fit of the line. In general, we would like to have a small value for se (which indicates that deviations from the line tend to be small) and a large value for r 2 (which indicates that the linear relationship explains a large proportion of the variability in the y values). It is possible to have a small se combined with a small r 2 or a large r 2 combined with a large se. Remember to consider both values. The value of the correlation coefficient as well as the values for the intercept and slope of the least-squares line can be sensitive to influential observations in the data set, particularly if the sample size is small. Because potentially influential observations are those whose x values are far away from most of the x values in the data set, it is important to look for such observations when examining the scatterplot. (Another good reason for always starting with a plot of the data!) If the relationship between two variables is nonlinear, it is preferable to model the relationship using a nonlinear model rather than fitting a line to the data. A plot of the residuals from a linear fit is particularly useful in determining whether a nonlinear model would be a more appropriate choice.
E X E RC I S E S 5 . 6 3 - 5 . 6 6 5.63 The “Admitted Students Highlights Report 2009” prepared by The College Board for Cal Poly San Luis Obispo summarizes responses to a survey completed by 2001 new students who enrolled at Cal Poly in fall 2008 and by 2000 students who were admitted to Cal Poly for the fall 2008 term but who enrolled at other universities. One question in the survey presented a list of “college images” (such as career-oriented and friendly) and asked students to indicate for each image whether or
Bold exercises answered in back
Data set available online
not they associated that image with Cal Poly. The percentage that associated an image with Cal Poly was recorded for enrolling students and for non-enrolling students for each image. For example, 61% of enrolling students but only 46% of non-enrolling students associated the image “career-oriented” with Cal Poly. The resulting data were used to construct a scatterplot that appeared in the report. The scatterplot is reproduced as Figure EX5.63 at the bottom of the following page.
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
288
Chapter 5 Summarizing Bivariate Data
The higher the speed, the better the fuel economy. The faster you drive, the more fuel efficient you become and the more gasoline you save.
a. What do you think the two dashed lines in the scatterplot represent? b. Write a short article appropriate for a student newspaper commenting on what can be learned from this scatterplot. You can assume that the scatterplot will appear with the article.
Notice that the only speeds that the letter writer provides data for are 17.5, 35, and 65 mpg. Studies of the relationship between y gas mileage and x speed have suggested that the relationship is not linear, and some have used a quadratic curve to describe the relationship between gas mileage and speed. Write a response to Mr. Cleary that explains how his three observed data points could still be consistent with the statement that higher highway speeds lead to reduced fuel efficiency. Include a graph to support your explanation.
5.64 The following is an excerpt from a letter to the editor written by Roger Cleary that appeared in the San Luis Obispo Tribune (September 16, 2008): The causes of poor fuel economy have nothing to do with higher highway speeds, notwithstanding all the press hoopla, including the July 19 Miami Herald claim that “There is no question that slower speeds will save gasoline,” and the July 3 statement by Drive Smarter Challenge vehicle director Deron Lovaas that, “I’m not sure whether most people make the connection between how fast they drive and how much fuel they use.” I decided to gather the speed facts for myself using my Chevy, which comes equipped with a fuel usage driver information center, real-time read out. At a road speed of 17.5 mph, it averages 10 mpg; at 35 mph, it averages 20 mpg; and at 65 mph, it averages 30 mpg, all testing done with engine speed standardized at 2000 rpm.
5.65 The paper “How Lead Exposure Relates to Temporal Changes in IQ, Violent Crime, and Unwed Pregnancy” (Environmental Research [2000]: 1–22) investigated whether childhood lead exposure is related to criminal behavior in young adults. Using historical data, the author paired y 5 assault rate (assaults per 100,000 people) for each year from 1964 to 1997 with a measure of lead exposure (tons of gasoline lead per 1000 people) 23 years earlier. For example, the lead exposure from 1974 was paired with the assault rate from 1997. The author chose to go back 23 years for lead exposure
Percentage of enrolling students marking image
100
80 Friendly Career-oriented Fun Comfortable
Highly respected
60
Challenging Intellectual Exciting Prestigious Personal Selective
40
20 Excitingly different Average 0
FIGURE EX5.63
Bold exercises answered in back
Isolated Back-up school
20 40 60 80 Percentage of nonenrolling students marking image
Data set available online
100
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
5.6 Interpreting and Communicating the Results of Statistical Analyses
2009 is used for the assault rate data. Also note that at the time the graph was constructed, assault rate data was only available through 1997. Spend a few minutes thinking about the information contained in this graph and then briefly explain what aspect of this graph accounts for the reported positive correlation between assault rate and lead exposure 23 years prior.
because the highest number of assaults are committed by people in their early twenties, and 23 years earlier would represent a time when those in this age group were infants. A least-squares regression line was used to describe the relationship between assault rate and lead exposure 23 years prior. Summary statistics given in the paper are 24.08 327.41 0.89
5.66 The following quote is from the paper “Evaluation of the Accuracy of Different Methods Used to Estimate Weights in the Pediatric Population” (Pediatrics [2009]: e1045–e1051):
Use the information provided to answer the following questions. a. What is the value of the correlation coefficient for x 5 lead exposure 23 years prior and y 5 assault rate? Interpret this value. Is it reasonable to conclude that increased lead exposure is the cause of increased assault rates? Explain. b. What is the equation of the least-squares regression line? Use the line to predict assault rate in a year in which gasoline lead exposure 23 years prior was y^ 5 10.8768 1 1.4604x 2 0.5 tons per 1000 people. 7.259x 2 1 9.2342x3 c. What proportion of year-to-year variability in assault rates can be explained by the relationship between assault rate and gasoline lead exposure 23 years earlier? d. The graph below appeared in the paper. Note that this is not a scatterplot of the (x, y) pairs—it is two separate time series plots. The time scale 1941, 1942, p , 1986 is the time scale used for the lead exposure data and the time scale 1964, 1965, p ,
As expected, the model demonstrated that weight increased with age, but visual inspection of an age versus weight plot demonstrated a nonlinear relationship unless infants and children were analyzed separately. The linear coefficient for age as a predictor of weight was 6.93 in infants and 3.1 to 3.48 in children. This quote suggests that when a scatterplot of weight versus age was constructed for all 1011 children in the study described in the paper, the relationship between y weight and x age was not linear. When the 1011 children were separated into two groups—infants (age birth to 1 year) and children (age 1 to 10 years)—and separate scatterplots were constructed, the relationship between weight and age appeared linear in each scatterplot. The slopes reported in the given quote (referred to as “the linear coefficient”) are expressed in kg/year. Briefly explain why the relationship between weight and age in the scatterplot for the combined group would appear nonlinear.
Tons of gasoline lead per 1000 population
1.5
500 450
1.2
400 350
0.9
300 250
0.6
200 Gasoline Lead
150
Aggravated Assault
0.3
100 50
Aggravated assaults per 100,000 inhabitants
intercept: slope r2
289
00 9 :2
00 6 19
86
00 3
:2
19
83
00 0
:2
19
80
99 7
:2
19
77
99 4
:1
19
74
99 1
:1
19
71
98 8
:1
:1 19
68
98 5 19
65
98 2
:1
:1 19
62
97 9 19
59
97 6
:1
:1 19
56
97 3 19
53
97 0
:1
:1 19
50
96 7 :1
47
44
19
41 19
19
:1
96 4
Lead year: Assault year (23 year lag)
Bold exercises answered in back
Data set available online
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
290
Chapter 5
Summarizing Bivariate Data
AC TI VI T Y 5. 1
Exploring Correlation and Regression Technology Activity (Applets)
Open the applet (available in CengageNOW at www .cengage.com/login) called CorrelationPoints. You should see a screen like the one shown below.
1. Using the mouse, you can click to add points to form a scatterplot. The value of the correlation coefficient for the data in the plot at any given time (once you have two or more points in the plot) will be displayed at the top of the screen. Try to add points to obtain a correlation coefficient that is close to 0.8. Briefly describe the factors that influenced where you placed the points. 2. Reset the plot (by clicking on the Reset bar in the lower left corner of the screen) and add points trying to produce a data set with a correlation that is close to 0.4. Briefly describe the factors that influenced where you placed the points. Now open the applet called RegDecomp. You should see a screen that looks like the one shown at the top of the next column.
AC TI VI T Y 5. 2
The black points in the plot represent the data set, and the blue line represents a possible line that might be used to describe the relationship between the two variables shown. The pink lines in the graph represent the deviations of points from the line, and the pink bar on the left-hand side of the display depicts the sum of squared errors for the line shown. 3. Using your mouse, you can move the line and see how the deviations from the line and the sum of squared errors change as the line changes. Try to move the line into a position that you think is close to the least-squares regression line. When you think you are close to the line that minimizes the sum of squared errors, you can click on the bar that says “Find Best Model” to see how close you were to the actual least-squares line.
Age and Flexibility
Materials needed: Yardsticks. In this activity, you will investigate the relationship between age and a measure of flexibility. Flexibility will be measured by asking a person to bend at the waist as far as possible, extending his or her arms toward the floor. Using a yardstick, measure the distance from the floor to the fingertip closest to the floor. 1. Age and the measure of flexibility just described will be measured for a group of individuals. Our goal is to determine whether there is a relationship between age and this measure of flexibility. What are two reasons why it would not be a good idea to use just
2. 3.
4.
5.
the students in your class as the subjects for your study? Working as a class, decide on a reasonable way to collect data on the two variables of interest. After your class has collected appropriate data, use them to construct a scatterplot. Comment on the interesting features of the plot. Does it look like there is a relationship between age and flexibility? If there appears to be a relationship between age and flexibility, fit a model that is appropriate for describing the relationship. In the context of this activity, write a brief description of the danger of extrapolation.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
Summary of Key Concepts and Formulas
291
Summary of Key Concepts and Formulas TERM OR FORMULA
COMMENT
Scatterplot
A graph of bivariate numerical data in which each observation (x, y) is represented as a point located with respect to a horizontal x axis and a vertical y axis.
Pearson’s sample correlation coefficient g zx zy r5 n21
A measure of the extent to which sample x and y values are linearly related; 1 r 1, so values close to 1 or 1 indicate a strong linear relationship.
Principle of least squares
The method used to select a line that summarizes an approximate linear relationship between x and y. The least-squares line is the line that minimizes the sum of the squared errors (vertical deviations) for the points in the scatterplot.
g 1x 2 x 2 1 y 2 y 2 5 b5 g 1x 2 x 2 2
1 g x2 1 g y2 n 1 g x2 2 g x2 2 n
g xy 2
The slope of the least-squares line.
a 5 y 2 bx
The intercept of the least-squares line.
Predicted (fitted) values y^ 1, y^ 2, . . . , y^ n
Obtained by substituting the x value for each observation in the data set into the least-squares line; y^ 1 5 a 1 bx1, . . . , y^ n 5 a 1 bxn
Residuals
Obtained by subtracting each predicted value from the corresponding observed y value: y1 2 y^ 1, . . . , yn 2 y^ n. These are the vertical deviations from the least-squares line.
Residual plot
Scatterplot of the (x, residual) pairs. Isolated points or a pattern of points in a residual plot are indicative of potential problems.
Residual (error) sum of squares SSResid 5 g 1 y 2 y^ 2 2
The sum of the squared residuals is a measure of y variation that cannot be attributed to an approximate linear relationship (unexplained variation).
Total sum of squares SSTo 5 g 1 y 2 y 2 2
The sum of squared deviations from the sample mean is a measure of total variation in the observed y values.
Coefficient of determination SSResid r2 5 1 2 SSTo Standard deviation about the least-squares line SSResid se 5 Å n22
The proportion of variation in observed y’s that can be explained by an approximate linear relationship. The size of a “typical” deviation from the least-squares line.
Transformation
A simple function of the x and/or y variable, which is then used in a regression.
Power transformation
An exponent, or power, p, is first specified, and then new (transformed) data values are calculated as transformed value (original value) p. A logarithmic transformation is identified with p 0. When the scatterplot of original data exhibits curvature, a power transformation of x and/or y will often result in a scatterplot that has a linear appearance.
Logistic regression function p 5
e a1bx 1 1 e a1bx
The graph of this function is an S-shaped curve. The logistic regression function is used to describe the relationship between probability of success and a numerical predictor variable.
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
292
Chapter 5 Summarizing Bivariate Data
Chapter Review Exercises 5.67 - 5.79 The accompanying data represent x amount of catalyst added to accelerate a chemical reaction and y 5 resulting reaction time:
5.67
x y
1 49
2 46
3 41
4 34
5 25
a. Calculate r. Does the value of r suggest a strong linear relationship? b. Construct a scatterplot. From the plot, does the word linear provide the most effective description of the relationship between x and y? Explain. The paper “A Cross-National Relationship
5.68
Between Sugar Consumption and Major Depression?” (Depression and Anxiety [2002]: 118–120) concluded that there was a correlation between refined sugar consumption (calories per person per day) and annual rate of major depression (cases per 100 people) based on data from six countries. The following data were read from a graph that appeared in the paper: Country
Sugar Consumption
Depression Rate
150 300 350 375 390 480
2.3 3.0 4.4 5.0 5.2 5.7
Korea United States France Germany Canada New Zealand
a. Compute and interpret the correlation coefficient for this data set. b. Is it reasonable to conclude that increasing sugar consumption leads to higher rates of depression? Explain. c. Do you have any concerns about this study that would make you hesitant to generalize these conclusions to other countries? The following data on x 5 score on a measure of test anxiety and y 5 exam score for a sample of n 5 9 students are consistent with summary quantities given in the paper “Effects of Humor on Test Anxiety and Performance” (Psychological Reports [1999]: 1203–1212):
5.69
x 23 14 14 0 17 20 20 15 21 y 43 59 48 77 50 52 46 51 51
a. Construct a scatterplot, and comment on the features of the plot. b. Does there appear to be a linear relationship between the two variables? How would you characterize the relationship? c. Compute the value of the correlation coefficient. Is the value of r consistent with your answer to Part (b)? d. Is it reasonable to conclude that test anxiety caused poor exam performance? Explain.
5.70 Researchers asked each child in a sample of 411 school-age children if they were more or less likely to purchase a lottery ticket at a store if lottery tickets were visible on the counter. The percentage that said that they were more likely to purchase a ticket by grade level are as follows (R&J Child Development Consultants, Quebec, 2001): Grade
Percentage That Said They Were More Likely to Purchase
6 8 10 12
32.7 46.1 75.0 83.6
a. Construct a scatterplot of y percentage who said they were more likely to purchase and x grade. Does there appear to be a linear relationship between x and y? b. Find the equation of the least-squares line. y^ 5 222.37 1 9.08x
5.71 Percentages of public school students in fourth grade in 1996 and in eighth grade in 2000 who were at or above the proficient level in mathematics were given in the article “Mixed Progress in Math” (USA Today, August 3, 2001) for eight western states: State Arizona California Hawaii Montana New Mexico Oregon Utah Wyoming
4th grade (1996)
8th grade (2000)
15 11 16 22 13 21 23 19
21 18 16 37 13 32 26 25
Higher values for x indicate higher levels of anxiety. Bold exercises answered in back
Data set available online
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
Chapter Review Exercises
293
a. Construct a scatterplot, and comment on any intera. One observation was (25, 70). What is the esting features. corresponding residual? b. Find the equation of the least-squares line that sumb. What is the value of the sample correlation marizes the relationship between x 5 1996 fourthcoefficient? grade math proficiency percentage and y 5 2000 c. Suppose that SSTo 5 2520.0 (this value was not eighth-grade math proficiency percentage. y^ 5 23.14 1 1.52x given in the paper). What is the value of se? c. Nevada, a western state not included in the data set, 5.74 The paper “Aspects of Food Finding by Winhad a 1996 fourth-grade math proficiency of 14%. tering Bald Eagles” (The Auk [1983]: 477–484) examWhat would you predict for Nevada’s 2000 eighthined the relationship between the time that eagles spend grade math proficiency percentage? How does your aerially searching for food (indicated by the percentage prediction compare to the actual eighth-grade value of eagles soaring) and relative food availability. The acof 20 for Nevada? companying data were taken from a scatterplot that ap5.72 The following table gives the number of organ peared in this paper. Let x denote salmon availability and transplants performed in the United States each year y denote the percentage of eagles in the air. from 1990 to 1999 (The Organ Procurement and x 0 0 0.2 0.5 0.5 1.0 Transplantation Network, 2003): y 28.2 69.0 27.0 38.5 48.4 31.1 Year
Number of Transplants (in thousands)
1 (1990) 2 3 4 5 6 7 8 9 10 (1999)
15.0 15.7 16.1 17.6 18.3 19.4 20.0 20.3 21.4 21.8
a. Construct a scatterplot of these data, and then find the equation of the least-squares regression line that describes the relationship between y number of transplants performed and x year. Describe how the number of transplants performed has changed over time from 1990 to 1999. b. Compute the 10 residuals, and construct a residual plot. Are there any features of the residual plot that indicate that the relationship between year and number of transplants performed would be better described by a curve rather than a line? Explain.
5.73 The paper “Effects of Canine Parvovirus (CPV) on Gray Wolves in Minnesota” (Journal of Wildlife Management [1995]: 565–570) summarized a regression of y 5 percentage of pups in a capture on x 5 percentage of CPV prevalence among adults and pups. The equation of the least-squares line, based on n 5 10 observations, was y^ 5 62.9476 2 0.54975x, with r 2 5 .57. Bold exercises answered in back
Data set available online
x y
1.2 26.9
1.9 8.2
2.6 4.6
3.3 7.4
4.7 7.0
6.5 6.8
a. Draw a scatterplot for this data set. Would you describe the pattern in the plot as linear or curved? b. One possible transformation that might lead to a straighter plot involves taking the square root of both the x and y values. Use Figure 5.38 to explain why this might be a reasonable transformation. c. Construct a scatterplot using the variables !x and !y. Is this scatterplot more nearly linear than the scatterplot in Part (a)? d. Using Table 5.5, suggest another transformation that might be used to straighten the original plot.
5.75 Data on salmon availability (x) and the percentage of eagles in the air ( y) were given in the previous exercise. a. Calculate the correlation coefficient for these data. b. Because the scatterplot of the original data appeared curved, transforming both the x and y values by taking square roots was suggested. Calculate the correlation coefficient for the variables !x and !y. How does this value compare with that calculated in Part (a)? Does this indicate that the transformation was successful in straightening the plot? 5.76 No tortilla chip lover likes soggy chips, so it is important to find characteristics of the production process that produce chips with an appealing texture. The accompanying data on x 5 frying time (in seconds) and y 5 moisture content (%) appeared in the paper, “Thermal and Physical Properties of Tortilla Chips as a
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
294
Chapter 5 Summarizing Bivariate Data
Function of Frying Time” (Journal of Food Processing and Preservation [1995]: 175–189):
simulation. This resulted in the following data and scatterplot:
Frying time (x): 5 Moisture 16.3 content ( y):
60 1.3
Firefighter x y
1 51.3 49.3
2 34.1 29.5
3 41.1 30.6
4 36.3 28.2
5 36.5 28.0
a. Construct a scatterplot of these data. Does the relationship between moisture content and frying time appear to be linear? b. Transform the y values using y 5 log(y) and construct a scatterplot of the (x, y) pairs. Does this scatterplot look more nearly linear than the one in Part (a)? c. Find the equation of the least-squares line that describes the relationship between y and x. d. Use the least-squares line from Part (c) to predict moisture content for a frying time of 35 minutes.
Firefighter x y
6 35.4 26.3
7 35.4 33.9
8 38.6 29.4
9 40.6 23.5
10 39.5 31.6
5.77
10 9.7
15 8.1
20 4.2
25 3.4
30 2.9
45 1.9
The article “Reduction in Soluble Protein and
Chlorophyll Contents in a Few Plants as Indicators of Automobile Exhaust Pollution” (International Journal of Environmental Studies [1983]: 239–244) reported the following data on x 5 distance from a highway (in meters) and y 5 lead content of soil at that distance (in parts per million):
Fire-simulation consumption
50 40 30 20 35
42
0.3 62.75
1 37.51
5 29.70
10 20.71
15 17.65
20 15.41
The regression equation is firecon = –11.4 + 1.09 treadcon
x y
25 14.15
30 13.50
40 12.11
50 11.40
75 10.85
100 10.85
Predictor Constant treadcon s = 4.70
5.78 An accurate assessment of oxygen consumption provides important information for determining energy expenditure requirements for physically demanding tasks. The paper “Oxygen Consumption During Fire Suppression: Error of Heart Rate Estimation” (Ergonomics [1991]: 1469–1474) reported on a study in which x 5 oxygen consumption (in milliliters per kilogram per minute) during a treadmill test was determined for a sample of 10 firefighters. Then y 5 oxygen consumption at a comparable heart rate was measured for each of the 10 individuals while they performed a fire-suppression Bold exercises answered in back
Data set available online
Treadmill consumption
a. Does the scatterplot suggest an approximate linear relationship? b. The investigators fit a least-squares line. The resulting Minitab output is given in the following:
x y
a. Use a statistical computer package to construct scatterplots of y versus x, y versus log(x), log( y) versus 1 1 log(x), and versus . y x b. Which transformation considered in Part (a) does the best job of producing an approximately linear relationship? Use the selected transformation to predict lead content when distance is 25 m.
49
Coef Stdev t-ratio p –11.37 12.46 –0.91 0.388 1.0906 0.3181 3.43 0.009 R-sq = 59.5% R-sq(adj) = 54.4%
Predict fire-simulation consumption when treadmill consumption is 40. c. How effectively does a straight line summarize the relationship? d. Delete the first observation, (51.3, 49.3), and calculate the new equation of the least-squares line and the value of r 2. What do you conclude? (Hint: For the original data, g x 5 388.8, g y 5 310.3, g xy 5 12,306.58, and g x 2 5 15,338.54, 2 g y 5 10,072.41.)
5.79 Consider the four (x, y) pairs (0, 0), (1, 1), (1, 1), and (2, 0). a. What is the value of the sample correlation coefficient r ? b. If a fifth observation is made at the value x 6, find a value of y for which r 0.5. c. If a fifth observation is made at the value x 6, find a value of y for which r 0.5. Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
295
Cumulative Review Exercises
Cumulative Review Exercises CR5.1 - CR5.19 CR5.1 The article “Rocker Shoe Put to the Test: Can it Really Walk the Walk as a Way to Get in Shape?” (USA Today, October 12, 2009) describes claims made by Skechers about Shape-Ups, a shoe line introduced in 2009. These curved-sole sneakers are supposed to help you “get into shape without going to the gym” according to a Skechers advertisement. Briefly describe how you might design a study to investigate this claim. Include how you would select subjects and what variables you would measure. Is the study you designed an observational study or an experiment?
CR5.2 Data from a survey of 1046 adults age 50 and older were summarized in the AARP Bulletin (November 2009). The following table gives relative frequency distributions of the responses to the question, “How much do you plan to spend for holiday gifts this year?” for respondents age 50 to 64 and for respondents age 65 and older. Construct a histogram for each of the two age groups and comment on the differences between the two age groups. (Notice that the interval widths in the relative frequency distribution are not the same, so you shouldn’t use relative frequency on the y-axis for your histograms.)
Amount Plan to Spend
Relative Frequency for Age Group 50 to 64
Relative Frequency for Age Group 65 and Older
.20 .13 .16 .12 .11 .28
.36 .11 .16 .10 .05 .22
less than $100 $100 to ,$200 $200 to ,$300 $300 to ,$400 $400 to ,$500 $500 to ,$1000
CR5.3 The graph in Figure CR5.3 appeared in the report “Testing the Waters 2009” (Natural Resources Defense Council). Spend a few minutes looking at the graph and reading the caption that appears with the graph. Briefly explain how the graph supports the claim that discharges of polluted storm water may be responsible for increased illness levels.
CR5.4 The cost of Internet access was examined in the report “Home Broadband Adoption 2009” (pewinternet.org). In 2009, the mean and median amount paid monthly for service for broadband users was reported as $39.00 and $38.00, respectively. For
FIGURE CR5.3 Influence of Heavy Rainfall on Occurrence of E. Coli Infections 100
60
Number of cases Rainfall
50
80
60 30 40
Rainfall (ml)
Number of cases
40
20 20
10
0 May 1
May 4
May 7
May 10
May 13
May 16
May 19
May 22
May 25
May 28
May 31
The graph shows the relationship between unusually heavy rainfall and the number of confirmed cases of E. coli infection that occurred during a massive disease outbreak in Ontario, Quebec, in May 2000. The incubation period for E.Coli is usually 3 to 4 days, which is consistent with the lag between extreme precipitation events and surges in the number of cases. Bold exercises answered in back
Data set available online
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent rights restrictions require it.
296
Chapter 5
Summarizing Bivariate Data
dial-up users, the mean and median amount paid monthly were $26.60 and $20.00, respectively. What do the values of the mean and median tell you about the shape of the distribution of monthly amount paid for broadband users? For dial-up users?
Victoria and Albert Museum (“Enigmas of Bidri,” Surface Engineering [2005]: 333–339), listed in increasing order:
CR5.5 Foal weight at birth is an indicator of health, so it is of interest to breeders of thoroughbred horses. Is foal weight related to the weight of the mare (mother)? The accompanying data are from the paper “Suckling
a. Construct a dotplot for these data. b. Calculate the mean and median copper content. c. Will an 8% trimmed mean be larger or smaller than the mean for this data set? Explain your reasoning.
Behaviour Does Not Measure Milk Intake in Horses” (Animal Behaviour [1999]: 673–678): Observation
Mare Weight (x, in kg)
Foal weight (y, in kg)
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
556 638 588 550 580 642 568 642 556 616 549 504 515 551 594
129 119 132 123.5 112 113.5 95 104 104 93.5 108.5 95 117.5 128 127.5
The correlation coefficient for these data is 0.001. Construct a scatterplot of these data and then write a few sentences describing the relationship between mare weight and foal weight that refer both to the value of the correlation coefficient and the scatterplot.
CR5.6 In August 2009, Harris Interactive released the results of the “Great Schools” survey. In this survey, 1086 parents of children attending a public or private school were asked approximately how much they had spent on school supplies over the last school year. For this sample, the mean amount spent was $235.20 and the median amount spent was $150.00. What does the large difference between the mean and median tell you about this data set? CR5.7 Bidri is a popular and traditional art form in India. Bidri articles (bowls, vessels, and so on) are made by casting from an alloy containing primarily zinc along with some copper. Consider the following observations on copper content (%) for a sample of Bidri artifacts in London’s Bold exercises answered in back
Data set available online
2.0 3.1 3.6
2.4 3.2 3.7
2.5 3.3 4.4
2.6 3.3 4.6
2.6 3.4 4.7
2.7 3.4 4.8
2.7 3.6 5.3
2.8 3.6 10.1
3.0 3.6
Medicare’s new medical plans offer a wide range of variations and choices for seniors when picking a drug plan (San Luis Obispo Tribune, November 25, 2005). The monthly cost for a stand-alone drug plan can vary from a low of $1.87 in Montana, Wyoming, North Dakota, South Dakota, Nebraska, Minnesota, and Iowa to a high of $104.89. Here are the lowest and highest monthly premiums for stand-alone Medicare drug plans for each state:
CR5.8
State
$ Low
$ High
Alabama Alaska Arizona Arkansas California Colorado Connecticut Delaware District of Columbia Florida Georgia Hawaii Idaho Illinois Indiana Iowa Kansas Kentucky Louisiana Maine Maryland Massachusetts Michigan Minnesota Mississippi Missouri Montana Nebraska Nevada
14.08 20.05 6.14 10.31 5.41 8.62 7.32 6.44 6.44 10.35 17.91 17.18 6.33 13.32 12.30 1.87 9.48 12.30 17.06 19.60 6.44 7.32 13.75 1.87 11.60 10.29 1.87 1.87 6.42
69.98 61.93 64.86 67.98 66.08 65.88 65.58 68.91 68.91 104.89 73.17 64.43 68.88 65.04 70.72 99.90 67.88 70.72 70.59 65.39 68.91 65.58 65.69 99.90 70.59 68.26 99.90 99.90 64.63 (continued)
Video Solution available
Copyright 2010 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part. Due to electronic rights, some third party content may be suppressed from the eBook and/or eChapter(s). Editorial review has deemed that any suppressed content does not materially affect the overall learning experience. Cengage Learning reserves the right to remove additional content at any time if subsequent r