Biostatistics For Dummies, 2nd Edition (True PDF

Biostatistics
2nd Edition
by Monika Wahi, MPH and

John C. Pezzullo, PhD
Biostatistics For Dummies®, 2nd Edition
Published by: John Wiley & Sons, Inc., 111 River Street, Hoboken, NJ 07030-5774, www.wiley.com
Copyright © 2024 by John Wiley & Sons, Inc. All rights reserved, including rights for text and data mining and training
of artificial technologies or similar technologies.
Media and software compilation copyright © 2024 by John Wiley & Sons, Inc. All rights reserved, including rights for
text and data mining and training of artificial technologies or similar technologies.
Published simultaneously in Canada
No part of this publication may be reproduced, stored in a retrieval system or transmitted in any form or by any
means, electronic, mechanical, photocopying, recording, scanning or otherwise, except as permitted under Sections
107 or 108 of the 1976 United States Copyright Act, without the prior written permission of the Publisher. Requests to
the Publisher for permission should be addressed to the Permissions Department, John Wiley & Sons, Inc., 111 River
Street, Hoboken, NJ 07030, (201) 748-6011, fax (201) 748-6008, or online at http://www.wiley.com/go/
permissions.
Trademarks: Wiley, For Dummies, the Dummies Man logo, Dummies.com, Making Everything Easier, and related
trade dress are trademarks or registered trademarks of John Wiley & Sons, Inc. and may not be used without written
permission. All other trademarks are the property of their respective owners. John Wiley & Sons, Inc. is not
associated with any product or vendor mentioned in this book.
LIMIT OF LIABILITY/DISCLAIMER OF WARRANTY: THE PUBLISHER AND THE AUTHOR MAKE NO

REPRESENTATIONS OR WARRANTIES WITH RESPECT TO THE ACCURACY OR COMPLETENESS OF THE CONTENTS
OF THIS WORK AND SPECIFICALLY DISCLAIM ALL WARRANTIES, INCLUDING WITHOUT LIMITATION WARRANTIES
OF FITNESS FOR A PARTICULAR PURPOSE. NO WARRANTY MAY BE CREATED OR EXTENDED BY SALES OR
PROMOTIONAL MATERIALS. THE ADVICE AND STRATEGIES CONTAINED HEREIN MAY NOT BE SUITABLE FOR
EVERY SITUATION. THIS WORK IS SOLD WITH THE UNDERSTANDING THAT THE PUBLISHER IS NOT ENGAGED
IN RENDERING LEGAL, ACCOUNTING, OR OTHER PROFESSIONAL SERVICES. IF PROFESSIONAL ASSISTANCE IS
REQUIRED, THE SERVICES OF A COMPETENT PROFESSIONAL PERSON SHOULD BE SOUGHT. NEITHER THE
PUBLISHER NOR THE AUTHOR SHALL BE LIABLE FOR DAMAGES ARISING HEREFROM. THE FACT THAT AN
ORGANIZATION OR WEBSITE IS REFERRED TO IN THIS WORK AS A CITATION AND/OR A POTENTIAL SOURCE OF
FURTHER INFORMATION DOES NOT MEAN THAT THE AUTHOR OR THE PUBLISHER ENDORSES THE INFORMATION
THE ORGANIZATION OR WEBSITE MAY PROVIDE OR RECOMMENDATIONS IT MAY MAKE. FURTHER, READERS
SHOULD BE AWARE THAT INTERNET WEBSITES LISTED IN THIS WORK MAY HAVE CHANGED OR DISAPPEARED
BETWEEN WHEN THIS WORK WAS WRITTEN AND WHEN IT IS READ.
For general information on our other products and services, please contact our Customer Care Department within
the U.S. at 877-762-2974, outside the U.S. at 317-572-3993, or fax 317-572-4002. For technical support, please visit
https://hub.wiley.com/community/support/dummies.
Wiley publishes in a variety of print and electronic formats and by print-on-demand. Some material included with
standard print versions of this book may not be included in e-books or in print-on-demand. If this book refers to
media such as a CD or DVD that is not included in the version you purchased, you may download this material at
http://booksupport.wiley.com. For more information about Wiley products, visit www.wiley.com.
Library of Congress Control Number: 2024939111
ISBN 978-1-394-25146-9 (pbk); ISBN 978-1-394-25147-6 (ebk); ISBN 978-1-394-25148-3 (ebk)

Contents at a Glance
Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Part 1: Getting Started with Biostatistics . . . . . . . . . . . . . . . . . . . . . . 5

CHAPTER 1: Biostatistics 101. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
CHAPTER 2: Overcoming Mathophobia: Reading and Understanding
Mathematical Expressions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
CHAPTER 3: Getting Statistical: A Short Review of Basic Statistics. . . . . . . . . . . . . . . . . 29
Part 2: Examining Tools and Processes . . . . . . . . . . . . . . . . . . . . . . . . 51

CHAPTER 4: Counting on Statistical Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
CHAPTER 5: Conducting Clinical Research. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
CHAPTER 6: Taking All Kinds of Samples. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
CHAPTER 7: Having Designs on Study Design. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
Part 3: Getting Down and Dirty with Data . . . . . . . . . . . . . . . . . . . . 99

CHAPTER 8: Getting Your Data into the Computer. . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
CHAPTER 9: Summarizing and Graphing Your Data. . . . . . . . . . . . . . . . . . . . . . . . . . . 111
CHAPTER 10: Having Confidence in Your Results. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
Part 4: Comparing Groups. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139

CHAPTER 11: Comparing Average Values between Groups . . . . . . . . . . . . . . . . . . . . . 141
CHAPTER 12: Comparing Proportions and Analyzing Cross-Tabulations. . . . . . . . . . 159
CHAPTER 13: Taking a Closer Look at Fourfold Tables. . . . . . . . . . . . . . . . . . . . . . . . . . 173
CHAPTER 14: Analyzing Incidence and Prevalence Rates in Epidemiologic Data. . . . 191
Part 5: Looking for Relationships with Correlation

and Regression. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
CHAPTER 15: Introducing Correlation and Regression. . . . . . . . . . . . . . . . . . . . . . . . . . 201
CHAPTER 16: Getting Straight Talk on Straight-Line Regression. . . . . . . . . . . . . . . . . . 213
CHAPTER 17: More of a Good Thing: Multiple Regression. . . . . . . . . . . . . . . . . . . . . . . 233
CHAPTER 18: A Yes-or-No Proposition: Logistic Regression . . . . . . . . . . . . . . . . . . . . . 249
CHAPTER 19: Other Useful Kinds of Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
CHAPTER 20: Getting the Hint from Epidemiologic Inference. . . . . . . . . . . . . . . . . . . . 291
Part 6: Analyzing Survival Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
CHAPTER 21: Summarizing and Graphing Survival Data. . . . . . . . . . . . . . . . . . . . . . . . 301
CHAPTER 22: Comparing Survival Times. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
CHAPTER 23: Survival Regression. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327
Part 7: The Part of Tens. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349

CHAPTER 24: Ten Distributions Worth Knowing. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
CHAPTER 25: Ten Easy Ways to Estimate How Many Participants You Need. . . . . . . 361
Index. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 369
Table of Contents
INTRODUCTION. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
About This Book. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Foolish Assumptions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
Icons Used in This Book. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Beyond the Book. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Where to Go from Here . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
PART 1: GETTING STARTED WITH BIOSTATISTICS. . . . . . . . . . . . 5
CHAPTER 1: Biostatistics 101. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

Brushing Up on Math and Stats Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
Doing Calculations with the Greatest of Ease. . . . . . . . . . . . . . . . . . . . . . 8
Concentrating on Epidemiologic Research . . . . . . . . . . . . . . . . . . . . . . . . 9
Drawing Conclusions from Your Data . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
Statistical estimation theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
Statistical decision theory. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
A Matter of Life and Death: Working with Survival Data. . . . . . . . . . . . 13
Getting to Know Statistical Distributions. . . . . . . . . . . . . . . . . . . . . . . . . 13
Figuring Out How Many Participants You Need. . . . . . . . . . . . . . . . . . . 14
CHAPTER 2: Overcoming Mathophobia: Reading and

Understanding Mathematical Expressions . . . . . . . . . . 15
Breaking Down the Basics of Mathematical Formulas . . . . . . . . . . . . . 16
Displaying formulas in different ways . . . . . . . . . . . . . . . . . . . . . . . . 16
Checking out the building blocks of formulas. . . . . . . . . . . . . . . . . . 17
Focusing on Operations Found in Formulas. . . . . . . . . . . . . . . . . . . . . . 18
Basic mathematical operations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
Powers, roots, and logarithms. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
Factorials and absolute values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
Simple and complicated formulas . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
Counting on Collections of Numbers. . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
One-dimensional arrays. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
Higher-dimensional arrays. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
Arrays in formulas. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
Sums and products of the elements of an array . . . . . . . . . . . . . . . 27
Table of Contents v
CHAPTER 3: Getting Statistical: A Short Review of
Basic Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
Taking a Chance on Probability. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
Thinking of probability as a number. . . . . . . . . . . . . . . . . . . . . . . . . . 30
Following a few basic rules of probabilities. . . . . . . . . . . . . . . . . . . . 30
Comparing odds versus probability. . . . . . . . . . . . . . . . . . . . . . . . . . 32
Some Random Thoughts about Randomness . . . . . . . . . . . . . . . . . . . . 33
Selecting Samples from Populations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
Recognizing that sampling isn’t perfect. . . . . . . . . . . . . . . . . . . . . . . 34
Digging into probability distributions. . . . . . . . . . . . . . . . . . . . . . . . . 35
Introducing Statistical Inference. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
Statistical estimation theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
Statistical decision theory. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
Honing In on Hypothesis Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
Getting the language down . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
Testing for significance. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
Understanding the meaning of “p value” as the result of a test. . . 42
Examining Type I and Type II errors. . . . . . . . . . . . . . . . . . . . . . . . . . 42
Grasping the power of a test. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
Going Outside the Norm with Nonparametric Statistics. . . . . . . . . . . . 47
PART 2: EXAMINING TOOLS AND PROCESSES. . . . . . . . . . . . . . . 51
CHAPTER 4: Counting on Statistical Software . . . . . . . . . . . . . . . . . . . . . . 53

Considering the Evolution of Statistical Software. . . . . . . . . . . . . . . . . . 54
Comparing Commercial to Open-Source Software . . . . . . . . . . . . . . . . 54
Checking Out Commercial Software. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
SAS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
SPSS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
Microsoft Excel. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
Online analytics platforms. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
Focusing on Open-Source and Free Software. . . . . . . . . . . . . . . . . . . . . 58
Open-source software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
Other free statistical software. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
Choosing Between Code-based and Non–Code-Based Methods . . . . 60
Storing Data in the Cloud. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
CHAPTER 5: Conducting Clinical Research. . . . . . . . . . . . . . . . . . . . . . . . . . . 61

Designing a Clinical Trial. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
Identifying aims, objectives, hypotheses, and variables . . . . . . . . . 62
Deciding who is eligible for the study. . . . . . . . . . . . . . . . . . . . . . . . . 64
Choosing the structure of a clinical trial. . . . . . . . . . . . . . . . . . . . . . . 65
Using randomization. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
Selecting the analyses to use. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
vi Biostatistics For Dummies

Determining how many participants to enroll in a clinical trial . . . 68
Assembling the study protocol . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
Carrying Out a Clinical Trial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
Protecting clinical trial participants. . . . . . . . . . . . . . . . . . . . . . . . . . . 71
Collecting and validating data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
Analyzing Your Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
Dealing with missing data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
Handling multiplicity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
CHAPTER 6: Taking All Kinds of Samples. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

Making Forgivable (and Non-Forgivable) Errors. . . . . . . . . . . . . . . . . . . 78
Framing Your Sample. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
Sampling for Success . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
Taking a simple random sample. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
Taking a stratified sample . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
Engaging in systematic sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
Sampling clusters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
Sampling at your convenience . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
Sampling in multiple stages. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
CHAPTER 7: Having Designs on Study Design. . . . . . . . . . . . . . . . . . . . . . . 87

Presenting the Study Design Hierarchy. . . . . . . . . . . . . . . . . . . . . . . . . . 88
Describing what we see . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
Getting analytical. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
Going from observational to experimental. . . . . . . . . . . . . . . . . . . . 90
Climbing the Evidence Pyramid. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
Starting at the base: Expert opinion. . . . . . . . . . . . . . . . . . . . . . . . . . 91
Making the case with case studies . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
Making statements about the population. . . . . . . . . . . . . . . . . . . . . 92
Going from case series to case-control . . . . . . . . . . . . . . . . . . . . . . . 93
Following a cohort over time. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
Advancing to the clinical trial stage. . . . . . . . . . . . . . . . . . . . . . . . . . . 97
Reaching the top: Systematic reviews and meta-analyses . . . . . . . 97
PART 3: GETTING DOWN AND DIRTY WITH DATA . . . . . . . . . . 99
CHAPTER 8: Getting Your Data into the Computer . . . . . . . . . . . . . . 101

Looking at Levels of Measurement. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
Classifying and Recording Different Kinds of Data. . . . . . . . . . . . . . . . 103
Dealing with free-text data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
Assigning participant study identification (ID) numbers. . . . . . . . . . 104
Organizing name and address data in the study ID crosswalk. . . 104
Collecting categorical data in your research database . . . . . . . . . 105
Table of Contents vii

Recording numerical data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
Entering date and time data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
Checking Your Entered Data for Errors . . . . . . . . . . . . . . . . . . . . . . . . . 109
Creating a File that Describes Your Data File . . . . . . . . . . . . . . . . . . . . 110
CHAPTER 9: Summarizing and Graphing Your Data . . . . . . . . . . . . . 111

Summarizing and Graphing Categorical Data. . . . . . . . . . . . . . . . . . . . 112
Summarizing Numerical Data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
Locating the center of your data. . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
Describing the spread of your data . . . . . . . . . . . . . . . . . . . . . . . . . 119
Numerically expressing the symmetry and shape
of the distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
Structuring Numerical Summaries into Descriptive Tables . . . . . . . . 123
Graphing Numerical Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
Showing the distribution with histograms. . . . . . . . . . . . . . . . . . . . 124
Summarizing grouped data with bars, boxes, and whiskers . . . . 126
Depicting the relationships between numerical variables
with other graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
CHAPTER 10: Having Confidence in Your Results. . . . . . . . . . . . . . . . . . 129

Feeling Confident about Confidence Interval Basics . . . . . . . . . . . . . . 130
Defining confidence intervals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
Understanding and interpreting confidence levels . . . . . . . . . . . . 131
Taking sides with confidence intervals. . . . . . . . . . . . . . . . . . . . . . . 132
Calculating Confidence Intervals. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
Before you begin: Formulas for confidence limits in
large samples. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
The confidence interval around a mean . . . . . . . . . . . . . . . . . . . . . 134
The confidence interval around a proportion. . . . . . . . . . . . . . . . . 135
The confidence interval around an event count or rate . . . . . . . . 136
Relating Confidence Intervals and Significance Testing. . . . . . . . . . . . 137
PART 4: COMPARING GROUPS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
CHAPTER 11: Comparing Average Values between Groups. . . . . . 141

Grasping Why Different Situations Need Different Tests . . . . . . . . . . 142
Comparing the mean of a group of numbers to
a hypothesized value . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
Comparing the mean of two groups of numbers. . . . . . . . . . . . . . 143
Comparing the means of three or more groups of numbers. . . . 143
Comparing means in data grouped on several
different variables. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
Adjusting for a confounding variable when
comparing means . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
Comparing means from sets of matched numbers. . . . . . . . . . . . 145
Comparing means of matched pairs . . . . . . . . . . . . . . . . . . . . . . . . 146
viii Biostatistics For Dummies

Using Statistical Tests for Comparing Averages . . . . . . . . . . . . . . . . . . 146
Surveying Student t tests. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
Assessing the ANOVA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
Running nonparametric tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
Estimating the Sample Size You Need for Comparing Averages . . . . 158
Using formulas for manual calculation . . . . . . . . . . . . . . . . . . . . . . 158
Software and web pages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
CHAPTER 12: Comparing Proportions and Analyzing

Cross-Tabulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
Examining Two Variables with the Pearson Chi-Square Test . . . . . . . 161
Understanding how the chi-square test works. . . . . . . . . . . . . . . . 161
Pointing out the pros and cons of the chi-square test. . . . . . . . . . 167
Modifying the chi-square test: The Yates
continuity correction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
Focusing on the Fisher Exact Test. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
Understanding how the Fisher Exact test works . . . . . . . . . . . . . . 169
Noting the pros and cons of the Fisher Exact test . . . . . . . . . . . . . 170
Calculating Power and Sample Size for Chi-Square and
Fisher Exact Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
CHAPTER 13: Taking a Closer Look at Fourfold Tables. . . . . . . . . . . . 173

Focusing on the Fundamentals of Fourfold Tables . . . . . . . . . . . . . . . 174
Choosing the Correct Sampling Strategy. . . . . . . . . . . . . . . . . . . . . . . . 175
Producing Fourfold Tables in a Variety of Situations. . . . . . . . . . . . . . 176
Describing the association between two binary variables . . . . . . 177
Quantifying associations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
Evaluating diagnostic procedures. . . . . . . . . . . . . . . . . . . . . . . . . . . 183
Investigating treatments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
Looking at inter- and intra-rater reliability. . . . . . . . . . . . . . . . . . . .188
CHAPTER 14: Analyzing Incidence and Prevalence Rates in

Epidemiologic Data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
Understanding Incidence and Prevalence. . . . . . . . . . . . . . . . . . . . . . . 192
Prevalence: The fraction of a population with a
particular condition. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
Incidence: Counting new cases. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
Understanding how incidence and prevalence are related . . . . . 194
Analyzing Incidence Rates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
Expressing the precision of an incidence rate. . . . . . . . . . . . . . . . . 194
Comparing incidences with the rate ratio. . . . . . . . . . . . . . . . . . . . 195
Calculating confidence intervals for a rate ratio. . . . . . . . . . . . . . . 196
Comparing two event rates. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
Comparing two event counts with identical exposure. . . . . . . . . . 197
Estimating the Required Sample Size. . . . . . . . . . . . . . . . . . . . . . . . . . . 198
Table of Contents ix
PART 5: LOOKING FOR RELATIONSHIPS WITH
CORRELATION AND REGRESSION. . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
CHAPTER 15: Introducing Correlation and Regression . . . . . . . . . . . 201

Correlation: Estimating How Strongly Two Variables
Are Associated. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
Lining up the Pearson correlation coefficient. . . . . . . . . . . . . . . . . 202
Analyzing correlation coefficients. . . . . . . . . . . . . . . . . . . . . . . . . . . 203
Regression: Discovering the Equation that Connects
the Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
Understanding the purpose of regression analysis. . . . . . . . . . . . 207
Talking about terminology and mathematical notation . . . . . . . . 208
Classifying different kinds of regression . . . . . . . . . . . . . . . . . . . . . 209
CHAPTER 16: Getting Straight Talk on Straight-Line

Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
Knowing When to Use Straight-Line Regression. . . . . . . . . . . . . . . . . . 213
Understanding the Basics of Straight-Line Regression . . . . . . . . . . . . 215
Running a Straight-Line Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
Taking a few basic steps. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
Walking through an example. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
Interpreting the Output of Straight-Line Regression. . . . . . . . . . . . . . 220
Seeing what you told the program to do. . . . . . . . . . . . . . . . . . . . . 220
Evaluating residuals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
Making your way through the regression table . . . . . . . . . . . . . . . 224
Wrapping up with measures of goodness-of-fit. . . . . . . . . . . . . . . 227
Scientific fortune-telling with the prediction formula . . . . . . . . . . 228
Recognizing What Can Go Wrong with Straight-Line Regression. . . . 229
Calculating the Sample Size You Need. . . . . . . . . . . . . . . . . . . . . . . . . . 230
CHAPTER 17: More of a Good Thing: Multiple Regression. . . . . . . . 233

Understanding the Basics of Multiple Regression . . . . . . . . . . . . . . . . 234
Defining a few important terms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
Being aware of how the calculations work . . . . . . . . . . . . . . . . . . . 235
Executing a Multiple Regression Analysis in Software. . . . . . . . . . . . . 236
Preparing categorical variables. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
Recoding categorical variables as numerical. . . . . . . . . . . . . . . . . . 237
Creating scatter charts before you jump into multiple
regression analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
Taking a few steps with your software. . . . . . . . . . . . . . . . . . . . . . . 241
Interpreting the Output of a Multiple Regression Analysis. . . . . . . . . 241
Examining typical multiple regression output. . . . . . . . . . . . . . . . . 241
Checking out optional output to request. . . . . . . . . . . . . . . . . . . . . 243
Deciding whether your data are suitable for
regression analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
Determining how well the model fits the data . . . . . . . . . . . . . . . . 244
x Biostatistics For Dummies

Watching Out for Special Situations that Arise in Multiple
Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
Synergy and anti-synergy. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
Collinearity and the mystery of the disappearing significance. . . 246
Calculating How Many Participants You Need . . . . . . . . . . . . . . . . . . . 247
CHAPTER 18: A Yes-or-No Proposition: Logistic Regression. . . . . . 249

Using Logistic Regression. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
Understanding the Basics of Logistic Regression. . . . . . . . . . . . . . . . . 251
Fitting a function with an S shape to your data . . . . . . . . . . . . . . . . . . 252
Handling multiple predictors in your logistic model . . . . . . . . . . . 255
Running a Logistic Regression Model with Software. . . . . . . . . . . . . . 256
Interpreting the Output of Logistic Regression. . . . . . . . . . . . . . . . . . . 257
Seeing summary information about the variables. . . . . . . . . . . . . 258
Assessing the adequacy of the model . . . . . . . . . . . . . . . . . . . . . . . 258
Checking out the table of regression coefficients. . . . . . . . . . . . . .259
Predicting probabilities with the fitted logistic formula. . . . . . . . . 260
Making yes or no predictions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .262
Heads Up: Knowing What Can Go Wrong with
Logistic Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
Don’t misinterpret odds ratios for numerical predictors . . . . . . . 267
Beware of the complete separation problem. . . . . . . . . . . . . . . . . 267
Figuring Out the Sample Size You Need for Logistic Regression . . . . 268
CHAPTER 19: Other Useful Kinds of Regression. . . . . . . . . . . . . . . . . . . . 271

Analyzing Counts and Rates with Poisson Regression. . . . . . . . . . . . . 271
Introducing the generalized linear model. . . . . . . . . . . . . . . . . . . . 272
Running a Poisson regression. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
Interpreting the Poisson regression output . . . . . . . . . . . . . . . . . . 275
Discovering other uses for Poisson regression. . . . . . . . . . . . . . . . 276
Anything Goes with Nonlinear Regression . . . . . . . . . . . . . . . . . . . . . . 279
Distinguishing nonlinear regression from other kinds . . . . . . . . . 279
Checking out an example from drug research. . . . . . . . . . . . . . . . 280
Running a nonlinear regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . 282
Interpreting the output. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
Using equivalent functions to fit the parameters
you really want. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
Smoothing Nonparametric Data with LOWESS. . . . . . . . . . . . . . . . . . . 286
Running LOWESS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
Adjusting the amount of smoothing. . . . . . . . . . . . . . . . . . . . . . . . . 289
CHAPTER 20: Getting the Hint from Epidemiologic Inference. . . 291

Staying Clearheaded about Confounding . . . . . . . . . . . . . . . . . . . . . . . 292
Avoiding overloading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
Adjusting for confounders. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 294
Understanding Interaction (Effect Modification) . . . . . . . . . . . . . . . . . 296
Table of Contents xi
Getting Casual about Cause. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
Rothman’s causal pie . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
Bradford Hill’s criteria of causality . . . . . . . . . . . . . . . . . . . . . . . . . . 298
PART 6: ANALYZING SURVIVAL DATA. . . . . . . . . . . . . . . . . . . . . . . 299
CHAPTER 21: Summarizing and Graphing Survival Data. . . . . . . . . 301

Understanding the Basics of Survival Data. . . . . . . . . . . . . . . . . . . . . . 302
Examining how survival times are intervals . . . . . . . . . . . . . . . . . . 302
Recognizing that survival times aren’t normally distributed. . . . . 303
Considering censoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
Looking at the Life-Table Method. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
Making a life table. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
Interpreting a life table. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
Graphing hazard rates and survival probabilities
from a life table. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312
Digging Deeper with the Kaplan-Meier Method. . . . . . . . . . . . . . . 313
Heeding a Few Guidelines for Life-Tables and the
Kaplan-Meier Method. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315
Recording survival times correctly . . . . . . . . . . . . . . . . . . . . . . . . . . 315
Miscoding censoring information. . . . . . . . . . . . . . . . . . . . . . . . . . . 316
CHAPTER 22: Comparing Survival Times . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317

Comparing Survival between Two Groups with the
Log-Rank Test. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
Understanding what the log-rank test is doing. . . . . . . . . . . . . . . . 319
Running the log-rank test on software. . . . . . . . . . . . . . . . . . . . . . . 320
Looking at the calculations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320
Assessing the assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323
Considering More Complicated Comparisons . . . . . . . . . . . . . . . . . . . 324
Estimating the Sample Size Needed for Survival Comparisons . . . . . 324
CHAPTER 23: Survival Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327

Knowing When to Use Survival Regression. . . . . . . . . . . . . . . . . . . . . . 328
Grasping the Concepts behind Survival Regression. . . . . . . . . . . . . . . 329
The steps to perform a PH regression. . . . . . . . . . . . . . . . . . . . . . . 330
Hazard ratios. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334
Executing a Survival Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335
Interpreting the Output of a Survival Regression. . . . . . . . . . . . . . . . . 337
Testing the validity of the assumptions. . . . . . . . . . . . . . . . . . . . . . 339
Checking out the table of regression coefficients. . . . . . . . . . . . . .340
Homing in on hazard ratios and their confidence intervals . . . . . 341
Assessing goodness-of-fit and predictive ability of the model. . . 342
Focusing on baseline survival and hazard functions. . . . . . . . . . . 342
xii Biostatistics For Dummies

How Long Have I Got, Doc? Constructing Prognosis Curves. . . . . . . . 343
Obtaining the necessary output . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343
Finding h. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345
Estimating the Required Sample Size for a Survival Regression . . . . 346
PART 7: THE PART OF TENS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349
CHAPTER 24: Ten Distributions Worth Knowing. . . . . . . . . . . . . . . . . . . 351

The Uniform Distribution. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352
The Normal Distribution. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
The Log-Normal Distribution. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
The Binomial Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 354
The Poisson Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 355
The Exponential Distribution. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 356
The Weibull Distribution. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 356
The Student t Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357
The Chi-Square Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 358
The Fisher F Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359
CHAPTER 25: Ten Easy Ways to Estimate How

Many Participants You Need. . . . . . . . . . . . . . . . . . . . . . . . . 361
Comparing Means between Two Groups . . . . . . . . . . . . . . . . . . . . . . . 362
Comparing Means among Three, Four, or Five Groups. . . . . . . . . . . . 362
Comparing Paired Values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363
Comparing Proportions between Two Groups. . . . . . . . . . . . . . . . . . . 363
Testing for a Significant Correlation. . . . . . . . . . . . . . . . . . . . . . . . . . . . 363
Comparing Survival between Two Groups . . . . . . . . . . . . . . . . . . . . . . 364
Scaling from 80 Percent to Some Other Power . . . . . . . . . . . . . . . . . . 365
Scaling from 0.05 to Some Other Alpha Level. . . . . . . . . . . . . . . . . . . . 365
Adjusting for Unequal Group Sizes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366
Allowing for Attrition. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366
INDEX. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 369
Table of Contents xiii

Introduction
B
iostatistics is the practical application of statistical concepts and techniques
to topics in the biology and life sciences fields. Because these are broad
fields, biostatistics covers a very wide area. It is used when studying many
types of experimental units, from viruses to trees to fleas to mice to people.
Biostatistics involves designing research studies, safely conducting human
research, collecting and verifying research data, summarizing and displaying the
data, and analyzing the data to answer research hypotheses and draw meaningful
conclusions.
It is not possible to cover all the subspecialties of biostatistics in one book, because
such a book would have to include chapters on molecular biology, genetics,
agricultural studies, animal research (both inside and outside the lab), clinical
trials, and epidemiological research. So instead, we focus on the most widely
applicable topics of biostatistics and on the topics that are most relevant to human
research based on a survey of graduate-level biostatistics curricula from major
universities.
About This Book

We wrote this book to be used as a reference. Our intention was for you to pull out
this book when you want information about a particular topic. This means you
don’t have to read it from beginning to end to find it useful. In fact, you can jump
directly to any part that interests you. We hope you’ll be inclined to look through
the book from time to time, open it to a page at random, read a page or two, and
get a useful reminder or pick up a new fact.
Only in a few places does this book provide detailed steps about how to perform a
particular statistical calculation by hand. Instruction like that may have been
necessary in the mid-1900s. Back then, statistics students spent hours in a com-
puting lab, which is a room that had an adding machine. Thankfully, we now have
statistical software to do this for us (see Chapter 4 for advice on choosing statistical
software). When describing statistical tests, our focus is always on the concepts
behind the method, how to prepare your data for analysis, and how to interpret
the results. We keep mathematical formulas and derivations to a minimum. We
Introduction 1
only include them when we think they help explain what’s going on. If you really
want to see them, you can find them in many biostatistics textbooks, and they’re
readily available online.
Because good study design is crucial for the success of any research, this book
gives special attention to the design of both epidemiologic studies and clinical tri-
als. We also pay special attention to providing advice on how to calculate the num-
ber of participants you need for your study. You will find easy-to-apply examples
of sample-size calculations in the chapters describing significance tests in
Parts 4, 5, and 6, and in Chapter 25.
Foolish Assumptions
We wrote this book to help several kinds of people. We assume you fall into one of
the following categories:
» Students at the undergraduate or graduate level who are taking a course in

biostatistics and want help with the topics they’re studying in class
» Professionals who have had no formal biostatistical training, and possibly no

statistical training at all, who now must analyze biological or research data as
part of their work
» Doctors, nurses, and other healthcare professionals who want to carry out
human research
If you’re interested in biostatistics, then you’re no dummy! But perhaps you

sometimes feel like a dummy when it comes to biostatistics, or statistics in gen-
eral, or even mathematics. Don’t feel bad. We both have felt that way many times
over the years. In fact, we still feel like that whenever we are propelled into an
area of biostatistics with which we are unfamiliar, because it is new to us. (If you
haven’t taken a basic statistics course yet, you may want to get Statistics For
Dummies by Deborah J. Rumsey, PhD — published by Wiley — and read parts of
that book first.)
What is important to keep in mind when learning biostatistics is that you don’t
have to be a math genius to be a good biostatistician. You also don’t need any
special math skills to be an excellent research scientist who can intelligently
design research studies, execute them well, collect and analyze data properly, and
draw valid conclusions. You just have to have a solid grasp of the basic concepts
and know how to utilize statistical software properly to obtain the output you need
and interpret it.
2 Biostatistics For Dummies

Icons Used in This Book
Icons are the little graphics in the margins of this book, and are used to draw your
attention to certain kinds of material. Here’s what they mean:
This icon signals information especially worth keeping in mind. Your main take-
aways from this book should be the material marked with this icon.
We use this icon to flag explanations of technical topics, such as derivations and
computational formulas that you don’t have to know to do biostatistics. They are
included to give you deeper insight into the material.
This icon refers to helpful hints, ideas, shortcuts, and rules of thumb that you can
use to save time or make a task easier. It also highlights different ways of thinking
about a topic or concept.
This icon alerts you to discussion of a controversial topic, a concept that is often
misunderstand, or a pitfall or common mistake to guard against in biostatistics.
Beyond the Book

In addition to the abundance of information and guidance related to using biosta-
tistics for analysis of research data that we provide in this book, you get access to
even more help and information online at Dummies.com. Check out this book’s
online Cheat Sheet. Just go to www.dummies.com and search for “Biostatistics For
Dummies Cheat Sheet.”
Where to Go from Here

You’re already off to a good start! You’ve read this introduction, so you have a
good idea of what this book is all about. For a more detailed list of topics, take a
look at the Contents at a Glance. This drills down into each part and shows you
what each chapter is all about. Finally, skim through the full-blown Table of Con-
tents, which drills further down into each chapter, showing you the headings for
the sections and subsections of that chapter.
Introduction 3
If you want to get the big picture of what biostatistics encompasses and the areas
of biostatistics covered in this book, then read Chapter 1. This is a top-level
overview of the book’s topics. Here are a few other special parts of this book you
may want to jump into first, depending on your interest:
» If you’re uncomfortable with mathematical notation, then Chapter 2 is the

place to start.
» If you want a quick refresher on basic statistics like what you would learn in
a typical introductory course, then read Chapter 3.
» You can get an introduction to human research and clinical trials in

Chapters 5, 7, and 20.
» If you want to learn about collecting, summarizing, and graphing data, jump
to Part 3.
» If you need to know about working with survival data, you can go right to
Part 6.
» If you’re puzzled about a particular statistical distribution function, then look

at Chapter 24.
» And if you need to calculate some quick sample-size estimates, turn to

Chapter 25.

1
Getting
Started with
Biostatistics
IN THIS PART . . .
Get comfortable with mathematical notation that

uses numbers, special constants, variables, and
mathematical symbols — a must for all you
mathophobes.
Review basic statistical concepts you may have

learned previously, such as probability, randomness,
populations, samples, statistical inference, and more.
IN THIS CHAPTER
» Getting up to speed on the

prerequisites for biostatistics
» Understanding the human research

environment
» Surveying the specific procedures

used to analyze biological data
» Estimating how many participants

you need
» Working with distributions
Chapter 1
Biostatistics 101
B
iostatistics deals with the design and execution of scientific studies involv-
ing biology, the acquisition and analysis of data from those studies, and the
interpretation and presentation of the results of those analyses. This book
is meant to be a useful and easy-to-understand companion to the more formal
textbooks used in graduate-level biostatistics courses. Because most of these
courses teach how to analyze data from epidemiologic studies and clinical trials,
this book focuses on that as well. In this first chapter, we introduce you to the
fundamentals of biostatistics.
Brushing Up on Math and Stats Basics

Chapters 2 and 3 are designed to bring you up to speed on the basic math and sta-
tistical background that’s needed to understand biostatistics and give you supple-
mentary information or context that you may find useful while reading the rest of
this book.
» Many people feel unsure of themselves when it comes to understanding

mathematical formulas and equations. Although this book contains fewer
CHAPTER 1 Biostatistics 101 7

formulas than many statistics books, we include them when they help
illustrate a concept or describe a calculation that’s simple enough to do by
hand. But if you’re a real mathophobe, you probably dread looking at any
chapter that has a math expression anywhere in it. That’s why we include
Chapter 2, “Overcoming Mathophobia” to show you how to read and
understand the basic mathematical notation we use in this book. We
cover everything from basic mathematical operations to functions
and beyond.
» If you’re in a graduate-level biostatistics course, you’ve probably already taken

one or two introductory statistics courses. But that may have been a while ago,
and you may feel unsure of your knowledge of the basic statistical concepts.
Or you may have little or no formal statistical training but now find yourself in
a work situation where you interact with clinical researchers, participate in the
design of research projects, or work with the results from biological research.
If so, read Chapter 3, which provides an overview of the fundamental concepts
and terminology of statistics. There, you get the scoop on topics such as
probability, randomness, populations, samples, statistical inference, accuracy,
precision, hypothesis testing, nonparametric statistics, and simulation
techniques.
Doing Calculations with

the Greatest of Ease
For instructional purposes, some chapters in this book include step-by-step
instructions for performing statistical tests and analyses by hand. We include
such instruction only to illustrate the concepts that are involved in the procedure
or to demonstrate calculations that are simple to do manually.
However, we demonstrate many of the statistical functions we talk about in this

book using R, which is a free, open-source software package. If you are in a class
and assigned a particular software package to use, you will have to use that soft-
ware for the course, which may be commercial software associated with a fee.
However, if you are learning on your own, you may choose to use open-source
software, which is free. Chapter 4 provides guidance on both commercial and
free software.
8 PART 1 Getting Started with Biostatistics

Concentrating on Epidemiologic Research
This book covers topics that are applicable to all areas of biostatistics, concentrat-
ing on methods that are especially relevant to epidemiologic research — studies
involving people. This includes clinical trials, which are experiments done to
develop therapeutic interventions such as drugs. Because policy in healthcare is
often based on the results from clinical trials, if you make mistake analyzing
clinical trial data, it can have disastrous and wide-ranging human and financial
consequences. Even if you don’t expect to ever work in a domain that relies heav-
ily on clinical trials (such as drug development research), ensuring that you have
a working knowledge of how to manage the statistical issues seen in clinical trials
is critical.
Three chapters discuss clinical trials:
» Chapter 5 describes the statistical aspects of clinical trials as three phases.

First, it covers the design phase, where a study protocol is written. Next, it
describes the execution phase, where data are collected, and efforts are made
to prevent invalid or missing data. In the final phase, data from the study are
analyzed and interpreted to answer the hypotheses.
» Chapter 7 presents epidemiologic study designs and explains the importance

of the clinical trial as a study design.
» Chapter 20 explains the role well-designed clinical trials play in accruing

evidence of causal inference in biostatistics.
Much of the work in biostatistics is using data from samples to make inferences
about the background population from which the sample was drawn. Now that we
have large databases, it is possible to easily take samples of data. Chapter 6
provides guidance on different ways to take samples of larger populations so you
can make valid population-based estimates from these samples. Sampling is
especially important when doing observational studies. While clinical trials
covered are experiments, where participants are assigned interventions, in obser-
vational studies, participants are merely observed, with data collected and statis-
tics performed to make inferences. Chapter 7 describes these observational study
designs, and the statistical issues that need to be considered when analyzing data
arising from such studies.
Data used in biostatistics are often collected in online databases, but some data
are still collected on paper. Regardless of the source of the data, they must be put
into electronic format and arranged in a certain way to be able to be analyzed
using statistical software. Chapter 8 is devoted to describing how to get your data
into the computer and arrange it properly so it can be analyzed correctly. It also

describes how to collect and validate your data. Then in Chapter 9, we show you
how to summarize each type of data and display it graphically. We explain how to
make bar charts, box-and-whiskers charts, and more.
Drawing Conclusions from Your Data

Most statistical analysis involves inferring, or drawing conclusions about the pop-
ulation at large based on your observations of a sample drawn from that popula-
tion. The theory of statistical inference is often divided into two broad sub-theories:
estimation theory and decision theory.
Statistical estimation theory

Chapter 10 deals with statistical estimation theory, which addresses the question of
how accurately and precisely you can estimate a population parameter from the
values you observe in your sample. For example, you may want to estimate the
mean blood hemoglobin concentration in adults with Type II diabetes, or the true
correlation coefficient between body weight and height in certain pediatric popu-
lations. Chapter 10 describes how to estimate these parameters by constructing a
confidence interval around your estimate. The confidence interval is the range that
is likely to include the true population parameter, which provides an idea of the
precision of your estimate.
Statistical decision theory

Much of the rest of this book deals with statistical decision theory, which is how to
decide whether some effect you’ve observed in your data reflects a real difference
or association in the background population or is merely the result of random
fluctuations in your data or sampling. If you measure the mean blood hemoglobin
concentration in two different samples of adults with Type II diabetes, you will
likely get a different number. But does this difference reflect a real difference
between the groups in terms of blood hemoglobin concentration? Or is this differ-
ence a result of random fluctuations? Statistical decision theory helps you decide.
In Part 4, we cover statistical decision theory in terms of comparing means and

proportions between groups, as well as understanding the relationship between
two or more variables.

Comparing groups
In Part 4, we show you different ways to compare groups statistically.
» In Chapter 11, you see how to compare average values between two or
more groups by using t tests and ANOVAs. We also describe their nonpara-
metric counterparts that can be used with skewed or other non-normally
distributed data.
» Chapter 12 shows how to compare proportions between two or more groups,

such as the proportions of patients responding to two different drugs, using
the chi-square and Fisher Exact tests on cross-tabulated (cross-tab) data.
» Chapter 13 focuses on one specific kind of cross-tab called the fourfold table,
which has exactly two rows and two columns. Because the fourfold table
provides the opportunity for some particularly insightful calculations, it’s
worth a chapter of its own.
» In Chapter 14, you discover how the terminology used in epidemiologic

studies is applied to specifically formatted fourfold tables to calculate
incidence and prevalence rates.
Looking for relationships between variables

Epidemiology and biostatistics are interested in causal inference, which means try-
ing to figure out what causes particular outcomes in biological research. While it
is possible to look at the relationship between two variables in a bivariate analysis,
regression analysis is the part of statistics that enables you to explore the rela-
tionship between multiple variables and one outcome in the same model so you
can evaluate their relative cause of the outcome. Here are some use-cases for
regression:
» You may want to know whether there’s a statistically significant association

between one or more variables and an outcome, even if there are other
variables in the model. You may ask: Does being overweight increase the
likelihood of getting liver cancer? Or: Is exercising fewer hours per week
associated with higher blood pressure measurements? In answering both
of those questions, you may want to control other variables known to
influence the outcome.
» You may want to develop a formula for predicting the value of a variable from
the observed values of one or more other variables. For example, you may
want to predict how long a newly diagnosed cancer patient may survive
based on their age, obesity status, and medical history.

» You may be fitting a theoretical formula to some data to estimate one of
the parameters appearing in that formula. An example of such a problem
is determining how fast the kidneys can remove a drug from the body, which
is called a terminal elimination rate constant. This can be estimated from
measurements of drug concentration in the blood taken at various times
after taking a dose of the drug.
Regression analysis can manage all these tasks and many more. Regression is so
important in biological research that all the chapters in Part 5 are focused on some
aspect of regression.
If you have never learned correlation and regression analysis, read Chapter 15,
which introduces these topics. We cover simple straight-line regression in
Chapter 16, which includes one predictor variable. We extend that to cover
multiple regression with more than one predictor variable in Chapter 17. These
three chapters deal with ordinary linear regression, where you’re trying to predict
the value of a numerical outcome variable from one or more other variables. An
example would be trying to predict mean blood hemoglobin concentration using
variables like age, blood pressure level, and Type II diabetes status. Ordinary
linear regression uses a formula that’s a simple summation of terms, each of
which consists of a predictor variable multiplied by a regression coefficient.
But in real-world biological and epidemiologic research, you encounter more

complicated relationships. Chapter 18 describes logistic regression, where the out-
come is the occurrence or non-occurrence of an event (such as being diagnosed
with Type II diabetes), and you want to predict the probability that the event will
occur. You also find out about several other kinds of regression in Chapter 19:
» Poisson regression, where the outcome is the number of events that occur in
an interval of time
» Nonlinear least-squares regression, where the relationship between the

predictors and numerical outcome can be more complicated than a simple
summation of terms in a linear model
» LOWESS curve-fitting, where you fit a custom function to describe your data
Finally, Part 5 ends with Chapter 20, which provides guidance on the mechanics
of regression modeling, including how to develop a modeling plan, and how to
choose variables to include in models.

A Matter of Life and Death: Working
with Survival Data
Sooner or later, everyone dies, and in biological research, it becomes especially
important to characterize that sooner-or-later part as accurately as possible using
survival analysis techniques. But characterizing survival can get tricky. It’s possi-
ble to say that patients may live an average of 5.3 years after they are diagnosed
with a particular disease. But what is the exact survival experience? Imagine you
do a study with patients who have this disease. You may ask: Do all patients tend
to live around five or six years, or do half the patients die within the first few
months, and the other half survive ten years or more? And what if some patients
live longer than the observational period of your study? How do you include them
in your analysis? And what about participants who stopped returning calls from
your study staff? You do not know if these dropouts went on to live or die. How do
you include their data in your analysis?
The need to study survival with data like these led to the development of survival
analysis techniques. But survival analysis is not only intended to study the
outcome of death. You can use survival analysis to study the time to the first
occurrence of non-death events as well, like remission or recurrence of cancer,
the diagnosis of a particular condition, or the resolution of a particular condition.
Survival analysis techniques are presented in Part 6.
Getting to Know Statistical Distributions

Statistics books always contain tables, so why should this one be any different?
Back in the not-so-good old days, when analysts had to do statistical calculations
by hand, they needed to use tables of the common statistical distributions to com-
plete the calculation of the significance test. They needed tables for the normal
distribution, Student t, chi-square, Fisher F, and others. Now, software does all
this for you, including calculating exact p values, so these printed tables aren’t
necessary anymore.
But you should still be familiar with the common statistical distributions that may
describe the fluctuations in your data, or that may be referenced in the course of
performing a statistical calculation. Chapter 24 contains a list of commonly used
distribution functions, with explanations of where you can expect to encounter
those distributions and what they look like. We also include a description of some
of their properties and how they’re related to other distributions. Some of them
are accompanied by a small table of critical values, corresponding to statistical
significance at α = 0.05.

Figuring Out How Many Participants
You Need
Of all the statistical challenges a researcher may encounter, none seems to instill
as much apprehension and insecurity as having to estimate the number of partici-
pants needed for a study. While smaller sample sizes mean less data collection
work, you want to make sure your target sample size is large enough so that in the
end, your study has sufficient power. You want to conduct a study with a high
probability of yielding a statistically significant result if the hypothesized effect is
truly present in the population.
Because sample-size estimation is such an important part of the design of any

research project, this book shows you how to make those estimates for the
situations you’re likely to encounter when doing biological research. As we
describe each statistical test in Parts 4, 5, 6, and 7, we explain how to estimate
the number of participants needed to provide sufficient power for that test. In
addition, Chapter 25 describes ten simple rules for getting a “quick and dirty”
estimate of the required sample size.

IN THIS CHAPTER
» Reading mathematical notation
» Understanding formulas and what

they mean
» Working with arrays (collections of

numbers)
Chapter 2
Overcoming
Mathophobia: Reading
and Understanding
Mathematical
Expressions
L
et’s face it: Many people fear math, and statistical calculations require math.
In this chapter, we help you become more comfortable with reading
mathematical expressions, which are combinations of numbers, letters, math
operations, punctuation, and grouping symbols. We also help you become more
comfortable with equations, which connect two expressions with an equal sign.
And we review formulas, which are equations designed for specific calculations.
(For simplicity, for the rest of the chapter, we use the term formula to refer to
expressions, equations, and formulas.) We also explain how to write formulas,
which you need to know in order to tell a computer how to do calculations with
your data.
CHAPTER 2 Overcoming Mathophobia: Reading and Understanding Mathematical Expressions 15

We start the chapter by showing you how to interpret the mathematical formulas
you encounter throughout this book. We don’t deconstruct the intricacies of com-
plicated mathematical operations. Instead, we explain how mathematical opera-
tions are indicated in this book. If you feel unsure of your grasp on algebra,
consider reviewing Algebra I For Dummies and Algebra II For Dummies, which are
both written by Mary Jane Sterling and published by Wiley.
Breaking Down the Basics of

Mathematical Formulas
One way to think of a mathematical formula is as a shorthand way to describe how
to do a certain calculation. Formulas are made up of numbers, constants, and
variables interspersed with symbols that indicate mathematical operations, punc-
tuation, and typographic effects. Formulas are constructed using relatively stan-
dardized rules that have evolved over centuries. In the following sections, we
describe two different kinds of formulas that you encounter in this book: typeset
and plain text. We also describe two of the building blocks from which formulas
are created: constants and variables.
Displaying formulas in different ways

Formulas can be expressed in print in two different formats: typeset format and
plain text format:
» A typeset format utilizes special symbols, and when printed, the formula is
spread out in a two-dimensional structure, like this:
n 2
i 1
xi m
SD
n 1
» A plain text format prints the formula out as a single line, which is easier to
type if you’re limited to the characters on a keyboard:
SD sqrt sum x[i] m ^ 2, i , 1, n / n 1
You must know how to read both types of formula displays — typeset and plain
text. The examples in this chapter show both styles. But you may never have to
construct a professional-looking typeset formula (unless you’re writing a book,
like we’re doing right now). On the other hand, you’ll almost certainly have to
write out plain text formulas as part of organizing, preparing, editing, and ana-
lyzing your data.

Checking out the building
blocks of formulas
No matter how they’re written, formulas are essentially recipes that tell you how
to calculate a result, or how a value is defined. To cook up your own result, you
need to know how to follow the recipe. When initially approaching a formula, it’s
helpful to start by examining the building blocks from which formulas are con-
structed. These include constants, which are numbers with specified values, and
variables, which represent quantities that can take on different values at different
times.
Constants
Constants are values that can be represented explicitly (using the numerals 0
through 9 with or without a decimal point), or symbolically (using a letter in the
Greek or Roman alphabet). Symbolic constants represent a particular value impor-
tant in mathematics, physics, or some other discipline, such as:
» The Greek letter π usually represents 3.14159 (plus a zillion more digits). This
Greek letter is spelled pi and pronounced pie, and represents the ratio of the
circumference of any circle to its diameter.
» The number 2.71828 (plus a zillion more digits) is represented by e (which is

italicized when written, and is pronounced like the letter “e”). Later in this
chapter, we describe one way e is used. You’ll see e in statistical formulas
throughout this book and in almost every other mathematical and statistical
textbook. Whenever you see an italicized e in this book, it refers to the
number 2.718 unless we explicitly say otherwise.
The official mathematical definition of e is: The value of the expression

1 1/n n , which approaches infinity as n gets larger and larger. Unlike π, e
has no simple geometrical interpretation. Here is an example used to help
learners envision e: Assume you put exactly one dollar in a bank account that’s
paying 100 percent annual interest, compounded continuously. After exactly
one year, your account will have e dollars in it. That includes the interest on
your original dollar, plus the interest on the interest — about $1.72 (to the
nearest penny) — added to the original dollar for a total of $2.72. (This is just
an example. We don’t think there is a single bank out there advertising annual
returns in terms of e!)
Mathematicians and scientists use lots of other specific Greek and Roman letters
as symbols for specific constants, but you need only a few of them in your biosta-
tistics work. π and e are the most common, and we define others in this book as
they come up in topics we present.

Variables
The term variable has two slightly different meanings:
» In mathematics and engineering, a variable is a symbol that represents

some quantity in a formula. It is usually a letter of the alphabet. You are
probably used to seeing variables like x and y in algebra, for example.
» In statistics and computer science, a variable is a name referring to a single

data value or an entire field, which is a column of data in a spreadsheet or
database. The variable name is made up of letters (like SBP for systolic blood
pressure), but may also contain numbers (such as SBP1, SBP2, and SBP3).
Technically, the variable name refers to a place in the computer’s memory
where the data value or field is stored. For example, a computer programmer
writing a statistical software program may ask if the variable SBP is greater
than or equal to 120 mmHg.
The names of variables may be written in uppercase or lowercase letters depend-

ing upon typographic conventions or preferences, or on the requirements of the
software being used.
Variables are always italicized in typeset formulas, but not in plain text formulas.
Focusing on Operations Found in Formulas

A formula tells you how the building blocks of numbers, constants, and variables
are to be combined. In other words, a formula is a recipe for the calculations
you’re supposed to carry out on these quantities. But formulas are not always easy
to read. A particular symbol — such as the minus sign — can be interpreted dif-
ferently, depending upon the context of the formula. Also, a particular mathemat-
ical operation like multiplication can be represented in different ways in a formula.
In the following sections we explain the basic mathematical operations you see in
formulas throughout this book and describe two types of equations you’ll encoun-
ter in statistical books and articles.
Basic mathematical operations

The four basic mathematical operations are addition, subtraction, multiplication,
and division (ah, yes — the basics you learned in elementary school). Different
symbols are associated with these operations, as you discover in the following
sections.

Addition and subtraction
Addition and subtraction are always indicated by the + and – symbols, respec-
tively, placed between two numbers or variables. Compared to the plus sign, the
minus sign can be tricky when it comes to interpreting it in a formula.
» A minus sign placed immediately before a number indicates a negative

quantity. For example, –5° indicates five degrees below 0, and –5 kg indicates
a weight loss of 5 kilograms.
» A minus sign placed immediately before a variable tells you to reverse the
sign of the value of the variable. Therefore, –x means that if x is positive, you
should now make it negative. But it also means that if x is negative, make it
positive (so, if x was –5 kg, then –x would be 5 kg). Used this way, the minus
sign is referred to as a unary operator because it’s acting on only one variable.
Multiplication
The word term is generic for an individual item or element in a formula. Multipli-
cation of terms is indicated in several ways, as shown in Table 2-1.
TABLE 2-1 Multiplication Options

What It Is Example Where It’s Used
Asterisk 2*5 Plain text formulas, but almost never in

typeset formulas
Cross 2 5 Typeset formula, between two variables or

two constants being multiplied together
Raised dot 2 5 Typeset formula
Term is immediately in front of a 2 5 3 16 Typeset formula

parenthesized expression
Brackets and curly braces 2 6 5 3 /2 20 Typeset formula containing “nested”

parentheses
Two or more terms running 2 r versus 2 r In typeset formulas only

together
You can put terms right next to each other to imply multiplication only when it’s
perfectly clear from the context of the formula that the authors are using only
single-letter variable names (like x and y), and that they’re describing calcula-
tions where it makes sense to multiply those variables together. In other words,
you can’t put numeric terms right after one another to imply multiplication,

meaning you can’t replace 5 × 3 with 53, because 53 is an actual number itself. And
you shouldn’t replace variables like length × width with lengthwidth, because it
looks like you’re referring to a single variable named lengthwidth.
Division
Like multiplication, division can be indicated in several ways:
» With a slash (/) in plain text formulas: Distance/Time
» With a division symbol (÷) in typeset formulas: Distance ÷ Time
» With a long horizontal bar in typeset formulas:

Distance
Time
Powers, roots, and logarithms

In the next section, we cover powers, roots, and logarithms, all three of which are
related to the idea of repeated multiplication.
Raising to a power
Raising to a power is a shorthand way to indicate repeated multiplication by the
same number. You indicate raising to a power by:
» Superscripting in typographic formulas, such as 5 3
» Using ** in plain text formulas, such as 5 **3
» Using ^ in plain text formulas, such as 5 ^ 3
All the preceding expressions are read as “five to the third power,” “five to the
power of three,” or “five cubed.” It says to multiply three fives together: 5 × 5 × 5,
which gives you 125.
Here are some other features of power:
» A power doesn’t have to be a whole number. You can raise a number to a

fractional power (such as 3.8). You can’t visualize this in terms of repeated
multiplications, but your scientific calculator can show you that 2.6 3.8 is equal
to approximately 37.748.

» A power can be negative. A negative power indicates the reciprocal of the
quantity, which is when you divide the quantity by 1 (meaning 1/x ). So x 1
means 1 divided by x, and in general, x n is the same as 1/x n (such as 2–3 = ½).
Remember the constant e (2.718. . .)? Almost every time you see e used in a for-
mula, it’s being raised to some power. This means you almost always see e with
an exponent after it. Raising e to a power is called exponentiating, and another way
of representing e x in plain text is exp(x). Remember, x doesn’t have to be a whole
number. By typing =exp(1.6) in the formula bar in Microsoft Excel (or doing the
equation on a scientific calculator), you see that exp(1.6) equals approximately
4.953. We talk more about exponentiating in other book sections, especially
Chapters 18 and 24.
Taking a root
Taking a root involves asking the power question backwards. In other words, we
ask: “What base number, when raised to a certain power, equals a certain num-
ber?” For example, “What number, when raised to the power of 2 (which is
squared), equals 100?” Well, 10 10 (also expressed 10 2) equals 100, so the square
root of 100 is 10. Similarly, the cube root of 1,000,000 is 100, because 100 100 100
(also expressed 100 3 ) equals a million.
Root-taking is indicated by a radical sign (√) in a typeset formula, where the term
from which we intend to take the root is located “under the roof” of the radical
sign, as 25 is shown here: 25 . If no numbers appear in the notch of the radical
sign, it is assumed we are taking a square root. Other roots are indicated by putt-
ing a number in the notch of the radical sign. Because 2 8 is 256, we say 2 is the
eighth root of 256, and we put 8 in the notch of the radical sign covering 256, like
this: 8 256 . You also can indicate root-taking by expressing it different ways used
in algebra: n x is equal to x 1/n and can be expressed as x ^ 1/n in plain text.
Looking at logarithms
In addition to root-taking, another way of asking the power question backwards
is by saying, “What exponent (or power) must I raise a particular base number to
in order for it to equal a certain number?” For root-taking, in terms of using a
formula, we specify the power and request the base. With logarithms, we specify
the base and request the power (or exponent).
For example, you may ask, “What power must I raise 10 to in order to get
1,000?” The answer is 3, because 10 3 1, 000. You can say that 3 is the logarithm
of 1,000 (for base 10), or, in mathematical terms: Log 10 1, 000 3. Simi-
larly, because 2 8 256, you say that Log 2 256 8. And because e 1.6 4.953,
then Log e 4.953 1.6 .

There can be logarithms to any base, but three bases occur frequently enough to
have their own nicknames:
» Base-10 logarithms are called common logarithms.
» Base-e logarithms are called natural logarithms.
» Base-2 logarithms are called binary logarithms.
The logarithmic function naming is inconsistent among different authors, pub-

lishers, and software writers. Sometimes Log means natural logarithm, and some-
times it means common logarithm. Often Ln is used for natural logarithm, and Log
is used for common logarithm. Names like Log10 and Log2 may also be used to
identify the base.
The most common kind of logarithm used in this book is the natural logarithm, so
in this book we always use Log to indicate natural (base-e) logarithms. When we
want to refer to common logarithms, we use Log 10 , and when referring to binary
logarithms, we use Log 2 .
An antilogarithm (usually shortened to antilog) is the inverse of a logarithm. As an

example of an antilog, if y is the log of x, then x is the antilog of y. For another
example, the base-10 logarithm of 1,000 is 3, so the base-10 antilog of 3 is 1,000.
Calculating an antilog is exactly the same as raising the base to the power of the
logarithm. That is, the base-10 antilog of 3 is the same as 10 raised to the power
of 3 (which is 10 3, or 1,000). Similarly, the natural antilog of any number is e
(2.718) raised to the power of that number. As an example, the natural antilog of
5 is e 5, or approximately 148.41.
Factorials and absolute values

So far we’ve covered mathematical operators that are written either between the
two numbers, which are the subject of the operation (such as the plus in 5 + 8), or
before the number it operates on if there is only one number (like the minus sign
used as a unary operator described earlier, as in –5°). Next we cover factorials and
absolute values, which are mathematical operators that have a unique format in
typeset expressions.
Factorials
Although a statistical formula may contain an exclamation point, that doesn’t
mean that you should sound excited when you read the formula aloud (although it
may be tempting to do so!). An exclamation mark (!) after a number is shorthand

for calculating that number’s factorial. To do that, you write down all the whole
numbers from 1 to the factorial number in a row, and then multiply them all
together. For example, the expression 5!, which is read as five factorial, means to
calculate 1 2 3 4 5 (which equals 120).
Even though standard keyboards have a ! key, most computer programs and
spreadsheets don’t let you use ! to indicate factorials. For example, to do the cal-
culation of 5! in Microsoft Excel, you use the formula =FACT(5).
Here are a few factorials fun facts:
» Factorials can be very large. For example, 10! is 3,628,800, and 170! is about
7.3 10 306 , which is close to the processing limits for many computers.
» 0! isn’t 0, but is actually 1. Actually, it’s the same as 1!, which is also 1. That may
not make obvious sense, but is true, so you can memorize it.
» The definition of factorial can be extended to fractions and even to negative

numbers. But good news! You don’t have to deal with those kinds of factorials
in this book.
Absolute values
The term absolute value refers to the value of a number when it is positive (mean-
ing it has no minus sign before it). You indicate absolute value by placing vertical
bars immediately to the left and right of the number. So |5.7| equals 5.7, and
|–5.7| also equals 5.7. Even though most keyboards have the | (pipe) symbol, the
absolute value is usually indicated in plain text formulas as abs(5.7).
Functions
In this book, a function is a set of calculations that accepts one or more numeric
values (called arguments) and produces a numeric result. Regardless of typeset or
plain text, a function is indicated in a formula by the function name followed by a
set of parentheses that contain the argument or arguments. Here’s an example of
the function square root of x: sqrt(x).
The most commonly used functions have been given standard names. The preced-
ing sections in this chapter covered some of these, including sqrt for square root,
exp for exponentiate, log for logarithm, ln for natural log, fact for factorial, and abs
for absolute value.

When writing formulas with functions using software, be aware that each soft-
ware may have rules about case-sensitivity. It may require all caps, all lowercase,
or first-letter capitalization. Make sure to check the software’s documentation for
guidance (Chapter 4 discusses different statistical software packages.)
Simple and complicated formulas

Simple formulas have one or two numbers and only one mathematical operator
(for example, 5 3 ). But most statistical formulas you’ll encounter are more
complicated, with two or more operators and variables.
Whether doing calculations manually or using software, you need to ensure that
you do your formula calculations in the correct order (called the order of operation).
If you evaluate the terms and operations in the formula in the wrong order, you
will get incorrect results. In a complicated formula, the order in which you evalu-
ate the terms and operations is governed by the interplay of several rules arranged
in a hierarchy. Most computer programs try to follow the customary conventions
that apply to typeset formulas, but you need to check software’s documentation to
be sure.
Here’s a typical set of operator hierarchy rules. Within each hierarchical level,
operations are carried out from left to right:
1. Evaluate any terms and operations within parentheses, brackets, curly

braces, or absolute-value bars first, including terms inside parentheses that
follow the name of a function. Please note that nested functions are evaluated
inside out, so additional parentheses may be needed to prevent any confusion.
2. Evaluate negation, factorials, powers, and roots.
3. Evaluate multiplication and division.
4. Evaluate addition and subtraction.
In a typeset fraction, evaluate terms and operations above the horizontal bar (the
numerator) first, then terms and operations below the bar (the denominator)
next. After that, divide the numerator by the denominator.
Equations
An equation has two expressions with an equal sign between them. Most equations
appearing in this book have a single variable name to the left of the equal sign and
a formula to the right, like this: SEM SD / N . This style of equation defines the
variable appearing on the left in terms of the calculations specified on the right.

In doing so, it also provides the “cookbook” instructions for calculating the result,
which in this case is the SEM for any values of SD and N.
The book also contains another type of equation that appears in algebra, asserting
that the terms on the left side of the equation are equal to the terms on the right.
For example, the equation x 2 3 x asserts that x is a number that, when added
to 2, produces a number that’s 3 times as large as the original x. Algebra teaches
you how to solve this expression for x, and it turns out that the answer is x 1.
Counting on Collections of Numbers

A variable can refer to one value or to a collection of values called arrays. Arrays
can come with one or more dimensions.
One-dimensional arrays
A one-dimensional array can be thought of as a list of values. For instance, you
may record a list of fasting glucose values (in milligrams per deciliter, mg/dL )
from five study participants as 86, 110, 95, 125, and 64. You could use the variable
name Gluc to refer to this array containing five numbers, or elements. Using the
term Gluc in a formula refers to the entire five-element array.
You can refer to one particular element of this array (meaning one glucose mea-
surement) in several ways. You can use the index of the array, which is the number
that indicates the position of the element to which you are referring in the array.
» In a typeset formula, indices are typically indicated using subscripts. For

example, Gluc 3 refers to the third element in the array (which would be 95 in
our example).
» In a plain text formula, indices are typically indicated using brackets (such as
Gluc[3]).
The index can be a variable like I, so Gluc[i] would refer to the ith element of the
array. The term ith means the variable would be allowed to take on any value
between 1 and the maximum number of elements in the array (which in this case
would be 5).
In some programming languages and statistical books and articles, the indices
start at 0 for the first element, 1 for the second element, and so on, which can be
confusing. In this book, all arrays are indexed starting at 1.

Higher-dimensional arrays
Two-dimensional arrays can be understood as a table of values with rows and
columns, like a block of cells in a spreadsheet. There are also higher-dimensional
arrays that can be thought of as a whole collection of tables. Suppose that you
measure the fasting glucose on five participants on each of three treatment days.
You could think of your 15 measurements being laid out in a table with five rows
and three columns. If you want to represent this entire table with a single variable
name like Gluc, you can use double-indexing, with the first index specifying the
participant (1 through 5), and the second index specifying the day of the measure-
ment (1 through 3). Under that system, Gluc[3,2] indicates the fasting glucose
measurement for participant 3 on day 2. To express the array as a formula, we
would use the expression Gluc[i,j], which specifies the fasting glucose for the ith
subject on the jth day.
Special terms may be used to refer to arrays with one or two dimensions:
» A one-dimensional array is also referred to as a vector. But this can be

confusing, because the term vector is also used in mathematics, physics, and
biology to refer to completely different concepts.
» A two-dimensional array is sometimes called a matrix (plural: matrices). To

some, this term implies we are using a set of mathematical rules called matrix
algebra, and that’s not entirely incorrect. Mathematical descriptions of
multiple regression (covered in Chapter 17 of this book) make extensive use
of matrix algebra. Also, computer software may refer to tabular objects with
the term matrix.
Arrays in formulas
If you see an array name in a formula without any subscripts, it usually means
that you have to evaluate the formula for each element of the array, and the result
is an array with the same number of elements. So, if Gluc refers to the array with
the five elements 86, 110, 95, 125, and 64, then the expression 2 × Gluc results in
an array with each element in the same order multiplied by two: 172, 220, 190,
250, and 128.
When an array name appears in a formula with subscripts, the meaning depends
upon the context. It can indicate that the formula is to be evaluated only for some
elements of the array, or it can mean that the elements of the array are to be com-
bined in some way before being used (as described in the next section).

Sums and products of the
elements of an array
This Greek letter ∑ is known in English as capital sigma. Though harmless, ∑
strikes terror into the hearts of many learners as they encounter it statistics books
and articles (not to mention its less common but even scarier cousin Π, also known
as capital pi). Uppercase sigma and pi — namely ∑ and Π — correspond to the
Roman letters S and P, which stand for Sum and Product, respectively. These
symbols are almost always used in front of variables and expressions that
represent arrays.
When you see ∑ in a formula, just think of it as saying “sum of.” Assuming an
array named Gluc that is comprised of the five elements 86, 110, 95, 125, and 64,
you can read the expression Gluc as “the sum of the Gluc array” or “sum of
Gluc.” To evaluate it, add all five elements together to get 86 110 95 125 64 ,
which equals 480.
Sometimes the ∑ notation is written in a more complex form, where the index
variable i is displayed under (or to the right of) the ∑ as a subscript of the array
name, like this: i
Gluci . Though its meaning is the same as Gluc , you would
read it as, “the sum of the Gluc array over all values of the index i” (which pro-
duces the same result as Gluc , which is 480). The subscripted ∑ form is helpful
in expressing multi-dimensional arrays, when you may want to sum over only
one of the dimensions. For example, if Ai,j is a two-dimensional array:
10 15 33
25 8 1
then i
Ai , j means that you should sum over the rows (the i subscript) to get the
one-dimensional array: 35, 23, and 34. Likewise, i
Ai , j means to sum across the
columns (j’) to get the one-dimensional array: 58, 34.
Finally, you may see the full-blown official mathematical ∑ in all its glory, like
this:
b
i a
Gluci ,
which reads “sum of the Gluc array over values of the index i going from a to b,
inclusive.” So if a was equal to 1, and b was equal to 5, the expression would
become:
5
i 1
Gluci ,

which is just another way of summing all the elements, producing 480. But if you
wanted to omit the first and last elements of the array from the sum, you could
write:
4
i 2
Gluci ,
This expression says to add up only Gluc2 + Gluc3 + Gluc4, to get 110 95 125 , which
would equal 330.
4
Gluci Gluc2 Gluc 3 Gluc 4 100 95 125 330
i 2
Π works just like ∑, except that you multiply instead of add:
5
Gluc Gluc Gluci 86 110 95 125 64 7,189,600,000
i i 1
SCIENTIFIC NOTATION: THE EASY WAY TO

WORK WITH REALLY BIG AND REALLY
SMALL NUMBERS
Statistical analyses can generate extremely large as well as extremely small numbers,
but humans are most comfortable working with numbers that are in the range of 10s,
100s or 1,000s. Numbers much smaller than 1 (like 0.0000000000005) or much larger
than 1,000 (like 5,000,000,000,000) are difficult for humans to comprehend. So for
humans, working with extremely large or extremely small numbers is difficult and error-
prone (as is working with certain humans).
Fortunately, to make it easier on all of us, we have scientific notation, which is a way to
represent very small or very large numbers to make the easier for humans to under-
stand. Here are three different ways to express the same number in scientific notation:
1.23 10 7 or 1.23E7, or 1.23e 7. All three mean “take the number 1.23, and then slide
the decimal point seven spaces to the right (adding zeros as needed).” To work this out
by hand, you could start by adding extra decimal places with zeros, like 1.2300000000.
Then, slide the decimal point seven places to the right to get 12300000.000 and clean it
up to get 12,300,000.
For very small numbers, the number after the E (or e) is negative, indicating that you
need to slide the decimal point to the left. For example, 1.23e–9 is the scientific notation
for 0.00000000123.
Note: Don’t be misled by the “e” that appears in scientific notation — it doesn’t stand for
the 2.718 constant. You should read it as “times ten raised to the power of.”

IN THIS CHAPTER
» Getting a handle on probability,

randomness, sampling, and inference
» Tackling hypothesis testing
» Understanding nonparametric
statistical tests
Chapter 3
Getting Statistical:
A Short Review of
Basic Statistics
T
his chapter provides an overview of basic concepts often taught in a one-
term introductory statistics course. These concepts form a conceptual
framework for the topics that we cover in greater depth throughout this
book. Here, you get the scoop on probability, randomness, populations, samples,
statistical inference, hypothesis testing, and nonparametric statistics.
Note: We only introduce the concepts here. They’re covered in much greater depth
in Statistics For Dummies and Statistics II For Dummies, both written by Deborah
J. Rumsey, PhD, and published by Wiley. Before you proceed, you may want to
skim through this chapter to review the topics we cover so you can fill any gaps in
your knowledge you may need in order to understand the concepts we introduce.
CHAPTER 3 Getting Statistical: A Short Review of Basic Statistics 29

Taking a Chance on Probability
Defining probability is hard to do without using another word that means the same
thing (or almost the same thing). Probability is the degree of certainty, the chance,
or the likelihood that an event will occur. Unfortunately, if you then try to define
chance or likelihood or certainty, you may wind up using the word probability in the
definition. No worries — we clear up the basics of probability in the following sec-
tions. We explain how to define probability as a number and provide a few simple
rules of probability. We also define odds, and compare odds to probability (because
they are not the same thing).
Thinking of probability as a number

Probability describes the relative frequency of the occurrence of a particular event,
such as getting heads on a coin flip or drawing the ace of spades from a deck of
cards. Probability is a number between 0 and 1, although in casual conversation,
you often see probabilities expressed as percentages. Probabilities are usually fol-
lowed by the word chance instead of probability. For example: If the probability of
rain is 0.7, you may hear someone say that there’s a 70 percent chance of rain.
Probabilities are numbers between 0 and 1 that can be interpreted this way:
» A probability of 0 (or 0 percent) means that the event definitely won’t occur.
» A probability of 1 (or 100 percent) means that the event definitely will occur.
» A probability between 0 and 1 (such as 0.7) means that — on average, over

the long run — the event will occur some predictable part of the time (such as
70 percent of the time).
The probability of one particular event happening out of N equally likely events
that could happen is 1/N. So with a deck of 52 different cards, the probability of
drawing any one specific card (such as the ace of spades) compared to any of the
other 51 cards is 1/52.
Following a few basic rules of probabilities

Here are three basic rules, or formulas, of probabilities. We call the first one the
not rule, the second one the and rule, and the third one the or rule. In the formulas
that follow, we use Prob as an abbreviation for probability, expressed as a fraction
(between 0 and 1).

Don’t use percentage numbers (0 to 100) in probability formulas.
Even though these rules of probabilities may seem simple when presented here,
applying them together in complex situations — as is done in statistics — can get
tricky in practice. Here are descriptions of the not rule, the and rule, and the or rule.
» The not rule: The probability of some event X not occurring is 1 minus the
probability of X occurring, which can be expressed in an equation like this:
Prob not X 1 Prob X
So if the probability of rain tomorrow is 0.7, then the probability of no rain

tomorrow is 1 – 0.7, or 0.3.
» The and rule: For two independent events, X and Y, the probability of event X
and event Y both occurring is equal to the product of the probability of each of
the two events occurring independently. Expressed as an equation, the and
rule looks like this:
Prob X and Y Prob X Prob Y
As an example of the and rule, imagine that you flip a fair coin and then draw
a card from a deck. What’s the probability of getting heads on the coin flip and
also drawing the ace of spades? The probability of getting heads in a fair coin
flip is 1/2, and the probability of drawing the ace of spades from a deck of
cards is 1/52. Therefore, the probability of having both of these events occur is
1/2 multiplied by 1/52, which is 1/104, or approximately 0.0096 (which is — as
you can see — very unlikely).
» The or rule: For two independent events, X and Y, the probability of X or Y (or
both) occurring is calculated by a more complicated formula, which can be
derived from the preceding two rules. Here is the formula:
Prob X or Y 1 1 Prob X 1 Prob Y
As an example, suppose that you roll a pair of six-sided dice. What’s the
probability of rolling a 4 on at least one of the two dice? For one die, there is a
1/6 chance of rolling a 4, which is a probability of about 0.167. (The chance of
getting any particular number on the roll of a six-sided die is 1/6, or 0.167.)
Using the formula, the probability of rolling a 4 on at least one of the two dice
is 1 1 0.167 1 0.167 , which works out to 1 0.833 0.833 , or 0.31,
approximately.
The and and or rules apply only to independent events. For example, if there is a 0.7
chance of rain tomorrow, you may make contingency plans. Let’s say that if it
does not rain, there is a 0.9 chance you will have a picnic rather than stay in a read
a book, but if it does rain, there is only a 0.1 chance you will have a picnic rather

than stay in a read a book. Because the likelihood of having a picnic is conditional
on whether or not it rains, raining and having a picnic are not independent events,
and these probability rules cannot apply.
Comparing odds versus probability

You see the word odds used a lot in this book, especially in Chapter 13, which is
about the fourfold cross-tab (contingency) table, and Chapter 18, which is about
logistic regression. The terms odds and probability are linked, but they actually
mean something different. Imagine you hear that a casino customer places a bet
because the odds of losing are 2-to-1. If you ask them why they are doing that,
they will tell you that such a bet wins — on average — one out of every three
times, which is an expression of probability. We will examine how this works
using formulas.
The odds of an event equals the probability of the event occurring divided by the
probability of that event not occurring. We already know we can calculate the prob-
ability of the event not occurring by subtracting the probability of the event occur-
ring from 1 (as described in the previous section). With that in mind, you can
express odds in terms of probability in the following formula:
Odds Probability/ 1 Probability
With a little algebra (which you don’t need to worry about), you can solve this
formula for probability as a function of odds:
Probability Odds/ 1 Odds
Returning to the casino example, if the customer says their odds of losing are
2-to-1, they mean 2/1, which equals 2. If we plug the odds of 2 into the second
equation, we get 2/(1+2), which is 2/3, which can be rounded to 0.6667. The cus-
tomer is correct — they will lose two out of every three times, and win one out of
every three times, on average.
Table 3-1 shows how probability and odds are related.
As shown in Table 3-1, for very low probabilities, the odds are very close to the prob-
ability. But as probability increases, the odds increase exponentially. By the time
probability reaches 0.5, the odds have become 1, and as probability approaches 1, the
odds become infinitely large! This definition of odds is consistent with its
common-language use. As described earlier with the casino example, if the odds
of a horse losing a race are 3:1, that means if you bet on this horse, you have three
chances of losing and one chance of winning, for a 0.75 probability of losing.

TABLE 3-1 The Relationship between Probability and Odds
Probability Odds Interpretation
1.0 Infinity The event will definitely occur.
0.9 9 The event will occur 90% of the time (it is nine times as likely to occur as
to not occur).
0.75 3 The event will occur 75% of the time (it is three times as likely to occur as
to not occur).
0.5 1.0 The event will occur about half the time (it is equally likely to occur or not occur).
0.25 0.3333 The event will occur 25% of the time (it is one-third as likely to occur as
to not occur).
0.1 0.1111 The event will occur 10% of the time (it is 1/9th as likely to occur as to not occur).
0 0 The event definitely will not occur.
Some Random Thoughts about

Randomness
When discussing probability, it is also important to define the word random. Like
the word probability, we use the word random all the time, and though we all have
some intuitive concept of it, it is hard to define with precise language. In statis-
tics, we often talk about random events and random variables. Random is a term
that applies to sampling. In terms of a sequence of random numbers, random
means the absence of any pattern in the numbers that can be used to predict what
the next number will be.
The important point about the term random is that you can’t predict a specific out-
come if a random element is involved. But just because you can’t predict a specific
outcome with random numbers doesn’t mean that you can’t make any predictions
about these numbers. Statisticians can make reasonably accurate predictions
about how a group of random numbers behave collectively, even if they cannot
predict a specific outcome when any randomness is involved.
Selecting Samples from Populations

Suppose that we want to know the average systolic blood pressure (SBP) of all the
adults in a particular city. Measuring an entire population is called doing a census,
and if we were to do that and calculate the average SBP in that city, we would have
calculated a population parameter.

But the idea of doing a census to calculate such a parameter is not practical. Even
if we somehow had a list of everyone in the city we could contact, it would be not
be feasible to visit all of them and measure their SBP. Nor would it be necessary.
Using inferential statistics, we could draw a sample from this population, measure
their SBPs, and calculate the mean as a sample statistic. Using this approach, we
could estimate the mean SBP of the population.
But drawing a sample that is representative of the background population depends

on probability (as well as other factors). In the following sections, we explain why
samples are valid but imperfect reflections of the population from which they’re
drawn. We also describe the basics of probability distributions. For a more exten-
sive discussion of sampling, see Chapter 6.
Recognizing that sampling isn’t perfect

As used in epidemiologic research, the terms population and sample can be defined
this way:
» Population: All individuals in a defined target population. For example, this

may be all individuals in the United States living with a diagnosis of Type II
diabetes.
» Sample: A subset of the target population actually selected to participate in a

study. For example, this could be patients in the United States living with
Type II diabetes who visit a particular clinic and meet other qualification
criteria for the study.
Any sample, no matter how carefully it is selected, is only an imperfect reflection

of the population. This is due to the unavoidable occurrence of random sampling
fluctuations called sampling error.
To illustrate sampling error, we obtained a data set containing the number of pri-
vate and public airports in each of the United States and the District of Columbia
in 2011 from Statista (available at https://www.statista.com/statistics/
185902/us-civil-and-joint-use-airports-2008/). We started by making a
histogram of the entire data set, which would be considered a census because it
contains the entire population of states. A histogram is a visualization to deter-
mine the distribution of numerical data, and is described more extensively in
Chapter 9. Here, we briefly summarize how to read a histogram:
» A histogram looks like a bar chart. It is specifically crafted to display a

distribution.

» The histogram’s y-axis represents the number (or frequency) of individuals in
the data that fall in the numerical ranges (known as classes) of the value being
charted, which are listed across the x-axis. In this case, the y-axis would
represent number of states falling in each class.
» This histogram’s x-axis represents classes, or numerical ranges of the value

being charted, which is in this case is number of airports.
We first made a histogram of the census, then we took four random samples of 20
states and made a histogram of each of the samples. Figure 3-1 shows the results.
FIGURE 3-1:
Distribution of
number of
private and public
airports in
2011 in the
population (of
50 states and the
District of
Columbia), and
four different
samples of
20 states from the
same population.
© John Wiley & Sons, Inc.
As shown in Figure 3-1, when comparing the sample distributions to the distribu-
tion of the population using the histograms, you can see there are differences.
Sample 2 looks much more like the population than Sample 4. However, they are
all valid samples in that they were randomly selected from the population. The
samples are an approximation to the true population distribution. In addition, the
mean and standard deviation of the samples are likely close to the mean and stan-
dard deviation of the population, but not equal to it. (For a refresher on mean and
standard deviation, see Chapter 9.) These characteristics of sampling error —
where valid samples from the population are almost always somewhat different
than the population — are true of any random sample.
Digging into probability distributions

As described in the preceding section, samples differ from populations because of
random fluctuations. Because these random fluctuations fall into patterns,

statisticians can describe quantitatively how these random fluctuations behave
using mathematical equations called probability distribution functions. Probability
distribution functions describe how likely it is that random fluctuations will
exceed any given magnitude. A probability distribution can be represented in sev-
eral ways:
» As a mathematical equation that calculates the chance that a fluctuation

will be of a certain magnitude. Using calculus, this function can be integrated,
which means turned into another related function that calculates the proba-
bility that a fluctuation will be at least as large as a certain magnitude.
» As a graph of the distribution, which looks and works much like a

histogram.
» As a table of values indicating how likely it is that random fluctuations will

exceed a certain magnitude.
In the following sections, we break down two types of distributions: those that
describe fluctuations in your data, and those that you encounter when performing
statistical tests.
Distributions that describe your data

Here are some common distributions that describe the random fluctuations found
in data analyzed by biostatisticians:
» Normal: The familiar, bell-shaped, normal distribution is probably the most

common distribution you will encounter. As an example, systolic blood
pressure (SBP) is found to follow a normal distribution in human populations.
» Log-normal: The log-normal distribution is also called a skewed distribution.

This distribution describes many laboratory results, such as enzymes and
antibody titers, where most of the population tests on the low end of the
scale. It is also the distribution seen for lengths of hospital stays, where most
stays are 0 or 1 days, and the rest are longer.
» Binomial: The binomial distribution describes proportions, and represents the

likelihood that a value will take one of two independent values (as whether an
event occurs or does not occur). As an example, in a class held regularly
where students can only pass or fail, the proportion who fail will follow a
binomial distribution.
» Poisson: The Poisson distribution describes the number of occurrences of

sporadic random events (rather than the binomial distribution, which is for
more common events). Examples of where the Poisson distribution is used in
biostatistics is where the events are not as common, such as deaths from
specific cancers each year.

Chapter 24 describes these and other distribution functions in more detail, and
you also encounter them throughout this book.
Distributions important to statistical testing

Some probability distributions don’t describe fluctuations in data values but
instead describe fluctuations in calculated values as part of a statistical test (when
you are calculating what’s called a test statistic). Distributions of test statistics
include the Student t, chi-square, and Fisher F distributions. Test statistics are
used to obtain the p values that result from the tests. See “Getting the language
down” later in this chapter for a definition of p values.
Introducing Statistical Inference

Statistical inference is where you draw conclusions (or infer) about a population
based on estimations from a sample from that population. The challenge posed by
statistical inference theory is to extract real information from the noise in our
data. This noise is made up of these random fluctuations as well as measurement
error. This very broad area of statistical theory can be subdivided into two topics:
statistical estimation theory and statistical decision theory.
Statistical estimation theory

Statistical estimation theory focuses how to improve the accuracy and precision of
metrics calculated from samples. It provides methods to estimate how precise
your measurements are to the true population value, and to calculate the range of
values from your sample that’s likely to include the true population value. The
following sections review the fundamentals of statistical estimation theory.
Accuracy and precision

Whenever you make an estimation or measurement, your estimated or measured
value can differ from the truth by being inaccurate, imprecise, or both.
» Accuracy refers to how close your measurement tends to come to the true
value, without being systematically biased in one direction or another. Such a
bias is called a systematic error.
» Precision refers to how close several replicate measurements come to each

other — that is, how reproducible they are.

In estimation, both random and systematic errors reduce precision and accuracy.
You cannot control random error, but you can control systematic error by improv-
ing your measurement methods. Consider the four different situations that can
arise if you take multiple measurements from the same population:
» High precision and high accuracy is an ideal result. It means that each
measurement you take is close to the others, and all of these are close to the
true population value.
» High precision and low accuracy is not as ideal. This is where repeat
measurements tend to be close to one another, but are not that close to the
true value. This situation can when you ask survey respondents to self-report
their weight. The average of the answers may be similar survey after survey,
but the answers may be inaccurately lower than truth. Although it is easy to
predict what the next measurement will be, the measurement is less useful if
it does not help you know the true value. This indicates you may want to
improve your measurement methods.
» Low precision and high accuracy is also not as ideal. This is where the
measurements are not that close to one another, but are not that far from the
true population value. In this case, you may trust your measurements, but
find that it is hard to predict what the next one will be due to random error.
» Low precision and low accuracy shows the least ideal result, which is a low
level of both precision and accuracy. This can only be improved through
improving measurement methods.
Sampling distributions and standard errors

The standard error (abbreviated SE) is one way to indicate the level of precision
about an estimate or measurement from a sample. The SE tells you how much the
estimate or measured value may vary if you were to repeat the experiment or the
measurement many times using a different random sample from the same popu-
lation each time, and recording the value you obtained each time. This collection
of numbers would have a spread of values, forming what is called the sampling
distribution for that variable. The SE is a measure of the width of the sampling
distribution, as described in Chapter 9.
Fortunately, you don’t have to repeat the entire experiment a large number of
times to calculate the SE. You can usually estimate the SE using data from a single
experiment by using confidence intervals.

Confidence intervals
An important application of statistical estimation theory in biostatistics is calculat-
ing confidence intervals. Confidence intervals provide another way to indicate the
precision of an estimate or measurement from a sample. A confidence interval (CI) is
an interval placed around an estimated value to represent the range in which you
strongly believe the true value for that variable lies. How wide you make this
interval is dependent on a numeric expression of how strongly you believe the true
value lie within it, which is called the confidence level (CL). If calculated properly,
your stated confidence interval should encompass the true value a percentage of
the time at least equal to the stated confidence level. In fact, if you are indeed
making an estimate, it is best practices to report that estimate along with confi-
dence intervals. As an example, you could express the 95 percent CI of the mean
ages of a sample of graduating master’s degree students from a university this
way: 32 years (95 percent CI 28 – 34 years).
At this point, you may be wondering how to calculate CIs. If so, turn to Chapter 10,
where we describe how to calculate confidence intervals around means, propor-
tions, event rates, regression coefficients, and other quantities you measure,
count, or calculate.
Statistical decision theory

Statistical decision theory is a large branch of statistics that includes many sub-
topics. It encompasses all the famous (and many not-so-famous) statistical tests
of significance, including the Student t tests and the analysis of variance (other-
wise known as ANOVA). Both t tests and ANOVAs are covered in Chapter 11. Statis-
tical decision theory also includes chi-square tests (explained in Chapter 12) and
Pearson correlation tests (included in Chapter 16), to name a few.
In its most basic form, statistical decision theory deals with using a sample to
make a decision as to whether a real effect is taking place in the background
population. We use the word effect throughout this book, which can refer to differ-
ent concepts in different circumstances. Examples of effects include the
following:
» The average value of a measurement may be different in one group

compared to another. For example, obese patients may have higher systolic
blood pressure (SBP) measurements on average compared to non-obese
patients. Another example is that the mean SBP of two groups of hyperten-
sive patients may be different because each group is using a different
drug — Drug A compared to Drug B. The difference between means in these
groups is considered the effect size.

» The average value of a measurement may be different from zero (or from
some other specified value). For example, the average reduction in pain level
measurement in surgery patients from post-surgery compared to 30 days
later may have an effect that is different from zero (or so we would hope)!
» Two numerical variables may be associated (also called correlated). For

example, the taller people are on average, the more they weigh. When two
variables like height and weight are associated in this way, the effect is called
correlation, and is typically quantified by the Pearson correlation coefficient
(described in Chapter 15).
Honing In on Hypothesis Testing

The theory of statistical hypothesis testing was developed in the early 20th
century. Among other uses, it was designed to apply the scientific method to data
sampled from populations. In the following sections, we explain the steps of
hypothesis testing, the potential results, and possible errors that can be made
when interpreting a statistical test. We also define and describe the relationships
between power, sample size, and effect size in testing.
Getting the language down

Here are some of the most common terms used in hypothesis testing:
» Null hypothesis (abbreviated H 0): The assertion that any apparent effect
you see in your data is not evidence of a true effect in the population, but is
merely the result of random fluctuations.
» Alternate hypothesis (abbreviated H 1 or H Alt): The assertion that there

indeed is evidence in your data of a true effect in the population over and
above what would be attributable to random fluctuations.
» Significance test: A calculation designed to determine whether H 0 can

reasonably explain what you see in your data or not.
» Significance: The conclusion that random fluctuations alone can’t account for
the size of the effect you observe in your data. In this case, H 0 must be false,
so you accept H Alt .
» Statistic: A number that you obtain or calculate from your sample.
» Test statistic: A number calculated from your sample that is part of perform-
ing a statistical test. It can be for the purpose of testing H 0 . In general, the test

statistic is usually calculated as the ratio of a number that measures the size
of the effect (the signal) divided by a number that measures the size of the
random fluctuations (the noise).
» p value: The probability or likelihood that random fluctuations alone (in the
absence of any true effect in the population) can produce the effect observed
in your sample (or, at least as large as the effect you observe in your sample).
The p value is the probability of random fluctuations making the test statistic
at least as large as what you calculate from your sample (or, more precisely,
at least as far away from H 0 in the direction of H Alt ).
» Type I error: Choosing that H Alt is correct when in fact, no true effect above
random fluctuations is present.
» Alpha (α): The probability of making a Type I error.
» Type II error: Choosing that H 0 is correct when in fact there is indeed a true
effect present that rises above random fluctuations.
» Beta (β): The probability of making a Type II error.
» Power: The same as 1 – β, which is probability of choosing H Alt as correct

when in fact there is a true effect above random fluctuations present.
Testing for significance

All the common statistical significance tests, including the Student t test, chi-
square, and ANOVA, work on the same general principle. They compare the size of
the effect seen in your sample against the size of the random fluctuations present
in your sample. We describe individual statistical significance tests in detail
throughout this book. Here, we describe the generic steps that underlie all the
common statistical tests of significance.
1. Reduce your raw sample data down into a single number called a test
statistic.
Each test statistic has its own formula, but in general, the test statistic repre-
sents the magnitude of the effect you’re looking for relative to the magnitude of
the random noise in your data. For example, the test statistic for the unpaired
Student t test for comparing means between two groups is calculated as a
fraction:
Mean of Group 1 Mean of Group 2

Student t statistic =
Standard Error of the Difference
The numerator is a measure of the effect, which is the mean difference

between the two groups. And the denominator is a measure of the random

noise in your sample, which is represented by the spread of values within each
group. Thinking about this fraction philosophically, you will notice that the
larger the observed effect is (numerator) relative to the amount of random
noise in your data (denominator), the larger the Student t statistic will be.
2. Determine how likely (or unlikely) it is for random fluctuations to

produce a test statistic as large as the one you actually got from
your data.
To do this, you use complicated formulas to generate the test statistic. Once
the test statistic is calculated, it is placed on a probability distribution. The
distribution describes how much the test statistic bounces around if only
random fluctuations are present (that is, if H 0 is true). For example, the Student
T statistic is placed on the Student T distribution. The result from placing the
test statistic on a distribution is known as the p value, which is described in the
next section.
Understanding the meaning of “p value”

as the result of a test
The end result of a statistical significance test is a p value, which represents the
probability that random fluctuations alone could have generated results. If that
probability is medium to high, the interpretation is that the null hypothesis, or H 0 ,
is correct. If that probability is very low, then the interpretation is that we reject the
null hypothesis, and accept the alternate hypothesis (H Alt ) as correct. If you find
yourself rejecting the null, you can say that the effect seen in your data is statisti-
cally significant.
How small should a p value be before we reject the null hypothesis? The technical
answer is this is arbitrary and depends on how much of a risk you’re willing to
take of being fooled by random fluctuations (that is, of making a Type I error). But
in practice, the value of 0.05 has become accepted as a reasonable criterion for
declaring significance, meaning we fail to reject the null for p values of 0.05 or
greater. If you adopt the criterion that p must be less than 0.05 to reject the null
hypothesis and declare your effect statistically significant, this is known as set-
ting alpha (α) to 0.05, and will establish your likelihood of making a Type I error
to no more than 5 percent.
Examining Type I and Type II errors

The outcome of a statistical test is a decision to either accept the H 0 , or reject H 0
in favor of H Alt . Because H 0 pertains to the population’s true value, the effect you
see in your sample is either true or false for the population from which you are

sampling. You may never know what that truth is, but an objective truth is out
there nonetheless.
The truth can be one of two answers — the effect is there, or the effect is not
there. Also, your conclusion from your sample will be one of two answers — the
effect is there, or the effect is not there.
These factors can be combined into the following four situations:
» Your test is not statistically significant, and H0 is true. This is an ideal

situation because the conclusion of your test matches the truth. If you were
testing the mean difference in effect between Drug A and Drug B, and in truth
there was no difference in effect, if your test also was not statistically signifi-
cant, this would be an ideal result.
» Your test is not statistically significant, but H0 is false. In this situation, the
interpretation of your test is wrong and does not match truth. Imagine testing
the difference in effect between Drug C and Drug D, where in truth, Drug C
had more effect than Drug D. If your test was not statistically significant, it
would be the wrong result. This situation is called Type II error. The probability
of making a Type II error is represented by the Greek letter beta (β).
» Your test is statistically significant, and HAlt is true. This is another

situation where you have an ideal result. Imagine we are testing the difference
in effect between Drug C and Drug D, where in truth, Drug C has more of an
effect. If the test was statistically significant, the interpretation would be to
reject H0, which would be correct.
» Your test is statistically significant, but HAlt is false. This is another

situation where your test interpretation does not match the truth. If there was
in truth no difference in effect between Drug A and Drug B, but your test was
statistically significant, it would be incorrect. This situation is called Type I
error. The probability of making a Type I error is represented by the Greek
letter alpha (α).
We discussed setting α = 0.05, meaning that you are willing to tolerate a Type I
error rate of 5 percent. Theoretically, you could change this number. You can
increase your chance of making a Type I error by increasing your α from 0.05 to a
higher number like 0.10, which is done in rare situations. But if you reduce your α
to number smaller than 0.05 — like 0.01, or 0.001 — then you run the risk of never
calculating a test statistic with a p value that is statistically significant, even if a
true effect is present. If α is set too low, it means you are being very picky about
accepting a true effect suggested by the statistics in your sample. If a drug really
is effective, you want to get a result that you interpret as statistically significant
when you test it. What this shows is that you need to strike a balance between the

likelihood of committing Type I and Type II errors — between the α and β error
rates. If you make α too small, β will become too large, and vice versa.
At this point, you may be wondering, “Is there any way to keep both Type I and
Type II error small?” The answer is yes, and it involves power, which is described
in the next section.
Grasping the power of a test

The power of a statistical test is the chance that it will come out statistically sig-
nificant when it should — that is, when the alternative hypothesis is really true.
Power is a probability and is very often expressed as a percentage. Beta (β) is the
chance of getting a nonsignificant result when the alternative hypothesis is true,
so you see that power and β are related mathematically: Power = 1 – β.
The power of any statistical test depends on several factors:
» The α level you’ve established for the test — that is, the chance you’re willing
to accept making a Type I error (usually 0.05)
» The actual magnitude of the effect in the population, relative to the amount of
noise in the data
» The size of your sample
Power, sample size, effect size relative to noise, and α level can’t all be varied
independently. They’re interrelated, because they’re connected and constrained
by a mathematical relationship involving all four quantities.
This relationship between power, sample size, effect size relative to noise, and α
level is often very complicated, and it can’t always be written down explicitly as a
formula. But the relationship does exist. As evidence of this, for any particular
type of test, theoretically, you can determine any one of the four quantities if you
know the other three. So for each statistical test, there are four different ways to
do power calculations, with each way calculating one of the four quantities from
arbitrarily specified values of the other three. (We have more to say about this in
Chapter 5, where we describe practical issues that arise during the design of
research studies.) In the following sections, we describe the relationships between
power, sample size, and effect size, and briefly review how you can perform power
calculations.

Power, sample size, and effect size relationships
The α level of a statistical test is usually set to 0.05 unless there are special con-
siderations, which we describe in Chapter 5. After you specify the value of α, you
can display the relationship between α and the other three variables — power,
sample size, and effect size — in several ways. The next three graphs show these
relationships for the Student t test as an example, because graphs for other sta-
tistical tests are generally similar to these:
» Power versus sample size, for various effect sizes: For all statistical tests,
power always increases as the sample size increases, if the other variables
including α level and effect size are held constant. This relationship is illus-
trated in Figure 3-2. “Eff” is the effect size — the between-group difference
divided by the within-group standard deviation.
Small samples will not be able to produce significant results unless the effect size
is very large. Conversely, statistical tests using extremely large samples including
many thousands of participants are almost always statistically significant unless
the effect size is near zero. In epidemiological studies, which often involve
hundreds of thousands of subjects, statistical tests tend to produce extremely
small (and therefore extremely significant) p values, even when the effect size is
so small that it’s of no practical importance (meaning it is clinically insignificant).
FIGURE 3-2:
The power of a
statistical test
increases as the
sample size and
the effect size
increase.
» Power versus effect size, for various sample sizes: For all statistical tests,
power always increases as the effect size increases, if other variables including
the α level and sample size are held constant. This relationship is illustrated
in Figure 3-3. “N” is the number of participants in each group.

FIGURE 3-3:
The power of a
statistical test
increases as the
effect size
increases.
For very large effect sizes, the power approaches 100 percent. For very small
effect sizes, you may think the power of the test would approach zero, but you
can see from Figure 3-3 that it doesn’t go down all the way to zero. It actually
approaches the α level of the test. (Keep in mind that the α level of the test is
the probability of the test producing a significant result when no effect is
truly present.)
» Sample size versus effect size, for various values of power: For all
statistical tests, sample size and effect size are inversely related, if other variables
including α level and power are held constant. Small effects can be detected
only with large samples, and large effects can often be detected with small
samples. This relationship is illustrated in Figure 3-4.
FIGURE 3-4:
Smaller effects
need larger
samples.

This inverse relationship between sample size and effect size takes on a very
simple mathematical form (at least to a good approximation): The required
sample size is inversely proportional to the square of the effect size that can
be detected. Or, equivalently, the detectable effect size is inversely propor-
tional to the square root of the sample size. So, quadrupling your sample size
allows you to detect effect sizes only one-half as large.
How to do power calculations

Power calculations can be an important step in the design of a research study
because they estimate how many individuals you will need in your sample to
achieve the objectives of your study. You don’t want your study to be underpow-
ered, because then it will have a high risk of missing real effects. You also don’t
want your study to be overpowered, because then it’s larger, costlier, and more
time-consuming than necessary. You need to include a power/sample-size calcu-
lation for research proposals submitted for funding and for any protocol you sub-
mit to a human research ethical review board for approval. You can perform power
calculations using several different methods:
» Computer software: The larger statistics packages such as SPSS, SAS, and R
enable you to perform a wide range of power calculations. Chapter 4
describes these different packages. There are also programs specially
designed for conducting power calculations, such as PS and G*Power, which
are described in Chapter 4.
» Web pages: Many of the more common power calculations can be performed
online using web-based calculators. An example of one of these is here:
https://clincalc.com/stats/samplesize.aspx.
» Rules of thumb: Some approximate sample-size calculations are simple

enough to do on a scrap of paper or even in your head! You find some of
these in Chapter 25.
Going Outside the Norm with

Nonparametric Statistics
All statistical tests are derived on the basis of some assumptions about your data.
Most of the classical significance tests, including Student t tests, analysis of var-
iance (ANOVA), and regression tests, assume that your data are distributed
according to some classical sampling distribution, which is also called a frequency
distribution. Most tests assume your data has a normal distribution (see

Chapter 24). Because the classic distribution functions are all written as mathe-
matical expressions involving parameters (like means and standard deviation),
they’re called parametric distribution functions.
Parametric tests assume that your data conforms to a parametric distribution func-
tion. Because the normal distribution is the most common statistical distribution,
the term parametric test is often used to mean a test that assumes normally dis-
tributed data. But sometimes your data don’t follow a parametric distribution. For
example, it may be very noticeably skewed, as shown in Figure 3-5a.
Sometimes, you may be able to perform a mathematical transformation of your

data to make it more normally distributed. For example, many variables that have
a skewed distribution can be turned into normally distributed numbers by taking
logarithms, as shown in Figure 3-5b. If, by trial and error, you can find some kind
of transformation that normalizes your data, you can run the classical tests on the
transformed data, as described in Chapter 9.
FIGURE 3-5:
Skewed data (a)
can sometimes
be turned into
normally
distributed data
(b) by taking
logarithms.
If you transform your data to get it to assume a normal distribution, any analyses
done on it will need to be “untransformed” to be interpreted. For example, if you
have a data set of patients with different lengths of stay in a hospital, you will
likely have skewed data. If you log-transform these data so that they are normally
distributed, then generate statistics (like calculate a mean), you will need to do an
inverse log transformation on the result before you interpret it.
But sometimes your data are not normally distributed, and for whatever reason,
you give up on trying to do a parametric test. Maybe you can’t find a good trans-
formation for your data, or maybe you don’t want to have to undo the transfor-
mation in order to do your interpretation, or maybe you simply have too small of

a sample size to be able to perceive a clear parametric distribution when you make
a histogram. Fortunately, statisticians have developed other tests that you can use
that are not based on the assumption your data are normally distributed, or have
any parametric distribution. Unsurprisingly, these are called nonparametric tests.
Most of the common classic parametric tests have nonparametric counterparts
you can use as an alternative. As you may expect, the most widely known and
commonly used nonparametric tests are those that correspond to the most widely
known and commonly used classical tests. Some of these are shown in Table 3-2.
TABLE 3-2 Nonparametric Counterparts of Classic Tests

Classic Parametric Test Nonparametric Equivalent
One-group or paired Student t test (see Chapter 11) Wilcoxon Signed-Ranks test
Two-group Student t test (see Chapter 11) Mann-Whitney U test
One-way ANOVA (see Chapter 11) Kruskal-Wallis test
Pearson Correlation test (see Chapter 15) Spearman Rank Correlation test
Most nonparametric tests involve first sorting your data values, from lowest to
highest, and recording the rank of each measurement. Ranks are like class ranks
in school, where the person with the highest grade point average (GPA) is ranked
number 1, and the person with the next highest GPA is ranked number 2 and so on.
Ranking forces each individual to be separated from the next by one unit of rank.
In data, the lowest value has a rank of 1, the next highest value has a rank of 2, and
so on. All subsequent calculations are done with these ranks rather than with the
actual data values. However, using ranks instead of the actual data loses informa-
tion, so you should avoid using nonparametric tests if your data qualify for
parametric methods.
Although nonparametric tests don’t assume normality, they do make certain

assumptions about your data. For example, many nonparametric tests assume
that you don’t have any tied values in your data set (in other words, no two
participants have exactly the same values). Most parametric tests incorporate

adjustments for the presence of ties, but this weakens the test and makes the
results less exact.
Even in descriptive statistics, the common parameters have nonparametric coun-

terparts. Although means and standard deviations can be calculated for any set of
numbers, they’re most useful for summarizing data when the numbers are nor-
mally distributed. When you don’t know how the numbers are distributed, medi-
ans and quartiles are much more useful as measures of central tendency and
dispersion (see Chapter 9 for details).

2
Examining Tools
and Processes
IN THIS PART . . .
Select statistical software to use from the commercial

and free choices available, and discover other ways to
do statistical calculations, such as using web-based
programs and specialized software for sample-size
calculations.
Understand human research — how biostatistics

influences the design and execution of clinical trials
and how treatments are developed and approved.
Grasp different ways of taking samples from

populations to increase the accuracy of the statistical
estimates you make.
Consider different study designs used in

epidemiologic research and see how research design
connects to the practice of biostatistics.
IN THIS CHAPTER
» Examining the evolution of statistical

software
» Surveying commercial, open source,

and free options
» Considering code-based versus

non–code-based software
» Storing data in the cloud
Chapter 4
Counting on Statistical
Software
B
efore statistical software, complex regressions we could do in theory were
too complicated to do manually using real datasets. It wasn’t until the
1960s with the development of the SAS suite of statistical software that
analysts were able to do these calculations. As technology advanced, different
types of software were developed, including open-source software and web-based
software.
As you may imagine, all these choices led to competition and confusion among
analysts, students, and organizations utilizing this software. Organizations
wonder what statistical packages to implement. Professors wonder which ones to
teach, and students wonder which ones to learn. The purpose of this chapter is
to help you make informed choices about statistical software. We describe and
provide guidance regarding the practical choices you have today among the
statistical software available. We discuss choosing between:
» Commercial software, such as SAS and SPSS
» Open-source software, such as R and Python
» Free software applications, such as G*Power and PS (Power and

Sample Size Calculation)
CHAPTER 4 Counting on Statistical Software 53

We also provide guidance on how to choose between code-based and non–code-
based software, and end by providing advice on cloud data storage.
Considering the Evolution of

Statistical Software
The first widespread commercial statistical software invented is called SAS, and it
is still used today. SAS was developed originally in the 1960s and 1970s to run on
mainframe computers. Around 2000, SAS was adapted to personal computers
(known as PC SAS), adding a user-friendly graphical user interface (GUI). During
the growth of SAS, other commercial statistical packages appeared, the most pop-
ular being IBM’s SPSS. SAS continues to be the go-to program for big data analy-
sis, where analysts can easily access large datasets from servers. In contrast, SPSS
continues to be used on a personal computer like PC SAS.
If you were to take a college statistics course in the year 2000, your course would
have likely taught either SAS or SPSS. Professors would have made either SPSS or
SAS available to you for free or for a nominal license fee from your college book-
store. If you take a college statistics course today, you may be in the same
situation — or, you may find yourself learning so-called open-source statistical
software packages. The most common are R and Python. This software is free to
the user and downloadable online because it is built by the user community, not a
company.
As the Internet evolved, more options became available for statistical software. In
addition to the existing stand-alone applications described earlier, specialized
statistical apps were developed that only perform one or a small collection of spe-
cific statistical functions (such as G*Power and PS, which are for calculating sam-
ple sizes). Similarly, web-based online calculators were developed, which are
typically programmed to do one particular function (such as calculate a chi-square
statistic and p value from counts of data, as described in Chapter 12). Some web
pages feature a collection of such calculators.
Comparing Commercial to
Open-Source Software
Before 2010, if an organization performed statistical analysis as part of its core
function, it needed to purchase commercial statistical software like SAS or
SPSS. Advantages of implementing commercial software include the ability to
54 PART 2 Examining Tools and Processes

perform many statistical functions, technical support from the software com-
pany, and the expectation that the software will remain in use in the future as the
company continues to support and upgrade it.
However, organizations today are hesitant to adopt commercial software when

they can instead use open-source software like R or Python. Admittedly, even
though it is free of charge, there are many downsides to open-source software.
First, you need to hire analysts who know how to use it so well that they can figure
out what to do when there’s a problem because open-source software does not
have tech support. Next, you need to hire a lot more analysts than you would with
commercial software because a lot of their work will be in trying to customize the
software for your use and keep it updated so that your organization runs smoothly.
So, why are new organizations today hesitant to adopt commercial software when
open-source software has so many downsides? The main reason is that the
old advantages of commercial software are not as true anymore. SAS and SPSS
are expensive programs, but they have much of the same functionality as open-
source R and Python, which are free. In some cases, analysts prefer the
open-source application to the commercial application because they can custom-
ize it more easily to their setting. Also, it is not clear that commercial software is
innovating ahead of open-source software. Organizations do not want to get
entangled with expensive commercial software that eventually starts to perform
worse than free open-source alternatives!
As a result, many organizations use both commercial and open-source statistical

software in integrated application pipelines. Therefore, it is important to be com-
fortable evaluating and using various commercial software, even if open-source
options are becoming more popular.
Checking Out Commercial Software

In the following sections, we discuss the most popular commercial statistical
software available currently.
SAS
SAS is the oldest commercial software currently available. It started out as having
two main components — Base SAS and SAS Stat — that provided the most used
statistical calculations. However, today, it has grown to include many additional
components and sublanguages. SAS has always been so expensive that only orga-
nizations with a significant budget can afford to purchase and use it. However,

because individual learners need to be able to practice SAS even if they cannot
afford it, SAS developed a free, online version called SAS OnDemand for Academics
(ODA) that is available at https://welcome.oda.sas.com.
Originally, SAS ran as a command-prompt software without a guided user inter-

face, or GUI, which came later in the 2000s when PC SAS was invented. In the
original SAS, the user would gain access to datasets in SAS format that resided on
a SAS server in the same environment. The user would write code files using SAS
code and run these files against the SAS data. This action would produce a log file
that explained how the code was executed and reported any errors. It would also
produce output that provided the results of the statistical procedures.
Today, the experience of using SAS has been modernized. In PC SAS and SAS ODA,
it is easy to view code, log, and output files in different windows and switch back
and forth between them. It is also easier to import data into and out of the SAS
environment and create integrated application pipelines involving the SAS envi-
ronment. The new commercial cloud-based version of SAS called Viya is intended
to be used with data stored in the cloud rather than on SAS servers (see the later
section “Storing Data in the Cloud” for more).
SAS is entrenched in some industries, such as pharmaceutical, insurance, and

banking, because SAS has historically been the only program powerful enough to
handle the size of their datasets. Those settings traditionally used SAS servers for
data storage. Now, this practice is being challenged because other analytic options
may look more appealing than what SAS has to offer (see the section “Focusing on
open-source and free software”). In addition, many companies are having trouble
maintaining their old-fashioned SAS servers and want to move their data to cloud
storage. These industries are looking for SAS users to help them modernize their
operations.
Students often find that SAS is challenging to learn when compared to other sta-
tistical software, especially open-source software. Why learn legacy commercial
software like SAS today, when it is so much harder to learn than other software?
The answer is that SAS is still standard software in some domains, such as phar-
maceutical research. This means that even if those organizations choose to even-
tually migrate away from SAS, they will need to hire SAS users to help with the
migration.
SPSS
SPSS was invented more recently than SAS and runs in a fundamentally different
way. SPSS does not expect you to have a data server the way SAS does. Instead,
SPSS runs as a stand-alone program like PC SAS, and expects you to import data
into it for analysis. Therefore, SAS is more likely to be used in a team environ-
ment, while SPSS tends to have individual users.

Like SAS, SPSS produces output, but unlike SAS, SPSS is typically manipulated by
the user through selections in menus rather than through writing code and run-
ning it. SPSS produces one long output file that includes all the output from each
SPSS session. In the output file, SPSS includes code it writes automatically from
the way you manipulate the menu. Therefore, like with SAS, it is possible to save
SPSS code files and output files and rerun the same code later. SPSS is available
from IBM’s website at www.ibm.com/products/spss-statistics/pricing.
Microsoft Excel
Microsoft Excel has been used in some domains for statistical calculations, but it
is difficult to use with large datasets. Excel has built-in functions for summariz-
ing data (such as calculating means and standard deviations talked about in
Chapter 9). It also has common probability distribution functions such as Student
t (Chapter 11) and chi-square (Chapter 12). You can even do straight-line regres-
sion (Chapter 16), as well as more extensive analyses available through add-ins.
These functions can come in handy when doing quick calculations or learning
about statistics, but using Excel for statistical projects evokes many challenges.
Using a spreadsheet for statistics means your data are stored in the same place as
your calculations, creating privacy concerns (and a mess!). So, while Excel can be
helpful mathematically — especially when making extra calculations based on
estimates in printed statistical output — it is not a good practice to use it for
extensive statistical projects.
Microsoft Excel is available in different formats, including both downloadable and

web based. Purchase it from Microsoft at www.microsoft.com.
Online analytics platforms

A more modern approach to statistical software is to create an online platform
known as an analytics suite that allows you to connect to data sources and conduct
analytics online. Here are a few popular online platforms:
» Tableau: Tableau is known for being able to provide real-time data-driven

graphical displays online, and organizations may adopt Tableau to develop
customized dashboards. It is available at www.tableau.com.
» GraphPad: This online platform provides analytics support, such as curve-

fitting, and provides a graphical suite called Prism. It is available at www.
graphpad.com.

There are both advantages and disadvantages to using these online commercial
platforms. Advantages include that online software tends to follow a cheaper sub-
scription paid monthly or annually, and you get continuous upgrades because the
software is web based. The main downside is these platforms have a high learning
curve and require a lot of work to fully adopt, so you have to ask yourself if it
makes sense with your project.
Focusing on Open-Source
and Free Software
Open-source software refers to software that has been developed and supported by
a user community. Although open-source software has licenses, they are typically
free but require you to adhere to certain policies when using the software. In this
section, we talk about the two most popular open-source statistical software
packages: R and Python.
Open-source software
The two most popular and extensive open-source statistical programs are R and
Python.
» R: R is statistical software that has been developed and is maintained by the R

user community. It has two interfaces: R GUI, which looks similar to PC SAS
and SPSS, and RStudio, which is an integrated development environment
(IDE). Analysts prefer to use RStudio when developing graphical displays for
the web, while R GUI is fine for most statistical work. To run R, you download
and install the base application. Then, for specified functions not included in
the base application, you install additional R packages. Like with PC SAS, in R,
you import or connect to datasets, develop and save code files to run on
those datasets, and produce output you can save. Base R, R packages, and
documentation are available on the Comprehensive R Archive Network
(CRAN) server at https://cran.r-project.org.
» Python: Python is an open-source programming language that is often used

to analyze data. As with R, Python is developed and maintained by its own
user community and runs in a similar way. Although you still develop code
that runs against datasets in the Python environment, the Python and R code
are different. Instead of packages as in R, Python has libraries. Python is
available at www.python.org/downloads.

Students often wonder what the differences are between R and Python, and which
one to learn. They are essentially the same, although scientific disciplines have
leaned toward adopting R, and engineering disciplines have leaned toward Python.
Many students find themselves easily learning both.
Other free statistical software

Other statistical software packages are free, but they are not technically open-
source — meaning they were not developed by an open-source community, and
they are not licensed the same way.
Software that performs many functions

This section provides examples of free software that performs many functions like
SAS and R.
» OpenStat and LazStats are free statistical programs developed by Dr. Bill
Miller that use menus that resemble SPSS. Dr. Miller provides several
excellent manuals and textbooks that support both programs. OpenStat and
LazStats are available at https://openstat.info.
» Epi Info was developed by the United States Centers for Disease Control to
acquire, manage, analyze, and display the results of epidemiological research.
What makes it different than other statistical software is that it contains
modules for creating survey forms and collecting data. Epi Info is available at
https://www.cdc.gov/epiinfo/index.html.
Software for calculating sample size

Biostatisticians frequently encounter the problem of estimating sample size. The
following are two free applications we recommend for performing sample-size
calculations:
» G*Power: G*Power was developed at the Universität Düsseldorf and is used

to estimate the sample size for many different types of tests. Throughout this
book, when we discuss sample-size calculations, we give you advice on how to
do them using G*Power. G*Power is available at www.psychologie.hhu.de/
arbeitsgruppen/allgemeine-psychologie-und-arbeitspsychologie/
gpower. To use the program, you download it from this website and install it
on your computer.

» PS (Power and Sample Size Calculation): The PS program was developed by
W.D. Dupont and W.D. Plummer at Vanderbilt University. Like G*Power, you
download the application from its website and install it on your computer. The
PS interface is similar to that of G*Power. PS is available at https://biostat.
app.vumc.org/wiki/Main/PowerSampleSize.
Choosing Between Code-based and

Non–Code-Based Methods
Most of the software mentioned up to this point in this chapter — including SAS,
SPSS, R and Python — use code files that can be saved and rerun on data at a later
date. These programs run fundamentally differently from programs such as
Microsoft Excel, where you can run statistics on data, but no code files are pro-
duced and saved. Also, when you use web-based calculators, specialized apps like
G*Power and PS for sample-size calculations, or online commercial platforms, no
code files are produced and saved.
This is an important issue in statistics. When no code files are produced or saved,
you have no record of the steps in your analysis. If you need to be able to repro-
duce your analysis, the only way to be sure of this is to use software that allows
you to save the code so you can run it again.
Storing Data in the Cloud

Cloud-based storage refers to storing large data files on a set of Internet servers
designed specifically for large data storage. Unlike old-fashioned stand-alone
servers in server rooms, cloud-based servers share storage space across the Inter-
net, providing instantaneous access and back-up capabilities. If you want to get rid
of an old-fashioned server in your server room (that could be a SAS server), you will
have to contract with a cloud-based storage company to use its space for your data.
Then, you will have to find a way to move your data from your server into your new
cloud storage. You will also have to make sure you want to have a long-term rela-
tionship with this company, so you don’t have to move your data out anytime soon.
Although moving data to the cloud may be an onerous task, you may not have any
choice, because physical storage space may be running out. Many new organiza-
tions start with cloud data storage for that reason. Once your data are stored in the
cloud, they are more easily accessed using online analytics platforms such as SAS
Viya and Tableau.

IN THIS CHAPTER
» Planning and carrying out an

experimental clinical trial study
» Protecting the participants
» Collecting, validating, and analyzing

research data
Chapter 5
Conducting Clinical
Research
T
his chapter provides a closer look at a special kind of human research —
the clinical trial. The purpose of a clinical trial is to test one or more
interventions, such as a medication or other product or action thought to be
therapeutic (such as drinking green tea or exercising). One of the important fea-
tures of a clinical trial is that it is an experimental study design, meaning that
participants in the study are assigned by the study staff which intervention to
take. Therefore, there are serious ethical considerations around clinical trials. On
the other hand, the clinical trial study design provides the highest quality
evidence you can obtain to determine whether or not an intervention actually
works, which is a form of causal inference. In this chapter, we cover approaches
to designing and executing a high-quality clinical trial and explain the ethical
considerations that go along with this.
Designing a Clinical Trial

Clinical trials should conform to the highest standards of scientific rigor, and that
starts with the design of the study. The following sections note some aspects of
good experimental design.
CHAPTER 5 Conducting Clinical Research 61

Identifying aims, objectives, hypotheses,
and variables
The aims or goals of a clinical trial are short general statements (often just one
statement) of the overall purpose of the experiment. For example, the aim of an
experiment may be “to assess whether drinking green tea every day improves
alertness in older adults.”
The objectives are much more specific than the aims. In a clinical trial, the objec-
tives usually refer to the effect of the intervention (treatment being tested) on
specific outcome variables at specific points in time in a group of a specific type of
study participants. In a drug trial, which is a type of experimental clinical research,
an efficacy study may have many individual efficacy objectives, as well as one or
two safety objectives, while a safety study may or may not have efficacy
objectives.
When designing a clinical trial, you should identify one or two primary objectives —
those that are most directly related to the aim of the study. This makes it easier to
determine whether the intervention meets the objectives once your analysis is
complete. You may then identify up to several dozen secondary objectives, which
may involve different variables or the same variables at different time points or in
different subsets of the study population. You may also list a set of exploratory
objectives, which are less important, but still interesting. Finally, if testing a risky
intervention (such as a pharmaceutical), you should list one or more safety
objectives (if this is an efficacy study) or some efficacy objectives (if this is a
safety study).
The objectives you select will determine what data you need to collect, so you have
to choose wisely to make sure all data related to those objectives can be collected
in the timeframe of your study. Also, these data will be processed from various
sources, including case report forms (CRFs), surveys, and centers providing labo-
ratory data. These considerations may limit the objectives you choose to study!
Drug clinical trials are usually efficacy studies. Here is an example of each type of
objectives you could have in an efficacy study:
» Primary efficacy objective: To compare the effect of new hypertension

(HTN) drug XYZ, relative to old drug ABC, on changes in systolic blood
pressure (SBP) from baseline to week 12, in participants with HTN.
» Secondary efficacy objective: To compare the effect of HTN drug XYZ,

relative to drug ABC, on changes in serum total cholesterol and serum
triglycerides from baseline to weeks 4 and 8, in participants with HTN.

» Exploratory efficacy objective: To compare the effect of drug XYZ, relative to
drug ABC, on changes in sexual function from baseline to weeks 4, 8, and 12,
in male and female subsets of participants with HTN.
» Safety objective: To evaluate the safety of drug XYZ, relative to drug ABC, in
terms of the occurrence of adverse events, changes from baseline in vital signs
such as temperature and heart rate, and changes in laboratory results of safety
panels (including tests on kidney and liver function), in participants with HTN.
For each of these objectives, it is important to specify the time range of participa-
tion subject to the analysis (such as the first week of the trial compared to other
time segments). Also, which groups are being compared for each objective should
be specified.
Hypotheses usually correspond to the objectives but are worded in a way that
directly relates to the statistical testing to be performed. So, the preceding pri-
mary objective may correspond to the following hypothesis: “The mean 12-week
reduction in SBP will be greater in the XYZ group than in the ABC group.” Alter-
natively, the hypothesis may be expressed in a more formal mathematical nota-
tion and as a null and alternate pair (see Chapters 2 and 3 for details on these
terms and the mathematical notation used):
H Null : XYZ ABC 0

H Alt : XYZ ABC 0
where mean of ( SBPWeek 12 SBPBaseline ).
In all types of human research, identifying the variables to collect in your study
should be straightforward after you’ve selected a study design and enumerated all
the objectives. In a clinical trial, you will need to operationalize the measurements
you need, meaning you will need to find a way to measure each concept specified
in the objectives. Measurements can fall into these categories:
» Administrative: This information includes data related to recruitment,

consent, and enrollment, as well as contact information for each participant.
You need to keep track of study eligibility documentation, as well as the date
of each visit, which study activities took place, and final status at end of study
(such as whether participants completed the study, dropped out of the study,
or any other outcome).
» Intervention-related: This includes data related to the intervention for each

participant, such as group assignment, dosing level, compliance, and adher-
ence measures. If you plan to assign a participant to a particular group but
they end up in another group, you need to keep track of both group
assignments.

» Outcome-related: This includes ensuring you are measuring both efficacy
and safety outcomes on a regular schedule that is documented. You may be
asking the participant to keep records, or they may need to be measured in
person (to obtain laboratory values, X-rays, and other scans, ECGs, and so on).
» Potential confounding variables: These variables are determined by way of

what is known about participants’ relationship with the intervention, outcome,
and study eligibility criteria. Typically, potential confounding variables include
basic demographic information such as date of birth, gender, and ethnicity.
Confounders could also be measured with questions posed to the participant
about health behaviors such as tobacco use, exercise patterns, and diet. There
are also questions about medical history, including current conditions, past
hospitalizations, family medical history, and current and past medication use.
Measurements such as height and weight as well as other physical measure-
ments can also be included.
Values that do not change over time, such as birth date and medical history, only
need to be recorded at the beginning of the study. In designs including follow-up
visits, values that change over time are measured multiple times over the duration
of data collection for the study. Depending on the research objectives, these could
include weight, medication use, and test results. Most of this data collection is
scheduled as part of study visits, and but some may be recorded only at unpredict-
able times, if at all (such as adverse events, and withdrawing from the study
before it is completed).
Deciding who is eligible for the study

Because you can’t examine the entire population for whom the intervention you’re
studying is intended, you must select a sample from that population. How you
filter in the right sample for your study is by explicitly defining the criteria a
potential participant has to meet to be eligible to be enrolled and maintained in
the study as a participant.
» Inclusion criteria are used during the screening process to identify

potential participants who are members of the population about whom you
want to draw conclusions. A reasonable inclusion criterion for a study of a
lipid-lowering intervention would be, “Participant must have a documented
diagnosis of hyperlipidemia, defined as total cholesterol 200 mg/dL and
LDL 130 mg/dL at screening.”
» Exclusion criteria are used to identify potential participants who do not fall in
the population being studied. They are also used to rule out participation by
individuals who are otherwise in the population being studied but should not

participate for practical reasons (such as safety, or risk of privacy breach).
A reasonable exclusion criterion for a study of a lipid-lowering treatment
would be, “Participants who are not willing to change their medication
during the study are not eligible to participate.”
» Withdrawal criteria apply to the follow-up portion of the study. They

describe situations that could arise during the study that would put the
participant in a state where participation should no longer take place. One
example would be that the participant is diagnosed with Alzheimer’s disease
during the study and can no longer make decisions on their own. A typical
withdrawal criterion may be, “If the participant no longer has decision-
making capacity, they will be withdrawn.”
Choosing the structure of a clinical trial

Many clinical trials include a comparison of two or more interventions. These
types of clinical trials typically have one of the following structures (or designs),
each of which has pros and cons:
» Parallel: In this clinical trial design, each participant receives one of the
interventions, and the groups are compared. Parallel designs are simpler,
quicker, and easier for each participant than crossover designs, but you need
more participants for the statistics to work out. Trials with very long treatment
periods usually have to be parallel.
» Crossover: In a crossover design, each clinical trial participant receives all

the interventions in sequence during consecutive treatment periods (called
phases) separated by washout intervals (lasting from several days to several
weeks). Crossover designs can be more efficient because each participant
serves their own control, eliminating inter-participant variability. But you can
use crossover designs only if you’re certain that at the end of each washout
period, the participant will have been restored to the same condition as at the
start of the study. This may be impossible for studies of progressive diseases,
like cancer or emphysema, or for drugs that last a long time in the body and
are hard to wash out, like SSRIs and marijuana.
Using randomization
Randomized controlled trials (RCTs) are the gold standard for clinical research (as
described in Chapters 7 and 20). In an RCT, the participants are randomly allocated

by intervention into treatment groups (in a parallel trial) or into treatment-
sequence groups (in a crossover design). Randomization provides several
advantages:
» It helps in reducing bias. It specifically helps to eliminate treatment bias, which

is where certain treatments are preferentially given to certain participants.
A clinician may feel inclined to assign a drug with fewer side effects to healthier
participants, but if participants are randomized, then this bias goes away.
Another important bias reduced by randomization is confounding, where the
treatment groups differ with respect to some characteristic that influences
the outcome.
» Randomization makes it easier to interpret the results of statistical testing.
» It facilitates blinding. Blinding (also called masking) refers to concealing the

identity of the intervention from both participants and researchers. There
are two types of blinding:
• Single-blinding: Where participants don’t know what intervention they’re

receiving, but the researchers do.
• Double-blinding: Where neither the participants nor the researchers know

which participants are receiving which interventions.
• Note: In all cases of blinding, for safety reasons, it is possible to unblind

individual participants, as at least one of the members of the research
team has the authority to unblind.
Blinding eliminates bias resulting from the placebo effect, which is where
participants tend to respond favorably to any treatment (even a placebo),
especially when the efficacy variables are subjective, such as pain level.
Double-blinding also eliminates deliberate and subconscious bias in the
investigator’s evaluation of a participant’s condition.
The simplest kind of randomization involves assigning each newly enrolled

participant to a treatment group by the flip of a coin or a similar method. But
simple randomization may produce an unbalanced pattern, like the one shown in
Figure 5-1 for a small study of 12 participants and two treatments: Drug (D) and
Placebo (P).
FIGURE 5-1:
Simple
randomization.

If you were hoping to have six participants in each group, you won’t be pleased if
you end up with three participants receiving the drug and nine receiving the
placebo, because it’s unbalanced. But unbalanced patterns like this arise quite
often from 12 coin flips. (Try it if you don’t believe us.) A better approach is to
require six participants in each group but shuffle those six Ds and six Ps around
randomly, as shown in Figure 5-2.
FIGURE 5-2:
Random
shuffling.
This arrangement is better because there are exactly six participants assigned to
drug and placebo each. But this particular random shuffle happens to assign more
drugs to the earlier participants and more placebos to the later participant (which
is just by chance). If the recruitment period is short, this would be perfectly fine.
However, if these 12 participants were enrolled over a period of five or six months,
seasonal effects may be mistaken for treatment effects, which is an example of
confounding.
To make sure that both treatments are evenly spread across the entire recruitment
period, you can use blocked randomization, in which you divide your subjects into
consecutive blocks and shuffle the assignments within each block. Often the block
size is set to twice the number of treatment groups. For instance, a two-group
study would use a block size of four. This is shown in Figure 5-3.
FIGURE 5-3:
Blocked
randomization.
You can create simple and blocked randomization lists in Microsoft Excel using
the RAND() built-in function to shuffle the assignments. You can also use the web
page at https://www.graphpad.com/quickcalcs/randomize1.cfm to generate
blocked randomization lists quickly and easily.
Selecting the analyses to use

You should select the appropriate analytic approach for each of your study hypoth-
eses based on the type of data involved, the structure of the study, and the require-
ments of the hypothesis. The rest of this book describes statistical methods to

analyze the kinds of data you’re likely to encounter in human research. Your
strategy is to apply them to a clinical trial design. In clinical trials, changes in
values of variables over time, and differences between treatments in crossover
studies are often analyzed by paired t tests and repeated-measures ANOVAs.
Differences between groups of participants in parallel studies are often analyzed

by unpaired t tests and ANOVAs. Often, final regression models are developed for
clinical trial interpretation because these can control for residual confounding
(which are covered in the chapters in Part 5). In longer clinical trials, time until
death (survival time) and the times to the occurrence of other endpoint events
(besides death) are analyzed by survival methods (Part 6 focuses on survival anal-
ysis methods).
Determining how many participants

to enroll in a clinical trial
Chapter 3 presents the concept of statistical power, and for a clinical trial, you
should enroll enough participants to provide sufficient statistical power when
testing the primary objective of the study. The specific way you calculate the
required sample size depends on the statistical test that’s used for the primary
hypothesis. Each chapter of this book that describes hypothesis tests also shows
how to estimate the required sample size for that test. To get quick sample-size
estimates, you can use G*Power (an application for sample-size calculations
described in Chapter 4), or you can use the formulas, tables, and charts in
Chapter 25 and on the book’s Cheat Sheet at www.dummies.com (just search for
“Biostatistics For Dummies Cheat Sheet”).
You must also allow some extra space in your target sample-size estimate for
some of the enrolled participants to drop out or otherwise not contribute the data
you need for your analysis. For example, suppose that you need full data from
64 participants for sufficient statistical power to answer your main objective. If
you expect a 15 percent attrition rate from the study, which means you expect only
85 percent of the enrolled participants to have analyzable data, then you need
to plan to enroll 64/0.84, or 76, participants in the study.
Assembling the study protocol

A study protocol (or just protocol) is a document that lays out exactly what you plan
to do to collect and analyze data in a research study. For ethical reasons, every
research study involving human participants should have a protocol, and for other
types of studies, having a protocol prepared before starting the research is

considered best practices. In a clinical trial, the protocol is especially important
because the participants are being assigned to interventions by the researcher,
and there is often double-blinding and randomization.
In terms of standard elements, a formal drug clinical trial protocol typically

contains these components:
» Title: A title conveys as much information about the trial as you can fit into
one sentence, including the protocol ID, name of the study, clinical phase, type
and structure of trial, type of randomization and blinding, name of the drug or
drugs being tested, treatment regimen, intended effect, and the population
being studied (which could include a reference to individuals with a particular
medical condition). A title can be quite long — this example title has all the
preceding elements:
Protocol BCAM521-13-01 (ASPIRE-2) — a Phase-IIa, double-blind, placebo-

controlled, randomized, parallel-group study of the safety and efficacy of three
different doses of AM521, given intravenously, once per month for six months,
for the relief of chronic pain, in adults with knee osteoporosis.
» Background information: This section includes information about the

disease for which the drug is an intended treatment. It includes the epidemiol-
ogy (the condition’s prevalence and impact), and its known physiology down
to the molecular level. It also includes a review of treatments currently
available (if any), and information about the drug or drugs being tested,
including mechanism of action, the results of prior testing, and known and
potential risks and benefits to participants.
» Rationale: The rationale for the study states why it makes sense to do this
study at this time and includes a justification for the choice of doses, how
the drug is administered (such as orally or intravenously), duration of drug
administration, and follow-up period.
» Aims, objectives, and hypotheses: We discuss these items in the earlier

section “Identifying aims, objectives, hypotheses, and variables.”
» Detailed descriptions of all inclusion, exclusion, and withdrawal criteria:

See the earlier section “Deciding who is eligible for the study” for more about
these terms.
» Design of the clinical trial: As described in the earlier section “Choosing the
structure of a clinical trial,” the clinical trial’s design defines its structure. This
includes the number of treatment groups as well as consecutive stages of the
study. These stages could include eligibility screening, washout, treatment,
follow-up, and so on. This section often includes a schematic diagram of the
structure of the study.

» Drug description: This description details each drug that will be administered
to the participants. This includes the chemical composition with the results of
chemical analysis of the drug, if available. It also includes instructions about
how to store, prepare, and administer the drug correctly.
» Blinding and randomization schemes: These schemes include descriptions

of how and when the study will be unblinded. This includes the emergency
unblinding of individual participants if necessary. See the earlier section
“Using randomization” for more information.
» Procedural descriptions: This section describes every procedure that will

be performed at every visit. These include administrative procedures, such
as enrollment and informed consent, and diagnostic procedures, such as
physical exams and measuring vital signs. It covers all activities where
data are collected from participants in the study.
» Safety considerations: These factors include the known and potential side
effects of each drug included. This section also includes the known and
potential side effects of each procedure in the study, including X-rays, MRI
scans, and blood draws. It also describes steps taken to minimize the risk to
the participants.
» Handling of adverse events: This section describes how adverse events will
be addressed should they occur during the study. It includes a description of
the data that will be recorded, including the nature of the adverse event,
severity, dates and times of onset and resolution, any medical treatment given
for the event, and whether or not the investigator thinks the event was related
to the study drug. It also explains how the research study will support the
participant after the adverse event.
» Definition of safety, efficacy, and other analytical populations: This

section includes definitions of safety and efficacy variables and endpoints. In
other words, this section defines variables or changes in variables that serve
as indicators of safety or efficacy.
» Planned enrollment and analyzable sample size: Justification for these

numbers must also be provided.
» Proposed statistical analyses: Some protocols describe, in detail, every

analysis for every objective. Others include only a summary and refer to a
separate Statistical Analysis Plan (SAP) document for details of the proposed
analysis. This section should also include descriptions of how missing data will
be handled analytically, adjustments for multiple testing to control Type I errors
(see Chapter 3), and whether any interim analyses are planned. If a separate
SAP is used, it typically contains a detailed description of all the calculations and
analyses that will be carried out on the data, including the descriptive summa-
ries of all data and the testing of all the hypotheses specified in the protocol.
The SAP also usually contains mock-ups called shells of all the tables, listings,
and figures (referred to as TLFs) that will be generated from the data.

It will also contain administrative details, like names and contact information for
the research team, a table of contents, a list of abbreviations, description of poli-
cies of data handling, and financing and insurance agreements.
Carrying Out a Clinical Trial

After you’ve designed your clinical trial and have described it in the protocol doc-
ument, it’s time to move on to the next step. The operational details will, of
course, vary from one study to another, but a few aspects apply to all clinical tri-
als. In any study involving human participants, the most important consideration
is protecting those participants from harm.
Since the end of World War II, international agreements have established ethical
guidelines for human research all over the world. Regardless of the country in
which you are conducting research, prior to beginning a human research study,
your protocol will need to be approved by at least one ethics board. Selection of
ethics boards for human research depends upon where the research is taking place
and what institutions are involved (see the later section “Working with Institu-
tional Review Boards”).
Protecting clinical trial participants

In any research involving humans, two issues are of utmost importance:
» Safety: Minimizing the risk of physical harm to the participant from the drug
or drugs being tested and from the procedures involved in the study
» Privacy/confidentiality: Ensuring that data collected during the study are not
breached (stolen) and are not made public in a way that identifies a specific
participant without the participant’s consent
The following sections describe some of the infrastructure that helps protect human
subjects.
Surveying regulatory agencies

In the United States, several government organizations oversee human subjects’
protection:
» Commercial pharmaceutical research is governed by the Food and Drug

Administration (FDA).

» Most academic biological research is sponsored by the National Institutes of
Health (NIH) and is governed by the Office for Human Research Protections
(OHRP).
Because research ethics are international, other countries have similar agencies,
so international clinical trial oversight can get confusing. An organization called
the International Conference on Harmonization (ICH) works to establish a set of
consistent standards that can be applied worldwide. The FDA and NIH have
adopted many ICH standards (with some modifications).
Working with Institutional Review Boards

For all protocols that describe research that could potentially be considered human
research, in the United States, an ethics board called an Institutional Review Board
(IRB) must review the protocol and approve it (or find it exempt from human
research laws) before you can start the research. You have to submit an applica-
tion along with the protocol and your plans for gaining informed consent of
potential participants to an IRB with jurisdiction over your research and ensure
that you are approved to proceed before you start.
Most medical centers and academic institutions (as well as some industry part-
ners) run their own IRBs with jurisdiction over research conducted at their insti-
tution. If you’re not affiliated with one of these centers or institutions (for
example, if you’re a freelance biostatistician or clinician), you may need the ser-
vices of a consulting or free-standing IRB. The sponsor of the research may recom-
mend or insist you use a particular IRB for the project.
Getting informed consent

An important part of protecting human participants is making sure that they’re
aware of the risks of a study before agreeing to participate in it. They also need to
be fully informed as to what will happen in the study before they can give consent
to participate. You must prepare an Informed Consent Form (ICF) describing, in
simple language, the nature of the study, why it is being conducted, what is being
tested, what procedures participants will undergo, and the risks and benefits of
participation. The ICF is used to guide study staff when they explain the study to
potential participants, who are then expected to sign it if they want to participate
in the study.
Potential participants must be told that they can refuse to participate with no
penalty to them, and if they join the study, they can withdraw at any time for any
reason without fear of retribution or the withholding of regular medical care. The
IRB can provide ICF templates with examples of their recommended or required
wording.

Prior to performing any procedures on a potential participant (including screen-
ing tests involved in determining eligibility), study staff must go through an
approved ICF document with the potential participant and give them time to read
it and decide whether or not they want to participate. If they choose to participate,
the ICF must be signed and witnessed. The signed ICFs must be retained as part
of the official documentation for the project, along with all data collected (includ-
ing laboratory reports, ECG tracings, and records of all test products administered
to the participants), as well as records of all procedures performed on partici-
pants. The sponsor, the regulatory agencies, the IRB, and other entities may ask
to review these documents at any time as part of oversight.
Considering data safety monitoring

boards and committees
For clinical trials of interventions that are likely to be of low risk (such as drinking
green tea), investigators are usually responsible for ensuring participant safety by
tracking unexpected adverse events, abnormal laboratory tests, and other red
flags during the course of the study. But for studies involving high-risk treat-
ments (like cancer chemotherapy trials), a separate data safety monitoring board or
committee (DSMB or DSMC) should be arranged. A DSMB is typically required by
the sponsor, the investigator, the IRB, or a regulatory agency. A DSMB typically
has about six members, including expert clinicians in the relevant area of research
and a statistician, who are external to the study staff. A DSMB meets at regular
intervals during the clinical trial to review the unblinded safety data acquired up
to that point. The committee is authorized to modify, suspend, or even terminate
a study if it has serious concerns about the safety of the participants.
Getting certified in human subjects protection

As you may have surmised from the preceding sections, clinical research involves
regulatory requirements that include penalties for noncompliance. Therefore, you
shouldn’t try to guess how to write a research protocol, obtain IRB approval, or
engage in any other research methods with which you are unfamiliar, and just
hope that everything goes well. You should ensure that you, along with any others
who may be assisting you, are properly trained in matters relating to human sub-
jects protection.
Fortunately, such training is readily available. Most hospitals and medical centers
provide training in the form of online courses, workshops, lectures, and other
resources. As you comply with ongoing IRB training, you receive a certification in
human subjects protection. Most IRBs and funding agencies require proof of cer-
tification from study staff. If you don’t have access to that training at your insti-
tution, you can get certified by taking an online tutorial offered by the NIH
(https://grants.nih.gov/policy/humansubjects/research/training-and-
resources.htm).

Collecting and validating data
Data in a clinical trial are typically collected digitally and manually. Examples of
data collected digitally include a participant filling out an online survey or a blood
pressure monitor collecting data from a participant. Data are collected manually
when they are written down on paper first, then undergo data entry to become
digital. Either way, some paper forms may be included, and many digital forms
are created for data entry in a clinical trial. These forms are for data entry of data
collected from various parts of the study, but in clinical trial lingo, they are all
referred to as case report forms, or CRFs.
» In the case of digitally collected data, the central analytic team will run
routines for validating the data. They will communicate with study staff if they
find errors and work them out.
» In the case of manually collected data, data entry into a digital format will be
required. Study staff typically are expected to log into an online database with
CRFs and do data entry from data collected on paper.
» The sponsor of the study will provide detailed data entry instructions and
training to ensure high-quality data collection and validation of the data
collected in the study.
Analyzing Your Data

In the following sections, we describe some general situations that come up in all
clinical research, regardless of what kind of analysis you use.
Dealing with missing data

Most clinical trials have incomplete data for one or more variables, which can be
a real headache when analyzing your data. The statistical aspects of missing data
are quite complicated, so you should consult a statistician if you have more than
just occasional, isolated missing values. Here we describe some commonly used
approaches for coping with missing data:
» Exclusion: Exclude a case from an analysis if any of the required variables for
that analysis is missing. This seems simple, but the downside to this approach
is it can reduce the number of analyzable cases, sometimes quite severely.
And if the result is missing for a reason that’s related to treatment efficacy,
excluding the case can bias your results.

» Imputation: Imputation is where you replace a missing value with a value
you impute, or create yourself. When analysts impute in a clinical trial, they
typically take the mean or median of all the available values for that variable
and fill that in for the missing variable. In reality, you have to keep the
original variable, and then save a separate, imputed variable so that you
can document the type of imputation applied. There are a lot of downsides
to imputation. If you are imputing a small number of values, it’s not worth it
because it adds bias. You may as well just exclude those cases. But if you
impute a large number of values, you are basically making up the data
yourself, adding more bias.
» Last Observation Carried Forward (LOCF): LOCF is a special case of

imputation. Sometimes during follow-up, one of a series of sequential
measurements on a particular participant is missing. For example, imagine
that there were supposed to be four weekly glucose values measured, and
you were missing a measurement only on week three. In that case, you could
use the most recent previous value in the series, which is the week two
measurement, to impute the week three measurement. This technique is
called Last Observation Carried Forward (LOCF) and is one of the most widely
used strategies. Although imputation adds bias, LOCF adds bias in the
conservative direction, making it more difficult to demonstrate efficacy. This
approach is popular with regulators, who want to put the burden of proof on
the drug and study sponsor.
Handling multiplicity
Every time you perform a statistical significance test, you run a chance of being
fooled by random fluctuations into thinking that some real effect is present in
your data when, in fact, none exists (review Chapter 3 for a refresher on statistical
testing). If you declare the results of the test are statistically significant, and in
reality they are not, you are committing Type I error. When you say that you require
p < 0.05 to declare statistical significance, you’re testing at the 0.05 (or 5 percent)
alpha (α) level. This is another way of saying that you want to limit your Type I
error rate to 5 percent. But that 5 percent error rate applies to each and every sta-
tistical test you run. The more analyses you perform on a data set, the more your
overall α level increases. If you perform two tests at α = 0.05, your chance of at
least one of them coming out falsely significant is about 10 percent. If you run 40
tests, the overall α level jumps to 87 percent! This is referred to as the problem of
multiplicity, or as Type I error inflation.
Chapter 11 covers dealing with multiplicity when making multiple comparisons.

One approach discussed in Chapter 11 is performing post-hoc tests following an
ANOVA for comparing several groups. Post-hoc tests incorporate a built-in
adjustment to keep the overall α at only 5 percent across all comparisons. This can

be especially important when conducting an interim analysis, or an analysis done
before the official end of study data collection. But when you’re testing different
hypotheses — like when comparing different variables at different time points
between different groups — you are faced with some difficult decisions to make
about reducing Type I error inflation.
In sponsored clinical trials, the sponsor and DSMB will weigh in on how they want
to see Type I error inflation controlled. If you are working on a clinical trial with-
out a sponsor, you should consult with another professional with experience in
developing clinical trial analyses to advise you on how to control your Type I error
inflation given the context of your study.
Each time an interim analysis is conducted, a process called data close-out must
occur. This creates a data snapshot, and the last data snapshot from a data close-
out process produces the final analytic dataset, or dataset to be used in all analyses.
Data close-out refers to the process where current data being collected are copied
into a research environment, and this copy is edited to prepare it for analysis.
These edits could include adding imputations, unblinding, or creating other vari-
ables needed for analysis. The analytic dataset prepared for each interim analysis
and for final analysis should be stored with documentation, as decisions about
stopping or adjusting the trial are made based on the results of interim analyses.

IN THIS CHAPTER
» Grasping the concept of statistical

error
» Setting up your sampling frame
» Executing a sampling strategy
Chapter 6
Taking All Kinds
of Samples
S
ampling — or taking a sample — is an important concept in statistics. As
described in Chapter 3, the purpose of taking a sample — or a group of indi-
viduals from a population — and measuring just the sample is so that you
do not have to conduct a census and measure the whole population. Instead, you
can measure just the sample and use statistical approaches to make inferences
about the whole, which is called inferential statistics. You can estimate a measure-
ment of the entire population, which is called a parameter, by calculating a statistic
from your sample.
Some samples do a better job than others at representing the population from
which they are drawn. We begin this chapter by digging more deeply into some
important concepts related to sampling. We then describe specific sampling
approaches and discuss their pros and cons.
CHAPTER 6 Taking All Kinds of Samples 77

Making Forgivable (and
Non-Forgivable) Errors
A central concept in statistics is that of error. In statistics, the term error some-
times means what you think it means — that a mistake has been made. In those
cases, the statistician should take steps to avoid the error. But other times in sta-
tistics, the term error refers to a phenomenon that is unavoidable, and as statisti-
cians, we just have to cope with it.
For example, imagine that you had a list of all the patients of a particular clinic
and their current ages. Suppose that you calculated the average age of the patients
on your list, and your answer was 43.7 years. That would be a population param-
eter. Now, let’s say you took a random sample of 20 patients from that list and
calculated the mean age of the sample, which would be a sample statistic. Do you
think you would get exactly 43.7 years? Although it is certainly possible, in all
likelihood, the mean of your sample — the statistic — would be a different num-
ber than the mean of your population — the parameter. The fact that most of the
time a sample statistic is not equal to the population parameter is called sampling
error. Sampling error is unavoidable, and as statisticians, we are forced to accept it.
Now, to describe the other type of error, let’s add some drama. Suppose that when
you went to take a sample of those 20 patients, you spilled coffee on the list so you
could not read some of the names. The names blotted out by the coffee were there-
fore ineligible to be selected for your sample. This is unfair to the names under the
coffee stain — they have a zero probability of being selected for your sample, even
though they are part of the population from which you are sampling. This is called
undercoverage, and is considered a type of non-sampling error. Non-sampling error
is essentially a mistake. It is where something goes wrong during sampling that
you should try to avoid. And unlike sampling error, undercoverage is definitely a
mistake you should avoid making if you can (like spilling coffee).
Framing Your Sample

In the previous example, the patient list is considered your sampling frame. A sam-
pling frame represents the practical representation of the population from which
you are literally drawing your sample. We described this list as a printout of
patient names and their ages. Suppose that after the list was printed, a few more
patients joined the clinic, and a few patients stopped using the clinic because they
moved away. This situation means that your sampling frame — your list — is not
a perfect representation of the actual population from which you are drawing
your sample.

If you omit population members from your sampling frame, you get undercover-
age, which is a form of non-sampling error (the type of error you want to avoid).
Also, if you accidentally include members in your sampling frame who are not
part of the population (such as patients who moved away from the clinic after you
printed your list), and they actually get sampled, you have another form of non-
sampling error. Non-sampling error can also creep in from making sloppy mea-
surements during data collection, or making poor choices when designing your
study. Chapter 8 provides guidance on how to minimize errors during data collec-
tion, and Chapters 5 and 7 provide advice on study design.
Another sampling-related vocabulary word is simulation. When talking about

sampling, a simulation refers to pretending to have data from an entire popula-
tion from which you can take samples, and then taking different samples to see
what happens when you analyze the data. That way, you can make sample statis-
tics while peeking at what the population parameters actually are behind the
scenes to see how they behave together.
One simulation you could do to illustrate sampling error in Microsoft Excel is to

create a column of 100 values that represent ages of imaginary patients at a clinic
as an entire population.
» If you calculated the mean of these 100 values, you would be doing a simula-
tion of the population parameter.
» If you randomly sampled 20 of these values and calculated the mean, you
would be doing a simulation of a sample statistic.
» If you compared your parameter to the statistic to see how close they were to
each other, you would be doing a simulation of sampling error.
So far we’ve reviewed several concepts related to the act of sampling. However, we
haven’t yet examined different sampling strategies. It matters how you go about
taking a sample from a population; some approaches provide a sample that is
more representative of the population than other approaches. In the next section,
we consider and compare several different sampling strategies.
Sampling for Success

As mentioned earlier, the purpose of taking measurements from a sample of a
population is so that you can use it to perform inferential statistics, which enables
you to make estimates about the population without having to measure the entire
population. Theoretically, you want the statistics from your sample to be as close
as possible to the population parameters you are trying to estimate. To increase

the likelihood that this happens, you should try your best to draw a sample that is
representative of the population.
You may be wondering, “What is the best way to draw a sample that is representa-
tive of the background population?” The honest answer is, “It depends on your
resources.” If you are a government agency, you can invest a lot of resources in
conducting representative sampling from a population for your studies. But if you
are a graduate student working on a dissertation, then based on resources availa-
ble, you probably have to settle for a sample that is not as representative of the
population as a government agency could afford. Nevertheless, you can still use
your judgment to make the wisest decisions possible about your sampling approach.
Taking a simple random sample

Taking a simple random sample (SRS) is considered a representative approach to
sampling from a background population. In an SRS, every member of the popula-
tion has an equal chance of being selected randomly and included in the sample.
As an example, recall the printout of the current patient list from a clinic dis-
cussed in the previous section. Considering that list a clinical population, imagine
that you used scissors to cut the list up so that each name was on its own slip of
paper, and then you put all the slips of paper into a hat. If you want to take an SRS
of 20 patients, you could randomly remove 20 names from the hat. The SRS would
be seen as a highly representative sample.
In practice, an SRS is usually taken using a computer so that you can take advan-
tage of a random number generator (RNG) (and do not have to cut up all that paper).
Imagine that the patient list from which you were sampling was not printed on
paper, but was instead stored in a column in a spreadsheet in Microsoft Excel. You
could use the following steps to take an SRS of 20 patients from this list using the
computer:
1. Create a column containing random numbers.
You could create another column in the spreadsheet called “Random” and
enter the following formula into the top cell in the column: =RAND(). If you drag
that cell down so that the entire column contains this command, you will see
that Excel populates each cell with a random number between 0 and 1. Each
time Excel evaluates, the random number gets recalculated.
2. Sort the list by the random number column.
3. Select the top 20 rows from the list.

This process ensures that your sample of 20 patients was taken completely at
random. Statistical packages like those described in Chapter 4 have RNG com-
mands similar to the one in Excel.
Learners sometimes think that as long as they sort a spreadsheet of data by a col-
umn containing any value and then select a sample of rows from the top, that they
have automatically obtained an SRS. This is not correct! If you think about it more
carefully, you will realize why. If you sort names alphabetically, you will see pat-
terns in names (such as religious names, or names associated with certain lan-
guages, countries, or ethnicities). If you sort by another identifying column, such
as email address or city of residence, you will again see patterns in the data. If you
attempt to take an SRS from such data, it will be biased, not random, and not be
representative. That is why it is important to use a column with an RNG in it for
sorting if you are taking an SRS electronically.
Taking an SRS intuitively seems like the optimal way to draw a representative
sample. However, there are caveats. In the previous example, you started with a
clinical population in the form of a printed or electronic list of patients from which
you could draw a sample. But what if you want to sample from patients presenting
to the emergency department during a particular period of time in the future?
Such a list does not exist. In a situation like that, you could use systematic sam-
pling, which is explained later in the section “Engaging in systematic
sampling.”
Another caveat of SRS is that it can miss important subgroups. Imagine that in
your list of clinic patients, only 10 percent were pediatric patients (defined as
patients under the age of 18 years). Because 10 percent of 20 is two, you may
expect that a random sample of 20 patients from a population where 10 percent
are pediatric would include two pediatric patients. But in practice, in a situation
like this, it would not be unusual for an SRS of 20 patients to include zero pediatric
patients. If your SRS needs to ensure representation by certain subgroups, then
you should consider using stratified sampling instead.
Taking a stratified sample

In the previous section, we discussed a scenario where 10 percent of the patients
of a clinic are pediatric patients, and taking a sample of 20 using an SRS from a list
of the clinic population runs the risk of not including any pediatric patients. If
pediatric patients were important to the study, then this problem can be solved
with stratified sampling. The word stratum refers to a layer (as you see in a layer
cake), and the word strata is the plural of stratum. Stratified sampling can be seen
as sampling from strata, or layers.

In our scenario, if you choose to draw a stratified sample by age groups, you would
first have to separate the list into a pediatric list and a list of everyone else. Then,
you could take an SRS from each. Because you are concerned about each stratum,
you could make a rule that even though pediatric patients make up only 10 percent
of the background population, you want them to make up 50 percent of your
sample. If you did that, then when you took your SRS, you would oversample from
the pediatric list and select 10, while also taking an SRS of 10 from the list of
everyone else.
Drawing a stratified sample requires you to weight your overall estimate, or else
it will be biased. As an example, imagine that 15 percent of pediatric patients had
an oral health condition, and 50 percent of the rest of the patients had an oral
health condition. In a stratified sample of 20 patients where you draw 10 from the
pediatric population and 10 from the rest of the population, because the pediatric
population is oversampled (because they only make up 10 percent of the back-
ground population but make up 50 percent of our sample), if weights are not
applied, the estimate of the percentage of the population with an oral health con-
dition would be artificially reduced. That is why it is necessary to apply weights to
overall estimates derived from a stratified sample.
If you are familiar with large epidemiologic surveillance studies such as the
National Health and Nutrition Examination Survey (NHANES) in the United States,
you may be aware that extremely complex stratified sampling is used in the design
and execution of such studies. Stratified sampling in these studies is unlike the
simple example described earlier, where the stratification involves only two age
groups. In surveillance studies like NHANES, there may be stratified sampling
based on many characteristics, including age, gender, and location of residence. If
you need to select factors on which to stratify, trying looking at what factors were
used for stratification in historical studies of the same population. The kind of
stratified sampling used in large-scale surveillance studies is reviewed later in
this chapter in the section “Sampling in multiple stages.”
Engaging in systematic sampling

Earlier you considered a scenario where a clinic had a printed list of the entire
population of patients from which an SRS could be drawn. But what if you want to
sample from the population of patients who present to a particular emergency
department tonight between 6 p.m. and midnight? There is no convenient list
from which to draw such a sample. In a scenario like this, even though you can’t
draw an SRS, you want to use a system for obtaining a sample such that it would
be representative of the underlying population. To do that, you could use system-
atic sampling.

Imagine you are surveying a sample of patients about their opinions of waiting
times at a particular emergency department, and you are doing this in the time
window of between 6 p.m. and midnight tonight. To take a systematic sample of
this population, follow these steps:
1. Select a small number.
This is your starting number. If you select three, this means that — starting at
6 p.m. — the first patient to whom you would offer your survey would be the
third one presenting to the emergency department.
2. Select another small number.
This is your sampling number. If you select five, then after the first patient to
whom you offered the survey, you would ask every fifth patient presenting
to the emergency department to complete your survey.
3. Continue sampling until you have the size sample you need (or the time
window expires).
Chapter 4 describes the software G*Power that can be used for making
sample-size calculations.
In systematic sampling, you are technically starting at a random individual, then

selecting every kth member of the population, where k stands for the sampling
number you selected.
Systematic sampling is not representative if there are any time-related cyclic pat-
terns that could confer periodicity onto the underlying data. For example, suppose
that it was known that most pediatric patients present to the emergency depart-
ment between 6 p.m. and 8 p.m. If you chose to collect data during this time win-
dow, even if you used systematic sampling, you would undoubtedly oversample
pediatric patients.
Sampling clusters
Another challenge you may face as a biostatistician when it comes to sampling
from populations occurs when you are studying an environmental exposure. The
term exposure is from epidemiology and refers to a factor hypothesized to have a
causal impact on an outcome (typically a health condition). Examples of environ-
mental exposures that are commonly studied include air pollution emitted from
factories, high levels of contaminants in an urban water system, and environ-
mental pollution and other dangers resulting from a particular event (such as a
natural disaster).

Consider the scenario where parents in a community are complaining that a local
factory is emitting pollutants that they believe is resulting in a higher rate of
leukemia being diagnosed in the community’s youth. To study whether the par-
ents are correct or not, you need to sample members of the population based on
their proximity to the factory. This is where cluster sampling comes in.
Planning to do cluster sampling geographically starts with getting an accurate

map of the area from which you are sampling. In the United States, each state is
divided up into counties, and each county is further subdivided into smaller
regions determined by the U.S. census. Other countries have similar ways their
maps can be divided along official geographic boundaries. In the scenario described
where a factory is thought to be polluting, the factory could be placed on the map
and lines drawn around the locations from which a sample should be drawn.
Different methodologies are used depending upon the specific study, but they
usually involve taking an SRS of regions and from the sampled regions known as
clusters, taking an SRS of community members for study participation.
But cluster sampling is not only done geographically. As another example, clus-
ters of schools may be selected based on school district, rather than geography,
and an SRS drawn from each school. The important takeaway from cluster sam-
pling is that it is a sampling strategy optimized for drawing a representative sam-
ple when studying an exposure known to be uneven across the population.
Sampling at your convenience

If you have read this chapter from the beginning until now, you may be feeling a
little exasperated. And that may be because all the sampling strategies we have
discussed so far — SRS, stratified sampling, systematic sampling, and cluster
sampling — involve a lot of work for the researcher. In an SRS, you need to have
a list of the population from which to draw, and in stratified sampling, you have
to know the value of the characteristics on which you want to stratify your sample.
Each of these features makes designing your sampling frame more complicated.
Thinking this way, both systematic sampling and cluster sampling also add com-
plexity to your sampling frame. In systematic sampling, whether you use a static
list or you sample in real time, you need to keep track of the details of your sam-
pling process. In cluster sampling, you may be using a map or system of group-
ings from which to sample, and that also involves a lot of recordkeeping. You may
be asking by now, “Isn’t there an easier way?”
Yes! There is an easier and more convenient way: convenience sampling. Conve-
nience sampling is what you probably think it is — taking a sample from a popu-
lation based on convenience. For example, when statistics professors want to

know what students think about a new policy on campus, they can just ask who-
ever is in their classes, as those students are a convenient sample of the student
population.
The problem is that the answer they get may be very biased. Most of the students
in their classes may come from the sciences, and those studying art or literature
may feel very differently about the same policy. Although our convenience
sample would be a valid sample of the background population of students, it would
be such a biased sample that the results would probably be rejected by the rest of
the faculty — especially those from the art and literature departments!
Given that the results from convenience samples are usually biased, you may
think that convenience sampling is not a good strategy. In actuality, convenience
sampling comes in handy if you have a relatively low-stakes research question.
Customer satisfaction surveys are usually done with convenience samples, such as
those placing an order on a restaurant’s app. It is simple to program such a survey
into an app, and if the food quality is terrific and the service terrible, it will be
immediately evident even from a small convenience sample of app users complet-
ing the survey.
While low-stakes situations are fine for convenience sampling, high-stakes

situations — like studying whether a new drug is safe and/or effective — require
study designs and sampling approaches completely focused on minimizing bias.
As with SRS, convenience sampling is prone to omitting important subgroups
from the sample. Minimizing bias through sampling and other strategies is cov-
ered in detail in Chapter 5, which examines clinical research and describes how
researchers must present a well-defined protocol that includes selection criteria,
a sampling plan, and an analytic plan that undergoes regulatory approval prior to
the commencement of research activities. Other strategies for minimizing bias are
presented in Chapters 7 and 20, which cover study designs and causal inference.
Sampling in multiple stages

When conducting large, epidemiologic surveillance studies, it is necessary to do
an especially good job of sampling, because governments use results from these
studies on which to base public policy. As an example, because being obese puts
community members at risk for serious health conditions, government public
health agencies have a vested interest in making accurate estimates of the rates of
obesity in their communities.
For this reason, to strive to obtain a representative sample, researchers designing

large epidemiologic surveillance studies use multi-stage sampling. Multi-stage
sampling is a general term for using multiple sampling approaches at different

stages as part of a strategy to obtain a representative sample. Figure 6-1 provides
a schematic describing the multi-stage sampling in the U.S. surveillance study
mentioned earlier, NHANES.
FIGURE 6-1:
Example of
multi-stage
sampling from
the National
Health and
Nutrition.
Examination
Survey (NHANES).
As shown in Figure 6-1, in NHANES, there are four stages of sampling. In the first
stage, primary sampling units, or PSUs, are randomly selected. The PSUs are made
up of counties, or small groups of counties together. Next, in the second stage,
segments — which are a block or group of blocks containing a cluster of
households — are randomly selected from the counties sampled in the first stage.
Next, in the third stage, households are randomly selected from segments. Finally,
in stage four, to select each actual community member who will be offered
participation in NHANES, an individual is randomly selected from each house
hold sampled in the third stage.
That is how a sample of 8,704 individuals participating in NHANES in 2017–2018

was selected to represent the population of the approximately 325 million people
living in the United States at that time. The good news is that biostatisticians work
on teams to develop a multi-stage sampling strategy — no one is expected to set
up something so complicated all by themselves.

IN THIS CHAPTER
» Grasping the hierarchy of study

designs in epidemiology
» Appreciating relative strength of

different study designs for causal
inference
» Getting the details about different

study designs
Chapter 7
Having Designs on
Study Design
B
iostatistics can be seen as the application of a set of tools to answer ques-
tions posed through human research. When studying samples, these tools
are used in conjunction with epidemiologic study designs in such a way as
to facilitate causal inference, or the ability to determine cause and effect. Some
study designs are better than others at facilitating causal inference. Nevertheless,
regardless of the study design selected, an appropriate sampling strategy and sta-
tistical analysis that complements the study design must be used in conjunction
with it.
In this chapter, we provide an overview of epidemiologic study designs and pres-

ent them in a hierarchy so that you can relate them to the biostatistical approaches
described in the different chapters of this book. We start by looking at broad study
design categories such as observational, experimental, descriptive, and analytic,
and move into descriptive study designs including expert opinion, case studies
and case series, ecological (correlational) studies, and cross-sectional studies. We
present analytic study designs next — case-control studies and longitudinal
cohort studies — which are superior to descriptive designs in terms of developing
evidence for causal inference, and then move into the highest-level designs: sys-
tematic reviews and meta-analyses.
CHAPTER 7 Having Designs on Study Design 87

For a deeper dive into epidemiologic study designs, we encourage you to read
Epidemiology For Dummies by Amal K. Mitra (Wiley) and pay special attention to
Chapters 16 and 17, which are about causal inference and study design.
Presenting the Study Design Hierarchy

Figure 7-1 illustrates the epidemiologic study designs in terms of their relation-
ship with each other in a hierarchy as they apply to human health research (not
animal research or other domains of human research like psychology). As shown
in Figure 7-1, human health research may be split into two types: observational
and experimental.
FIGURE 7-1:
Study design
hierarchy.
Observational research is where humans are studied in terms of their health and
behavior, but they are not assigned to do any particular behavior as part of the
study — they are just observed. For example, imagine that a sample of women
were contacted by phone and asked about their use of birth control pills. In this
case, researchers are observing their behavior with regard to birth control pills.

This is in contrast to experimental research, which is where the researcher assigns
research participants to engage in certain behaviors as part of an intervention
being studied. Imagine that you were studying a new birth control pill on the mar-
ket and you wanted to know whether taking it changed a woman’s lipid profile,
which is determined by a laboratory test. You may do an experiment where you
assign research participants to take the pill and then evaluate their lipid profiles.
The difference between observational and experimental research is important

because there is a heavier ethical obligation in experimental research when the
researcher is assigning an intervention. Also, because of such ethical consider-
ations in addition to time, cost, and other limitations, there are far more observa-
tional studies than experimental studies.
Note that there are only two entries under the experimental category in
Figure 7-1 — small-scale experiments in laboratory settings and clinical trial
designs. The example of assigning research participants to take a new birth
control pill and then testing their lipid profiles is an example of a small-scale
experiment in a laboratory setting. We describe clinical trials later in this chapter
in “Advancing to the clinical trial stage.”
Observational studies, on the other hand, can be further subdivided into two
types: descriptive and analytic. Descriptive study designs include expert opinion,
case studies, case series, ecologic (correlational) studies, and cross-sectional
studies. These designs are called descriptive study designs because they focus on
describing health in populations. (We explain what this means in “Describing
What We See.”) In contrast to descriptive study designs, there are only two types
of analytic study designs: longitudinal cohort studies and case-control studies.
Unlike descriptive studies, analytic studies are designed specifically for causal
inference. These are described in more detail in the section, “Getting
Analytical.”
Describing what we see

As shown in Figure 7-1, there are two types of observational studies: descriptive
and analytic. Descriptive study designs focus on describing patterns of human
health and disease in populations, usually as part of surveillance, which is the act
of quantifying patterns of health and disease in populations. Cross-sectional is
one descriptive study design used in surveillance to produce incidence and
prevalence rates of conditions or behaviors (see Chapter 14). For example, results
from cross-sectional surveillance studies tell us that approximately 25 percent of
women aged 15 to 44 who currently use contraception in the United States choose

the birth control pill as their method of choice. While descriptive study designs are
necessary in a practical sense, they are poor at developing evidence for causal
inference, so they are considered inferior to analytic study designs.
Getting analytical
Analytic designs include longitudinal cohort studies and case-control studies.
These are the strongest observational study designs for causal inference. Longitu-
dinal cohort studies are used to study causes of more common conditions, like
hypertension (HTN). It is called longitudinal because follow-up data are collected
over years to see which members of the sample, or cohort, eventually get the out-
come, and which members do not. (In a cohort study, none of the participants has
the condition, or outcome, when they enter the study.) The cohort study design is
described in more detail under the section, “Following a cohort over time.”
Case-control studies are used when the outcome is not that common, such as liver
cancer. In the case of rare conditions, first a group of individuals known to have
the rare condition (cases) is identified and enrolled in the study. Then, a compa-
rable group of individuals known to not have the rare condition is enrolled in the
study as controls. The case-control study design is described in greater detail
under the section “Going from case series to case-control.”
Going from observational to experimental

You may notice in Figure 7-1 that observational studies (which are either descrip-
tive or analytic) comprise most of the figure. Experimental studies — where par-
ticipants are assigned to engage in certain behaviors or interventions — are less
common than observational studies because they have ethical concerns, and are
often expensive and complex. However, experimental studies benefit from gener-
ating the highest level of evidence for causal inference — much higher than
observational studies.
Climbing the Evidence Pyramid

Each of the study designs discussed in the previous sections generates a particular
level of evidence for causal inference. These levels of evidence may be arranged in
a pyramid. As shown in Figure 7-2, the study designs with the strongest evidence
for causal inference are at the top of the pyramid, and those with the weakest are
at the bottom.

FIGURE 7-2:
Levels
of evidence in
study designs.
Starting at the base: Expert opinion

At the base of the study design evidence pyramid shown Figure 7-2 is a descriptive
study design: expert opinion. When the condition of dementia was first identified,
few clinicians had seen patients with dementia. These clinicians served as experts
who would write about the dementia patients they treated and share their experi-
ences at medical conferences. This is what is meant by expert opinion, and while
it is helpful when conditions are first identified, expert opinion is considered a
very weak descriptive study design.
Making the case with case studies

Also at the base of Figure 7-2 are case studies and case series. To develop an
understanding of dementia when it was first identified, clinicians treating patients
needed to study them. They would write up case studies or case reports on individual
patients describing their symptoms and providing the best descriptive evidence as
possible. If the clinician was able to identify more than one patient, they could
write about a series of patients, which is known as a case series. While case studies
and case series are helpful for researchers when a condition is first identified, they
are considered as providing very weak evidence to use for causal inference.

Making statements about the population
Ecologic studies (also called correlational studies) and cross-sectional studies
appear in the next level up from expert opinion and case studies and case series.
Though ecologic studies are still descriptive designs that provide weak evidence,
they have the advantage of having potentially very large samples.
In ecologic studies, the experimental units are often entire populations (such of a
region or country). For example, in a study presented in Chapter 16 of Epidemiology
For Dummies by Amal K. Mitra (Wiley), the experimental unit is a country, and
15 countries were included in the analysis. The exposure being investigated is fat
intake from diet (which was operationalized as average saturated fat intake as a
percentage of energy in the diet). The outcome was deaths from coronary heart
disease (CHD), operationalized as 50-year CHD deaths per 1,000 person-years
(see Chapter 15 for more about rates in person-years). Figure 7-3 presents the
results in the form of a scatter plot.
FIGURE 7-3:
Ecologic study
results.
As shown in Figure 7-3, the country’s average value of the outcome (rate of CHD
deaths) is plotted on the y-axis because it’s the outcome. The exposure, average
dietary fat intake for the country, is plotted on the x-axis. The 15 countries in the
study are plotted according to their x-y coordinates. Notice that the United States
is in the upper-right quadrant of the scatter plot because it has high rates of both
the exposure and outcome. The strong, positive value of correlation coefficient r
(which is 0.92) indicates that there is a strong positive bivariate association
between the exposure and outcome, which is weak evidence for causality (flip to
Chapter 15 for more on correlation).

But the problem with ecologic studies is that the experimental unit is a whole
population — not an individual. What if the individuals in the United States who
ate low-fat diets were actually the ones to die of CHD? And what if the ones who
ate high-fat diets were more likely to die of something else? Attributing the
behavior of a group to an individual is called the ecologic fallacy, and can be a prob-
lem with interpreting results like the ones shown in Figure 7-3.
That is why we also have cross-sectional studies, where the experimental unit is
an individual, not a population. A cross-sectional study takes measurements of
individuals at one point in time — either through an in-person hands-on exami-
nation, or by survey (over the phone, Internet, or in person). The National Health
and Nutrition Examination Survey (NHANES) is a cross-sectional surveillance
effort done by the U.S. government on a sample of residents every year. NHANES
makes many measurements relevant to human health in the United States, includ-
ing dietary fat intake as well as status of many chronic diseases including CHD. If
an analysis of cross-sectional data like NHANES found that there was a strong
positive association between high dietary fat intake and a CHD diagnosis in the
individuals participating, it would still be weak evidence for causation, but would
be stronger than what was found in the ecologic study presented in Figure 7-3.
Going from case series to case-control

The reason that there are two types of analytic study designs — case-control
studies and cohort studies — is that cohort study designs do not work for statisti-
cally rare conditions. We use the term statistically rare because if someone you love
gets cancer, cancer does not seem very rare. Yet, if you enroll a cohort of thou-
sands of individuals including your loved one (who is free of cancer) in a cohort
study and measure this cohort yearly to see who is diagnosed with cancer, it would
take many years to get enough outcomes to be able to develop the regression
models (like the ones described in Chapters 16 through 23) that would be neces-
sary for causal inference. So for statistically rare conditions like the various can-
cers, you use the case-control design.
You can use a fourfold, or 2x2, table to better understand how case-control studies
are different from cohort studies. (Refer to Chapters 13 and 14 for more about 2x2
tables.) As shown in Figure 7-4, the 2x2 table cells are labeled relative to exposure
status (the rows) and outcome or disease status (the columns). For the columns,
D+ stands for having the disease (or outcome), and D– means not having the dis-
ease or outcome. Also, for the rows, E+ means having the exposure, and E– means
not having the exposure. Cell a includes the counts of individuals in the study who
were positive for both the exposure and outcome, and cell d includes the counts of
individuals who were negative for both the exposure and outcome (a and d are
concordant cells because the exposure and outcome statuses agree). In the

discordant cells, where the exposure and outcome do not agree, is b, which repre-
sents the count of those positive for the exposure but negative for the outcome,
and c, which represents the count of those negative for the exposure but positive
for the outcome.
FIGURE 7-4:
2x2 table cells.
The 2x2 table shown in Figure 7-4 is generic — meaning it can be filled in with
data from a cross-sectional study, a case-control study, a cohort study, or even a
clinical trial (if you replace the E+ and E– entries with intervention group assign-
ment). How the results are interpreted from the 2x2 table depend upon the under-
lying study design. In the case of a cross-sectional study, an odds ratio (OR) could
be calculated to quantify the strength of association between the exposure and
outcome (see Chapter 14). However, any results coming from a 2x2 table do not
control for confounding, which is a bias introduced by a nuisance variable associ-
ated with the exposure and the outcome, but not on the causal pathway between
the exposure and outcome (more on confounding in Chapter 20).
Imagine that you were examining the cross-sectional association between having
the exposure of obesity (yes/no), and having the outcome of HTN (yes/no). House-
hold income may be a confounding variable, because lower income levels are
associated with barriers to access to high-quality nutrition that could prevent
both obesity and HTN. However, in a bivariate analysis like is done in a 2x2 table,
there is no ability to control for confounding. To do that, you need to use a regres-
sion model like the ones described in Chapters 15 through 23.
So how would you use a 2x2 table for a case-control study on a statistically rare
condition like liver cancer? Suppose that patients thought to have liver cancer are
referred to a cancer center to undergo biopsies. Those with biopsies that are posi-
tive for liver cancer are placed in a registry. Suppose that in 2023 there were
30 cases of liver cancer found at this center that were placed in the registry. This
would be a case series. Imagine that you had a hypothesis — that high levels of
alcohol intake may have caused the liver cancer. You could interview the cases to
determine their exposure status, or level of alcohol intake before they were diag-
nosed with liver cancer. Imagine that 10 of the 30 reported high alcohol intake.
You will see that as some evidence for your hypothesis.

But you could not do causal inference unless you had a comparable comparison
group without liver cancer so that you could fill out your 2x2 table. Imagine that
you went back to the cancer center and were able to contact and enroll 30 patients
who had liver biopsies but were found not to have liver cancer to serve as controls.
Suppose that you interviewed this group and discovered that only two of them
reported high levels of alcohol intake. You could develop the 2x2 table like the one
shown in Figure 7-5.
FIGURE 7-5:
Example of
a typical
case-control
study 2x2 table.
As shown in Figure 7-5, what is important is not the 2x2 table itself, but the order
in which the counts are filled in. Notice that at the beginning of the study, you
already knew the case total was 30, and you had determined that your control total
would be 30 (although you are allowed to sample more controls if you want in a
case-control study).
The correct measure of relative risk to present for a case-control study is the OR (as
described in Chapter 14). It is important to acknowledge that when you present an
OR from a case-control study, you interpret it as an exposure OR, not an outcome
or disease OR. (It is also acceptable to present an OR in a cross-sectional study, but
in that case, you are presenting an outcome or disease OR.)
In a case-control study, because the condition is rare, you are sampling on the
outcome and calculating the likelihood that the cases compared to controls were
exposed. This study design is seen as extremely biased, which is why cohort stud-
ies are preferred, and are at a higher level of evidence. However, case-control
study designs are necessary for rare diseases.
Following a cohort over time

In the previous section, we pointed out that case-control studies are used for
rare diseases. Therefore, case-control studies do not have large sample sizes,
which is evident in Figure 7-5 where the total sample size is 60. In contrast,
cohort studies are used for studying common conditions, such as HTN, so they

include very large sample sizes — but they still use the same 2x2 table for inter-
pretation. Figure 7-6 offers an example of what a 2x2 table for a cohort study
might look like where the exposure is high alcohol intake and the outcome is HTN.
FIGURE 7-6:
Example of a
typical cohort
study 2x2 table.
As shown in Figure 7-6, the total number of participants is large. This cohort of
600 participants could have been naturally sampled from the population, or they
could be stratified by exposure, meaning that the study design could require a cer-
tain number of participants to be exposed and to be unexposed. Imagine that you
insisted that 300 of your participants have high alcohol intake, and 300 have low
alcohol intake. It may be harder to recruit for the study, but you would be sure to
have enough exposed participants for your statistics to work out. In the case of
Figure 7-6, 210 exposed and 390 unexposed participants were enrolled.
In cohort studies, all the participants are examined upon entering the study, and
those with the outcome are not allowed to participate. Therefore, at the beginning
of the study, all 600 of the participants did not have the outcome, which is HTN.
A cohort study is essentially a series of cross-sectional studies on the same cohort
called waves. The first wave is baseline, when the participants enter the study (all
of whom do not have the outcome). Baseline values of important variables are
measured (and criteria about baseline values may be used to set inclusion criteria,
such as minimum age for the study). Subsequent waves of cross-sectional data
collection take place at regular time intervals (such as every year or every two
years). Changes in measured baseline values are tracked over time, and subgroups
of the cohort are compared in terms of outcome status. Figure 7-6 shows the
exposure status from baseline, and the outcome status from the first wave.
Because the exposure is measured in a cohort study before any participants get the
outcome, it is considered the highest level of evidence among the observational
study designs. It is far less biased than the case-control study design. Several
measures of relative risk can be used to interpret a cohort study, including the OR,
risk ratio, and incidence rate (see Chapter 14).

Advancing to the clinical trial stage
Higher up the pyramid of evidence shown in Figure 7-2 are experiments. Not all
experiments are at such a high level of evidence — only high-quality clinical tri-
als. These are experiments, not observational studies. This is where the researcher
assigns the participants to engage in a particular behavior or intervention during
the study. There are different types of clinical trials as described in Chapter 5;
however, the highest-quality trials use both double-blinding and randomization.
Double-blinding is where both the researcher and the participant do not know
whether the participant was assigned to an active intervention (one being stud-
ied), or a control intervention. Randomization is where participants are randomly
assigned to groups (so there is no bias in selecting participants for each group).
It is possible to use a 2x2 table to analyze the results of a high-quality clinical trial
as long as the rows are replaced with the intervention groups. You can report the
same measure of relative risk as for a cohort study; however, the difference is that
the high-quality clinical trial would be seen as having much less bias than the
cohort study — and stronger causal evidence.
Reaching the top: Systematic reviews

and meta-analyses
Imagine a scenario where a new drug for HTN was developed, and several clinical
trials were conducted to see whether this drug was better than the most popular
current drug used for HTN. How would we be able to know whether, on balance,
the new drug was actually better when we have so many different clinical trials on
the same drug with different results?
We could ask a similar question about observational studies as well. Imagine that
multiple case-control studies were conducted to determine whether having liver
cancer was associated with the exposure of having high prediagnosis alcohol
intake. What is the overall answer? Does high alcohol intake cause liver cancer or
not? You could also imagine that multiple cohort studies could be conducted
examining association between the exposure of high alcohol intake and develop-
ing the outcome of HTN. How would the results of these cohort studies be taken
together to answer the question of whether high alcohol intake actually
causes HTN?
The answer to this question are systematic reviews and meta-analyses. In a sys-
tematic review, researchers set up inclusion and exclusion criteria for reports of
studies. Included in those criteria are requirements for a certain study design. For

interventions, randomized clinical trials containing a control group (called
randomized controlled trials, or RCTs) are usually required, but for other exposures,
either case-control or cohort study designs are required. With respect to
medications, RCTs are required as part of regulatory approval for distribution (see
Chapter 5), so expect to see meta-analyses arising from results from clinical
trials. In a systematic review, the studies included are compared and summarized
in a table, but their numerical estimates coming from their results are not com-
bined. The meta-analysis is the same as a systematic review except the numerical
estimates coming from the results reported are combined statistically to produce
an overall estimate based on the studies included. Systematic reviews and meta-
analyses are described in more detail in Chapter 20.
If you are looking for the highest quality of evidence right now about a current
treatment or exposure and outcome, read the most recent systematic reviews
and meta-analyses on the topic. If there aren’t any, it may mean that the treat-
ment, exposure, or outcome is new, and that there are not a lot of high quality
observational or experimental studies published on the topic yet.

3
Getting Down
and Dirty
with Data
IN THIS PART . . .
Collect and validate your data, avoiding common

pitfalls up front that can cause trouble later on.
Summarize your data in informative tables and display

them in easy-to-understand graphs.
Understand the concepts of accuracy, which is how

close a sample statistic is to the corresponding
population parameter, and precision, which refers to
how close replicate values of the sample statistic are
to each other.
Calculate standard errors and confidence intervals

for the sample estimates you can calculate from
your data.
IN THIS CHAPTER
» Understanding levels of
measurement (nominal, ordinal,
interval, and ratio)
» Defining and entering different kinds

of data into your research database
» Making sure your data are accurate
» Creating a data dictionary to describe

the data in your database
Chapter 8
Getting Your Data into
the Computer
B
efore you can analyze data, you have to collect it and get it into the
computer in a form that’s suitable for analysis. Chapter 5 describes this
process as a series of steps — figuring out what data you need and how they
are structured, creating data entry forms and computer files to hold your data, and
entering and validating your data.
In this chapter, we describe a crucially important component of that process,

which is storing the data properly in your research database. Different kinds of
data can be represented in the computer in different ways. At the most basic level,
there are numerical values and classifications, and most of us can immediately tell
the two apart — you don’t have to be a math genius to recognize “age” as numer-
ical data, and “occupation” as categorical information.
So why are we devoting a whole chapter to describing, entering, and checking dif-
ferent types of data? It turns out that the topic of data storage is not quite as trivial
as it may seem at first. You need to be aware of some important details or you may
wind up collecting your data the wrong way and finding out too late that you can’t
run the appropriate analysis. This chapter starts by explaining the different levels
of measurement, and shows you how to define and store different types of data. It
also suggests ways to check your data for errors, and explains how to formally
describe your database so that others are able to work with it if you’re not available.
CHAPTER 8 Getting Your Data into the Computer 101

Looking at Levels of Measurement
Around the middle of the 20th century, the idea of levels of measurement caught
the attention of biological and social-science researchers and, in particular, psy-
chologists. One classification scheme, which has become widely used (at least in
statistics textbooks), recognizes four levels at which variables can be measured:
nominal, ordinal, interval, and ratio:
» Nominal variables are expressed as mutually exclusive categories, like

country of origin (United States, China, India, and so on), type of care provider
(nurse, physician, social worker, and so on), and type of bacteria (such as
coccus, bacillus, rickettsia, mycoplasma, or spirillum). Nominal indicates that
the sequence in which you list the different categories is purely arbitrary. For
example, listing type of care provider as nurse, physician, and social worker
is no more or less natural than listing them as social worker, nurse, and
physician.
» Ordinal data have categorical values (or levels) that fall naturally into a logical
sequence, like the severity of cancer (Stages I, II, III, and IV), or an agreement
scale (often called a Likert scale) with levels of strongly disagree, somewhat
disagree, neither agree nor disagree, somewhat agree, or strongly agree. Note
that the levels are not necessarily equally spaced with respect to the concep-
tual difference between levels.
» Interval data represents numerical measurements where, unlike with ordinal

classifications, the difference (or interval) between two numbers is a meaning-
ful measure in terms of being equally spaced, but the zero point is completely
arbitrary and does not denote the complete absence of what you’re measur-
ing. For example, a change from 20 to 25 degrees Celsius represents the
same amount of temperature increase as a change from 120 to 125 degrees
Celsius. But 0 degrees Celsius is purely arbitrary — it does not represent the
total absence of temperature; it’s simply the temperature at which water
freezes (or, if you prefer, ice melts).
» Ratio data, unlike interval data, does have a true zero point. The numerical
value of a ratio variable is directly proportional to how much there is of what
you’re measuring, and a value of zero means there’s nothing at all. Income
and systolic blood pressure are good examples of ratio data; an individual
without a job may have zero income, which is not as bad as having a systolic
blood pressure of 0 mmHg, because then that individual would no longer be
alive!
Statisticians may pontificate about levels of measurement excessively, pointing out

cases that don’t fall neatly into one of the four levels and bringing up various
counterexamples. Nevertheless, you need to be aware of the concepts and termi-
nology in the preceding list because you’ll see them in statistics textbooks and
102 PART 3 Getting Down and Dirty with Data

articles, and because teachers love to include them on tests. The level of measure-
ment of variables impacts how and to what precision data are collected. Other
level-of-measurement considerations include minimizing the data collected to
only what is needed, which also reduces data-privacy concerns and cost. And,
more practically, knowing the level of measurement of a variable can help you
choose the most appropriate way to analyze that variable.
Classifying and Recording

Different Kinds of Data
Although you should be aware of the four levels of measurement described in the
preceding section, you also need to be able to classify and deal with data in a more
pragmatic way. The following sections describe various common types of data
you’re likely to encounter in the course of clinical and other research. We point
out some considerations you need to think through before you start collecting
your data.
Making bad decisions (or avoiding making decisions) about exactly how to repre-
sent the data values in your research database can mess it up, and quite possibly
doom the entire study to eventual failure. If you record the values to your variables
the wrong way in your data, it may take an enormous amount of additional effort
to go back and fix them, and depending upon the error, a fix may not even be
possible!
Dealing with free-text data

It’s best to limit free-text variables that are difficult to box into one of the four
levels of measurement, such as participant comments or write-in fields for Other
choices in a questionnaire. Basically, you should only collect free-text variables
when you need to record verbatim what someone said or wrote. Don’t use
free-text fields as a lazy-person’s substitute for what should be precisely defined
categorical data. Doing any meaningful statistical analysis of free-text fields is
generally very difficult, if not impossible.
You should also be aware that most software has field-length limitations for text
fields. Although commonly used statistical programs like Microsoft Excel, SPSS,
SAS, R, and Python may allow for long data fields, this does not excuse you from
designing your study so as to limit collection of free-text variables. Flip to
Chapter 4 for an introduction to statistical software.

Assigning participant study
identification (ID) numbers
Every participant in your study should have a unique participant study identifier
(typically called a study ID). The study ID is present in the participant’s data and
is used for identifying the participant on study materials (for example, laboratory
specimens sent for analysis). You may need to combine two variables to create a
unique identifier. In a single-site study that is carried out at only one geographical
location, the study ID can be a whole number that is two- to four-digits long. It
doesn’t have to start at 1; it can start at 100 if you want all the ID numbers to be
three-digits long without leading zeros. In multi-site studies that are carried out at
several locations (such as different clinics or labs), the number often follows some
logic. For example, it could have two parts, such as a site number and a local study
ID number separated by a hyphen (for example, 03-104), which is where you need
two variables to get a unique ID.
Organizing name and address data

in the study ID crosswalk
A research database should not include private identifying information for the
participant, such as the participant’s full name and home address. Yet, these data
need to be accessible to study staff to facilitate the research. Private data like this
is typically stored in a spreadsheet called a study ID crosswalk. This spreadsheet
keeps a link (or crosswalk) between the participant’s study ID and their private
data not to be stored in the research database. When you store names in the study
ID crosswalk, choose one of the following formats so that you can easily sort par-
ticipants into alphabetical order, or use the spreadsheet to facilitate study
mailings:
» A single variable: Last, First Middle (like Smith, John A)
» Two columns: One for Last, another for First and Middle
You may also want to include separate fields to hold prefixes (Mr., Mrs., Dr., and
so on) and suffixes (Jr., III, PhD, and so forth).
Addresses should be stored in separate fields for street, city, state (or province),
ZIP code (or comparable postal code).

Collecting categorical data in
your research database
Setting up your data collection forms and database tables for categorical data
requires more thought than you may expect. You may assume you already know
how to record and enter categorical data. You just type in the values — such as
“United States,” “nurse,” or “Stage I” — right? Wrong! (But wouldn’t it be nice
if it were that simple?) The following sections look at some of the issues you have
to address when storing categorical values as research data.
Carefully coding categories

The first issue you need to decide is how to code the categories. How are you going
to store the values in the research database? Do you want to enter the type of care
provider as nurse, physician, or social worker; or as N, P, or SW; or as 1 = nurse, 2 =
physician, and 3 = social worker; or in some other manner? Most modern statistical
software can analyze categorical data with any of these representations, but it is
easiest for the analyst if you code the variables using numbers to represent the
categories. Software like SPSS, SAS, and R lets you specify a connection between
number and text (for example, attaching a label to 1 to make it display Nurse on
statistical output) so you can store categories using a numerical code while also
displaying what the code means on statistical output. In general, best practices are
to set conventions and be consistent, and make sure the content and meaning of
each variable is documented. You can also attach variable labels.
Nothing is worse than having to deal with a data set in which a categorical variable
has been stored with numerical codes, but there is no key to the codes and the
person who created the data set is no longer available. This is why maintaining a
data dictionary — described later in this chapter in “Creating a File that Describes
Your Data File” — is a critical step for ensuring you analyze your research data
properly.
Microsoft Excel doesn’t care whether you type a word or a number in a cell, which
can create problems when storing data. You can enter Type of Caregiver as N for the
first subject, nurse for the second, NURSE for the third, 1 for the fourth, and Nurse
for the fifth, and Excel won’t stop you or throw up an error. Statistical programs
like R would consider each of these entries as a separate, unique category. Even
worse, you may inadvertently add a blank space in the cell before or after the text,
which will be considered yet another category. Details such as case-sensitivity of
character values (meaning patterns of being upper or lowercase) can impact que-
ries. In Excel, avoid using autocomplete, and enter all levels of categorical vari-
ables as numerical codes (which can be decoded using your data dictionary).

Dealing with more than two levels in a category
When a categorical variable has more than two levels (like the Type of Caregiver or
Likert agreement scale examples we describe in the earlier section “Looking at Lev-
els of Measurement”), data storage gets even more interesting. First, you have to
ask yourself, “Is this variable a Choose only one or Choose all that apply variable?”
The coding is completely different for these two kinds of multiple-choice
variables.
You handle the Choose only one situation just as we describe for Type of Caregiver in
the preceding section — you establish numeric code for each alternative. For the
Likert scale example, if the item asked about patient satisfaction, you could have a
categorical variable called PatSat, with five possible values: 1 for strongly disagree,
2 for somewhat disagree, 3 for neither agree nor disagree, 4 for somewhat agree,
and 5 for strongly agree. And for the Type of Caregiver example, if only one kind of
caregiver is allowed to be chosen from the three choices of nurse, physician, or
social worker, you can have a categorical variable called CaregiverType with three
possible values: 1 for nurse, 2 for physician, and 3 for social worker. Depending
upon the study, you may also choose to add a 4 for other, and a 9 for unknown
(9, 99, and 999 are codes conventionally reserved for unknown). If you find
unexpected values, it is important to research and document what these mean to
help future analysts encountering the same data.
But the situation is quite different if the variable is Choose all that apply. For the
Type of Caregiver example, if the patient is being served by a team of caregivers,
you have to set up your database differently. Define separate variables in the data-
base (separate columns in Excel) — one for each possible category value. Imagine
that you have three variables called Nurse, Physician, and SW (the SW stands for
social worker). Each variable is a two-value category, also known as a two-state
flag, and is populated as 1 for having the attribute and 0 for not having the attrib-
ute. So, if participant 101’s care team includes only a physician, participant 102’s
care team includes a nurse and a physician, and participant 103’s care team
includes a social worker and a physician, the information can be coded as shown
in the following table.
Subject Nurse Physician SW
101 0 1 0
102 1 1 0
103 0 1 1
If you have variables with more than two categories, missing values theoretically
can be indicated by leaving the cell blank, but blanks are difficult to analyze in
statistical software. Instead, categories should be set up for missing values so they
can be part of the coding system (such as using a numerical code to indicate

unknown, refused, or not applicable). The goal is to make sure that for every cate-
gorical variable, a numerical code is entered and the cell is not left blank.
Never try to cram multiple choices into one column! For example, don’t enter 1, 2
into a cell in the CaregiverType column to indicate the patient has a nurse and phy-
sician. If you do, you have to painstakingly split your single multi-valued column
into separate two-state flag columns (described earlier) before you analyze the
data. Why not do it right the first time?
Recording numerical data

For numerical data (meaning interval and ratio data), the main issue is how much
precision to record. Recording a numeric value to as many decimals as you have
available is usually best. For example, if a scale can measure body weight to the
nearest tenth of a kilogram, record it in the database to that degree of precision.
You can always round off to the nearest kilogram later if you want, but you can
never “unround” a number to recover digits you didn’t record. So it’s best to
record values in your data from measurement instruments to the degree of preci-
sion provided.
Along the same lines, don’t group numerical data into intervals when recording it.
If you know the age to the nearest year, don’t record Age in 10-year intervals (such
as 20 to 29, 30 to 39, 40 to 49, and so on). You can always have the computer do
that kind of grouping later, but you can never recover the age in years if all you
record is the decade.
Some statistical programs let you store numbers in different formats. The pro-
gram may refer to these different storage modes using arcane terms for short, long,
or very long integers (whole numbers) or single-precision (short) or double-precision
(long) floating point (fractional) numbers. Each type has its own limits, which may
vary from one program to another or from one kind of computer to another. For
example, a short integer may be able to represent only whole numbers within the
range from 32,768 to 32.767, whereas a double-precision floating-point number
could easily handle a number like 1.23456789012345 10 250 . Excel has no trouble
storing numerical data in any of these formats, so to make these choices, it is best
to study the statistical program you will use to analyze the data. That way, you can
make rules for storing the data in Excel that make it easy for you to analyze the
data once it is imported into the statistical program.
Following are issues to consider with respect to numerical variables in Excel:
» Don’t put two numbers (such as a blood pressure reading of 135 / 85 mmHg)
into one column of data. Excel won’t complain about it, but it will treat it as

text because of the embedded “/”, rather than as numerical data. Instead,
create two separate variables and enter each number into the appropri-
ate variable.
» When recording multiple types of measurements (such as days, weeks,

months, and years), use two columns to record the data (such as time and type).
In the first column, store the value of the variable, and in the second column,
store a code to indicate the type (such as 1 = days, 2 = weeks, 3 = months, and
4 = years). As an example, “3 weeks” would be entered as a 3 in the time column
and 2 in the type column.
Missing numerical data requires a little more thought than missing categorical
data. Some researchers use 99 (or 999, or 9999) to indicate a missing value in
categorical data, but this approach should not be used for numeric data (because
the statistical program will see these values as actual measured values, and not
codes for missing data). The simplest technique for indicating missing numerical
data is to leave it blank. Most software treats blank cells as missing data in a cal-
culation, but this changes depending on the software, so it’s important to confirm
missing values handling in your analysis.
Entering date and time data

Now we’re going to tell you something that sounds like we’re contradicting the
advice we just gave you (but, of course, we’re not!). Most statistical software
(including Microsoft Excel) can represent dates and times as a single variable (an
“instant” on a continuous timeline), so take advantage of that if you can. In Excel,
you can enter the date and time as one variable (for example, 07/15/2020 08:23),
not as a separate date variable and a time variable. This method is especially use-
ful when dealing with events that take place over a short time interval (like events
occurring during a surgical procedure). It is important to collect all potential start
and end dates so any duration during the study can be calculated.
Some programs may store a date and time as a Julian Date, whose zero occurred at
noon, Greenwich Mean Time, on Jan. 1, 4713 BC. (Nothing happened on that date;
it’s purely a numerical convenience.)
What if you don’t know the day of the month? This happens a lot with medical
history items; a participant may say, “I got the flu in September 2021.” Most
software (including Excel) insists that a date variable be a complete date, and
won’t accept just a month and a year. In this case, a business rule is created to set
the day (as either the 1st, 15th, or last day of the month). Similarly, if both the
month and day are missing, you can set up a business rule to estimate both.

If you impute a date, just create a new column with the imputed date, because you
want to be cautious. Make sure to keep the original partial date for traceability.
Any date imputation should be consistent with the study protocol, and not bias the
results. Completely missing dates should be left blank, as statistical software
treats blank cells as missing data.
Because of the way most statistics programs store dates and times, they can easily
calculate intervals between any two points in time by simple subtraction. It is best
practices to store raw dates and times, and let the computer calculate the intervals
later (rather than calculate them yourself). For example, if you create variables for
date of birth (DOB) and a visit date (VisDt) in Excel, you can calculate an accurate
age at the time of the visit with this formula:
Age (VisDt DOB ) / 365.25
Checking Your Entered Data for Errors

After you’ve entered all your data into the computer, there are a few things you
can do to check for errors:
» Examine the smallest and largest values in numerical data: Have the
software show you the smallest and largest values for each numerical variable.
This check can often catch decimal-point errors (such as a hemoglobin value of
125 g/dL instead of 12.5 g/dL) or transposition errors (for example, a weight of
517 pounds instead of 157 pounds).
» Sort the values of variables: If your program can show you a sorted list of all
the values for a variable, that’s even better — it often shows misclassified
categories as well as numerical outliers.
» Search for blanks and commas: You can have Excel search for blanks
in category values that shouldn’t have blanks, or for commas in numeric
variables. Make sure the “Match entire cell contents” option is deselected in
the Find and Replace dialog box (you may have to click the Options button to
see the check box). This operation can also be done using statistical software.
Be wary if there a large number of missing values, because this could indicate
a data collection problem.
» Tabulate categorical variables: You can have your statistics program tabulate
each categorical variable (showing you the frequency each different category
occurred in your data). This check usually finds misclassified categories. Note
that blanks and special characters in character variables may cause incorrect
results when querying, which is why it is important to do this check.

» Spot-checking data entry: If doing data entry from forms or printed material,
choose a percentage to double-check (for example, 10 percent of the forms
you entered). This can help you tell if there are any systematic data entry
errors or missing data.
Creating a File that Describes

Your Data File
Every research database, large or small, simple or complicated, should include a
data dictionary that describes the variables contained in the database. It is a neces-
sary part of study documentation that needs to be accessible to the research team.
A data dictionary is usually set up as a table (often in Excel), where each row pro-
vides documentation for each variable in the database. For each variable, the dic-
tionary should contain the following information (sometimes referred to as
metadata, which means “data about data”):
» A variable name (usually no more than ten characters) that’s used when
telling the software what variables you want it to use in an analysis
» A longer verbal description of the variable in a human-readable format (in

other words, a person reading this description should be able to understand
the content of the variable)
» The type of data (text, categorical, numerical, date/time, and so on)
• If numeric: Information about how that number is displayed (how many

digits are before and after the decimal point)
• If date/time: How it’s formatted (for example, 12/25/13 10:50pm or

25Dec2013 22:50)
• If categorical: What codes and descriptors exist for each level of the
category (these are often called picklists, and can be documented on a
separate tab in an Excel data dictionary)
» How missing values are represented in the database (99, 999, “NA,”
and so on)
Database programs like SQL and statistical programs like SAS often have a func-
tion that can output information like this about a data set, but it still needs to be
curated by a human. It may be helpful to start your data dictionary with such out-
put, but it is best to complete it in Excel. That way, you can add the human cura-
tion yourself to the Excel data dictionary, and other research team members can
easily access the data dictionary to better understand the variables in the database.

IN THIS CHAPTER
» Representing categorical data
» Characterizing numerical variables
» Putting numerical summaries into

tables
» Displaying numerical variables with

bars and graphs
Chapter 9
Summarizing and
Graphing Your Data
A
large study can involve thousands of participants, hundreds of variables,
and millions of individual data points. You need to summarize this ocean
of individual values for each variable down to a few numbers, called
summary statistics, that give readers an idea of what the whole collection of
numbers looks like — that is, how they’re distributed.
When presenting your results, you usually want to arrange these summary
statistics into tables that describe how the variables change over time or differ
between categories, or how two or more variables are related to each other. And,
because a picture really is worth a thousand words, you will want to display these
distributions, changes, differences, and relationships graphically. In this chapter,
we show you how to summarize and graph both categorical and numerical data.
Note: This chapter doesn’t cover time-to-event (survival) data, which is the topic
of Chapter 22.
CHAPTER 9 Summarizing and Graphing Your Data 111

Summarizing and Graphing
Categorical Data
A categorical variable is summarized by tallying the number of participants in
each category and expressing this number as a count. You might also compute a
percentage of the total number of participants in all categories combined. So a
sample of 422 participants can be summarized by health insurance type, as shown
in Table 9-1.
TABLE 9-1 Study Participants Categorized by Health Insurance Type

Health Insurance Type Count Percent of Total
Commercial 128 30.3%
Public 141 33.4%
Military 70 16.6%
Other 83 19.7%
Total 422 100%
The joint distribution of participants between two categorical variables is sum-

marized by a cross-tabulation (or cross-tab). Table 9-2 shows an example of a
cross-tab of the same participants in our example with type of health insurance
on one axis, and urban-rural classification of their residence on the other.
TABLE 9-2 Cross-Tabulation of Participants by Two

Categorical Variables
Health Insurance Type
Commercial Public Military Other Total
Urban-Rural Rural 60 60 34 42 196

Classification of
Residence Urban 68 81 36 41 226
Total 128 141 70 83 422

After looking at the frequencies in Table 9-2, you may be curious about the per-
centages, which would make these numbers more comparable. But a cross-tab
can get very cluttered if you try to include them, as there are different types: the
column percentage, the row percentage, and the total percentage. For example,
the 60 rural residents with commercial health insurance in Table 9-2 comprise
46.9 percent of all participants with commercial health insurance, because
60 divided by the total number with commercial health insurance, which is 128
(the column total), equals 46.9 percent.
Groups are often compared across columns, and if that is the intention, column
percentages should be displayed. But if you divide these same 60 rural residents
with commercial insurance by their row total of 169 rural residents, you find they
make up 30.6 percent of all rural residents, which is a row percentage. And if you
go on to divide these 60 participants by the total sample size of the study, which
is 422, you find that they make up 14.2 percent of all participants in the study.
Categorical data are typically displayed graphically as frequency bar charts and as
pie charts:
» Frequency bar charts: Displaying the spread of participants across

the different categories of a variable is commonly done by a bar chart
(see Figure 9-1a). Generally, statistical programs are used to make bar charts.
To create a bar chart manually from a tally of participants in each category,
you draw a graph containing one vertical bar for each category, making the
height proportional to the number of participants in that category.
» Pie charts: Pie charts indicate the relative number of participants in each
category by the angle of a circular wedge, which can also be considered more
deliciously as a piece of the pie. To create a pie chart manually, you multiply
the percentage of participants in each category by 360, which is the number
of degrees of arc in a full circle, and then divide by 100. By doing that, you are
essentially figuring out what proportion of the circle to devote to that pie
piece. Next, you draw a circle with a compass, and then split it up into wedges
using a protractor — remember from high school math? Trust us, it’s easier to
use statistical software.
Most scientific writers recommend the usage of bar charts over pie charts. They
express more information in a smaller space, and allow for more accurate visual
comparisons.

FIGURE 9-1:
A frequency
bar chart (a)
and pie chart (b).
Summarizing Numerical Data

Summarizing a numerical variable isn’t as simple as summarizing a categorical
variable. The summary statistics for a numerical variable should convey how the
individual values of that variable are distributed across your sample in a concise
and meaningful way. These summary statistics should give you some idea of the
shape of the true distribution of that variable in the population from which you
draw your sample (read Chapter 3 and Chapter 6 to refresh your memory about
sampling). That true population distribution can have almost any shape, includ-
ing the typical shapes shown in Figure 9-2: normal, skewed, pointy-topped, and
bimodal (two-peaked).
FIGURE 9-2:
Four different
shapes of
distributions:
normal (a),
skewed (b),
pointy-topped (c),
and bimodal
(two-peaked) (d).

How can you convey a visual picture of what the true distribution may look like by
using just a few summary numbers? By reporting values of measures of some
important characteristics of these distributions, so that the reader can infer the
shape. This is similar to learning that one Olympic ice skater scored an average of
9.0 compared to another who scored an average of 5.0. You will not know what the
skate routines looked like unless you watch them, but the score will already tell
you that if you were to watch them, you would expect to see that the one that
scored 9.0 was executed in a more visually pleasing way than the one that
scored 5.0.
Frequency distributions have names for their important characteristics, including:
» Center: Where along the distribution of the values do the numbers tend
to center?
» Dispersion: How much do these numbers spread out?
» Symmetry: If you were to draw a vertical line down the middle of the
distribution, does the distribution shape appear as if the vertical line is a
mirror, reflecting an identical shape on both sides? Or do the sides look
noticeably different — and if so, how?
» Shape: Is the top of the distribution nicely rounded, or pointier, or flatter?
Like using average skating scores to describe the visual appeal of an Olympic skate
routine, to describe a distribution you need to calculate and report numbers that
measure each of these four characteristics. These characteristics are what we
mean by summary statistics for numerical variables.
Locating the center of your data

When you start exploring a set of numbers, an important first step is to determine
what value they tend to center around. This characteristic is called, intuitively
enough, central tendency. Many statistical textbooks describe three measures of
central tendency: mean (which is the same as average), median, and mode. You may
assume these are the three optimal measures to describe a distribution (because
they all begin with m and are easy to remember). But all three have limitations,
especially when dealing with data obtained from samples in human research, as
described in the following sections.
Arithmetic mean
The arithmetic mean, also commonly called the mean (or the average), is the most
familiar and most often quoted measure of central tendency. Throughout this
book, whenever we use the two-word term the mean, we’re referring to the

arithmetic mean. (There are several other kinds of means besides the arithmetic
mean, which we describe later in this chapter.)
The mean of a sample is often denoted by the symbol m or by placing a horizontal

bar over the name of the variable, like X . The mean is obtained by adding up the
values and dividing by the sample size — meaning how many there are. (If you are
using software for this, make sure missing values are excluded, or the equation
will not compute.) Here’s a small sample of numbers — the diastolic blood pres-
sure (DBP) values of seven study participants (in mmHg) arranged in increasing
numerical order: 84, 84, 89, 91, 110, 114, and 116. For the DBP sample:
Arithmetic Mean 84 84 89 91 110 114 116 / 7

688 / 7 98.3 approximately
You can write the general formula for the arithmetic mean of N number of values
contained in the variable X in several ways:
N
Xi Xi
i 1 i
X
Arithmetic Mean m X
N N N
See Chapter 2 for a refresher on mathematical notation and formulas, including

how to interpret the various forms of the summation symbol Σ (the Greek capital
sigma). In the rest of this chapter, we use the simplest form, meaning the form
without the i subscripts that refer to specific elements of an array, whenever
possible.
Some statistical books use the notation such that capital X and capital N refer to
census parameters, and lowercase versions of those to refer to sample statistics.
In this book, we make it clear each time we present this notation whether we are
talking about a census or a sample.
Median
Like the mean, the median is a common measure of central tendency. In fact, it
could be argued that the median is the only one of the three that really takes the
word central seriously.
The median of a sample is the middle value in the sorted (ordered) set of numbers.
By definition, half of the numbers are smaller than the median, and half are larger.
The median of a population frequency distribution function (like the curves shown
in Figure 9-2) divides the total area under the curve into two equal parts: Half of
the area under the curve (AUC) lies to the left of the median, and half lies to
the right.
Consider the sample of diastolic blood pressure (DBP) measurements from seven
study participants from the preceding section. If you arrange the values in order

from lowest to highest mmHg, you can list them as 84, 84, 89, 91, 110, 114,
and 116. There are seven values, and 91 is the fourth of the seven sorted values, so
that is the median. Three DBPs in the sample are smaller than 91 mmHg, and
three are larger than 91 mmHg. If you have an even number of values, the median
is the average of the two middle values. So imagine that you add a value of 118
mmHg to the top of your list, so you now have eight values. To get the median, you
would make an average of the fourth and fifth value, which would be (91 + 110)/2
= 100.5 mmHg (don’t be thrown off by the 0.5).
Statisticians often say that they prefer the median to the mean because the median
is much less strongly influenced by extreme outliers than the mean. For example,
if the largest value for DBP had been very high — such as 150 mmHg instead of
116 mmHg — the mean would have jumped from 98.3 mmHg up to 103.1 mmHg.
But in the same case, the median would have remained unchanged at 91. Here’s an
even more extreme example: If a multibillionaire were to move into a certain
state, the mean family net worth in that state might rise by hundreds of dollars,
but the median family net worth would probably rise by only a few cents (if it were
to rise at all). This is why you often hear the median rather than mean income in
reports comparing income across regions.
Mode
The mode of a sample of numbers is the most frequently occurring value in the
sample. One way to remember this is to consider that mode means fashion in
French, so the mode is the most popular value in the data set. But the mode has
several issues when it comes to summarizing the centrality of observed values for
continuous numerical variables. Often there are no exact duplicates, so there is no
mode. If there are any exact duplicates, they usually are not in the center of the
data. And if there is more than one value that is duplicated the same number of
times, you will have more than one mode.
So the mode is not a good summary statistic for sampled data. But it’s useful for
characterizing a population distribution, because it’s the value where the peak of
the distribution function occurs. Some distribution functions can have two peaks
(a bimodal distribution), as shown earlier in Figure 9-2d, indicating two distinct
subpopulations, such as the distribution of age of death from influenza in many
populations, where we see a mode in young children, and another mode in older
adults.
Considering some other “means” to

measure central tendency
Several other kinds of means are useful measures of central tendency in certain
circumstances. They’re called means because they all calculated using the same

approach. The difference is that each type of mean adds a slightly different twist
to the basic mathematical process.
INNER MEAN
The inner mean (also called the trimmed mean) of N numbers is calculated by
removing the lowest value (the minimum) and the highest value (the maximum),
and calculating the arithmetic mean of the remaining N – 2 inner values. For the
sample of seven values of DBP from study participants from the example used
earlier in this chapter (which were 84, 84, 89, 91, 110, 114, and 116 mmHg), you
would drop the minimum and the maximum to compute the inner
mean: 84 89 91 110 114 / 5 488 / 5 97.6.
An inner mean that is even more inner can be calculated by making an even
stricter rule. The rule could be to drop the two (or more) of the highest and two (or
more) of the lowest values from the data, and then calculate the arithmetic mean
of the remaining values. In the interest of fairness, you should always chop the
same number of values from the low end as from the high end. Like the median
(discussed earlier in this chapter), the inner mean is more resistant to extreme
values called outliers than the arithmetic mean.
GEOMETRIC MEAN
The geometric mean (often abbreviated GM) can be defined by two different-
looking formulas that produce exactly the same value. The basic definition has
this formula:
N
Geometric Mean GM IIX
We describe the product symbol Π (the Greek capital pi) in Chapter 2. This formula
is telling you to multiply the values of the N observations together, and then take
the Nth root of the product. Using the numbers from the earlier example (where
you had DBP data on seven participants, with the values 84, 84, 89, 91, 110, 114,
and 116 mmHg), the equation looks like this:
7 7
GM 84 84 89 91 110 114 116 83,127, 648,746,160, 93.4
Even with technology, this formula is computationally challenging. By using log-

arithms (which turn multiplications into additions and roots into divisions), you
can develop a numerically stable alternative formula, which is:
log( X ) log( X )
log( GM ) ,or GM antilog
N N
This formula may look complicated, but it really just says, “The geometric mean
is the antilog of the mean of the logs of the values in the sample.” In other words,

to calculate the GM using this formula, you take the log of each value in your
sample, then average all those logs together, and then take the antilog of that
average. You can choose to use either natural or common logarithms, but make
sure that whatever you choose, you use same type of antilog. (Flip to Chapter 2 for
the basics of logarithms.)
Describing the spread of your data

After central tendency (described earlier in “Locating the center of your data”),
the second most important set of summary statistics for numerical values refers
to how tightly or loosely they tend to cluster around a central value, meaning how
they are dispersed. There are several common measures of dispersion, as you find
out in the following sections.
Standard deviation, variance, and

coefficient of variation
The standard deviation (usually abbreviated SD, sd, or just s) of a set of numerical
values tells you how much the individual values tend to differ from the mean in
either direction (see “Locating the center of your data” for a discussion of the
mean). The SD is calculated as follows:
( di )2
i
SD sd s where d i Xi X
N 1
This formula is saying that you calculate the SD of a set of N numbers by first
subtracting the mean from each value ( X i ) to get the deviation (d i ) of each value
from the mean. Then, you take the square each of these deviations and add up the
d i2 terms. After that, you divide that number by N – 1, and finally, you take
the square root of that number to get your answer, which is the SD.
For the sample of diastolic blood pressure (DBP) measurements for seven study
participants in the example used earlier in this chapter, where the values are 84,
84, 89, 91, 110, 114, and 116 mmHg and the mean is 98.3 mmHg, you calculate the
SD as follows:
( 84 98.3 ) 2 ( 84 98.3 ) 2 ... (116 98.3 ) 2

SD 14.4
7 1
Several other useful measures of dispersion are related to the SD:
» Variance: The variance is just the square of the SD. For the DBP example, the
variance 14.4 2 207.36.

» Coefficient of variation: The coefficient of variation (CV) is the SD divided by
the mean. For the DBP example, CV 14.4 / 98.3 0.1465, or 14.65 percent.
Range
The range of a set of values is the minimum value subtracted from the maximum
value:
Range maximum value minimum value
Consider the example from the preceding section, where you had DBP measure-
ments from seven study participants (which were 84, 84, 89, 91, 110, 114, and
116 mmHg). The minimum value is 84, the maximum value is 116, and the range
is 32 (equal to 116 84).
Centiles
The basic idea of the median is that ½ (half) of your numbers are less than the
median, and the other ½ are greater than the median. This concept can be
extended to other fractions besides ½.
A centile (also referred to as percentile) is a value that a certain percentage of the

values are less than. For example, ¼ of the values are less than the 25th centile
(and ¾ of the values are greater). The median is just the 50th centile. The 25th,
50th, and 75th centiles are called the first, second, and third quartiles, respectively,
and are used often. There are other sets of centiles, such as deciles, which break at
every ten percentiles, that are used less often.
As we explain in the earlier section “Median,” if the sorted sequence of your

numerical variable has no middle value, you have to calculate the median as the
average of the two middle numbers. The same situation comes up in calculating
centiles, but there are different ways that statistical software does the calculation.
Fortunately, the different formulas they use give nearly the same result.
The inter-quartile range (IQR) is the difference between the 25th and 75th centiles
(the first and third quartiles).
Numerically expressing the symmetry

and shape of the distribution
In the following sections, we discuss two summary statistics used to describe
aspects of the symmetry and shape of the distribution of values of numerical vari-
ables (pictured earlier in Figure 9-2).

Skewness
Skewness refers to the left-right symmetry of the distribution. Figure 9-3 illus-
trates some examples.
FIGURE 9-3:
Distributions
can be left-
skewed (a),
symmetric (b), or
right-skewed (c).
Figure 9-3b shows a symmetrical distribution. If you look back to Figures 9-2a
and 9-2c, which are also symmetrical, they look like the vertical line in the
center is a mirror reflecting perfect symmetry, so these have no skewness. But
Figure 9-2b has a long tail on the right, so it is considered right skewed (and if you
flipped the shape horizontally, it would have a long tail on the left, and be consid-
ered left-skewed, as in Figure 9-3a).
How do you express skewness in a summary statistic? The most common skew-
ness coefficient, often represented by the Greek letter γ (lowercase gamma), is
calculated by averaging the cubes (third powers) of the deviations of each point
from the mean and scaling by the SD. Its value can be positive, negative, or zero.
Here is how to interpret the skewness coefficient (γ):
» A negative γ indicates left-skewed data (Figure 9-3a).
» A zero γ indicates unskewed data (Figures 9-2a and 9-2c, and Figure 9-3b).
» A positive γ indicates right-skewed data (Figures 9-2b and 9-3c).
Notice that in Figure 9-3a, which is left-skewed, the γ = –0.7, and for Figure 9-3c,
which is right-skewed, the γ = 0.7. And for Figure 9-3b — the symmetrical
distribution — the γ = 0, but this almost never happens in real life. So how large
does γ have to be before you suspect real skewness in your data? A rule of thumb
for large samples is that if γ is greater than 4 / N , your data are probably skewed.
Kurtosis
Kurtosis is a less-used summary statistic of numerical data, but you still need to
understand it. Take a look at the three distributions shown in Figure 9-4, which

all have the same mean and the same SD. Also, all three have perfect left-right
symmetry, meaning they are unskewed. But their shapes are still very different.
Kurtosis is a way of quantifying these differences in shape.
FIGURE 9-4:
Three
distributions:
leptokurtic (a),
normal (b), and
platykurtic (c).
A good way to compare the kurtosis of the distributions in Figure 9-4 is through
the Pearson kurtosis index. The Pearson kurtosis index is often represented by the
Greek letter k (lowercase kappa), and is calculated by averaging the fourth powers
of the deviations of each point from the mean and scaling by the SD. Its value can
range from 1 to infinity and is equal to 3.0 for a normal distribution. The excess
kurtosis is the amount by which k exceeds (or falls short of) 3.
One way to think of kurtosis is to see the distribution as a body silhouette. If you
think of a typical distribution function curve as having a head (which is near the
center), shoulders on either side of the head, and tails out at the ends, the term
kurtosis refers to whether the distribution curve tends to have
» A pointy head, fat tails, and no shoulders, which is called leptokurtic, and is
shown in Figure 9-4a (where k 3).
» An appearance of being normally distributed, as shown in Figure 9-4b

(where k 3 ).
» Broad shoulders, small tails, and not much of a head, which is called
platykurtic. This is shown in Figure 9-4c (where k 3 ).
A very rough rule of thumb for large samples is that if k differs from 3 by more than
8 / N , your data have abnormal kurtosis.

Structuring Numerical Summaries
into Descriptive Tables
Now you know how to calculate the basic summary statistics that convey the gen-
eral idea of how a set of numerical values is distributed. So which summary sta-
tistics do you report? Generally, you select a few of the most useful summary
statistics in summarizing your particular data set, and arrange them in a concise
way. Many biostatisticians choose to report N, mean, SD, median, minimum, and
maximum, and arrange them something like this:
mean SD N
median minimum maximum
Consider the example used earlier in this chapter of seven measures of diastolic
blood pressure (DBP) from a sample of study participants (with the values of 84,
84, 89, 91, 110, 114, and 116 mmHg), where you calculated all these summary sta-
tistics. Remember not to display decimals beyond what were collected in the orig-
inal data. Using this arrangement, the numbers would be reported this way:
98.3 14.4 7
91 84 116
The real utility of this kind of compact summary is that you can place it in each
cell of a table to show changes over time and between groups. For example, a
sample of systolic blood pressure (SBP) measurements taken from study partici-
pants before and after treatment with two different hypertension drugs (Drug A
and Drug B) can be summarized concisely, as shown in Table 9-3.
TABLE 9-3 Systolic Blood Pressure Treatment Results

Before Treatment After Treatment Change
Median Median Median

Mean ± SD (N) (min – max) Mean ± SD (N) (min – max) Mean ± SD (N) (min – max)
Drug 138.7 ± 10.3 (40) 139.5 121.1 ± 13.9 (40) 121.5 -17.6 ± 8.0 (40) –17.5 (–34 – 4)
A (117 – 161) (85 – 154)
Drug 141.0 ± 10.8 (40) 143.5 141.0 ± 15.4 (40) 142.5 -0.1 ± 9.9 (40) 1.5 (–25 – 18)
B (111 – 160) (100 – 166)

Table 9-3 shows that Drug A tended to lower blood pressure by about 18 mmHg.
For Drug A, mean SBP changed from 139 to 121 mmHg from before to after treat-
ment, whereas the Drug B group produced no noticeable change in blood pressure
because it stayed around 141 mmHg from pretreatment to post-treatment. All
that’s missing are some p values to indicate the significance of the changes over
time within each group and of the differences between the groups. We show you
how to calculate those in Chapter 11.
Graphing Numerical Data

Displaying information graphically is a central part of interpreting and commu-
nicating the results of scientific research. You can easily spot subtle features in a
graph of your data that you’d never notice in a table of numbers. Entire books have
been written about graphing numerical data, so we only give a brief summary of
some of the more important points here.
Showing the distribution with histograms

Histograms are bar charts that show what fraction of the participants have values
falling within specified intervals called classes. The main purpose of a histogram is
to show you how the values of a numerical value are distributed. This distribution
is an approximation of the true population frequency distribution for that vari-
able, as shown in Figure 9-5.
FIGURE 9-5:
Population
distribution of
systolic blood
pressure (SBP)
measurements in
mmHg (a) and
distribution of a
sample from that
population (b).

The smooth curve in Figure 9-5a shows how SBP values are distributed in an infi-
nitely large population. The height of the curve at any SBP value is proportional to
the fraction of the population in the immediate vicinity of that SBP. This curve has
the typical bell shape of a normal distribution.
The histogram in Figure 9-5b indicates how the SBP measurements of 60 study
participants randomly sampled from the population might be distributed. Each
bar represents an interval or class of SBP values with a width of ten mmHg. The
height of each bar is proportional to the number of participants in the sample
whose SBP fell within that class.
Log-normal distributions
Because a sample is only an imperfect representation the population, determining
the precise shape of a distribution can be difficult unless your sample size is very
large. Nevertheless, a histogram usually helps you spot skewed data, as shown in
Figure 9-6a. This kind of shape is typical of a log-normal distribution
(Chapter 25), which is a distribution you often see when analyzing biological
measurements, such as lab values. It’s called log-normal because if you take a
logarithm (of any type) of each data value, the resulting logs will have a normal
distribution, as shown in Figure 9-6b.
FIGURE 9-6:
Log-normal
data are
skewed (a), but
the logarithms
are normally
distributed (b).
Because distributions are so important to biostatistics, it’s a good practice to pre-

pare a histogram for every numerical variable you plan to analyze. That way, you
can see whether it’s noticeably skewed and, if so, whether a logarithmic transfor-
mation makes the distribution normal enough so you can use statistics intended
for normal distributions on your data.

If you can’t find any transformation that makes your data look even approxi-
mately normal, then you have to analyze your data using nonparametric methods,
which don’t assume that your data are normally distributed.
Summarizing grouped data with bars,

boxes, and whiskers
Sometimes you want to show how a numerical variable differs from one group of
participants to another. For example, blood levels of a certain cardiovascular
enzyme vary among the cardiology patients at four different clinics: Clinic A, B, C,
and D. Two types of graphs are commonly used for this purpose: bar charts and
box-and-whiskers plots.
Bar charts
One simple way to display and compare the means of several groups of data is
with a bar chart, like the one shown in Figure 9-7a. Here, the bar height for each
group of patients equals the mean (or median, or geometric mean) value of the
enzyme level for patients at the clinic represented by the bar. And the bar chart
becomes even more informative if you indicate the spread of values for each clini-
cal sample by placing lines representing one SD above and below the tops of the
bars, as shown in Figure 9-7b. These lines are always referred to as error bars,
which is an unfortunate choice of words that can cause confusion when error bars
are added to a bar chart. In this case, error refers to statistical error (described in
Chapter 6).
FIGURE 9-7:
Bar charts
showing mean
values (a) and
standard
deviations (b).
But even with error bars, a bar chart still doesn’t provide a picture of the distribu-
tion of enzyme levels within each group. Are the values skewed? Are there outliers?
Imagine that you made a histogram for each subgroup of patients — Clinic A,
Clinic B, Clinic C, and Clinic D. But if you think about it, four histograms would take
up a lot of space. There is a solution for this! Keep reading to find out what it is.

Box-and-whiskers charts
The box-and-whiskers plot (or B&W, or just box plot) plot uses very little space to
display a lot of information about the distribution of numbers in one or more
groups of participants. A box plot of the same enzyme data used in Figure 9-7 is
shown in Figure 9-8a.
FIGURE 9-8:
Box-and-whiskers
charts: no-frills (a)
and with variable
width and
notches (b).
Looking at Figure 9-8a, you notice the box plot for each group has the following
parts:
» A box spanning the interquartile range (IQR), extending from the first quartile
of the variable to the third quartile, thus encompassing the middle 50 percent
of the data.
» A thick horizontal line, drawn at the median, which is also the 50th centile. If
this falls in the middle of the box, your data are not skewed, but if it falls on
either side, be on the lookout for skewness.
» Lines called whiskers extending out to the farthest data point that’s not more
than 1.5 times the IQR away from the box, and terminate with a horizontal bar
on each side.
» Individual points lying outside the whiskers, which are considered outliers.
Box plots provide a useful visual summary of the distribution of each subgroup for
comparison, as shown in Figure 9-8a. As mentioned earlier, a median that’s not
located near the middle of the box indicates a skewed distribution.
Some software draws the different parts of a box plot according to different rules,
so you should always check your software’s documentation before you present a
box plot so you can describe your box plot accurately.

Software can provide various enhancements to the basic box plot. Figure 9-8b
illustrates two such embellishments you may consider using:
» Variable width: The widths of the bars can be scaled to indicate the relative
size of each group.
» Notches: The box can have notches that indicate the uncertainty in the
estimation of the median. If two groups have non-overlapping notches, they
probably have significantly different medians.
Depicting the relationships between

numerical variables with other graphs
We started this chapter by developing summary statistics and making graphs of
one numeric variable at a time. One example was where we took seven measure-
ments of diastolic blood pressure (DBP) from a group of study participants and
developed summary statistics. This is called a univariate analysis because it only
concerns one variable. But in the example of box plots in the preceding section, we
conducted a bivariate analysis because we were looking at the relationship between
two variables in a sample of patients from four different clinics. The two variables
were enzyme levels, and source clinic (Clinic A, B, C, or D). We could have done
another bivariate analysis looking at two continuous variables (such as two differ-
ent enzyme levels in participants) using a scatter plot, which is covered thor-
oughly in Chapter 16.
This chapter focused on univariate and bivariate summary statistics and graphs
that can be developed to help you and others better understand your data. But
many research questions are actually answered using multivariate analysis, which
allows for the control of confounders. Being able to control for confounders is one
of the main reasons biostatisticians opt for regression analysis, which we describe
in Part 5 and Chapter 23. In these chapters, we cover the appropriate summary
statistics and graphical techniques for showing relationships between variables
when setting up multivariate regression models.

IN THIS CHAPTER
» Investigating the basics of confidence

intervals
» Calculating confidence intervals for

several different statistics
» Linking significance testing to

confidence intervals
Chapter 10
Having Confidence
in Your Results
I
n Chapter 3, we describe how statistical inference relies on both accuracy and
precision when making estimates from your sample. We also discuss how the
standard error (SE) is a way to indicate the level of precision of your sample
statistic, but that SE is only one way of expressing the preciseness of your statis-
tic. In this chapter, we focus on another way — through the use of a confidence
interval (CI).
We assume that you’re familiar with the concepts of populations, samples, and
statistical estimation theory (see Chapters 3 and 6 if you’re not), and that you
know what SEs are (read Chapter 3 if you don’t). Keep in mind that when you
conduct a human research study, you’re typically enrolling a sample of study par-
ticipants drawn from a hypothetical population. For example, you may enroll a
sample of 50 adult diabetic patients who agree to be in your study as participants,
but they represent the hypothetical population of all adults with diabetes (for
details about sampling, turn to Chapter 6). Any numerical estimate you observe
from your sample is a sample statistic. A statistic is a valid but imperfect estimate
of the corresponding population parameter, which is the true value of that quantity
in the population.
CHAPTER 10 Having Confidence in Your Results 129

Feeling Confident about Confidence
Interval Basics
The main part of this chapter is about how to calculate confidence intervals
(Cis) around the sample statistics you get from research samples. But first, it’s
important for you to be comfortable with the basic concepts and terminology
related to CIs.
Defining confidence intervals

Informally, a confidence interval indicates a range (or interval) of numerical values
that’s likely to encompass the true value. More formally, the CI around your sam-
ple statistic is calculated in such a way that it has a specified likelihood of includ-
ing or containing the value of the corresponding population parameter.
The SE is usually written after a sample mean with a ± (read “plus or minus”)
symbol followed by the number representing the SE. As an example, you may
express a mean and SE blood glucose level measurement from a sample of adult
diabetics as 120 ± 3 mg/dL. By contrast, the CI is written as a pair of numbers —
known as confidence limits (CLs) — separated by a dash. The CI for the sample
mean and SE blood glucose could be expressed like this: 114 – 126 mg/dL. Notice
that 120 mg/dL — the mean — falls in the middle of the CI. Also, note that the
lower confidence limit (LCL) is 114 mg/dL, and the upper confidence limit (UCL) is
126 mg/dL. Instead of LCL and UCL, sometimes abbreviations are used, and are
written with a subscript L or U (as in CL L or CL U) indicating the lower and upper
confidence limits, respectively.
Although SEs and CIs are both used as indicators of the precision of a numerical
quantity, they differ in what they are intending to describe (the sample or the
population):
» A SE indicates how much your observed sample statistic may fluctuate if the
same study is repeated a large number of times, so the SE intends to describe
the sample.
» A CI indicates the range that’s likely to contain the true population parameter,
so the CI intends to describe the population.

If you want to have a more precise estimate of your population parameter from
your sample statistic, it’s best if the SEs are small and the CIs narrow. One impor-
tant property of both CIs and SEs is that how big they are varies inversely with the
square root of the sample size. For example, if you were to blow up your sample
size — let’s pretend to quadruple it — it would cut the size of the SE and the width
of the CI in half! This square root law is one of the most widely applicable rules in
all of statistics, and is the reason why you often hear researchers trying to find
ways to increase the sample size in their studies. In practice, a reasonable sample
size is reached based on budget and historical studies, because including the whole
population is usually not possible (or necessary).
Understanding and interpreting

confidence levels
The probability that the CI encompasses the true value of the population parame-
ter is called the confidence level of the CI. You can calculate a CI for any confidence
level, but the most commonly seen value is 95 percent. Whenever you report a CI,
you must state the confidence level. As an example, let’s restate our CI from the
analysis of mean blood glucose levels in a sample of adult diabetics to express that
we used the 95 percent confidence level: 95 percent CI = 114 – 126 mg/dL.
In general, higher confidence levels correspond to wider confidence intervals (so

you can have greater confidence that the interval encompasses the true value),
and lower confidence level intervals are narrower. As an example, a 90 percent CI
for the same data is a smaller range (115–125 mg/dL) and the 99 percent CI is a
larger range (112–128 mg/dL).
Although a 99 percent CI may be attractive, it can be hard to achieve in practice

because an exponentially larger sample is needed (as described earlier in this sec-
tion). Also, the wide range it provides can be relatively unhelpful. While dropping
to a 90 percent CI would reduce the range and sample size needed, having only
90 percent confidence that the true value is in the range is also not very helpful.
This may be why there seems to be an industry standard to use the 95 percent
confidence level when calculating and reporting CIs.
The confidence level is sometimes abbreviated CL, just like the confidence limit,
which can be confusing. Fortunately, the distinction is usually clear from the con-
text in which CL appears. When it’s not clear, we spell out what CL stands for.

FEEL CONFIDENT: DON’T
LIVE ON AN ISLAND!
Imagine that you enroll a sample of participants from some defined population in a
study and obtain a sample statistic. As an example, you calculate a mean blood glucose
level from a sample of 50 adult diabetics representing the background population of all
adult diabetics. Assume you calculate a 95 percent CI around this statistic, and then you
assert that you are 95 percent confident that your CI contains the true population value.
But what does that even mean? How can anyone be 95 percent confident? What does
that feel like?
There is a popular simulation to illustrate the interpretation of CIs and help learners
understand what it is like to be 95 percent confident. Imagine that you have a Microsoft
Excel spreadsheet, and you make up an entire population of 100 adult diabetics (maybe
they live on an island?). You make up a blood glucose measurement for each of them
and type it into the spreadsheet. Then, when you take the average of this entire column,
you get the true population parameter (in our simulation). Next, randomly choose a
sample of 50 measurements from your population of 100, and calculate a sample mean
and a 95 percent CI. Your sample mean will probably be different than the population
parameter, but that’s okay — that’s just sampling error.
Here’s where the simulation gets hard. You actually have to take 100 samples of 50. For
each sample, you need to calculate the mean and 95 percent CI. You may find yourself
making a list of the means and CIs from your 100 samples on a different tab in the
spreadsheet. Once you are done with that part, go back and refresh your memory as to
what the original population parameter really is. Get that number, then review all 100
CIs you calculated from all 100 samples of 50 you took from your imaginary population.
Because you made 95 percent CIs, 95 out of your 100 CIs will contain the true popula-
tion parameter (and 5 of them won’t)! This simulation is a way of demonstrating a proof
of the central limit theorem (CLT), and helps learners understand what it means to be
95 percent confident about their CI.
Taking sides with confidence intervals

As demonstrated in the simulation described in the sidebar “Feel Confident: Don’t
Live on an Island!”, 95 percent CIs contain the true population value 95 percent of
the time, and fail to contain the true value the other 5 percent of the time. Usually,
95 percent confidence limits are calculated to be balanced, so that the 5 percent
failures are split evenly. This means that the true population parameter is actually
less than the lower confidence limit 2.5 percent of the time, and it is actually
greater than the upper confidence limit 2.5 percent of the time. This is called a
two-sided, balanced CI.

In some situations, you may want all the failures to be on one side. In other words,
you want a one-sided confidence limit. Cars that run on gasoline may have a dec-
laration by their manufacturer that they go an average distance of at least 40 miles
per gallon (mpg). If you were to test this by keeping track of distance traveled and
gas usage on a sample of car trips, you may only be concerned if the average was
below the lower confidence limit, but not care if it was above the upper confidence
limit. This makes the boundary on one side infinite (which would really save you
money on gas!). For example, from the results of your study, you could have an
observed value of 45 mpg, with a one-sided confidence interval that goes from
42 mpg to plus infinity mpg!
In biostatistics, it is traditional to always use two-way CIs rather than one-way

CIs, as these are seen as most conservative.
Calculating Confidence Intervals

Although an SE and a CI are different calculations intended to express different
information, they are related in that the SE is used in the CI calculation. SEs and
CIs are calculated using different formulas (depending on the type of sample sta-
tistic for which you are calculating the SE and CI). In the following sections, we
describe methods of calculating SEs and CIs for commonly used sample
statistics.
Before you begin: Formulas for confidence

limits in large samples
Most of the methods we describe in the following sections are based on the
assumption that your sample statistic has a sampling distribution that’s approxi-
mately normal (Chapter 3 covers sampling distributions). There are strong theo-
retical reasons to assume a normal or nearly normal sampling distribution if you
draw a large enough samples.
For any normally distributed sample statistic, the lower and upper confidence
limits can be calculated from the observed value of the statistic (V) and standard
error (SE) of the statistic:
CL L V k SE
CLU V k SE

As you can see, CI calculations include a k x SE component, which is both added to
and subtracted from the estimate to get the limits. This component is called the
margin of error (ME).
Confidence limits computed this way are often referred to as normal-based,

asymptotic, or central-limit-theorem (CLT) confidence limits. The value of k in the
formulas depends on the desired confidence level and can be obtained from a table
of critical values for the normal distribution. Table 10-1 lists the k values for some
commonly used confidence levels.
TABLE 10-1 Multipliers for Normal-Based Confidence Intervals

Confidence Level Tail Probability k Value
50% 0.50 0.67
80% 0.20 1.28
90% 0.10 1.64
95% 0.05 1.96
98% 0.02 2.33
99% 0.01 2.58
For the most commonly used confidence level, 95 percent, k is 1.96, or approxi-
mately 2. This leads to the very simple approximation that 95 percent upper con-
fidence limit is about two SEs above the value, and the lower confidence limit is
about two SEs below the value.
The confidence interval around a mean

Suppose that you enroll a sample of 25 adult diabetics (N = 25) as participants in
a study, and find that they have an average fasting blood glucose level of 130 mg/
dL with a standard deviation (SD) of ±40 mg/dL. What is the 95 percent confi-
dence interval around that 130 mg/dL estimated mean?
To calculate the confidence limits around a mean using the formulas in the pre-
ceding section, you first calculate the SE, which in this case is the standard error
of the mean (SEM). The formula for the SEM is SEM SD / N , where SD is the SD
of the sample values, and N is the number of values included in the calculation.
For the fasting blood glucose study sample, where your SD was 40 mg/dL and your
sample size was 25, the SEM is SEM 40 / 25 , which is equal to 40/5, or 8 mg/dL.

Using k = 1.96 for a 95 percent confidence level (from Table 10-1), the sample
mean of 130 mg/dL, and the SD you just calculated of 8 mg/dL, you can compute
the lower and upper confidence limits around the mean using these formulas:
CL L 130 1.96 8 114.3
CLU 130 1.96 8 145.7
On the basis of your calculations, you would report your result this way: mean
glucose = 130 mg/dL (95 percent CI = 114 – 116 mg/dL).
Please note that you should not report numbers to more decimal places than their
precision warrants. In this example, the digits after the decimal point are practi-
cally meaningless, so the numbers are rounded off.
A version of the formula in the preceding section is designed to be utilized with

smaller samples, and uses k values derived from a table of critical values of the
Student t distribution. To calculate CIs this way, you need to know the number of
degrees of freedom (df). For a mean value, the df is always equal to N – 1, so in our
case, df = 25 – 1 = 24. Using a Student t table (see Chapter 24), you can find that
the Student-based k value for a 95 percent confidence level and 24 degrees of
freedom is equal to 2.06, which is a little bit larger than the normal-based k value
of 1.96. Using this k value instead of 1.96, you can calculate the 95 percent confi-
dence limits as 113.52 mg/dL and 146.48 mg/dL, which happen to round off to the
same whole numbers as the normal-based confidence limits. Generally, you don’t
have to use these more-complicated Student-based k values unless your N is quite
small (say, less than 25).
The confidence interval around

a proportion
If you were to conduct a study by enrolling and measuring a sample of 100 adult
patients with diabetes, and you found that 70 of them had their diabetes under
control, you’d estimate that 70 percent of the population of adult diabetics has
their diabetes under control. What is the 95 percent CI around that 70 percent
estimate?
There are multiple approximate formulas for CIs around an observed proportion,
which are also called binomial CIs. Let’s start by unpacking the simplest method
for calculating binomial CIs, which is based on approximating the binomial distri-
bution using a normal distribution (see Chapter 25). The N is the denominator of
the proportion, and you should only use this method when N is large (meaning at
least 50). You should also only use this method if the proportion estimate is not
very close to 0 or 1. A good rule of thumb is the proportion estimate should be
between 0.2 and 0.8.

Using this method, you first calculate the SE of the proportion using this formula:
SE p(1 p ) / N where p stands for proportion. Next, you use the normal-based
formulas in the earlier section “Before you begin: Formulas for confidence limits
in large samples” to calculate the ME and the confidence limits.
Using the numbers from the sample of 100 adult diabetics (of whom 70 have their
diabetes under control), you have p 0.7 andN 100 . Using those numbers, the SE
for the proportion is 0.7(1 0.7 ) / 100 or 0.046. From Table 10-1, k is 1.96 for 95
percent confidence limits. So for the confidence limits, CL L 0.7 1.96 0.046 and
CLU 0.7 1.96 0.046. If you calculate these out, you get a 95 percent CI of 0.61 to
0.79 (around the original estimate of 0.7). To express these fractions as percent-
ages, you report your result this way: “The percentage of adult diabetics in the
sample whose diabetes was under control was 70 percent (95 percent CI = 61 – 79
percent).”
The confidence interval around

an event count or rate
Suppose that you learned that at a large hospital, there were 36 incidents of
patients having a serious fall resulting in injury in the last three months. If that’s
the only incident report data you have to go on, then your best estimate of the
monthly serious fall rate is simply the observed count (N), divided by the length
of time (T) during which the N counts were observed: 36/3, or 12.0 serious falls per
month. What is the 95 percent CI around that estimate?
There are many approximate formulas for the CIs around an observed event count
or rate, which is also called a Poisson CI. The simplest method to calculate a Pois-
son CI is based on approximating the Poisson distribution by a normal distribu-
tion (see Chapter 24). It should be used only when N is large (at least 50). You first
calculate the SE of the event rate using this formula: SE N / T . Next, you use
the normal-based formulas in the earlier section “Before you begin: Formulas for
confidence limits in large samples” to calculate the lower and upper confidence
limits.
Using the numbers from hospital falls example, N 36 and T =3, so the SE for the
event rate is 36 / 3, which is the same as the square root of 2, which is 1.41.
According to Table 10-1, k is 1.96 for 95 percent CLs. So CLL = 12.0 – 1.96 × 1.41 and
CLU = 12.0 + 1.96 × 1.41, which works out to 95 percent confidence limits of 9.24 and
14.76. You report your result this way: “The serious fall rate was 12.0 (95 percent
CI = 9.24 – 14.76) per month.”
To calculate the CI around the event count itself, you estimate the SE of the count
N as SE N , then calculate the CI around the observed count using the formulas

in the earlier section “Before you begin: Formulas for confidence limits in large
samples.” So the SE of the 36 observed serious falls in a three-month period is
simply 36 , which equals 6.0. So for the confidence limits, we have
CL L 36.0 1.96 6.0 and CLU = 36.0 + 1.96 × 6.0. In this case, the ME is 11.76,
which works out to a 95 percent CI of 24.2 to 47.8 serious falls in the three-
month period.
Many other approximate formulas for CIs around observed event counts and rates
are available, most of which are more reliable when your N is small. These formu-
las are too complicated to attempt by hand, but fortunately, many statistical pack-
ages can do these calculations for you. Your best bet is to get the name of the
formula, and then look in the documentation for the statistical software you’re
using to see if it supports a command for that particular CI formula.
Relating Confidence Intervals

and Significance Testing
In Chapter 3, we introduce the concepts and terminology of significance testing,
and in Chapters 11 through 14, we describe specific significance tests. If you read
these chapters, you may have come to the correct conclusion that it is possible to
assess statistical significance by using CIs. To do this, you first select a number
that measures the amount of effect for which you are testing (known as the effect
size). This effect size can be the difference between two means or the difference
between two proportions. The effect size can also be a ratio, such as the ratio of
two means, or other ratios that provide a comparison, such as an odds ratio, a
relative risk ratio, or a hazard ratio (to name a few). The complete absence of any
effect corresponds to a difference of 0, or a ratio of 1, so we call these the “no-
effect” values.
The following statements are always true:
» If the 95 percent CI around the observed effect size includes the no-effect value,
then the effect is not statistically significant. This means that if the 95 percent
CI of a difference includes 0 or of a ratio includes 1, the difference is not large
enough to be statistically significant at α = 0.05, and we fail to reject the null.
» If the 95 percent CI around the observed effect size does not include the
no-effect value, then the effect is statistically significant. This means that if the
95 percent CI of a difference is entirely above or entirely below 0, or is entirely
above or entirely below 1 with respect to a ratio, the difference is statistically
significant at α = 0.05, and we reject the null.

The same kind of correspondence is true for other confidence levels and
significance levels. For example, a 90 percent confidence level corresponds to
the α = 0.10 significance level, and a 99 percent confidence level corresponds
to the α = 0.01 significance level, and so on.
So you have two different but related ways to estimate if an effect you see in your
sample is a true effect. You can use significance tests, or else you can use CIs.
Which one is better? Even though the two methods are consistent with one
another, in biostatistics, we are encouraged for ethical reasons to report the CIs
rather than the result of significant tests.
» The CI around the mean effect clearly shows you the observed effect size, as
well as the size of the actual interval (indicating your level of uncertainty about
the effect size estimate). It tells you not only whether the effect is statistically
significant, but also can give you an intuitive sense of whether the effect is
clinically important, also known as clinically significant.
» In contrast, the p value is the result of the complex interplay between the
observed effect size, the sample size, and the size of random fluctuations.
These are all boiled down into a single p value that doesn’t tell you whether
the effect was large or small, or whether it’s clinically significant or negligible.

4
Comparing
Groups
IN THIS PART . . .
Compare averages between two or more groups using

t tests, ANOVAs, and nonparametric tests.
Compare proportions between two or more groups

using the chi-square and Fisher Exact tests.
Analyze fourfold tables (also known as 2x2 cross-tabs)

to estimate relative risk using odds ratios, cumulative
incidence ratios, and other useful measures of
association.
Compare event rates, also known as person-time data,

and estimate prevalence and incidence rates.
Infer causal relationships between the predictors

and the outcome from the results of your regression
model.
IN THIS CHAPTER
» Determining which tests should be

used in different situations
» Preparing your data, running tests,

and interpreting the output
» Estimating the sample size you need

to compare average values
Chapter 11
Comparing Average
Values between Groups
C
omparing average values between groups of numbers is part of almost all
biostatistical analyses, and over the years, statisticians have developed
dozens of tests for this purpose. These tests include several different fla-
vors of the Student t test, analyses of variance (ANOVA), and a dizzying collection
of tests named after the men who popularized them, including Welch, Wilcoxon,
Mann-Whitney, and Kruskal-Wallis, to name just a few. The multitude of tests is
enough to make your head spin, which leaves many researchers with the uneasy
feeling that they may be using the wrong statistical test on their data.
In this chapter, we guide you through the menagerie of statistical tests for com-
paring groups of numbers. We start by explaining why there are so many tests
available, then guide you as to which ones are right for which situations. Next, we
show you how to execute these tests using R software, and how to interpret the
output. We focus on tests that are usually provided by modern statistical programs
(like those discussed in Chapter 4, which also explains how to install and get
started with R).
CHAPTER 11 Comparing Average Values between Groups 141

Grasping Why Different Situations
Need Different Tests
You may wonder why there are so many tests for such a simple task as comparing
averages. Well, “comparing averages” doesn’t refer to a specific situation. It’s a
broad term that can apply to different situations where you are trying to compare
averages. These situations can differ from each other on the basis of these and
other factors, which are listed here in order of most to least common:
» Within or between: You could be testing differences within groups or

differences between groups.
» Number of time points: You could be testing differences occurring at one

point or over a number of time points.
» Number of groups: You could be testing differences between two groups or

between three or more groups.
» Distribution of outcome: Your outcome measurement could follow the

normal distribution or some other distribution (see Chapter 3).
» Variation: You could be testing the differences in variation or spread across

groups (see Chapter 3).
These different factors can occur in any and all combinations, so there are a lot of
potential scenarios. In the following sections, we review situations you may fre-
quently encounter when analyzing biological data, and advise you as to how to
select the most appropriate testing approach given the situation.
Comparing the mean of a group of numbers

to a hypothesized value
Sometimes you have a measurement from the literature (called a historical control)
that provides a hypothesized value of your measurement, and you want to statis-
tically compare the average of a group to this mean. This situation is common
when you are comparing a value that was calculated based on statistical norms
derived based on the population (such as the IQ test, where 100 has been scaled to
be the population mean).
Typically, comparing a group mean to a historical control warrants using the one-
group Student t test that we describe in the later section “Surveying Student t tests.”
For data that are not normally distributed, the Wilcoxon Signed-Ranks (WSR) test
can be used instead, although it is not used often so we do not cover it in this
chapter. (If you need a review on what normally distributed means, see Chapter 3.)
142 PART 4 Comparing Groups

Comparing the mean of two
groups of numbers
Comparing the mean of two groups of numbers is probably the most common
situation encountered in biostatistics. You may be comparing mean levels of a
protein that is a hypothesized disease biomarker between a group of patients
known to have the disease and a group of healthy controls. Or, you may be com-
paring a measurement of drug efficacy between two groups of patients with the
same condition who are taking two different drugs. Or, you may be comparing
measurements of breast cancer treatment efficacy in women on one health insur-
ance plan compared to those on another health insurance plan.
Such comparisons are generally handled by the famous unpaired or “independent

sample” Student t test (usually just called the t test) that we describe later in the sec-
tion “Surveying Student t tests.” Importantly, the t test is based on two assump-
tions about the distribution of the measurement value being tested in the two
groups:
» The values must be normally distributed (called the normality assumption).

For data that are not normally distributed, instead of the t-test, you can use
the nonparametric Wilcoxon Sum-of-Ranks test (also called the Mann-Whitney U
test and the Mann-Whitney test). We demonstrate the Wilcoxon Sum-of-Ranks
test later in this chapter in the section “Running nonparametric tests.”
» The standard deviation (SD) of the values must be close for both groups
(called the equal variance assumption). As a reminder, the SD is the square root
of the variance. To remember why accounting for variation is important in
sampling, review Chapter 3. Also, Chapter 9 provides more information about
the importance of SD. If the two groups you are comparing have very different
SDs, you should not use a Student t test, because it may not give reliable
results, especially if you are also comparing groups of different sizes. A rule of
thumb is that one group’s SD divided by another group’s SD should not be
more than 1.5 to quality for a Student t test. If you feel your data do not qualify,
you can use an alternative called the Welch test (also called the Welch t test, or
the unequal-variance t test). As you see later in this chapter under “Surveying
Student t tests,” because the Welch test accounts for both equal and unequal
variance, it is the only one that is included in R statistical software.
Comparing the means of three or

more groups of numbers
Comparing the means of three or more groups of numbers is an obvious extension
of the two-group comparison in the preceding section. For example, you may have

recorded some biological measurement, like a value indicating level of response to
treatment among three diagnostic groups (such as mild, moderate, and severe
periodontitis). A comparison of the means of three or more groups is handled by
the analysis of variance (ANOVA), which we describe later in this chapter under
“Assessing the ANOVA.” When there is one grouping variable, like severity of per-
iodontitis, you have a one-way ANOVA. If the grouping variable has three levels
(like mild, moderate, and severe periodontitis), it’s called a one-way, three-level
ANOVA.
The null hypothesis of the one-way ANOVA is that all the groups have the same
mean. The alternative hypothesis is that at least one group has a mean that is
statistically significantly different from at least one of the other groups. The
ANOVA produces a single p value, and if that p is less than your chosen criterion
(typically α = 0.05), you conclude that at least one of the means must be statisti-
cally significantly different from at least one of the other means. (For a refresher
on hypothesis testing and p values, see Chapter 3.) But the problem with ANOVA
is that if it is statistically significant, it doesn’t tell you which groups have means
that are statistically significantly different. If you have a statistically significant
ANOVA, you have to follow-up with one or more so-called post-hoc tests (described
later under “Assessing the ANOVA”), which test for differences between the
means of each pair of groups in your ANOVA.
You can also use the ANOVA to compare just two groups. However, this one-way,
two-level ANOVA produces exactly the same p value as the classic unpaired equal-
variance Student t test.
Comparing means in data grouped

on several different variables
The ANOVA is a very flexible method in that it can accommodate comparing means
across several grouping variables at once. As an example, you could use an ANOVA
for comparing treatment response among participants with different levels of the
condition (such as mild, moderate, and severe periodontitis), who come from dif-
ferent clinics (such as Clinic A and Clinic B), and have undergone different treat-
ment approaches (such as using mouthwash or not). An ANOVA involving three
different grouping variables is called a three-way ANOVA (and compares at a more
granular level).
In ANOVA terminology, the term way refers to how many grouping variables are
involved, and the term level refers to the number of different levels within any one
grouping variable.

Like the t test, the ANOVA also assumes that the value you are comparing follows
a normal distribution, and that the SDs of the groups you are comparing are sim-
ilar. If your data are not normally distributed, you can use the nonparametric
Kruskal-Wallis test instead of the one-way ANOVA, which we demonstrate later in
the section “Running nonparametric tests.”
Adjusting for a confounding variable

when comparing means
Sometimes you are aware the variable you are comparing, such as reduction in
blood pressure, is influenced by not only a treatment approach (such as drug A
compared to drug B), but also by other confounding variables (such as age,
whether the patient has diabetes, whether the patient smokes tobacco, and so on).
These confounders are considered nuisance variables because they have a known
impact on the outcome, and may be more prevalent in some groups than others.
If a large proportion of the group on drug A were over age 65, and only a small
proportion of those on drug B were over age 65, older age would have an influence
on the outcome that would not be attributable to the drug. Such a situation
would be confounded by age. (See Chapter 20 for a comprehensive review of
confounding.)
When you are comparing means between groups, you are doing a bivariate com-
parison, meaning you are only involving two variables: the group variable and the
outcome. Adjusting for confounding must be done through a multivariate analysis
using regression.
Comparing means from sets

of matched numbers
Often when biostatisticians consider comparing means between two or more
groups, they are thinking of independent samples of data. When dealing with study
participants, independent samples means that the data you are comparing come
from different groups of participants who are not connected to each other statisti-
cally or literally. But in some scenarios, your intention is to compare means from
matched data, meaning some sort of pairing exists in the data. Here are some
common examples of matched data:
» The values come from the same participants, but at two or more different
times, such as before and after some kind of treatment, intervention, or event.
» The values come from a crossover clinical trial, in which the same participant
receives two or more treatments at two or more consecutive phases of the trial.

» The values come from two or more different participants who have been
paired, or matched, in some way as part of the study design. For example,
in a study of participants who have Alzheimer’s disease compared to healthy
participants, investigators may choose to age-match each Alzheimer’s patient
to a healthy control when they recruit so both groups have the same age
distribution.
Comparing means of matched pairs

If you have paired data, you must use a paired comparison. Paired comparisons are
usually handled by the paired student t test that we describe later in this chapter
under “Surveying Student t tests.” If your data aren’t normally distributed, you
can use the nonparametric Wilcoxon Signed-Ranks test instead.
The paired Student t test and the one-group Student t test are actually the same
test. When you run a paired t test, the statistical software first calculates the dif-
ference between each pair of numbers. If comparing a post-treatment value to a
pretreatment value, the software would start by subtracting one value from the
other for each participant. Finally, the software would run a test to see if those
mean differences were statistically significantly different from the hypothesized
value of 0 using a one-group test.
Using Statistical Tests for

Comparing Averages
Now that you have reviewed the different types of comparisons, you can continue
to consider the basic concepts behind them as you dig more deeply. In this section,
we discuss executing these tests in statistical software and interpreting the out-
put. We do that with several tests, including Student t tests, the ANOVA, and non-
parametric tests.
We opted not to clutter this chapter with pages of mathematical formulas for the
following tests because based on our own experience, we believe you’ll probably
never have to do one of these tests by hand. If you really want to see the formulas,
we recommend putting the name of the test in quotes in a search engine and look-
ing on the Internet.

Surveying Student t tests
In this section, we present the general approach to conducting a Student t test.
We walk through the computational steps common to the different kinds of t
tests, including one-group, paired, and independent. As we do that, we explain
the computational differences between the different test types. Finally, we
demonstrate how to run the t tests using open source software R, and explain
how to interpret the output (see Chapter 4 for more information about getting
started with R).
Understanding the general approach to a t test

As reviewed earlier, t tests are designed to compare two means only. If you mea-
sure the means of two groups, you see that they almost always come out to be
different numbers. The Student t tests are intended to answer the question, Is the
observed difference in means larger than what you would expect from random fluctua-
tions alone? The different t tests take the same general approach to answer this
question, using the following steps:
1. Calculate the difference (D) between the mean values you are comparing.
2. Calculate the precision of the difference, which is the magnitude of the

random fluctuations in that difference.
For the t test, calculate the standard error (SE) of that difference (see
Chapter 10 for a refresher on SE).
3. Calculate the test statistic, which in this case is t.
The test statistic expresses the size of the D relative to the size of its SE. That
is: t D / SE .
4. Calculate the degrees of freedom (df) of the t statistic.
df is a tricky concept, but is easy to calculate. For t, the df is the total number of
observations minus the number of means you calculated from those
observations.
5. Use the t and df to calculate the p value.
The p value is the probability that random fluctuations alone could produce a t
value at least as large as the value you just calculated based upon the Student t
distribution.
The Student t statistic is always calculated using the general equation D/SE. Each
specific type of t test we discussed earlier — including one-group, paired,
unpaired, and Welch — calculates D, SE, and df slightly differently. These different
calculations are summarized in Table 11-1.

TABLE 11-1 How t Tests Calculate Difference, Standard Error, and
Degrees of Freedom
Unpaired t
One-Group Paired Equal Variance Welch t Unequal Variance
D Difference between mean Mean of Difference between Difference between means of

of observations and a paired means of the two the two groups
hypothesized value (h) differences groups
SE SE of the observations SE of paired SE of difference, based SE of difference, from SE of

differences on a pooled estimate of each mean, by propagation of
SD within each group errors
df Number of Number of Total number of “Effective” df, based on the

observations – 1 pairs – 1 observations – 2 size and SD of the two groups
Executing a t test
Statistical software packages contain commands that can execute (or run) t tests
(see Chapter 4 for more about these packages). The examples presented here use
R, and in this section, we explain the data structure required for running the var-
ious t tests in R. For demonstration, we use data from the National Health and
Nutrition Examination Survey (NHANES) from 2017–2020 file (available at wwwn.
cdc.gov/nchs/nhanes/continuousnhanes/default.aspx?Cycle=2017-2020).
» For the one-group t test, you need the column of data containing the
variable whose mean you want to compare to the hypothesized value (H), and
you need to know H. R and other software enable you to specify a value for H
and assumes 0 if you don’t specify anything. In the NHANES data, the fasting
glucose variable is LBXGLU, so the R code to test the mean fasting glucose
against a maximum healthy level of 100 mg/dL in an R dataframe named
GLUCOSE is t.test(GLUCOSE$LBXGLU, mu = 100).
» For the paired t test, you need two columns of data representing the pair of
numbers you want to enter into the paired t test. For example, in NHANES,
systolic blood pressure (SBP) was measured in the same participant twice
(variables BPXOSY1 and BPXOSY2). To compare these with a paired t test in an
R dataframe named BP, the code is t.test(BP$BPXOSY1, BP$BPXOSY2, paired =
TRUE).
» For the independent t test, you need to have one column coded as the
grouping variable (preferable with a two-state flag coded as 0 and 1), and
another column with the value you want to test. We created a two-state flag in
the NHANES data called MARRIED where 1 = married and 0 = all other marital
statuses. To compare mean fasting glucose level between these two groups in
a dataframe named NHANES, we used this code: t.test(NHANES$LBXGLU ~
NHANES$MARRIED).

Interpreting the output from a t test
Listing 11-1 is the output from a one-sample t-test, where we tested the mean
fasting glucose in the NHANES participants against the hypothesized mean of 100
mg/dL:
LISTING 11-1: R Output from a One-Sample Student t Test

> t.test(GLUCOSE$LBXGLU, mu = 100)
One Sample t-test
data: GLUCOSE$LBXGLU
t = 21.209, df = 4743, p-value < 2.2e-16
alternative hypothesis: true mean is not equal to 100
95 percent confidence interval:
110.1485 112.2158
sample estimates:
mean of x
111.1821
The R output starts by stating what test was run and what data were used, and
then reports the t statistic (21.209), the df (4743), and the p value, which is writ-
ten in scientific notation: < 2.2e–16. If you have trouble interpreting this notation,
just remove the < and then copy and paste the rest of the number into a cell in
Microsoft Excel. If you do that, you will see in the formula bar that the number
resolves to 0.00000000000000022 — which is a very low p value! The shorthand
used for this in biostatistics is p < 0.0001, meaning it is sufficiently small. Because
of this small p value, we reject the null hypothesis and say that the mean glucose
of NHANES participants is statistically significantly different from 100 mg/dL.
But in what direction? For that, it is necessary to read down further in the R out-
put, under 95 percent confidence interval. It says the interval is 110.1485 mg/dL to
112.2158 mg/dL (if you need a refresher on confidence intervals, read Chapter 10).
Because the entire interval is greater than 100 mg/dL, you can conclude that the
NHANES mean is statistically significantly greater than 100 mg/dL.
Now, let’s examine the output from the paired t test of SBP measured two times
in the same participant, which is shown in Listing 11-2.

LISTING 11-2: R Output from a Paired Student t Test
> t.test(BP$BPXOSY1, BP$BPXOSY2, paired = TRUE)
Paired t-test
data: BP$BPXOSY1 and BP$BPXOSY2

t = 4.3065, df = 10325, p-value = 1.674e–05
alternative hypothesis: true mean difference is not equal to 0
0.1444651 0.3858467
sample estimates:
mean difference
0.2651559
Notice a difference between the output shown in Listings 11-1 and 11-2. In
Listing 11-1, the third line of output says, “alternative hypothesis: true mean is
not equal to 100.” That is because we specified the null hypothesis of 100 when we
coded the one-sample t test. Because we did a paired t test in Listing 11-2, this
null hypothesis now concerns 0 because we are trying to see if there is a statisti-
cally significant difference between the first SBP reading and the second in the
same individuals. Why should they be very different at all? In Listing 11-2, the p
value is listed as 1.674e–05, which resolves to 0.00001674 (to be stated as p <
0.0001). We were surprised to see a statistically significant difference! The output
says that the 95 percent confidence interval of the difference is 0.1444651 mmHg
to 0.3858467 mmHg, so this small difference may be statistically significant while
not being clinically significant.
Let’s examine the output from our independent t test of mean fasting glucose
values in NHANES participants who were married compared to participants with
all other marital statuses. This output is shown in Listing 11-3.
LISTING 11-3: R Output from an Independent t Test

> t.test(NHANES$LBXGLU ~ NHANES$MARRIED)
Welch Two Sample t-test
data: NHANES$LBXGLU BY NHANES$MARRIED

t = –4.595, df = 4731.2, p-value = 4.439e–06
alternative hypothesis: true difference in means between group $
–6.900665 –2.773287
sample estimates:
mean in group 0 mean in group 1
108.8034 113.6404

Importantly, at the top of Listing 11-3, notice that it says “Welch Two Sample
t-test.” This is because R insists on using Welch’s test instead of the Student t test
for independent t tests because Welch’s test accounts for unequal variance (as
well as equal variance) between groups, as discussed earlier. In the output under
the alternative hypothesis, notice that it says R is testing whether the true differ-
ence in means between group 0 and group 1 is not equal to 0 (remember, 1 = married
and 0 = all other marital statuses). R calculated a p value of 4.439e–06, which
resolves to 0.000004439 — definitely p < 0.0001! The groups are definitely statis-
tically significantly different when it comes to average fasting glucose.
But which group is higher? Well, for that, you can look at the last line of the out-
put, where it says that the mean in group 0 (all marital statuses except married)
is 108.8034 mg/dL, and the mean in group 1 (married) is 113.6404 mg/dL. So does
getting married raise your fasting glucose? Before you try to answer that, please
make sure you read up on confounding in Chapter 20!
But what if you just wanted to know if the variance in the fasting glucose mea-
surement in the married group was equal or unequal to the other group, even
though you were doing a Welch test that accommodates both? For that, you can
do an F test. Because we are not sure which group’s fasting glucose would be
higher, we choose a two-sided F test and use this code: var.test(LBXGLU ~
MARRIED, NHANES, alternative = "two.sided"), which produces the output shown in
Listing 11-4.
LISTING 11-4: R Output from an F Test

> var.test(LBXGLU ~ MARRIED, NHANES, alternative = "two.sided")
F test to compare two variances
data: LBXGLU by MARRIED

F = 0.97066, num df = 2410, denom df = 2332, p-value = 0.4684
alternative hypothesis: true ratio of variances is not equal to$
0.8955321 1.0520382
sample estimates:
ratio of variances
0.9706621
As shown in Listing 11-4, the p value on the F test is 0.4684. As a rule of thumb:
» If p 0.05, you would assume equal variances.
» If p 0.05, you would assume unequal variances.

In this case, because the p value is greater than 0.05, equal variances can be
assumed, and these data would qualify for the classic Student t test. As described
earlier, R gets around this by always using the Welch’s t test, which accommo-
dates both unequal and equal variances.
Assessing the ANOVA

In this section, we present the basic concepts underlying the analysis of variance
(ANOVA), which compares the means of three or more groups. We also describe
some of the more popular post-hoc tests used to follow a statistically significant
ANOVA. Finally, we show you how to run commands to execute an ANOVA and
post-hoc tests in R, and interpret the output.
Grasping how the ANOVA works

As described earlier in “Surveying Student t tests,” it is only possible to run a t
test on two groups. This is why we demonstrated the t test comparing married
NHANES participants (M) to all other marital statuses (OTH). We were testing the
null hypothesis M – OTH = 0 because we were only allowed to compare two groups!
So when comparing three groups, such as married (M), never married (NM), and
all others (OTH), it’s natural to think of pairing up the groups and running three t
tests (meaning testing M – NM, then testing M – OTH, then testing NM – OTH). But
running an exhaustive set of two-group t tests increases the likelihood of Type I
error, which is where you get a statistically significant comparison that is just by
chance (for a review, read Chapter 3). And this is just with three groups!
The general rule is that N groups can be paired up in N N 1 / 2 different ways,

so in a study with six groups, you’d have 6 5 / 2, or 15 two-group comparisons,
which is way too many.
The term one-way ANOVA refers to an ANOVA with only one grouping variable in
it. The grouping variable usually has three or more levels because if it has only
two, most analysts just do a t test. In an ANOVA, you are testing how spread out
the means of the various levels are from each other. It is not unusual for students
to be asked to calculate an ANOVA manually in a statistics class, but we skip that
here and just describe the result. One result derived from an ANOVA calculation is
expressed in a test statistic called the F ratio (designated simply as F). The F is the
ratio of how much variability there is between the groups relative to how much
variability there is within the groups. If the null hypothesis is true, and no true
difference exists between the groups (meaning the average fasting glucose in
M = NM = OTH), then the F ratio should be close to 1. Also, F’s sampling fluctua-
tions should follow the Fisher F distribution (see Chapter 24), which is actually a
family of distribution functions characterized by the following two numbers seen
in the ANOVA calculation:

» The numerator degrees of freedom: This number is often designated as df N
or df1, which is one less than the number of groups.
» The denominator degrees of freedom: This number is designated as df D or

df2 , which is the total number of observations minus the number of groups.
The p value can be calculated from the values of F, df1, and df2 , and the software
performs this calculation for you. If the p value from the ANOVA is statistically
significant — less than 0.05 or your chosen α level — then you can conclude that
the group means are not all equal and you can reject the null hypothesis. Techni-
cally, what that means is that at least one mean was so far away from another
mean that it made the F test result come out far away from 1, causing the p value
to be statistically significant.
Picking through post-hoc tests

Suppose that the ANOVA is not statistically significant (meaning F was larger than
0.05). It means that there is no point in doing any t tests, because all the means
are close to each other. But if the ANOVA is statistically significant, we are left
with the question: Which group means are higher or lower than others? Answering
that question requires us to do post-hoc tests, which are t tests done after an ANOVA
(post hoc is Latin for “after this”).
Although using post-hoc tests can be helpful, controlling Type I error is not that
easy in reality. There can be issues with the data that may make you not trust the
results of your post-hoc tests, such having too many levels to the group you are
testing in your ANOVA, or having one or more of the levels with very few partici-
pants (so the results are unstable). Still, if you have a statistically significant
ANOVA, you should do post-hoc t tests, just so you know the answer to the ques-
tion stated earlier.
It’s okay to do these post-hoc tests; you just have to take a penalty. A penalty is
where you deliberately make something harder for yourself in statistics. In this
case, we take a penalty by making it deliberately harder to conclude a p value on a
t test is statistically significant. We do that by adjusting the α to be lower than
0.05. How much we adjust it depends on the post-hoc test we choose.
» The Bonferroni adjustment uses this calculation to determine the new,

lower alpha: α/N, where N is the number of groups. As you can tell, the
Bonferroni adjustment is easy to do manually! In the case of our three marital
groups (M, NM, and OTH), our adjusted Bonferroni α would be 0.05/3, which is
0.016. This means that for a post-hoc t test of average fasting glucose
between two of the three marital groups, the p value would not be inter-
preted as significant unless the it was less than 0.016 (which is a tougher

criterion than only having to be less than 0.05). Even though the Bonferroni
adjustment is easy to do by hand, because most analysts use statistical
packages when doing these calculations, it is not used very often in practice.
» Tukey’s HSD (“honestly” significant difference) test adjusts α in a different

way than Bonferroni. It is intended to be used when there are equally-sized
groups in each level of the variable (also called balanced groups).
» The Tukey-Kramer test is a generalization of the original Tukey’s HSD test

to designed to handle different-sized (also called unbalanced) groups. Since
Tukey-Kramer also handles balanced groups, in R statistical software, only
the Tukey-Kramer test is available, and not Tukey’s HSD test (as demon-
strated later in this chapter in the section “Executing and interpreting
post-hoc t tests”).
» Scheffe’s test compares all pairs of groups, but also lets you bundle certain
groups together if doing so makes physical sense. For example, if you have
two treatment groups and a control group (such as Drug A, Drug B, and
Control), you may want to determine whether either drug is different from the
control. In other words, you may want to test Drug A and Drug B as one group
against the control group, in which case you use Scheffe’s test. Scheffe’s test is
the safest to use if you are worried your analysis may be suffering from Type I
error because it is the most conservative. On the other hand, it is less
powerful than the other tests, meaning it will miss a real difference in your
data more often than the other tests.
Running an ANOVA
Running a one-way ANOVA in R is similar to running an independent t test (see
the earlier section “Executing a t test”). However, in this case, we save the results
as an object, and then run R code on that object to get the output of our results.
Let’s turn back to the NHANES data. First, we need to prepare our grouping vari-
able, which is the three-level variable MARITAL (where 1 = married, 2 = never mar-
ried, and 3 = all over marital statuses). Next, we identify our dependent variable,
which is our fasting glucose variable called LBXGLU. Finally, we employ the aov
command to run the ANOVA in R, and save the results in an object called
GLUCOSE_aov. We use the following code: GLUCOSE_aov <- aov(LBXGLU ~
as.factor(MARITAL), data = NHANES). (The reason we have to use the as.factor com-
mand on the MARITAL variable is to make R handle it as an ordinal variable in the
calculation, not a numeric one.) Next, we can get our output by running a summary
command on this object using this code: summary(GLUCOSE_aov).

Interpreting the output of an ANOVA
We describe the R output here, but output from other statistical packages will have
similar information. The output begins with the variance table (or simply the
ANOVA table). You can tell it is a table because it looks like it has a column with no
heading followed by columns with the following headings: Df (for df), Sum Sq (for
the sum of squares), Mean Sq (mean square), F value (value of F statistic), and
Pr(>F) (p value for the F test). You may recall that in order for an ANOVA test to be
statistically significant at α = 0.05, the p value on the F must be < 0.05. It is easy
to identify that F = 12.59 on the output because it is labeled F value. But the p value
on the F is labeled Pr(>F), and that’s not very obvious. As you saw before, the p
value is in scientific notation, but resolves to 0.00000353, which is < 0.05, so it is
statistically significant.
If you use R for this, you will notice that at the bottom of the output it says Signif.
codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1. This is R explaining its coding sys-
tem for p values. It means that if a p value in output is followed by three asterisks,
this is a code for < 0.001. Two asterisks is a code for p < 0.01, and one asterisk
indicates p < 0.05. A period indicates p < 0.1, and no notation indicates the p value
is greater than or equal to 0.1 — meaning by most standards, it is not statistically
significant at all. Other statistical packages often use similar coding to make it
easy for analysts to pick out statistically significant p values in the output.
Several statistical packages that do ANOVAs offer one or more post-hoc tests as
optional output, so programmers tend to request output for both ANOVAs and
post-hoc tests, even before they know whether the ANOVA is statistically signifi-
cant or not, which can be confusing. ANOVA output from other software can
include a lot of extra information, such as a table of the mean, variance, standard
deviation, and count of the observations in each group. It may also include a test
for homogeneity of variances, which tests whether all groups have nearly the same
SDs. In R, the ANOVA output is very lean, and you have to request information like
this in separate commands.
Executing and interpreting post-hoc t tests

In the previous example, the ANOVA was statistically significant, so it qualifies for
post-hoc pairwise t tests. Now that we are at this step, we need to select which
adjustment to use. We already have an idea of what would happen if we used the
Bonferroni adjustment. We’d have to run t tests like we did before, only this time
we’d have to use the three-level MARITAL variable and run three t tests: One with
M and NM, a second with M and OTH, and a third with M and OTH. For each p value
we got, we would have to compare it to the adjusted Bonferroni α of 0.016 instead
of 0.05. By evaluating each p value, you can determine which pairs of groups are
statistically significantly different using the Bonferroni adjustment.

But Bonferroni is not commonly used in statistical software. In R, the most com-
mon post-hoc adjustments employed are Tukey-Kramer (using the TukeyHSD
command) and Scheffe (using the ScheffeTest command from the package Desc-
Tools). The reason why the Tukey HSD is not available in R is that the Tukey-
Kramer can handle both balanced and unbalanced groups. In the case of marital
statuses and fasting glucose levels in NHANES, the Tukey-Kramer is probably the
most appropriate test because we do not need the special features of the Scheffe
test. However, we explain the output anyway so that you can understand how to
interpret it.
To run the Tukey-Kramer test in R, we use the following code: TukeyHSD(GLUCOSE_

aov, conf.level=.95). Notice that the code refers to the ANOVA object we made previ-
ously called GLUCOSE_aov. The Tukey-Kramer output begins by restating the
test, and the contents of the ANOVA object GLUCOSE_aov.
Next is a table (also known as a matrix) with five columns. The first column does
not have a heading, but indicates which levels of MARITAL are being compared in
each row (for example, 2-1 means that 1 = M is being compared to 2 = NM). The
column diff indicates the mean difference between the groups being compared,
with lwr and upr referring to the lower and upper 95 percent confidence limits of
this difference, respectively. (R is using the 95 percent confidence limits because
we specified conf.level = .95 in our code.) Finally, in the last column labeled p adj is
the p value for each test. As you can see by the output, using the Tukey-Kramer
test and α = 0.05, M and NM are statistically significantly different (p = 0.0000102),
and OTH and M are statistically significantly different (p = 0.0030753), but NM and
OTH are not statistically significantly different (p = 0.1101964).
When doing a set of post-hoc tests in any software, the output will be formatted
as a table, with each comparison listed on its own row, and information about the
comparison listed in the columns.
In a real scenario, after completing your post-hoc test, you would stop here and
interpret your findings. But because we want to explain the Scheffe test, we can
take an opportunity compare what we find when we run that one, too. Let’s start
by loading the DescTools package using the R code library(DescTools) (Chapter 4
explains how to use packages in R). Next, let’s try the Scheffe test by using the
following code on our existing ANOVA object: ScheffeTest(GLUCOSE_aov).
The Scheffe test output is arranged in a similar matrix, but also includes R’s sig-
nificance codes. This time, according to R’s coding system, M and NM are statisti-
cally significantly different at p < 0.001, and M and OTH are statistically significantly
different at p < 0.01. Although the actual numbers are slightly different, the inter-
pretation is the same as what you saw using the Tukey-Kramer test.

Sometimes you may not know which post-hoc test to select for your ANOVA. If
you have an advisor or you are on a research team, you should discuss which one
is best. However, if you run the one you select and do not trust the results, it’s not
a bad idea to run the other ones and keep track of the results. The p values will
always come out different, but if the interpretation changes — meaning different
comparisons are statistically significant, depending upon what test you choose —
you may want to rethink doing post-hoc tests. This means that the results you are
getting are unstable.
Running nonparametric tests

As a reminder, the Wilcoxon Sum-of-Ranks test is the nonparametric alternative
to the t test, which you can use if your data do not follow a normal distribution.
Like with the t test, you can run a Wilcoxon Sum-of-Ranks test in R with options
that gives you results if you are doing a paired t test. But to simply repeat the
independent t test we did earlier comparing mean fasting glucose in married
NHANES participants compared to all other marital statuses, you would run this
code: wilcox.test(NHANES$LBXGLU ~ NHANES$MARRIED).
The Kruskal-Wallis test is a nonparametric ANOVA alternative. Like the ANOVA,

you can use the Kruskal-Wallis to test whether the mean fasting glucose is equal
in the three-level marital status variable MARITAL. The R code for the Kruskal-
Wallis test is different from the ANOVA code because it does not require you to
produce an object for the summary statistics. The following code prints the results
to the output: kruskal.test(LBXGLU ~ MARITAL, data = NHANES).
Nonparametric tests don’t compare group means or test for a nonzero mean dif-
ference. Rather, they compare group medians, or they deal with ranking the order
of variables and analyze those ranks. Because it this, the output from R and other
programs will likely focus on reporting the p value of the test.
Only use a nonparametric test if you are absolutely sure your data do not qualify
for a parametric test (meaning t test, ANOVA, and others that require a particular
distribution). Parametric tests are more powerful. In the NHANES example, the
data would qualify for a parametric test; we only showed you the code for non-
parametric tests as an example.

Estimating the Sample Size You Need
for Comparing Averages
There are several ways to estimate the sample size you need in order to be able to
detect if there is a significant result on a t test or an ANOVA. (Check out Chapter 3
for a refresher on the concepts of power and sample size.)
Using formulas for manual calculation

Chapter 25 provides a set of formulas that let you estimate how many participants
you need for several kinds of t tests and ANOVAs. As with all sample-size
calculations, you need to be prepared to specify two parameters: the effect size of
importance, which is the smallest between-group difference that’s worth knowing
about, and the amount of random variability in your data, expressed as the
within-group SD. If you plug these values into the formulas in Chapter 25, you
can calculate desired sample size.
Software and web pages

All the modern statistical programs covered in Chapter 4 provide power and
sample-size calculations for most standard statistical tests. As described in
Chapter 4, G*Power is menu-driven, and can be used for sample size calculations
for many tests, including t tests and ANOVAs. If you are using G*Power, to
estimate sample size for t tests, choose t tests from the test family drop-down
menu, and for ANOVA, choose F tests. Then, from the statistical test drop-down
menu, choose the test you plan to use and set type of power analysis to “A priori:
Compute required sample size – given α, power, and effect size.” Then enter the
parameters and click determine to calculate the sample size.
In terms of web pages, the website https://statpages.info lists several dozen

web pages that perform power and sample-size calculations for t tests and
ANOVAs.

IN THIS CHAPTER
» Testing for association between

categorical variables with the
Pearson chi-square and Fisher Exact
tests
» Estimating sample sizes for tests of

association
Chapter 12
Comparing Proportions
and Analyzing
Cross-Tabulations
S
uppose that you are studying pain relief in patients with chronic arthritis.
Some are taking nonsteroidal anti-inflammatory drugs (NSAIDs), which are
over-the-counter pain medications. But others are trying cannabidiol (CBD),
a new potential natural treatment for arthritis pain. You enroll 100 chronic arthri-
tis patients in your study and you find that 60 participants are using CBD, while
the other 40 are using NSAIDs. You survey them to see if they get adequate pain
relief. Then you record what each participant says (pain relief or no pain relief).
Your data file has two dichotomous categorical variables: the treatment group (CBD
or NSAIDs), and the outcome (pain relief or no pain relief).
You find that 10 of the 40 participants taking NSAIDs reported pain relief, which
is 25 percent. But 33 of the 60 taking CBD reported pain relief, which is 55 percent.
CBD appears to increase the percentage of participants experiencing pain relief by
30 percentage points. But can you be sure this isn’t just a random sampling
fluctuation?
CHAPTER 12 Comparing Proportions and Analyzing Cross-Tabulations 159

Data from two potentially associated categorical variables is summarized as a
cross-tabulation, which is also called a cross-tab or a two-way table. Because we are
studying the association between two variables, this is a form of bivariate analysis.
The rows of the cross-tab represent the different categories (or levels) of one vari-
able, and the columns represent the different levels of the other variable. The cells
of the table contain the count of the number of participants with the indicated
levels for the row and column variables. If one variable can be thought of as the
“cause” or “predictor” of the other, the cause variable becomes the rows, and the
“outcome” or “effect” variable becomes the columns. If the cause and outcome
variables are both dichotomous, meaning they have only two levels (like in this
example), then the cross-tab has two rows and two columns. This structure con-
tains four cells containing counts, and is referred to as a 2-by-2 (or 2 × 2) cross-
tab, or a fourfold table. Cross-tabs are displayed with an extra row at the bottom
and an extra column at the right to contain the sums of the cells in the rows and
columns of the table. These sums are called marginal totals, or just marginals.
Comparing proportions based on a fourfold table is the simplest example of test-

ing the association between two categorical variables. More generally, the vari-
ables can have any number of categories, so the cross-tab can be larger than
2 × 2, with multiple rows and many columns. But the basic question to be answered
is always the same: Is the spread of numbers across the columns so different from one
row to the next that the numbers can’t be explained away as random fluctuations?
Another way of asking the same question is: Is being a member of a particular row
associated with being a member of a particular column?
In this chapter, we describe two tests you can use to answer this question: the
Pearson chi-square test, and the Fisher Exact test. We also explain how to esti-
mate power and sample sizes for the chi-square and Fisher Exact tests.
Like with other statistical tests, you can run all the tests in this chapter from
individual-level data in a database, where there is one record per participant. But
the tests in this chapter can also be executed using data that has already been
summarized in the form of a cross-tab:
» Most statistical software is set up to work with individual-level data. In that

case, your data file needs to have two columns for the association you want to
test: one containing the categorical variable representing the treatment group
(or whatever category is on the y-axis), and one containing the categorical
variable representing the outcome. If you have the correct columns, all you
have to do is tell the statistical software you are using which test or tests you
want to run, and which variables to use in the test.

» Most statistical software is also set up so that you can do these tests using
summarized data (rather than individual-level data), so long as you set an
option in your programming when running the tests. In contrast, online
calculators that execute these tests expect you to have already cross-tabulated
the data. These calculators usually present a screen showing an empty table,
and you enter the counts into the table’s cells to run the calculation.
Examining Two Variables with the

Pearson Chi-Square Test
The most commonly used statistical test of association between two categorical
variables is called the chi-square test of association developed by Karl Pearson
around the year 1900. It’s called the chi-square test because it involves calculating
a number called a test statistic that fluctuates in accordance with the chi-square
distribution. Many other statistical tests also use the chi-square distribution, but
the test of association is by far the most popular. In this book, whenever we refer
to a chi-square test without specifying which one, we are referring to the Pearson
chi-square test of association between two categorical variables. (Please note that
some books use the notation X2 or x2 instead of saying the term chi-square.)
Understanding how the

chi-square test works
You don’t have to understand the equations behind the chi-square test if you have
a computer to do them, which is optimal, though it is possible to calculate the test
manually. This means you technically don’t have to read this section. But we
encourage you to do so anyway, because we think you’ll have a better appreciation
for the strengths and limitations of the test if you know its mathematical under-
pinnings. Here, we walk you through conducting a chi-square test manually
(which is possible to do in Microsoft Excel).
Calculating observed and expected counts

All statistical significance tests start with a null hypothesis (H 0 ) that asserts that no
real effect is present in the population, and any effect you think you see in your
sample is due only to random fluctuations. (See Chapter 3 for more information.)
The H 0 for the chi-square test asserts that there’s no association between the
levels of the row variable and the levels of the column variable, so you should
expect the relative spread of cell counts across the columns to be the same for
each row.

Figure 12-1 shows how this works out for the observed data taken from the exam-
ple in this chapter’s introduction. You can see from the marginal “Total” row that
the overall rate of pain relief (for both groups combined) is 43/100, or 43 percent.
FIGURE 12-1:
The observed
results comparing
CBD to NSAIDs
for the treatment
of pain from
chronic arthritis.
Figure 12-1 presents the actual data you observed from your survey, where the
observed counts are placed in each of the four cells. As part of the chi-square test
statistic calculation, you now need to calculate an expected count for each cell. This
is done by taking the product of the row and column marginals and dividing them
by the total. So, to determine the expected count in the CBD/pain relief cell, you
would multiply 43 (row marginal) by 60 (column marginal), then divide this by
100 (total) which comes out 25.8. Figure 12-2 presents the fourfold table with the
expected counts in the cells.
FIGURE 12-2:
Expected cell
counts if the null
hypothesis is true
(there is no
association
between either
drug and the
outcome).
The reason you need these expected counts is that they represent what would
happen under the null hypothesis (meaning if the null hypothesis were true). If
the null hypothesis were true:
» In the CBD-treated group, you’d expect about 25.8 participants to experience

pain relief (43 percent of 60), with the remaining 34.2 reporting no pain relief.

» In the NSAIDs-treated group, you’d expect about 17.2 participants to feel pain
relief (43 percent of 40) with the remaining 22.8 reporting no pain relief.
As you can see, this expected table assumes that you still have the overall pain
relief rate of 43 percent, but that you also have the pain relief rates in each group
equal to 43 percent. This is what would happen under the null hypothesis.
Now that you have observed and expected counts, you’re no doubt curious as to
how each cell in the observed table differs from its companion cell in the expected
table. To get these numbers, you can subtract each expected count from the
observed count in each cell to get a difference table (observed – expected), as shown
in Figure 12-3.
FIGURE 12-3:
Differences
between
observed and
expected cell
counts if the null
hypothesis is
true.
As you review Figure 12-3, because you know the observed and expected tables in
Figures 12-1 and 12-2 always have the same marginal totals by design, you should
not be surprised to observe that the marginal totals in the difference table are all
equal to zero. All four cells in the center of this difference table have the same
absolute value (7.2), with a plus and a minus value in each row and each column.
The pattern just described is always the case for 2 2 tables. For larger tables, the
difference numbers aren’t all the same, but they always sum up to zero for each
row and each column.
The values in the difference table in Figure 12-3 show how far off from H 0 your
observed data are. The question remains: Are those difference values larger than
what may have arisen from random fluctuations alone if H 0 is really true? You
need some kind of measurement unit by which to judge how unlikely those differ-
ence values are. Recall from Chapter 10 that the standard error (SE) expresses the
general magnitude of random sampling, so looking at the SE as a type of mea-
surement unit is a good way for judging the size of the differences you may expect
to see from random fluctuations alone. It turns out that it is easy to approximate
the SE of the differences because this is approximately equal to the square root of

the expected counts. The rigorous proof behind this is too complicated for most
mathophobes (as well as some normal people) to understand. Nevertheless, a
simple informal explanation is based on the idea that random event occurrences
typically follow the Poisson distribution for which the SE of the event count equals
the square root of the expected count (as discussed in Chapter 10).
Summarizing and combining scaled differences

For the upper-left cell in the cross-tab (CBD–treated participants who experience
pain relief), you see the following:
» The observed count (Ob) is 33.
» The expected count (Ex) is 25.8.
» The difference (Diff) is 33 25.8, or 7.2.
» The SE of the difference is 25.8 or 5.08
You can “scale” the Ob-Ex difference (in terms of unit of SE) by dividing it by the
SE measurement unit, getting the ratio Diff / SE 7.2 / 5.08, or 1.42. This means
that the difference between the observed number of CBD-treated participants who
experience pain relief and the number you would have expected if the CBD had no
effect on survival is about 1.42 times as large as you would have expected from
random sampling fluctuations alone. You can do the same calculation for the other
three cells and summarize these scaled differences. Figure 12-4 shows the differ-
ences between observed and expected cell counts, scaled according to the esti-
mated standard errors of the differences.
FIGURE 12-4:
Differences
between
observed and
expected cell
counts.
The next step is to combine these individual scaled differences into an overall
measure of the difference between what you observed and what you would have
expected if the CBD or NSAID use really did not impact pain relief differentially.
You can’t just add them up because the negative and positive differences would
cancel each other out. You want all differences (positive and negative) to contrib-
ute to the overall measure of how far your observations are from what you expected
under H 0 .

Instead of summing the differences, statisticians prefer to sum the squares of dif-
ferences, because the squares are always positive. This is exactly what’s done in
the chi-square test. Figure 12-5 shows the squared scaled differences, which are
calculated from the observed and expected counts in Figures 12-1 and 12-2 using
2
the formula Ob – Ex /Ex (rather than by squaring the rounded-off numbers in
Figure 12-4, which would be less accurate).
FIGURE 12-5:
Components of
the chi-square
statistic: squares
of the scaled
differences.
You then add up these squared scaled differences: 2.01 1.52 3.01 2.27 8.81
to get the chi-square test statistic. This sum is an excellent test statistic to mea-
sure the overall departure of your data from the null hypothesis:
» If the null hypothesis is true (use of CBD or NSAID does not impact pain relief
status), this statistic should be quite small.
» If one of the levels of treatment has a disproportionate association with the

outcome (in either direction), it will affect the whole table, and the result will
be a larger test statistic.
Determining the p value

Now that you calculated the test statistic, the only remaining task before interpre-
tation is to determine the p value. The p value represents the probability that ran-
dom fluctuations alone, in the absence of any true effect of CBD or NSAIDs on pain
relief, could lead to a value of 8.81 or greater for this test statistic. (We introduce p
values in Chapter 3.) Once again, the rigorous proof is very complicated, so we
present an informal explanation:
When the expected cell counts are very large, the Poisson distribution becomes
very close to a normal distribution (see Chapter 24 for more on the Poisson distri-
bution). If the H 0 is true, each scaled difference should be an approximately nor-
mally distributed random variable with a mean of zero and a standard deviation
of 1. The mean is zero because you subtract the expected value from the observed
value, and the standard deviation is 1 because it is divided by the SE. The sum of
the squares of one or more normally distributed random numbers is a number

that follows the chi-square distribution (also covered in Chapter 24). So the test
statistic from this test should follow the chi-square distribution. Now it is obvious
why it is named the chi-square test! The next step is to obtain the p value for the
test statistic. To do that manually, you would look up the test statistic (which is
8.81 in our case) in a chi-square table.
In actuality, the chi-square distribution refers to a family of distributions. Which

chi-square distribution you are using depends upon a number called the degrees of
freedom, abbreviated d.f. or df or by the Greek lowercase letter nu (v) (in this book
we use df). The df is a measure of the probability of independence between the value
of the predictor (row) variable and value of the column (outcome) variable.
How would you calculate the df for a chi-square test? The answer is it depends on
the number of rows in the cross-tab. For the 2 2 cross-tab (fourfold table) in this
example, you added up the four values in Figure 12-5, so you may think that you
should look up the 8.81 chi-square value with 4 df. But you’d be wrong. Note the
italicized word independence in the preceding paragraph. And keep in mind that
the differences (Ob – Ex ) in any row or column always add up to zero. The four
terms making up the 8.81 total aren’t independent of each other. It turns out that
the chi-square test statistic for a fourfold table has only 1 df, not 4. In general, an
N-by-M table, with N rows, M columns, and therefore N M cells, has only
N 1 M 1 df because of the constraints on the row and column sums. In our
case, N — which is the number of rows — is 2, so N-1 is 1. Also, M — which is the
number of columns — is 2, so M-1 is 1 also (and 1 times 1 is 1). Don’t feel bad if
this wrinkle caught you by surprise — even Karl Pearson who invented the
chi-square test got that part wrong!
So, if you were to manually look up the chi-square test statistic of 8.81 in a
chi-square table, you would have to look under the distribution for 1 df to find out
the p value. Alternatively, if you got this far and you wanted to use the statistical
software R to look up the p value, you would use the following code: pchisq(8.81, 1,
lower.tail = FALSE). Either way, the p value for chi-square = 8.81, with 1 df, is 0.003.
This means that there’s only a 0.003 probability that random fluctuations could
produce the effect seen, where CBD performs so differently than NSAIDs with
respect to pain relief in chronic arthritis patients. A 0.003 probability is the same
as 1 chance in 333 (because 1/0.003 333), meaning very unlikely, but not impos-
sible. So, if you set α = 0.05, because 0.003 < 0.05, your conclusion would be that
in the chronic arthritis patients in our sample, whether the participant took CBD
or NSAIDs was statistically significantly associated with whether or not they felt
pain relief.

Putting it all together with some
notation and formulas
The calculations of the Pearson chi-square test can be summarized concisely
using the cell-naming conventions shown in Figure 12-6, along with the standard
summation notation described in Chapter 2.
FIGURE 12-6:
A general way of
naming the cells
of a cross-tab
table.
Using these conventions, the basic formulas for the Pearson chi-square test are as
follows:
Ri C j
» Expected values: Ex i , j
T
, i 1, 2, ... N ; j 1, 2, ... M
N M ( Ob Ex i , j ) 2
» Chi-square statistic: 2 i, j
Ex i , j
i 1j 1
» Degrees of freedom: df (N 1)(M 1)
where i and j are array indices that indicate the row and column, respectively, of
each cell.
Pointing out the pros and cons

of the chi-square test
The Pearson chi-square test is very popular for several reasons:
» It’s easy! The calculations are simple to do manually in Microsoft Excel

(although this is not recommended because the risk of making a typing
mistake is high). As described earlier, statistical software packages like
the ones discussed in Chapter 4 can perform the chi-square test for both
individual-level data as well as summarized cross-tabulated data. Also,
several websites can perform the test, and the test has been implemented
on smartphones and tablets.

» It’s flexible! The test works for tables with any number of rows and columns,
and it easily handles cell counts of any magnitude. Statistical software can
usually complete the calculations quickly, even on big data sets.
But the chi-square test has some shortcomings:
» It’s not an exact test. The p value it produces is only approximate, so using
p 0.05 as your criterion for statistical significance (meaning setting α = 0.05)
doesn’t necessarily guarantee that your Type I error rate will be only
5 percent. Remember, your Type I error rate is the likelihood you will claim
statistical significance on a difference that is not true (see Chapter 3 for an
introduction to Type I errors). The level of accuracy of the statistical signifi-
cance is high when all the cells in the table have large counts, but it becomes
unreliable when one or more cell counts is very small (or zero). There are
different recommendations as to the minimum counts you need per cell in
order to confidently use the chi-square test. A rule of thumb that many
analysts use is that you should have at least five observations in each cell of
your table (or better yet, at least five expected counts in each cell).
» It’s not good at detecting trends. The chi-square test isn’t good at detecting
small but steady progressive trends across the successive categories of an
ordinal variable (see Chapter 4 if you’re not sure what ordinal is). It may give a
significant result if the trend is strong enough, but it’s not designed specifically
to work with ordinal categorical data. In those cases, you should use a
Mantel-Haenszel chi-square test for trend, which is outside the scope of this
book.
Modifying the chi-square test: The

Yates continuity correction
There is a little drama around the original Pearson chi-square of association test
that needs to be mentioned here. Yates, who was a contemporary of Pearson,
developed what is called the Yates continuity correction. Yates argued that in the
special case of the fourfold table, adding this correction results in more reliable
p values. The correction consists of subtracting 0.5 from the magnitude of the
(Ob – Ex ) difference before squaring it.
Let’s apply the Yates continuity correction for your analysis of the sample data in
the earlier section “Understanding how the chi-square test works.” Take a look at
Figure 12-3, which has the differences between the values in the observed and
expected cells. The application of the Yates correction changes the 7.20 (or –7.20)
difference in each cell to 6.70 (or –6.70). This lowers the chi-square value from

8.81 down to 7.63 and increases the p value from 0.0030 to 0.0057, which is still
very significant — the chance of random fluctuations producing such an apparent
effect in your sample is only about 1 in 175 (because 1/0.0057 175 ).
Even though the Yates correction to the Pearson chi-square test is only applicable
to the fourfold table (and not tables with more rows or columns), some statisti-
cians feel the Yates correction is too strict. Nevertheless, it has been automatically
built into statistical software like R, so if you run a Pearson chi-square using most
commercial software, it automatically uses the Yates correction when analyzing a
fourfold table (see Chapter 4 for a discussion of statistical software).
Focusing on the Fisher Exact Test

The Pearson chi-square test described earlier isn’t the only way to analyze cross-
tabulated data. Remember that one of the cons was that it is not an exact test?
Famous but controversial statistician R. A. Fisher invented another test in the
1920s that gives the exact p value for tables that can handle very small cell counts
(even cell counts of zero!). Not surprisingly, this test is called the Fisher Exact test
(also sometimes referred to Fisher’s exact test, or just Fisher).
Understanding how the Fisher

Exact test works
Like with the chi-square, you don’t have to know the details of the Fisher Exact
test to use it. If you have a computer do the calculations for you (which we always
recommend), you technically don’t have to read this section. But we encourage
you to read this section anyway so you’ll have a better appreciation for the
strengths and limitations of this test.
This test is conceptually pretty simple. Instead of taking the product of the mar-
ginals and dividing it by the total for each cell as is done with the chi-square test
statistic, Fisher exact test looks at every possible table that has the same marginal
totals as your observed table. You calculate the exact probability (Pr) of getting
each individual table using a formula that, for a fourfold table (using the notation
for Figure 12-6), is
( R1!)( R2!)(C 1!)(C 2!)

Pr
(Ob 1,1!)(Ob 1,2!)(Ob 2,1!)( Ob2,2!)(T !)

Those exclamation points indicate calculating the factorials of the cell counts (see
Chapter 2). For the example in Figure 12-1, the observed table has a probability of
(60!)(40!)(43!)(57!)
Pr 0.00196
(33!)(27!)(10!)(30!)(100!)
Other possible tables with the same marginal totals as the observed table have
their own Pr values, which may be larger than, smaller than, or equal to the Pr
value of the observed table. The Pr values for all possible tables with a specified set
of marginal totals always add up to exactly 1.
The Fisher Exact test p value is obtained by adding up the Pr values for all tables
that are at least as different from the H 0 as your observed table. For a fourfold
table that means adding up all the Pr values that are less than (or equal to) the Pr
value for your observed table.
For the example in Figure 12-1, the p value comes out to 0.00385, which means
that there’s only 1 chance in 260 (because 1/0.00385 260) that random fluctua-
tions could have produced such an apparent effect in your sample.
Noting the pros and cons of

the Fisher Exact test
The big advantages of the Fisher Exact test are as follows:
» It gives the exact p value.
» It is exact for all tables, with large or small (or even zero) cell counts.
Why do people still use the chi-square test, which is approximate and doesn’t
work for tables with small cell counts? Well, there are several problems with the
Fisher Exact test:
» The Fisher calculations are a lot more complicated, especially for tables larger
than 2 2. Many statistical software packages either don’t offer the Fisher Exact
test or offer it only for fourfold tables. Even if they offer it, you may execute
the test and find that it fails to finish the test, and you have to break into the
program to stop the procedure. Also, some interactive web pages perform the
Fisher Exact test for fourfold tables (including www.socscistatistics.com/
tests/fisher/default2.aspx). Only the major statistical software packages
(like SAS, SPSS, and R, described in Chapter 4) offer the Fisher Exact test for
tables larger than 2 2 because the calculations are so intense. For this reason,
the Fisher Exact test is only practical for small cell counts.

» The calculations can become numerically unstable for large cell counts, even
in a 2 2 table. The equations involve the factorials of the cell counts and
marginal totals, and these can get very large — even for modest sample
sizes — often exceeding the largest number that a computer program can
handle. Many programs and web pages that offer the Fisher Exact test for
fourfold tables fail with data from more than 100 subjects.
» Another issue is — like the chi-square test — the Fisher Exact test is not for
detecting gradual trends across ordinal categories.
Calculating Power and Sample Size for

Chi-Square and Fisher Exact Tests
Note: The basic ideas of power and sample-size calculations are described in
Chapter 3, and you should review that information before going further here.
Earlier in the section “Examining Two Variables with the Pearson Chi-Square
Test,” we used an example of an observational study design in which study par-
ticipants were patients who chose which treatment they were using. In this sec-
tion, we use an example from a clinical trial study design in which study
participants are assigned to a treatment group. The point is that the tests in this
section work on all types of study designs.
Let’s calculate sample size together. Suppose that you’re planning a study to test
whether giving a certain dietary supplement to a pregnant woman reduces her
chances of developing morning sickness during the first trimester of pregnancy,
which is the first three months. This condition normally occurs in 80 percent of
pregnant women, and if the supplement can reduce that incidence rate to only
60 percent, it would be considered a large enough reduction to be clinically sig-
nificant. So, you plan to enroll a group of pregnant women who are early in their
first trimester and randomize them to receive either the dietary supplement or a
placebo that looks, smells, and tastes exactly like the supplement. You will ran-
domly assign each participant to the either the supplement group or the placebo
group in a process called randomization. The participants will not be told which
group they are in, which is called blinding. (There is nothing unethical about this
situation because all participants will agree before participating in the study that
they would be willing to take the product associated with each randomized group,
regardless of the one to which they are randomized.)
You’ll have them take the product during their first trimester, and you’ll survey
them to record whether they experience morning sickness during that time (using

explicit criteria for what constitutes morning sickness). Then you’ll tabulate the
results in a 2 × 2 cross-tab. The table will look similar to Figure 12-1, but instead
will say “supplement” and “placebo” as the label on the two rows, and “did” and
“did not” experience morning sickness as the headings on the two columns. And
you’ll test for a significant effect with a chi-square or Fisher Exact test. So, your
sample size calculation question is: How many subjects must you enroll to have at least
an 80 percent chance of getting p 0.05 on the test if the supplement truly can reduce the
incidence from 80 percent to 60 percent?
You have several ways to estimate the required sample size. The most general and
most accurate way is to use power/sample-size software such as G*Power, which
is described in detail in Chapter 4. Or you can use the online sample-size calcula-
tor at https://clincalc.com/stats/samplesize.aspx, which produces the
same results.
You need to enroll additional subjects to allow for possible attrition during the
study. If you expect x percent of the subjects to drop out, your enrollment should be:
Enrollment 100 Analyzable Number / (100 x)
So, if you expect 15 percent of enrolled subjects to drop out and therefore be unan-
alyzable, you need to enroll 100 × 197/(100 – 15), or about 231 participants.

IN THIS CHAPTER
» Beginning with the basics of fourfold

tables
» Digging into sampling strategies for

fourfold tables
» Using fourfold tables in different

scenarios
Chapter 13
Taking a Closer Look
at Fourfold Tables
I
n Chapter 12, we show you how to compare proportions between two or more
groups with a cross-tab table. In general, a cross-tab shows the relationship
between two categorical variables. Each row of the table represents one partic-
ular category of one of the variables, and each column of the table represents one
particular category of the other variable. The table can have two or more rows and
two or more columns, depending on the number of different categories or levels
present in each of the two variables. (To refresh your memory about categorical
variables, read Chapter 8.)
Imagine that you are comparing the performance of three treatments (Drug A,
Drug B, and Drug C) in patients who could have four possible outcomes: improved,
stayed the same, got worse, or left the study due to side effects. In such a case,
your treatment variable would have three levels so your cross-tab would have
three rows, and your outcome variable would have four levels so your cross-tab
would have four columns.
But this chapter only focuses on the special case that occurs when both categorical
variables in the table have only two levels. Other words for two-level variables are
dichotomous and binary. A few examples of dichotomous variables are hyperten-
sion status (hypertension or no hypertension), obesity status (obese or not obese),
and pregnancy status (pregnant or not pregnant). The cross-tab of two
CHAPTER 13 Taking a Closer Look at Fourfold Tables 173

dichotomous variables has two rows and two columns. Because a 2 × 2 cross-tab
table has four cells, it’s commonly called a fourfold table. Another name you may
see for this table is a contingency table.
Chapter 12 includes a discussion of fourfold tables, and all that is included in

Chapter 12 applies not only to fourfold tables but also to larger cross-tab tables.
But because the fourfold table plays a pivotal role in public health with regard to
certain calculations used commonly in epidemiology and biostatistics, it warrants
a chapter all its own — this one! In this chapter, we describe several common
research scenarios in which fourfold tables are used, which are: comparing pro-
portions, testing for association, evaluating exposure and outcome associations,
quantifying the performance of diagnostic tests, assessing the effectiveness of
therapies, and measuring inter-rater and intra-rater reliability. In each scenario,
we describe how to calculate several common measures called indices (singular:
index), along with their confidence intervals. We also describe ways of sampling
called sampling strategies (see Chapter 6 for more on sampling).
Focusing on the Fundamentals

of Fourfold Tables
In contemplating statistical testing in a fourfold table, consider the process. As
described in Chapter 12, you first formulate a null hypothesis (H0) about the four-
fold table, set the significance level (such as α = 0.05), calculate a test statistic,
find the corresponding p value, and interpret the result. With a fourfold table, one
obvious test to use is the chi-square test (if necessary assumptions are met). The
chi-square test evaluates whether membership in a particular row is statistically
significantly associated with membership in a particular column. The p value on
the chi-square test is the probability that random fluctuations alone, in the
absence of any real effect in the population, could have produced an observed
effect at least as large as what you saw in your sample. If the p value is less than
α (which is 0.05 in your scenario), the effect is said to be statistically significant, and
the null is rejected. Assessing significance using a chi-square test is the most
common approach to testing a cross-tab of any size, including a fourfold table.
But fourfold tables can serve as the basis for developing other metrics besides chi-
square tests that can be useful in other ways, which are discussed in this chapter.
In the rest of this chapter, we describe many useful calculations that you can
derive from the cell counts in a fourfold table. The statistical software that cross-
tabulates your raw data can provide these indices depending upon the commands
it has available (see Chapter 4 for a review of statistical software). Thankfully (and
uncharacteristically), unlike in most chapters in this book, the formulas for many
indices derived from fourfold tables are simple enough to do manually with a

calculator (or using Microsoft Excel). All you need are the counts or frequencies of
each of the four cells. For these indices, you can also use a web page for calcula-
tion, which is available here: https://statpages.info/ctab2x2.html. This
chapter demonstrates how to calculate these indices in R (a free, open-source
software described in Chapter 4).
Like any other value you calculate from a sample, an index calculated from a four-
fold table is a sample statistic, which is an estimate of the corresponding population
parameter. A good researcher always wants to quote the precision of that estimate.
In Chapter 10, we describe how to calculate the standard error (SE) and confidence
interval (CI) for sample statistics such as means and proportions. Likewise, in this
chapter, we show you how to calculate the SE and CI for the various indices you
can derive from a fourfold table.
Though an index itself may be easy to calculate manually, its SE or CI usually is

not. Approximate formulas are available for some of the more common indices.
These formulas are usually based on the fact that the random sampling fluctua-
tions of an index (or its logarithm) are often nearly normally distributed if the
sample size is large enough. We provide approximate formulas for SEs where
they’re available, and demonstrate how to calculate them in R when possible.
For consistency, all the formulas in this chapter refer to the four cell counts of the
fourfold table, and the row totals, column totals, and grand total, in the same
standard way (see Figure 13-1). This convention is used in many online resources
and textbooks.
FIGURE 13-1:
These
designations for
cell counts and
totals are used
throughout this
chapter.
Choosing the Correct Sampling Strategy

In this section, we assume you are designing a cross-sectional study (see
Chapter 7 for a review of study design terminology). Using such a design, though
you could not assess cause-and-effect, you could evaluate the association between
an exposure (hypothesized cause) and outcome. For example, you may hypothe-
size that being obese (exposure) causes a patient to develop hypertension

(abbreviated HTN, outcome) over time. However, in a cross-sectional study,
exposure and outcome are measured at the same time, so you can only look for
associations. If your exposure and outcome are binary (such as obese: yes/no and
HTN: yes/no), you can use a fourfold table for this evaluation. But you would have
to develop a sampling strategy that would support your analytic plan.
The example given here is for the use of a fourfold table to interpret a cross-
sectional study. If you have heard of a fourfold table being analyzed as part of a
cohort study or longitudinal study, that is referring to a series of cross-sectional
studies done over time to the same group or cohort. Each round of data collection
is called a wave, and fourfold tables can be developed cross-sectionally (using data
from one wave), or longitudinally (using data from two waves).
As described in Chapter 6, you could try simple random sampling (SRS), but this
may not provide you with a balanced number of participants who are positive for
the exposure compared to negative to the exposure. If you are worried about this,
you could try stratified sampling on the exposure (such as requiring half the sam-
ple to be obese, and half the sample to not be obese). Although other sampling
strategies described in Chapter 6 could be used, SRS and stratified sampling are
the most common to use in cross-sectional study. Why is your sampling strategy
so important? As you see in the rest of this chapter, some indices are meaningful
only if the sampling is done accordingly so as to support a particular study design.
Producing Fourfold Tables in

a Variety of Situations
Fourfold tables can arise from a number of different scenarios, including the
following:
» Comparing proportions between two groups (see Chapter 12)
» Testing whether two binary variables are associated
» Assessing associations between exposures and outcomes
» Evaluating diagnostic procedures
» Evaluating therapies
» Evaluating inter-rater reliability
Note: These scenarios can also give rise to tables larger than 2 2, and fourfold
tables can arise in other scenarios besides these.

Describing the association between
two binary variables
Suppose you conduct a cross-sectional study by enrolling a random sample of
60 adults from the local population as participants in your study with the hypoth-
esis that being obese is associated with having HTN. For the exposure, suppose
you measure their height and weight, and use these values to calculate their body
mass index (BMI). You then use their BMI to classify them as either obese or non-
obese. For the outcome, you also measure their blood pressure in order to catego-
rize them as having HTN or not having HTN. This is simple random sampling
(SRS), as described in the earlier section “Choosing the Correct Sampling Strat-
egy.” You can summarize your data in a fourfold table (see Figure 13-2).
The table in Figure 13-2 indicates that more than half of the obese participants
have HTN and more than half of the non-obese participants don’t have HTN — so
there appears to be a relationship between being a membership in a particular row
and simultaneously being a member of a particular column. You can show this
apparent association is statistically significant in this sample using either a Yates
chi-square or a Fisher Exact test on this table (as we describe in Chapter 12). If you
do these tests, your p values will be p 0.016 and p 0.013 , respectively, and at
α = 0.05, you will be comfortable rejecting the null.
FIGURE 13-2:
A fourfold table
summarizing
obesity and
hypertension
in a sample of
60 participants.
But when you present the results of this study, just saying that a statistically sig-
nificant association exists between obesity status and HTN status isn’t enough.
You should also indicate how strong this relationship is and in what direction it
goes. A simple solution to this is to present the test statistic and p value to repre-
sent how strong the relationship is, and to present the actual results as row
or column percentages to indicate the direction. For the data in Figure 13-2, you
could say that being obese was associated with having HTN, because 14/21 =
66 percent of obese participants also had HTN, while only 12/39 = 31 percent of
non-obese participants had HTN.

Quantifying associations
How strongly is an exposure associated with an outcome? If you are considering
this question with respect to exposures and outcomes that are continuous vari-
ables, you would try to answer it with a scatter plot and start looking for correla-
tion and linear relationships, as discussed in Chapter 15. But in our case, with a
fourfold table, you are essentially asking: how strongly are the two levels of the
exposure represented in the rows associated with the two levels of the outcome
represented in the columns? In the case of a cohort study — where the exposure
is measured in participants without the outcome who are followed longitudinally
to see if they get the outcome — you can ask if the exposure was associated with
risk of the outcome or protection from the outcome. In a cohort design, you could
ask, “How much does being obese increase the likelihood of getting HTN?” You
can calculate two indices from the fourfold table that describe this increase, as you
discover in the following sections.
Relative risk and the risk ratio

In a cohort study, you seek to quantify the amount of risk (or probability) for the
outcome that is conferred by having the exposure. The risk of getting a negative
outcome is estimated as the fraction of participants who experience the outcome
during follow-up (because in a cohort design, all participants do not have the out-
come when they enter the study). Another term for risk is cumulative incidence rate
(CIR). You can calculate the CIR for the whole study, as well as separately for each
stratum of the exposure (in our case, obese and nonobese). Using the notation in
Figure 13-1, the CIR for participants with the exposure is a / r 1. For the example
from Figure 13-2, it’s 14 / 21, which is 0.667 (66.7 percent). And for those without
the exposure, the CIR is represented by c / r 2 . For this example, the CIR is calcu-
lated as 12 / 39 , which is 0.308 (30.8 percent).
The term exposure specifies a hypothesized cause of an outcome. If it is found that

a certain exposure typically causes risk for an outcome, it is called a risk factor, and
if it is found to confer protection, it is called a protective factor. Higher education
has been found to be a protective factor against many negative outcomes (such as
most injuries), and obesity has been found to be a risk factor for many negative
outcomes (such as HTN and Type II diabetes).
The term relative risk refers to the amount of risk one group has relative to another.
This chapter discusses different measures of relative risk that are to be used with
different study designs. It is important to acknowledge here that technically, the
term risk can only apply to cohort studies because you can only be at risk if you
possess the exposure but not the outcome for some period of time in a study, and
only cohort studies have this design feature. However, the other study designs —
including cross-sectional and case-control — intend to estimate the relative risk

you would get if you had done a cohort study, so they produce estimates of relative
risk (see Chapter 7).
In a cohort study, the measure of relative risk used is called the risk ratio (also
called cumulative incidence ratio). To calculate the risk ratio, first calculate the
CIR in the exposed, calculate the CIR in the unexposed, then take a ratio of the CIR
in the exposed to the CIR in the unexposed. This formula could be expressed
as: RR a / r1 / c / r 2 .
For this example, as calculated earlier, the CIR for the exposed was 0.667, and the
CIR for the unexposed was 0.308. Therefore, the risk-ratio calculation would be
0.667 / 0.308, which is 2.17. So, in this cohort study, obese participants were slightly
more than twice as likely to be diagnosed as having HTN during follow-up than
non-obese subjects.
Calculating a risk ratio as a measure of relative risk is appropriate for a cohort

study. However, there are restrictions when creating measures of relative risk for
cross-sectional and case-control designs. In a cross-sectional study, you would
not calculate a CIR in the exposed and the unexposed because the exposure and
outcome are measured at the same time, so there’s no time for any participants to
experience any risk during the study. Instead, you would calculate the prevalence
of the outcome in the exposed — which is a / r 1 — and the prevalence of the out-
come in the unexposed — which is c / r 2 (read Chapter 14 to for a discussion of
prevalence). Notice that even though we use different wording, these are the same
formulas as for the CIR. Then, instead of a risk ratio, in a cross-sectional study
you would use a prevalence ratio, which is calculated the same way as the risk ratio:
PR = (a/r1)/(c/r2).
In a case-control study, for a measure of relative risk, you must use the odds ratio
(discussed later in the section “Odds ratio”). You cannot use the risk ratio or
prevalence ratio in a case-control study. The odds ratio can also be used as a mea-
sure of relative risk in a cross-sectional study, and can technically be used in a
cohort study, although the preferred measure is the risk ratio.
Let’s go back to discussing the risk ratio. You can calculate an approximate
95 percent confidence interval (CI) around the observed risk ratio using the fol-
lowing formulas, which assume that the logarithm of the risk ratio is normally
distributed:
1. Calculate the standard error (SE) of the log of risk ratio using the follow-
ing formula:
SE b / ( a r 1) d / ( c r 2 )

2. Calculate Q with the following formula: Q e 1.96 SE where Q is simply a
convenient intermediate quantity that will be used in the next part of
the calculation, and e is the mathematical constant 2.718.
3. Find the lower and upper limits of the CI with the following formula:
RR
Q to ( RR Q )
95% CI
For confidence levels other than 95 percent, replace the z-score of 1.96 in Step 2
with the corresponding z-score shown in Table 10-1 of Chapter 10. As an example,
for 90 percent confidence levels, use 1.64, and for 99 percent confidence levels,
use 2.58.
For the example in Figure 13-2, you calculate 95 percent CI around the observed
risk ratio as follows:
1. SE 7 / (14 21) 27 / (12 39) , which is 0.2855.
2. Q e 1.96 0.2855
, which is 1.75.
3. The 95% CI 2.17

1.75
to (2.17 1.75), which is 1.24 to 3.80.
Using this formula, the risk ratio would be expressed as 2.17, 95 percent CI 1.24
to 3.80.
You could also use R to calculate a risk ratio and 95 percent CI for the fourfold
table in Figure 13-2 with the following steps:
1. Create a matrix.
Create a matrix called obese_HTN with this code: obese_HTN <-

matrix(c(14,12,7,27),nrow = 2, ncol = 2).
2. Load a library.
For many epidemiologic calculations, you can use the epitools package in R and
use a command from this package to calculate the risk ratio and 95 percent
CI. Load the epitools library with this command: library(epitools).
3. Run the command on the matrix.
In this case, run the riskratio.wald command on the obese_HTN matrix you
created in Step 1: riskratio.wald(obese_HTN).
The output is shown in Listing 13-1.

LISTING 13-1: R output from risk ratio calculation on data from Figure 13-2
> riskratio.wald(obese_HTN)
$data
Outcome
Predictor Disease1 Disease2 Total
Exposed1 14 7 21
Exposed2 12 27 39
Total 26 34 60
$measure
risk ratio with 95% C.I.
Predictor estimate lower upper
Exposed1 1.000000 NA NA
Exposed2 2.076923 1.09512 3.938939
$p.value
two-sided
Predictor midp.exact fisher.exact chi.square
Exposed1 NA NA NA
Exposed2 0.009518722 0.01318013 0.00744125
$correction
[1] FALSE
attr(,”method”)
[1] "Unconditional MLE & normal approximation (Wald) CI"
>
Notice that the output is organized under the following headings: $data, $measure,
$p.value, and $correction. Under the $measure section is a centered title that says
risk ratio with 95% C.I. — which is more than a hint! Under that is a table with the
following column headings: Predictor, estimate, lower, and upper. The estimate col-
umn has the risk ratio estimate (which you already calculated by hand and rounded
off to 2.17). The lower and upper columns have the confidence limits, which R
calculated as 1.09512 (round to 1.10) and 3.938939 (round to 3.94), respectively.
You may notice that because R used a slightly different SE formula than our man-
ual calculation, R’s CI was slightly wider.
Odds ratio
The odds of an event occurring is the probability of it happening divided by the
probability of it not happening. Assuming you use p to represent a probability, you
could write the odds equation this way: p / (1 p ). In a fourfold table, you would
represent the odds of the outcome in the exposed as a/b. You would also represent
the odds of the outcome in the unexposed as c/d.

Let’s apply this to the scenario depicted in Figure 13-2. This is a study of 60 indi-
viduals, where the exposure is obesity status (yes/no) and the outcome is HTN
status (yes/no). Using the data from Figure 13-2, the odds of having the outcome
for exposed participants would be calculated as a / b , which would be 14 / 7 , which
is 2.00. And the odds of having the outcome in the unexposed participants is c/d,
which would be 12 / 27, which is 0.444.
Odds have no units. They are not expressed as percentages. See Chapter 3 for a
more detailed discussion of odds.
When considering cross-sectional and cohort studies, the odds ratio (OR) repre-
sents the ratio of the odds of the outcome in the exposed to the odds of the out-
come in the unexposed. In case-control studies, because of the sampling approach,
the OR represents the ratio of the odds of exposure among those with the outcome
to the odds of exposure among those without the outcome. But because any four-
fold table has only one OR no matter how you calculate it, the actual value of the
OR stays the same, but how it is described and interpreted depends upon the study
design.
Let’s assume that Figure 13-2 presents data on a cross-sectional study, so we will
look at the OR from that perspective. Because you calculate the odds in the exposed
as a/b, and the odds in the unexposed as c/d, the odds ratio is calculated by divid-
ing a/b by b/c like this: OR a/b / c/d .
For this example, the OR is (14 / 7 ) / (12 / 27 ), which is 2.00 / 0.444 , which is 4.50. In
this sample, assuming a cross-sectional study, participants who were positive for
the exposure had 4.5 times the odds of also being positive for the outcome com-
pared to participants who were negative for the exposure. In other words, obese
participants had 4.5 times the odds of also having HTN compared to non-obese
participants.
You can calculate an approximate 95 percent CI around the observed OR using the
following formulas, which assume that the logarithm of the OR is normally
distributed:
1. Calculate the standard error of the log of the OR with the following
formula:
SE 1/ a 1/ b 1/ c 1/ d
2. Calculate Q with the following formula: Q e 1.96 SE, where Q is simply a

convenient intermediate quantity that will be used in the next part of
the calculation, and e is the mathematical constant 2.718.

3. Find the limits of the confidence interval with the following formula:
OR
Q to ( OR Q)
95% CI
Like with the risk ratio CI, for confidence levels other than 95 percent, replace the
z-score of 1.96 in Step 2 with the corresponding z-score shown in Table 10-1 of
Chapter 10. As an example, for 90 percent confidence levels, use 1.64, and for
99 percent confidence levels, use 2.58.
For the example in Figure 13-2, you calculate 95 percent CI around the observed
OR as follows:
1. SE 1 / 14 1 / 7 1 / 12 1 / 27 , which is 0.5785.
2. Q e 1.96 0.5785
, which is 3.11.
3. 95% CI 4.50
3.11
to (4.50 3.11), which is 1.45 to 14.0.
Using these calculations, the OR is estimated as 4.5, and the 95 percent CI as 1.45
to 14.0.
To do this operation in R, you would follow the same steps as listed at the end of
the previous section, except in Step 3, the command you’d run on the matrix is
oddsratio.wald() using this code: oddsratio.wald(obese_HTN). The output is laid out
the same way as shown in Listing 13-1, with a $measure section titled odds ratio
with a 95% C.I. In that section, it indicates that the lower and upper confidence
limits are 1.448095 (rounded to 1.45) and 13.98389 (rounded to 13.98), respec-
tively. This time, R’s estimate of the 95 percent CI was close to the one you got
with your manual calculation, but slightly narrower.
A wide 95 percent CI is the sign of an unstable (and not very useful) estimate.
Consider a 95 percent CI for an OR that goes from 1.45 to 14.0. If you are interpret-
ing the results of a cohort study, you are saying that obesity could increase the
odds of getting HTN by as little as 1.45, or as much as 14! Most researchers try to
solve this problem by increasing their sample size to reduce the size of their SE,
which will in turn reduce the width of the CI.
Evaluating diagnostic procedures

Many diagnostic procedures provide a positive or negative test result — such as a
COVID-19 test. Ideally, this result should correspond to the true presence or
absence of the medical condition for which the test was administered — meaning
a positive COVID-19 test should mean you have COVID-19, and a negative test
should mean you do not. The true presence or absence of a medical condition is

best determined by some gold standard test that the medical community accepts as
perfectly accurate in diagnosing the condition. For COVID-19, the polymerase
chain reaction (PCR) test is considered a gold standard test because of its high
level of accuracy. But gold standard diagnostic procedures (like PCR tests) can be
time-consuming and expensive, and in the case of invasive procedures like biop-
sies, they may be very unpleasant for the patient. Therefore, quick, inexpensive,
and relatively noninvasive screening tests are very valuable, even if they are not
perfectly accurate. They just need to be accurate enough to help filter in the best
candidates for a gold standard diagnostic test.
Most screening tests produce some false positive results, which is when the result
of the test is positive, but the patient is actually negative for the condition. Screen-
ing tests also produce some false negative results, where the result is negative in
patients where the condition is present. Because of this, it is important to know
false positive rates, false negative rates, and other features of screening tests to
consider their level of accuracy in your interpretation of their results.
You usually evaluate a new, experimental screening test for a particular medical
condition by administering the new test to a group of participants. These partici-
pants include some who have the condition and some who do not. For all the
participants in the study, their status with respect to the particular medical condi-
tion has been determined by the gold standard method, and you are seeing how
well your new, experimental screening test compares. You can then cross-
tabulate the new screening test results against the gold standard results repre-
senting the true condition in the participants. You would create a fourfold table in
a framework as shown in Figure 13-3.
FIGURE 13-3:
This is how data
are summarized
when evaluating
a proposed new
diagnostic
screening test.

Imagine that you are conducting a study at a primary care clinic. In the study, you
administer a newly developed home pregnancy test to 100 women who come to a
primary care appointment suspecting that they may be pregnant. This is conve-
nience sampling from a population defined as “all women who think they may be
pregnant,” which is the population to whom a home pregnancy test would be
marketed. At the appointment, all the participants would be given the gold stan-
dard pregnancy test, so by the end of the appointment, you would know their true
pregnancy status according to the gold standard, as well as what your new home
pregnancy test result said. Your results would next be cross-tabulated according
to the framework shown in Figure 13-4.
FIGURE 13-4:
Results from a
study of a new
experimental
home pregnancy
test.
The structure of the table in Figure 13-4 is important because if your results are
arranged in that way, you can easily calculate at least five important characteris-
tics of the experimental test (in our case, the home pregnancy test) from this
table: accuracy, sensitivity, specificity, positive predictive value (PPV), and nega-
tive predictive value (NPV). We explain how in the following sections.
Overall accuracy
Overall accuracy measures how often a test result comes out the same as the gold
standard result (meaning the test is correct). A perfectly accurate test never pro-
duces false positive or false negative results. In Figure 13-4, cells a and d represent
correct test results, so the overall accuracy of the home pregnancy test is a d /t .
Using the data in Figure 13-4, accuracy 33 51 / 100 , which is 0.84, or
84 percent.
Sensitivity and specificity

A perfectly sensitive test never produces a false negative result for an individual
with the condition. Conversely, a perfectly specific test never produces a false posi-
tive result for an individual negative for the condition. The goal of developing a
screening test is to balance the sensitivity against the specificity of the test based

on the context of the condition, to optimize the accuracy and get the best of both
worlds while minimizing both false positive and false negative results.
In a screening test with a sensitivity of 100 percent, the test result is always posi-
tive whenever the condition is truly present. In other words, the test will identify
all individuals who truly have the condition. When a perfectly sensitive test comes
out negative, you can be sure the person doesn’t have the condition. You calculate
sensitivity by dividing the number of true positive cases by the total number of
cases where the condition was truly present: a/c1 (that is, true positive/all
present). Using the data in Figure 13-4, sensitivity 33 / 37, which is 0.89. This
means that the home test comes out positive in only 89 percent of truly pregnant
women, and the other 11 percent were really pregnant, but had a false positive
result in the test.
A perfectly specific test never produces a false positive result for an individual
without the condition. In a test that has a specificity of 100 percent, whenever the
condition is truly absent, the test always has a negative result. In other words, the
test will identify all individuals who truly do not have the condition. When a per-
fectly specific test comes out positive, you can be sure the person has the condi-
tion. You calculate specificity by dividing the number of true negative cases by the
total number of cases where the condition was truly absent: d / c2 (that is, true
negative/all not present). Using the data in Figure 13-4, specificity 51/63 , which
is 0.81. This means that among the women who were not pregnant, the home test
was negative only 81 percent of the time, and 11 percent of women who were truly
negative tested as positive. (You can see why it is important to do studies like this
before promoting the use of a particular screening test!)
But imagine you work in a lab that processes the results of screening tests, and
you do not usually have access to the gold standard results. You may ask the ques-
tion, “How likely is a particular screening test result to be correct, regardless of
whether it is positive or negative?” When asking this about positive test results,
you are asking about positive predictive value (PPV), and when asking about neg-
ative test results, you are asking about negative predictive value (NPV). These are
covered in the following sections.
Sensitivity and specificity are important characteristics of the test itself. Observe
that the answers depend on the prevalence of the condition in the background
population. If the study population were older women, then the prevalence of
being pregnant would be lower, and that would impact the sensitivity and speci-
ficity. The prevalence will also impact the PPV and NPV, which we discuss in the
next section. For these reasons, it is important to use natural sampling in such a
study design.

Positive predictive value and negative
predictive value
The positive predictive value (PPV, also called predictive value positive) is the frac-
tion of all positive test results that are true positives. In the case of the pregnancy
test scenario, the PPV would be the fraction of the time a positive screening test
result means that the woman is truly pregnant. PPV is the likelihood that a posi-
tive test result is correct. You calculate PPV as a / r 1. For the data in Figure 13-4,
the PPV is 33 / 45, which is 0.73. So, if the pregnancy test result is positive, there’s
a 73 percent chance that the woman is truly pregnant.
The negative predictive value (NPV, also called predictive value negative) is the
fraction of all negative test results that are true negatives. In the case of the preg-
nancy test scenario, the NPV is the fraction of the time a negative screening test
results means the woman is truly not pregnant. NPV is the likelihood a negative
test result is correct. You calculate NPV as d / r 2. For the data in Figure 13-4, the
NPV is 51 / 55 , which is 0.93. So, if the pregnancy test result is negative, there’s a
93 percent chance that the woman is truly not pregnant.
Investigating treatments
In conditions where there are no known treatments, one of the simplest ways to
investigate a new treatment (such as a drug or surgical procedure) is to compare
it to a placebo or sham condition using a clinical trial study design. Because many
forms of dementia have no known treatment, it would be ethical to compare new
treatments for dementia to placebo or a sham treatment in a clinical trial. In those
cases, patients with the condition under study would be randomized (randomly
assigned) to an active group (taking the real treatment) and a control group (that
would receive the sham treatment), as randomization is a required feature of clin-
ical trials. Because some of the participants in the control group may appear to
improve, it is important that participants are blinded as to their group assign-
ment, so that you can tell if outcomes are actually improved in the treatment
compared to the control group.
Suppose you conduct a study where you enroll 200 patients with mild dementia
symptoms, then randomize them so that 100 receive an experimental drug
intended for mild dementia symptoms, and 100 receive a placebo. You have the
participants take their assigned product for six weeks, then you record whether
each participant felt that the product helped their dementia symptoms. You tabu-
late the results in a fourfold table, like Figure 13-5.

FIGURE 13-5:
Comparing a
treatment to a
placebo.
According to the data in Figure 13-5, 70 percent of participants taking the new
drug report that it helped their dementia symptoms, which is quite impressive
until you see that 50 percent of participants who received the placebo also reported
improvement. When patients report therapeutic effect from a placebo, it’s called
the placebo effect, and it may come from a lot of different sources, including the
patient’s expectation of efficacy of the product. Nevertheless, if you conduct a
Yates chi-square or Fisher Exact test on the data (as described in Chapter 12)
at α = 0.05, the results show treatment assignment was statistically signifi-
cantly associated with whether or not the participant reported a treatment effect
(p 0.006 by either test).
Looking at inter- and intra-rater reliability

Many measurements in epidemiologic research are obtained by the subjective
judgment of humans. Examples include the human interpretation of X-rays, CAT
scans, ECG tracings, ultrasound images, biopsy specimens, and audio and video
recordings of the behavior of study participants in various situations. Human
researchers may generate quantitative measurements, such as determining the
length of a bone on an ultrasound image. Human researchers may also generate
classifications, such as determining the presence or absence of some atypical fea-
ture on an ECG tracing.
Humans who perform such determinations in studies are called raters because
they are assigning ratings, which are values or classifiers that will be used in the
study. For the measurements in your study, it is important to know how consis-
tent such ratings are among different raters engaged in rating the same item. This
is called inter-rater reliability. You will also be concerned with how reproducible
the ratings are if one rater were to rate the same item multiple times. This is called
intra-rater reliability.

When considering the consistency of a binary rating (like yes or no) for the same
item between two raters, you can estimate inter-rater reliability by having each
rater rate the same group of items. Imagine we had two raters rate the same 50
scans as yes or no in terms of whether each scan showed a tumor or not. We cross-
tabbed the results and present them in Figure 13-6.
FIGURE 13-6:
Results of two
raters reading the
same set of
50 specimens
and rating each
specimen yes
or no.
Looking at Figure 13-6, cell a contains a count of how many scans were rated
yes — there is a tumor — by both Rater 1 and Rater 2. Cell b counts how many
scans were rated yes by Rater 1 but no by Rater 2. Cell c counts how many scans
were rated no by Rater 1 and yes by Rater 2, and cell d shows where Rater 1 and
Rater 2 agreed and both rated the scan no. Cells a and d are considered concordant
because both raters agreed, and b and c are discordant because both raters
disagreed.
Ideally, all the scans would be counted in concordant cells a or d of Figure 13-6,
and discordant cells b and c would contain zeros. A measure of how close the data
come to this ideal is called Cohen’s Kappa, and is signified by the Greek lowercase
kappa: κ. You calculate kappa as: 2( ad bc ) / ( r 1 c2 r 2 c1).
For the data in Figure 13-6, 2 22 16 5 7 / 27 21 23 29 ,

which is 0.5138. How is this interpreted?
If the raters are in perfect agreement, then κ = 1. If you generate completely ran-
dom ratings, you will see a κ = 0. You may think this means κ takes on a positive
value between 0 and 1, but random sampling fluctuations can actually cause κ to
be negative. This situation can be compared to a student taking a true/false test
where the number of wrong answers is subtracted from the number of right
answers as a penalty for guessing. When calculating κ, getting a score less than
zero indicates the interesting combination of being both incorrect and unfortu-
nate, and is penalized!

So, how do you interpret a κ of 0.5138? There’s no universal agreement as to an
acceptable value for . One common convention is that values of κ less than 0.4 are
considered poor, those between 0.4 and 0.75 are acceptable, and those more than
0.75 are excellent. In this case, our raters may be performing acceptably.
For CIs forκ, you won’t find an easy formula, but the fourfold table web page
(https://statpages.info/ctab2x2.html) provides approximate CIs. For the
preceding example, the 95 percent CI is 0.202 to 0.735. This means that for your
two raters, their agreement was 0.514 (95 percent CI 0.202 to 0.735), which sug-
gests that the agreement level was acceptable.
You can construct a similar table to Figure 13-6 for estimating intra-rater reli-
ability. You would do this by having one rater rate the same groups of scans in two
separate sessions. In this case, in the table in Figure 13-6, you’d replace the by
Rater with in Session in the row and column labels.

IN THIS CHAPTER
» Determining and expressing the

prevalence of a condition
» Calculating incidence rates and rate

ratios, along with their standard
errors
» Comparing incidence rates between

two populations
» Estimating sample size needed to

compare incidence rates
Chapter 14
Analyzing Incidence and
Prevalence Rates in
Epidemiologic Data
E
pidemiology is the study of the causes of health and disease in human
populations. It is sometimes defined as characterizing the three Ds — the
distribution and determinants of human disease (although epidemiology
technically also concerns more positive outcomes, such as human health and
wellness). This chapter describes two concepts central to epidemiology: prevalence
and incidence. Prevalence and incidence are also frequently encountered in other
areas of human research as well. We describe how to calculate incidence rates and
prevalence proportions. Then we concentrate on the analysis of incidence. (For an
introduction to prevalence and to learn how to calculate prevalence ratios, see
Chapter 13.) Later in this chapter, we describe how to calculate confidence inter-
vals around incidence rates and rate ratios, and how to compare incidence rates
between two populations.
CHAPTER 14 Analyzing Incidence and Prevalence Rates in Epidemiologic Data 191

Understanding Incidence and Prevalence
Incidence and prevalence are two related but distinct concepts. In the following
sections, we define each of these concepts and provide examples. After that, we
describe the relationship between incidence and prevalence.
Prevalence: The fraction of a population

with a particular condition
The prevalence of a condition in a population is the proportion of the population
that has that condition at any given moment. It’s calculated by creating a fraction
with a numerator and a denominator. The denominator is the total population eli-
gible to have the condition. The numerator is the number of individuals from the
population who have the condition at a given time. If you divide this numerator by
this denominator, you will calculate the prevalence of the condition in that
population.
Prevalence can be expressed as a decimal fraction, a percentage, or a rate per so

many (usually per 1,000, per 10,000, or per 100,000). For example, a 2021 survey
found that 11.6 percent of the U.S. adult population has Type II diabetes. But a
rarer outcome — such as a monthly hospitalization rate for those suffering from
influenza — may be expressed as 31.7 per 100,000. The prevalence is expressed as
the result of a calculation from this fraction, but stated as a rate so that it is easy
to envision. It would be hard to envision that 0.0317 percent of influenza sufferers
were hospitalized in one month. On the other hand, it is much easier to envision
almost 32 people from a town with a population of 100,000 being hospitalized in
one month — provided you also envision that everyone in the town had influenza.
Because prevalence is a proportion, it’s analyzed in exactly the same way as any
other proportion. The standard error (SE) of a prevalence ratio can be estimated
by the formula in Chapter 13. Confidence intervals (CIs) for a prevalence estimate
can be obtained from exact methods based on the binomial distribution or from
formulas based on the normal approximation to the binomial distribution. Also,
prevalence can be compared between two or more populations using the chi-
square or Fisher Exact test. For this reason, the remainder of this chapter focuses
on how to analyze incidence rates.
Incidence: Counting new cases

The incidence of a condition is the rate at which new cases of that condition appear
in a population. Incidence is generally expressed as an incidence rate (R), which —
like prevalence — is a fraction. The numerator for incidence is defined as the
number of observed events (N) in a particular time period. (Consider an event to

mean that a member of the population goes from not having the condition to hav-
ing the condition.) Take note that while incidence expresses the number of new
cases of the condition in the numerator, in contrast, prevalence includes all
cases — both new and existing — in the numerator. The denominator for inci-
dence is defined as the number of individuals in the population who could have
had the event multiplied by the interval of time being used. This is also called time
exposed or exposure (E). So, the equation for incidence is the number of observed
events divided by the exposure, which is (E): R = N/E.
Exposure is measured in units of person-time, such as person-days or person-

years. Incidence rates are expressed as the number of cases per unit of person-
time. The unit of person-time is used so that the incidence rate can at least be the
size of a whole number so it is easier to interpret and compare.
The incidence rate should be estimated by counting events over a narrow enough
interval of time so that the number of observed events is a small fraction of the
total population studied. One year is narrow enough for calculating incidence of
Type II diabetes in adults because 0.02 percent of the adult population develops
diabetes in a year. However, one year isn’t narrow enough to be useful when con-
sidering the incidence of an acute condition like influenza. In influenza and other
infectious diseases, the intervals of interest would be in terms of daily, weekly,
and monthly trends. It’s not very helpful to know that 30 percent of the popula-
tion came down with influenza in a one-year period.
Consider City XYZ, which has a population of 300,000 adults. None of them has
been diagnosed with Type II diabetes. Suppose that in 2023, 30 adults from City
XYZ were newly diagnosed with Type II diabetes. The incidence of adult Type II
diabetes in City XYZ would be calculated with a numerator of 30 cases and a
denominator of 300,000 adults in one year. Using the incidence formula, this
works out to 0.0001 new cases per person-year. As described before, in epidemi-
ology, rates are reconfigured to have at least whole numbers so that they are eas-
ier to interpret and envision. For this example, you could express City XYZ’s 2023
adult Type II diabetes incidence rate as 1 new case per 10,000 person-years, or as
10 new cases per 100,000 person-years.
Now imagine another city — City ABC — has a population of 80,0000 adults, and
like with City XYZ, none of them had ever been diagnosed with Type II diabetes.
Now, assume that in 2023, 24 adults from City ABC were newly diagnosed with
Type II diabetes. City ABC’s 2023 incidence rate would be calculated as 24 cases in
80,000 individuals in one year, which works out to 24 / 80, 000 or 0.0003 new cases
per person-year. To make the estimate comparable to City XYZ’s estimate, let’s
express City ABC’s estimate as 30 new cases per 100,000 person-years. So, the
2023 adult Type II diabetes incidence rate in City ABC — which is 30 new cases per
100,000 person-years — is three times as large as the 2023 adult Type II diabetes

incidence rate for City XYZ, which is 10 new cases for 100,000 person-years.
(Looks like City ABC’s public health department needs to get advice from City
XYZ!)
Understanding how incidence and

prevalence are related
From the definitions and examples in the preceding sections, you see that inci-
dence and prevalence are two related but distinct concepts. The incidence rate
tells you how fast new cases of some condition arise in a population, and preva-
lence tells you what fraction of the population has that condition at any moment.
You may expect that conditions with higher incidence rates would have higher
prevalence than conditions with lower incidence rates. This is true with common
chronic conditions, such as hypertension. But if a condition is acute — including
infectious diseases, such as influenza and COVID-19 — the duration of the condi-
tion may be short. In such a scenario, a high incidence rate may not be paired with
a high prevalence. Relatively rare chronic diseases of long duration — such as
dementia — have low yearly incidence rates, but as human health improves and
humans live longer on average, the prevalence of dementia increases.
Analyzing Incidence Rates

The preceding sections show you how to calculate incidence rates and express
them in larger units that are easier to envision. But, as we emphasize in Chapter 10,
whenever you report an estimate you’ve calculated, you should also indicate the
level of precision of that estimate. How precise are those incident rates? And
how can you tell when the difference between two incidence rates is statistically
significant? The next sections show you how to calculate standard errors (SEs)
and confidence intervals (CIs) for incidence rates, and how to compare incidence
rates between two populations.
Expressing the precision of an

incidence rate
The precision of an incidence rate (R) is expressed using a confidence interval (CI).
The SE of R typically is not reported, because the event rate usually isn’t normally
distributed. The SE is computed only as part of the CI calculation.

Random fluctuations in R are attributed entirely to fluctuations in the event
count (N). We are assuming the exposure (described earlier in this chapter as the
person-time in the denominator, abbreviated as E) is known exactly — or at least,
much more precisely than N. Therefore, the CI for the event rate is based on the CI
for N. Here’s how you calculate the CI for R:
1. Calculate the confidence interval (CI) for N.
Chapter 11 provides approximate SE and CI formulas based on the normal

approximation to the Poisson distribution (see Chapter 24). These approxima-
tions are reasonable when N is large — meaning N ≥ 50 events:
95% CI = N 1.96 N to N 1.96 N
2. Divide the lower and upper confidence limits for N by the exposure (E).
The answer is the CI for the incidence rate R.
Earlier in the chapter, we describe City ABC, which had a population of 80,000
adults without a diagnosis of Type II diabetes. In 2023, 24 new diabetes cases were
identified in adults in City ABC, so the event count (N) is 24, and the exposure (E)
is 80,000 person-years (because we are counting 80,000 persons for one year).
Even though 24 is not that large, let’s use this example to demonstrate calculating
a CI for R. The incidence rate (R) is N /E , which is 24 per 80,000 person-years, or
30 per 100,000 person-years. How precise is this incidence rate?
To answer this, first, you should find the confidence limits for N. Using the
approximate formula, the 95 percent CI around the event count of 24 is
24 1.96 24 to 24 + 1.96 24 , or 14.4 to 33.6 events. Next, you divide the lower and
upper confidence limits of N by the exposure using these formulas: 14.4/80,000 =
0.00018 for the lower limit, and 33.6/80,000 = 0.00042 for the upper limit. Finally,
you can express these limits as 18.0 to 42.0 events per 100,00 person-years — the
CI for the incidence rate. Your interpretation would be that City ABC’s 2023 inci-
dence rate for Type II diabetes in adults was 30.0 (95 percent CI 18.0 to 42.0) per
100,000 person-years.
Comparing incidences with the rate ratio

When comparing incidence rates between two populations, you should calculate a
rate ratio (RR) by dividing one incidence rate by the other. So for two groups with
event counts N 1 and N 2, exposures E 1 and E 2, and incidence rates R1 and R2, respec-
tively, you calculate the RR for Group 2 relative to Group 1 as a reference, like this:
R2 N 2 /E 2
RR
R1 N 1 /E 1

Let’s revisit the example of 2023 incidence of Type II diabetes in adults in City XYZ
compared to City ABC. For City XYZ, you have N1 = 30 and E1 = 300,000. For City
ABC, you have N1 = 24 and E2 = 80,000. The RR for City ABC relative to City XYZ is
RR 24 / 80, 000 / 30/ 300, 000 , or 3.0, indicating that City ABC has three times the
adult Type II diabetes incidence in 2023 compared to City XYZ. You could calculate
the difference R2 R1 between two incidence rates if you wanted to, but in epi-
demiology, RRs are used much more often than rate differences.
Calculating confidence intervals

for a rate ratio
Whenever you report an RR you’ve calculated, you should also indicate how pre-
cise it is. The exact calculation of a CI around RR is quite difficult, but if your
observed event counts are large enough (meaning ≥ 10), then the following
approximate formula for the 95 percent CI around an RR works reasonably well:
1/N 1 1/N 2
95% CI RR /Q to RR Q where Q e 1.96 .
For other confidence levels, you can replace the 1.96 in the Q formula with the
appropriate critical z value for the normal distribution.
So, for the 2023 adult Type II diabetes example, you would setN 1 30, N 2 24 , and
1.96 1/ 24 1/ 30
RR = 3.0. The equation would be Q e , so the 95 percent lower and
upper confidence limits would be 3.0 /1.71 and 3.0 1.71, meaning the CI of the RR
would be from 1.75 to 5.13. You would interpret this by saying that that 2023 RR for
adult Type II diabetes incidence is 3.0 times the rate in City ABC compared to City
XYZ (95 percent CI 1.75 to 5.13).
Comparing two event rates

The examples in this chapter have compared incidence (or event) rates of adult
Type II diabetes in 2023 between City XYZ and City ABC. These two event rates are
represented as R1for City XYZ, andR2 for City ABC. They are based on City XYZ hav-
ing an N 1 of 30 events and City ABC having an N 2 of 24 events, and on exposures
E 1 and E 2 for City XYZ and City ABC, respectively. The difference in event rates
between City XYZ and City ABC can be tested for significance by calculating the
95 percent CI around the RR, and observing whether that CI includes the value
of 1.0. Because the RR is a ratio, having 1.0 included in the CI indicates that
City XYZ’s and City ABC’s rates could be identical. If the 95 percent CI around the

RR includes 1, the RR isn’t statistically significantly different from 1, so the two
rates aren’t significantly different from each other (assuming α = 0.05). But if the
95 percent CI is either entirely above or entirely below 1.0, the RR is statistically
significantly different from 1, so the two rates are significantly different from
each other (assuming α = 0.05).
For the City ABC and City XYZ adult Type II diabetes 2023 rate comparison, the
observed RR was 3.0, with a 95 percent confidence interval of 1.75 to 5.13. This CI
does not include 1.0 — in fact, it is entirely above 1.0. So, the RR is significantly
greater than 1, and you would conclude that City ABC has a statistically significantly
higher adult Type II diabetes incidence rate than City XYZ (assuming α = 0.05).
Comparing two event counts

with identical exposure
If — and only if — the two exposures (E 1 and E 2) are identical, there’s an extremely
simple rule for testing whether two event counts (N 1 and N 2) are significantly dif-
2
N1 N 2
ferent from each other at the level of α = 0.05: If 4 , then the Ns are
N1 N 2
statistically significantly different (at α = 0.05).
To interpret the formula into words, if the square of the difference is more than
four times the sum, then the event counts are statistically significantly different
at α = 0.05. The value of 4 in this rule approximates 3.84, the chi-square value
corresponding to p = 0.05.
Imagine you learned that in City XYZ, there were 30 fatal car accidents in 2022. In
the following year, 2023, you learned City XYZ had 40 fatal car accidents. You may
wonder: Is driving in City XYZ getting more dangerous every year? Or was the observed
increase from 2022 to 2023 due to random fluctuations? Using the simple rule, you
2
can calculate 30 – 40 / 30 40 100 /70 1.4, which is less than 4. Having
30 events — which in this case are fatal car accidents — isn’t statistically signifi-
cantly different from having 40 events in the same time period. As you see from
the result, the increase of 10 in one year is likely statistical noise. But had
the number of events increased more dramatically — say from 30 to 50 events —
the increase would have been statistically significant. This is because
2
30 50 / 30 50 400/ 80 5.0, which is greater than 4.

Estimating the Required Sample Size
As in all sample-size calculations, you need to specify the desired statistical power
and the α level of the test. Let’s set power to 80 percent and α to 0.05, as these are
common settings. When comparing event rates (R1 and R2) between two groups
with R1 as the reference group, you must also specify:
» The expected rate in the reference group (R1)
» The effect size of importance, expressed as the rate ratio RR R2 /R1
» The expected ratio of exposure in the two groups E 2 /E 1
For example, suppose that you’re designing a study to test whether rotavirus gas-
troenteritis has a higher incidence in City XYZ compared to City ABC. You’ll enroll
an equal number City XYZ and City ABC residents, and follow them for one year to
see whether they get rotavirus. Suppose that the one-year incidence of rotavirus
in City XYZ is 1 case per 100 person-years (an incidence rate of 0.01 case per
patient-year, or 1 percent per year). You want to have an 80 percent likelihood of
getting a statistically significant result assuming p = 0.05 (you want to set power
at 80 percent and α = 0.05). When comparing the incidence rates, you are only
concerned if they differ by more than 25 percent, which translates to a RR of 1.25.
This means you expect to see 0.01 × 1.25 = 0.0125 cases per patient-year in City ABC.
If you want to use G*Power to do your power calculation (see Chapter 4), under
Test family, choose z tests for population-level tests. Under Statistical test, choose
Proportions: Difference between two independent proportions because the two rates are
independent. Under Type of power analysis, choose A priori: Compute required
sample size – given α, power and effect size, and under the Input Parameters section,
choose two tails so you can test if one is higher or lower than the other. Set Propor-
tion p1 to 0.01 (to represent City XYZ’s incidence rate), Proportion p2 to 0.0125 (to
represent City ABC’s expected incidence rate), α err prob (α) to 0.05, and Power
(1-β err prob) (power) to 0.8 for 80 percent, and keep a balanced Allocation ration
N2/N1 of 1. After clicking Calculate, you’ll see you need at least 27,937 person-
years of observation in each group, meaning observing 57,000 participants over a
one-year study. The shockingly large target sample size illustrates a challenge
when studying incidence rates of rare illnesses.

5
Looking for
Relationships
with Correlation
and Regression
IN THIS PART . . .
Understand correlation, which is the strength and

direction of the relationship between two variables,
and regression, which is a set of techniques for
describing the relationship between two variables.
Get a handle on straight-line regression, which is the

simplest kind of regression.
Do multiple regression analysis, which is regression

when there’s more than one predictor variable.
Find out about other useful kinds of regression you

encounter in biological research, such as Poisson
regression and nonlinear regression.
IN THIS CHAPTER
» Getting a handle on correlation

analysis
» Understanding the many kinds of

regression analysis
Chapter 15
Introducing Correlation
and Regression
C
orrelation, regression, curve-fitting, model-building — these terms all
describe a set of general statistical techniques that deal with the relation-
ships among variables. Introductory statistics courses usually present only
the simplest form of correlation and regression, equivalent to fitting a straight
line to a set of data. But in the real world, correlations and regressions are seldom
that simple — statistical problems may involve more than two variables, and the
relationship among them can be quite complicated.
The words correlation and regression are often used interchangeably, but they refer
to two different concepts:
» Correlation refers to the strength and direction of the relationship between

two variables, or among a group of variables.
» Regression refers to a set of techniques for describing how the values of a

variable or a group of variables may cause, predict, or be associated with the
values of another variable.
CHAPTER 15 Introducing Correlation and Regression 201

You can study correlation and regression for many years and not master all of it.
In this chapter, we cover the kinds of correlation and regression most often
encountered in biological research and explain the differences between them. We
also explain some terminology used throughout Parts 5 and 6.
Correlation: Estimating How Strongly

Two Variables Are Associated
Correlation refers to the extent to which two variables are related. In the following
sections, we describe the Pearson correlation coefficient and discuss ways to analyze
correlation coefficients.
Lining up the Pearson correlation

coefficient
The Pearson correlation coefficient is represented by the symbol r and measures
the extent to which two variables (X and Y) tend to lie along a straight line when
graphed. If the variables have no relationship, r will be 0, and the points will be
scattered across the graph. If the relationship is perfect the points will lie exactly
along a straight line, and r will either be:
» 1: If the variables have a direct or positive relationship, meaning when one

goes up, the other goes up, or
» –1: If the variables have an inverse or negative relationship, meaning when one
goes up, the other goes down
Correlation coefficients can be positive (indicating upward-sloping data) or nega-

tive (indicating downward-sloping data). Figure 15-1 shows what several differ-
ent values of r look like.
Note: The Pearson correlation coefficient measures the extent to which the points
lie along a straight line. If your data follow a curved line, the r value may be low or
zero, as shown in Figure 15-2. All three graphs in Figure 15-2 have the same
amount of random scatter in the points, but they have quite different r values.
Pearson r is based on a straight-line relationship and is too small (or even zero) if
the relationship is nonlinear. So, you shouldn’t interpret r 0 as evidence of lack
of association or independence between two variables. It could indicate only the
lack of a straight-line relationship between the two variables.
202 PART 5 Looking for Relationships with Correlation and Regression

FIGURE 15-1:
100 data points,
with varying
degrees of
correlation.
FIGURE 15-2:
Pearson r is
based on a
straight-line
relationship.
Analyzing correlation coefficients

In the following sections, we show the common kinds of statistical analyses that
you can perform on correlation coefficients.
Testing whether r is statistically significantly

different from zero
Before beginning your calculations for correlation coefficients, remember that the
data used in a correlation — the “ingredients” to a correlation — are the values
of two variables referring to the same experimental unit. An example would be
measurements of height (X) and weight (Y) in a sample of individuals. Because
your raw data (the X and Y values) always have random fluctuations due to either
sampling error or measurement imprecision, a calculated correlation coefficient is
also subject to random fluctuations.
Even when X and Y are completely independent, your calculated r value is almost
never exactly zero. One way to test for a statistically significant association
between X and Y is to test whether r is statistically significantly different from zero
by calculating a p value from the r value (see Chapter 3 for a refresher on p values).
The correlation coefficient has a strange sampling distribution, so it is not useful

for statistical testing. Instead, the quantity t can be calculated from the observed

correlation coefficient r, based on N observations, by the formula
t r/ 1 r 2 / N 2 . Because t fluctuates in accordance with the Student t dis-
tribution with N 2 degrees of freedom (df), it is useful for statistical testing (see
Chapter 11 for more about t).
For example, if r 0.500 for a sample of 12 participants, then t 0.5/

1 0.5 2 / 12 2 , which works out to t 1.8257, with 10 degrees of freedom. You
can use the online calculator at https://statpages.info/pdfs.html and calcu-
late the p by entering the t and df values. You can also do this in R by using
the code:
2 * pt(q = 1.8257, df = 10, lower.tail = FALSE).
Either way, you get p 0.098, which is greater than 0.05. At α = 0.05, the r value
of 0.500 is not statistically significantly different from zero (see Chapter 12 for
more about α).
How precise is an r value?

You can calculate confidence limits around an observed r value using a somewhat
roundabout process. The quantity z, calculated by the Fisher z transformation
1
z log 1 r / 1 r , is approximately normally distributed with a standard
2
deviation of 1/ N 3 . Therefore, using the formulas for normal-based confidence
intervals (see Chapter 10), you can calculate the lower and upper 95 percent con-
fidence limits around z : z Lower z 1.96 / N 3 and Z Upper z 1.96 / N 3 . You
can turn these into the corresponding confidence limits around r by the reverse of
the z transformation: r e 2 x 1 / e 2 z 1 for z z Lower and z z Upper .
Here are the steps for calculating 95 percent confidence limits around an observed
r value of 0.05 for a sample of 12 participants (N = 12):
1. Calculate the Fisher z transformation of the observed r value:
1
z log 1 0.5 / 1 0.5 0.549
2
2. Calculate the lower and upper 95 percent confidence limits for z:
z Lower 0.549 1.96 / 12 3 0.104

z Upper 0.549 1.96 / 12 3 1.203

3. Calculate the lower and upper 95 percent confidence limits for r:
( 0.104 ) ( 0.104 )
rLower (e2 1)/( e 2 1) 0.104
1.203 1.203
rUpper (e2 1)/( e 2 1) 0.835
Notice that the 95 percent confidence interval goes from –0.104 to 0.835 , a range
that includes the value zero. This means that the true r value could indeed be zero,
which is consistent with the non-significant p value of 0.098 that you obtained
from the significance test of r in the preceding section.
Determining whether two r values are

statistically significantly different
Suppose that you have two correlation coefficients and you want to test whether
they are statistically significantly different. It doesn’t matter whether the two r
values are based on the same variables or are from the same group of participants.
Imagine that a significance test for comparing two correlation coefficient values
(which we will call r1 and r2 ) that were obtained from N 1 and N 2 participants,
respectively. You can utilize the Fisher z transformation to get z 1 and z 2 . The dif-
ference ( z 1 z 2) has a standard error (SE) of SE z 2 z1 1/ N 1 3 1/ N 2 3 .
You obtain the test statistic for the comparison by dividing the difference by its
SE. You can convert this to a p value by referring to a table (or web page) of the
normal distribution.
For example, if you want to compare an r1 value of 0.4 based on an N1 of 100 par-
ticipants with an r2 value of 0.6 based on an N2 of 150 participants, you perform the
following steps:
1. Calculate the Fisher z transformation of each observed r value:
1
z1 log 1 0.4 / 1 0.4 0.424
2
1
z2 log 1 0.6 / 1 0.6 0.693
2
2. Calculate the (z 2 z1) difference:
0.693 0.424 0.269
3. Calculate the SE of the (z 2 z1) difference:
SE z 2 z1 1/ 100 3 1/ 150 3 0.131

4. Calculate the test statistic:
0.269/0.131 2.05
5. Look up 2.05 in a normal distribution table or web page such as https://

statpages.info/pdfs.html (or edit and run the R code provided earlier
in “Testing whether r is statistically significantly different from zero”),
and observe that the p value is 0.039 for a two-sided test.
A two-sided test is used when you’re interested in knowing whether either r is

larger than the other. The p value of 0.039 is less than 0.05, meaning that the
two correlation coefficients are statistically significantly different from each
other at α = 0.05.
Determining the required sample

size for a correlation test
If you are planning to conduct a study where the outcome is a correlation between
two variables designated X and Y, you need to be sure to enroll a large enough
sample so that if the correlation is indeed statistically significant, you have enough
sample for r to show it. As described in Chapter 11 with the t test and the ANOVA,
the sample size can be estimated through one big equation, where you plug the
estimated effect size along with the α and power you select into an equation, and
calculate the sample size (see Chapter 3 for the scoop on effect size and selecting
α and power).
For a sample-size calculation for a correlation coefficient, you need to plug in the
following design parameters of the study into the equation:
» The desired α level of the test: The p value that’s considered significant
when you’re testing the correlation coefficient (usually 0.05).
» The desired power of the test: The probability of rejecting the null hypoth-
esis if the alternative hypothesis is true (usually set to 0.8 or 80 percent).
» The effect size of importance: The smallest r value that is considered

practically important, or clinically significant. If the true r is less than this value,
then you don’t care whether the test comes out significant, but if r is greater
than this value, you want to get a significant result.
It may be challenging to select an effect size, and context matters. One approach
would be to start by referring to Figure 15-1 to select a potential effect size, then
do a sample-size calculation and see the result. If the result requires more sam-
ples than you could ever enroll, then try making the effect size a little larger and
redoing the calculation until you get a more reasonable answer.

You can use software like G*Power (see Chapter 4) to perform the sample-size
calculation. If you use G*Power:
1. Under Test Family, choose t-tests.
2. Under Statistical Test, choose Correlation: Point Biserial model.
3. Under Type of Power Analysis, choose A Priori: Compute required sample

size – given α, power, and effect size.
4. Under Tail(s), because either r could be greater, choose two.
5. Under Effect Size, which is the expected difference between r1 and r2, enter the
effect size you expect.
6. Under α err prob, enter 0.05.
7. Under Power (1-β err prob), enter 0.08.
8. Click Calculate.
The answer will appear under Total sample size. As an example, if you enter these
parameters and an effect size of 0.02, the total sample size will be 191.
Regression: Discovering the Equation

that Connects the Variables
As described earlier, correlation assesses the relationship between two continuous
numeric variables (as compared to categorical variables, as described in
Chapter 8). This relationship can also be evaluated with regression analysis to
provide more information about how these two variables are related. But perhaps
more importantly, regression is not limited to continuous variables, nor is it
limited to only two variables. Regression is about developing a formula that
explains how all the variables in the regression are related. In the following sec-
tions, we explain the purpose of regression analysis, identify some terms and
notation typically used, and describe common types of regression.
Understanding the purpose

of regression analysis
You may wonder how fitting a formula to a set of data can be useful. There are
actually many uses. With regression, you can

» Test for a significant association or relationship between two or more
variables. The process is similar to correlation, but is more generalized to
produce a unique equation or formula relating to the variables.
» Get a compact representation of your data. A well-fitting regression model

succinctly summarizes the relationships between the variables in your data.
» Make precise predictions, or prognoses. With a properly fitted survival

function (see Chapter 23), you can generate a customized survival curve for a
newly diagnosed cancer patient based on that patient’s age, gender, weight,
disease stage, tumor grade, and other factors to predict how long they will
live. A bit morbid, perhaps, but you could certainly do it.
» Do mathematical manipulations easily and accurately on a fitted function

that may be difficult or inaccurate to do graphically on the raw data. These
include making estimates within the range of the measured values (called
interpolation) as well as outside the measured values (called extrapolation, and
considered risky). You may also want to smooth the data, which is described in
Chapter 19.
» Obtain numerical values for the parameters that appear in the regres-
sion model formula. Chapter 19 explains how to make a regression model
based on a theoretical rather than known statistical distribution (described in
Chapter 3). Such a model is used to develop estimates like the ED50 of a drug,
which is the dose that produces one-half the maximum effect.
Talking about terminology and

mathematical notation
A regression model is a formula that describes how one variable, the dependent vari-
able, depends on one or more other variables, and on one or more parameters.
(While it is technically possible to have more than one dependent variable in a
model, a discussion of this type of regression is outside the scope of this book.)
The dependent variable is also called the outcome, and the other variables are
called independent variables or predictors. Parameters refer to the other terms that
appear in the formula that make the function come as close as possible to the
observed data which are determined by the statistical software you are using.
If you have only one independent variable, it’s often designated by X, and the
dependent variable is designated by Y. If you have more than one independent
variable, variables are usually designated by letters toward the end of the alphabet
(W, X, Y, Z). Parameters are often designated by letters toward the beginning of the

alphabet (a, b, c, d). There’s no consistent rule regarding uppercase versus lower-
case letters.
Sometimes a collection of predictor variables is designated by a subscripted vari-

able ( X 1 , X 2 and so on) and the corresponding coefficients by another subscripted
variable (b1 , b 2, and so on).
In mathematical texts, you may see a regression model with three predictors
written in one of several ways, such as
»Z a bX cY dV (different letters for each variable and parameter)
»Y b0 b1 X 1 b2 X 2 (using a general subscript-variable notation)
In practical work, using the actual names of the variables from your data and
using meaningful terms for parameters is easiest to understand and least error-
prone. For example, consider the equation for the first-order elimination of an
injected drug from the blood, Conc Conc0 e ke Time. This form, with its short but
meaningful names for the two variables, Conc (blood concentration) and Time
(time after injection), and the two parameters, Conc 0 (concentration at Time 0)
and ke (elimination rate constant), would probably be more meaningful to a reader
than Y a e -b X
.
Classifying different kinds of regression

You can classify regression on the basis of
» How many predictors or independent variables appear in the model
» The type of data of the outcome variable
» What mathematical form to which the data appear to conform
There are different terms for different types of regression. In this book, we refer
to regression models with one predictor in the model as simple regression, or uni-
variate regression. We refer to regression models with multiple predictors as mul-
tivariate regression.
In the next section, we explain how the type of outcome variable determines
which regression to select, and after that, we explain how the mathematical form
of the data influences the type of regression you choose.

Examining the outcome variable’s type of data
Here are the different regressions we cover in this book by type of outcome
variable:
» Ordinary regression (also called linear regression) is used when the outcome
is a continuous variable whose random fluctuations are governed by the
normal distribution (see Chapters 16 and 17).
» Logistic regression is used when the outcome variable is a two-level or

dichotomous variable whose fluctuations are governed by the binomial
distribution (see Chapter 18).
» Poisson regression is used when the outcome variable is the number of

occurrences of a sporadic event whose fluctuations are governed by the
Poisson distribution (see Chapter 19).
» Survival regression when the outcome is a time to event, often called a survival
time. Part 6 covers the entire topic of survival analysis, and Chapter 23 focuses
on regression.
Figuring out what kind of function is being fitted

Another way to classify different types of regression analysis is according to
whether the mathematical formula for the model is linear or nonlinear in the
parameters.
In a linear function, you multiply each predictor variable by a parameter and then
add these products to give the predicted value. You can also have one more param-
eter that isn’t multiplied by anything — it’s called the constant term or the inter-
cept. Here are some linear functions:
»Y a bX
»Y a bX cX 2 dX 3
»Y a bX cLog (W ) dX /Cos( Z )
In these examples, Y is the dependent variable or the outcome, and X, W, and Z are
the independent variables or predictors. Also, a, b, c, and d are parameters.

The predictor variables can appear in a formula in nonlinear form, like squared or
cubed, inside functions like Log and Sin, and multiplied by each other. But as long
as the coefficients appear only in a linear way, the function is still considered lin-
ear in the parameters. By that, we mean each coefficient is multiplied by a term
involving predictor variables, with the terms added together in a linear equation.
A nonlinear function is anything that’s not a linear function. For example:
Y a/ b e –c X
is nonlinear in the parameters, because the parameter b is in the denominator of

a fraction, and the parameter c is in an exponent. The parameter a appears in a
linear form, but if any of the parameters appear in a nonlinear way, the function
is said to be nonlinear in the parameters. Nonlinear regressions are covered in
Chapter 19.

IN THIS CHAPTER
» Determining when to use straight-

line regression
» Running a straight-line regression

and making sense of the output
» Examining results for issues and

problems
» Estimating needed sample size for

straight-line regression
Chapter 16
Getting Straight Talk on
Straight-Line Regression
C
hapter 15 refers to regression analyses in a general way. This chapter
focuses on the simplest type of regression analysis: straight-line regression.
You can visualize it as fitting a straight line to the points in a scatter plot
from a set of data involving just two variables. Those two variables are generally
referred to as X and Y. The X variable is formally called the independent variable (or
the predictor or cause). The Y variable is called the dependent variable (or the out-
come or effect).
Knowing When to Use Straight-Line

Regression
You may see straight-line regression referred to in books and articles by several
different names, including linear regression, simple linear regression, linear univariate
regression, and linear bivariate regression. This abundance of references can be con-
fusing, so we always use the term straight-line regression.
CHAPTER 16 Getting Straight Talk on Straight-Line Regression 213

Straight-line regression is appropriate when all of these things are true:
» You’re interested in the relationship between two — and only two —

numerical variables. At least one of them must be a continuous variable
that serves as the dependent variable (Y).
» You’ve made a scatter plot of the two variables and the data points seem to
lie, more or less, along a straight line (as shown in Figures 16-1a and 16-1b).
You shouldn’t try to fit a straight line to data that appears to lie along a curved
line (as shown in Figures 16-1c and 16-1d).
» The data points appear to scatter randomly around the straight line over the
entire range of the chart, with no extreme outliers (as shown in Figures 16-1a
and 16-1b).
FIGURE 16-1:
Straight-line
regression is
appropriate for
both strong and
weak linear
relationships
(a and b), but not
for nonlinear
(curved-line)
relationships
(c and d).

You should proceed with straight-line regression when one or more of the follow-
ing are true:
» You want to test whether there’s a statistically significant association between

the X and Y variables.
» You want to know the value of the slope and/or intercept (also referred to as
the Y intercept) of a line fitted through the X and Y data points.
» You want to be able to predict the value of Y if you know the value of X.
Understanding the Basics of

The formula of a straight line can be written like this: Y a bX . This formula
breaks down this way:
» Y is the dependent variable (or outcome).
» X is the independent variable (or predictor).
» a is the intercept, which is the value of Y when X 0.
» b is the slope, which is the amount Y changes when X increases by 1.
In straight-line regression, our goal is to develop the best-fitting line for our data.
Using least-squares as a guide, the best-fitting line through a set of data is the
one that minimizes the sum of the squares (SSQ) of the residuals. Residuals are the
vertical distances of each point from the fitted line, as shown in Figure 16-2.
FIGURE 16-2:
On average, a
good-fitting line
has smaller
residuals than a
bad-fitting line.

For curves, finding the best-fitting curve is a very complicated mathematical
problem. What’s nice about the straight-line regression is that it’s so simple that
you can calculate the least-squares parameters from explicit formulas. If you’re
interested (or if your professor insists that you’re interested), we present a gen-
eral outline of how those formulas are derived.
Think of a set of data containing X i and Yi , in which i is an index that identifies

each observation in the set, as described in Chapter 2. From those data, SSQ can be
calculated like this:
SSQ ( a bX i Yi ) 2
i
If you’re good at first-semester calculus, you can find the values of a and b that
minimize SSQ by setting the partial derivatives of SSQ with respect to a and b
equal to 0. If you stink at calculus, trust that this leads to these two simultaneous
equations:
a( N ) b( Y ) ( Y )
a( X ) b( X 2 ) ( XY )
where N is the number of observed data points.
These equations can be solved for a and b:
( Y )( X 2 ) ( X )(( XY )
a
( N )(( X 2 ) ( X ) 2
( XY ) ( a )( X )
b
( X 2)
See Chapter 2 if you don’t feel comfortable reading the mathematical notations or
expressions in this section.
Running a Straight-Line Regression

Even if it is possible, it is not a good idea to calculate regressions manually or with
a calculator. You’ll go crazy trying to evaluate all those summations and other
calculations, and you’ll almost certainly make a mistake somewhere in your
calculations.

Fortunately, most statistical packages can perform a straight-line regression.
Microsoft Excel has built-in functions for calculating the slope and intercept of
the least-squares straight line. You can also find straight-line regression web
pages (several are listed at https://statpages.info). If you use R, you can
explore using the lm() command. (See Chapter 4 for an introduction to statistical
software.) In the following sections, we list the basic steps for running a straight-
line regression, complete with an example.
Taking a few basic steps

The exact steps you take to run a straight-line regression depend on what soft-
ware you’re using, but here’s the general approach:
1. Structure your data into the proper form.
Usually, the data consist of two columns of numbers, one representing the
independent variable and the other representing the dependent variable.
2. Tell the software which variable is the independent variable and which
one is the dependent variable.
Depending on the software, you may type in the variable names, or pick them
from a menu or list in your file.
3. If the software offers output options, tell it that you want it to output
these results:
• Graphs of observed and calculated values
• Summaries and graphs of the residuals
• Regression table
• Goodness-of-fit measures
4. Execute the regression in the software (tell it to run the regression).
Then go look for the output. You should see the output you requested in
Step 3.
Walking through an example

To see how to run a straight-line regression and interpret the output, we use the
following example throughout the rest of this chapter.
Consider how blood pressure (BP) is related to body weight. It may be reasonable
to suspect that people who weigh more have higher BP. If you test this hypothesis

on people and find that there really is an association between weight and BP, you
may want to quantify that relationship. Maybe you want to say that every extra
kilogram of weight tends to be associated with a certain amount of increased
BP. Even though you are testing an association, the reality is that you believe that
as people weigh more, it causes their BP to go up — not the other way around. So,
you would characterize weight as the independent variable (X), and BP as the
dependent variable (Y). The following sections take you through the steps of gath-
ering data, creating a scatter plot, and interpreting the results.
Gathering the data

Suppose that you recruit a sample of 20 adults from a particular clinical popula-
tion to participate in your study (see Chapter 6 for more on sampling). You weigh
them and measure their systolic BP (SBP) as a measure of their BP. Table 16-1
shows a sample of weight and SBP data from 20 participants. Weight is recorded
in kilograms (kg), and SBP is recorded in the strange-sounding units of millime-
ters of mercury (mmHg).
TABLE 16-1 Weight and Blood Pressure Data

Participant Study ID Body Weight (kg) SBP (mmHg)
1 74.4 109
2 85.1 114
3 78.3 94
4 77.2 109
5 63.8 104
6 77.9 132
7 78.9 127
8 60.9 98
9 75.6 126
10 74.5 126
11 82.2 116
12 99.8 121
13 78.0 111

Participant Study ID Body Weight (kg) SBP (mmHg)
14 71.8 116
15 90.2 115
16 105.4 133
17 100.4 128
18 80.9 128
19 81.8 105
20 109.0 127
Creating a scatter plot

It’s not easy to identify patterns and trends between weight and SBP by looking at
a table like Table 16-1. You get a much clearer picture of how the data are related
if you make a scatter plot. You would place weight, the independent variable, on
the X axis, and SBP, the dependent variable, on the Y axis, as shown in
Figure 16-3.
FIGURE 16-3:
Scatter plot of
SBP versus body
weight.
Examining the results

In Figure 16-3, you can see the following pattern:
» Low-weight participants have lower SBP, which is represented by the

points near the lower-left part of the graph.

» Higher-weight participants have higher SBP, which is represented by the
points near the upper-right part of the graph.
You can also tell that there aren’t any higher-weight participants with a very low
SBP, because the lower-right part of the graph is rather empty. But this relation-
ship isn’t completely convincing, because several participants in the lower weight
range of 70 to 80 kg have SBPs over 125 mmHg.
A correlation analysis (described in Chapter 15) will tell you how strong this type
of association is, as well as its direction (which is positive in this case). The results
of a correlation analysis help you decide whether or not the association is likely
due to random fluctuations. Assuming it is not, proceeding to a regression analy-
sis provides you with a mathematical formula that numerically expresses the
relationship between the two variables (which are weight and SBP in this example).
Interpreting the Output of

In the following sections, we walk you through the printed and graphical output
of a typical straight-line regression run. Its looks will vary depending on your
software. The output in this chapter was generated using R (see Chapter 4 to get
started with R). But regardless of the software you use, you should be able to pro-
gram the regression so the following elements appear on your output:
» A statement of what you asked the program to do (the code you ran for the
regression)
» A summary of the residuals, including graphs that display the residuals and
help you assess whether they’re normally distributed
» The regression table (providing the results of the regression model)
» Measures of goodness-of-fit of the line to the data
Seeing what you told the program to do

In the example data set, the SBP variable is named BPs, and the weight variable is
named wgt. In Figure 16-4, the first two lines produced by the statistical software
reprint the code you ran. The code says that you wanted to fit a linear formula
where the software estimates the parameters to the equation SBP = weight based on
your observed SBP and weight values. The code used was lm(formula = BPs ~ Wgt).

FIGURE 16-4:
Sample straight-
line regression
output from R.
The actual equation of the straight line is not SBP = weight, but more accurately
SBP = a + b × weight, with the a (intercept) and b (slope) parameters having been
left out of the model. This is a reminder that although the goal is to evaluate the
SBP = weight model conceptually, in reality, this relationship will be numerically
different with each data set we use.
Evaluating residuals
The residual for a point is the vertical distance of that point from the fitted line.
It’s calculated as Residual Y ( a b X ) , where a and b are the intercept and
slope of the fitted straight line, respectively.
Most regression software outputs several measures of how the data points scatter
above and below the fitted line, which provides an idea of the size of the residuals
(see “Summary statistics for the residuals” for how to interpret these measures).
The residuals for the sample data are shown in Figure 16-5.
FIGURE 16-5:
Scattergram of
SBP versus
weight, with the
fitted straight line
and the residuals
of each point
from the line.

Summary statistics for the residuals
If you read about summarizing data in Chapter 9, you know that the distribution
of values from a numerical variable are reported using summary statistics, such as
mean, standard deviation, median, minimum, maximum, and quartiles. Summary
statistics for residuals are what you should expect to find in the residuals section
of your software’s output. Here’s what you see in Figure 16-4 at the top under
Residuals:
» The minimum and maximum values: These are labeled as Min and Max,
respectively, and represent the two largest residuals, or the two points that lie
farthest away from the least-squares line in either direction. The minimum is
negative, indicating it is below the line, while the positive maximum is above
the line. The minimum is almost 21 mmHg below the line, while the maximum
lies about 17 mmHg above the line.
» The first and third quartiles: These are labeled IQ and 3Q on the output.
Looking under IQ, which is the first quartile, you can tell that about 25 percent
of the data points (which would be 5 out of 20) lie more than 4.7 mmHg below
the fitted line. For the third quartile results, you see that another 25 percent
lie more than 6.5 mmHg above the fitted line. The remaining 50 percent of the
points lie within those two quartiles.
» The median: Labeled Median on the output, a median of –3.4 tells you that half
of the residuals, which is 10 of the 20 data points, are less than –3.4, and half are
greater than –3.4. The negative sign means the median lies below the fitted line.
Note: The mean isn’t included in these summary statistics because the mean of
the residuals is always exactly 0 for any kind of regression that includes an
intercept term.
The residual standard error, often called the root-mean-square (RMS) error in regres-
sion output, is a measure of how tightly or loosely the points scatter above or
below the fitted line. You can think of it as the standard deviation (SD) of the resid-
uals, although it’s computed in a slightly different way from the usual SD of a set
of numbers. RMS uses N 2 instead of N 1 in the denominator of the SD
formula. At the bottom of Figure 16-4, Residual standard error is expressed as
9.838 mmHg. You can think of it as another summary statistic for residuals
Graphs of the residuals

Most regression programs will produce different graphs of the residuals if
requested in code. You can use these graphs to assess whether the data meet the
criteria for executing a least-squares straight-line regression. Figure 16-6 shows
two of the more common types of residual graphs. The one on the left is called a
residuals versus fitted graph, and the one on the right is called a normal Q-Q graph.

FIGURE 16-6:
The residuals
versus fitted (a)
and normal (b)
Q-Q graphs help
you determine
whether your
data meets the
requirements for
straight-line
regression.
As stated at the beginning of this section, you calculate the residual for each point
by subtracting the predicted Y from the observed Y. As shown in Figure 16-6, a
residuals versus fitted graph displays the values of the residuals plotted along the Y
axis and the predicted Y values from the fitted straight line plotted along the X
axis. A normal Q-Q graph shows the standardized residuals, which are the residuals
divided by the RMS value, along the Y axis, and theoretical quantiles along the X
axis. Theoretical quantiles are what you’d expect the standardized residuals to be if
they were exactly normally distributed.
Together, the two graphs shown in Figure 16-6 provide insight into whether your
data conforms to the requirements for straight-line regression:
» Your data must lie above and below the line randomly across the whole
range of data.
» The average amount of scatter must be fairly constant across the whole range
of data.
» The residuals should be approximately normally distributed.
You need years of experience examining residual plots before you can interpret
them confidently, so don’t feel discouraged if you can’t tell whether your data
complies with the requirements for straight-line regression from these graphs.
Here’s how we interpret them, allowing for other biostatisticians to disagree:
» We read the residuals versus fitted chart in Figure 16-6 to show the points
lying equally above and below the fitted line, because this appears true
whether you’re looking at the left, middle, or right part of the graph.

» Figure 18-6 shows most of the residuals lie within 10 mmHg of the line, but
several larger residuals appear to be where the SBP is around 115 mmHg.
We think this is a little suspicious. If you see a pattern like this, you should
examine your raw data to see whether there are unusual values associated
with these particular participants.
» If the residuals are normally distributed, then in the normal Q-Q chart in
Figure 16-6, the points should lie close to the dotted diagonal line, and
shouldn’t display any overall curved shape. Our opinion is that the points
follow the dotted line pretty well, so we’re not concerned about lack of
normality in the residuals.
Making your way through

the regression table
When doing regression, it is common to focus on the table of regression coeffi-
cients. It’s likely where you look first when interpreting your results, and where
you concentrate most of your attention. When you run a straight-line regression,
statistical programs typically produce a table of regression coefficients that looks
much like the one in Figure 16-4 under the heading Coefficients.
For straight-line regression, the coefficients table has two rows that correspond
with the two parameters of the straight line:
» The intercept row: This row is labeled (Intercept) in Figure 16-4, but can be
labeled Intercept or Constant in other software.
» The slope row: This row is usually labeled with the name of the independent
variable in your data, so in Figure 16-4, it is named Wgt. It may be labeled Slope
in some programs.
The table has several different columns, depending on the software. In

Figure 16-4, the columns are Estimate, Std. Error, t value, and Pr (>|t|). How to inter-
pret the results from these columns is discussed in the next section.
The values of the coefficients (the

intercept and the slope)
The first column of the table of regression coefficients usually shows the values of
the slope and intercept of the fitted straight line (labeled Estimate in Figure 16-4).
Other column headings include Coefficient or the single letter B or C (in uppercase
or lowercase), depending on the software.

Looking at the rows in Figure 16-4, the intercept (labeled (Intercept)) is the pre-
dicted value of Y when X is equal to 0, and is expressed in the same units of mea-
surement as the Y variable. The slope (labeled Wgt) is the amount the predicted
value of Y changes when X increases by exactly one unit of measurement, and is
expressed in units equal to the units of Y divided by the units of X.
In the example shown in Figure 16-4, the estimated value of the intercept is
76.8602 mmHg, and the estimated value of the slope is 0.4871 mmHg/kg.
» The intercept value of 76.9 mmHg means that a person who weighs 0 kg is
predicted to have a SBP of about 77 mmHg. But nobody weighs 0 kg! The
intercept in this example (and in many straight-line relationships in biology)
has no physiological meaning at all, because 0 kg is completely outside the
range of possible human weights.
» The slope value of 0.4871 mmHg/kg does have a real-world meaning. It means
that every additional 1 kg of weight is associated with a 0.4871 mmHg increase
in SBP. If we multiply both estimates by 10, we could say that every additional
10 kg of body weight is associated with almost a 5 mmHg SBP increase.
The standard errors of the coefficients

The second column in the regression table often contains the standard errors of
the estimated parameters. In Figure 16-4, it is labeled Std. Error, but it could be
stated as SE or use a similar term. We use SE to mean standard error for the rest
of this chapter.
Because data from your sample always have random fluctuations, any estimate
you calculate from your data will be subject to random fluctuations, whether it is
a simple summary statistic or a regression coefficient. The SE of your estimate
tells you how precisely you were able to estimate the parameter from your data,
which is very important if you plan to use the value of the slope (or the intercept)
in a subsequent calculation.
Keep these facts in mind about SE:
» SEs always have the same units as the coefficients themselves. In the
example shown in Figure 16-4, the SE of the intercept has units of mmHg,
and the SE of the slope has units of mmHg/kg.
» Round off the estimated values. It is not helpful to report unnecessary

digits. In this example, the SE of the intercept is about 14.7, so you can say that
the estimate of the intercept in this regression is about 77 15 mmHg. In the
same way, you can say that the estimated slope is 0.49 0.18 mmHg/kg.

When reporting regression coefficients in professional publications, you may
state the SE like this: “The predicted increase in systolic blood pressure with
weight (±1 SE) was 0.49 0.18 mmHg/kg.”
If you know the value of the SE, you can easily calculate a confidence interval (CI)
around the estimate (see Chapter 10 for more information on CIs). These expres-
sions provide a very good approximation of the 95 percent confidence limits
(abbreviated CL), which mark the low and high ends of the CI around a regression
coefficient:
Lower 95% CL Coefficient 2 SE
Upper 95% CL Coefficient 2 SE
More informally, these are written as 95% CI coefficient 2 SE .
So, the 95 percent CI around the slope in our example is calculated as

0.49 2 0.176 , which works out to 0.49 0.35 , with the final confidence limits
of 0.14 to 0.84 mmHg. If you submit a manuscript for publication, you may express
the precision of the results in terms of CIs instead of SEs, like this: “The predicted
increase in SBP as a function of body weight was 0.49 mmHg/kg
95% CI : 0.14 0.84 .”
The Student t value

In most output, there is a column in the regression table that shows the ratio of
the coefficient divided by its SE. This column is labeled t value in Figure 16-4, but
it can be labeled t or other names. This column is not very useful. You can think of
this column as an intermediate quantity in the calculation of what you’re really
interested in, which is the p value for the coefficient.
The p value
A column in the regression tables (usually the last one) contains the p value,
which indicates whether the regression coefficient is statistically significantly
different from 0. In Figure 16-4, it is labeled Pr |t| , but it can be called a variety
of other names, including p value, p, and Signif.
In Figure 16-4, the p value for the intercept is shown as 5.49e 05, which is equal
to 0.0000549 (see the description of scientific notation in Chapter 2). Assuming
we set α at 0.05, the p value is much less than 0.05, so the intercept is statistically
significantly different from zero. But recall that in this example (and usually in
straight-line regression), the intercept doesn’t have any real-world importance.
It’s equals the estimated SBP for a person who weighs 0 kg, which is nonsensical,
so you probably don’t care whether it’s statistically significantly different from
zero or not.

But the p value for the slope is very important. Assuming α = 0.05, if it’s less than
0.05, it means that the slope of the fitted straight line is statistically significantly
different from zero. This means that the X and Y variables are statistically signifi-
cantly associated with each other. A p value greater than 0.05 would indicate that
the true slope could equal zero, and there would be no conclusive evidence for a
statistically significant association between X and Y. In Figure 18-4, the p value for
the slope is 0.0127, which means that the slope is statistically significantly differ-
ent from zero. This tells you that in your model, body weight is statistically sig-
nificantly associated with SBP.
If you want to test for a significant correlation between two variables at α = 05, you
can look at the p value for the slope of the least-squares straight line. If it’s less
than 0.05, then the X and Y variables are also statistically significantly correlated.
The p value for the significance of the slope in a straight-line regression is always
exactly the same as the p value for the correlation test of whether r is statistically
significantly different from zero, as described in Chapter 15.
Wrapping up with measures

of goodness-of-fit
The last few lines of output in Figure 16-4 contain several indicators of how
well the straight line represents the data. The following sections describe this part
of the output.
The correlation coefficient

Most straight-line regression programs provide the classic Pearson r correlation
coefficient between X and Y (see Chapter 15 for details). But the program may pro-
vide you the correlation coefficient in a roundabout way by outputting r 2 rather
than r itself. In Figure 16-4, at the bottom under Multiple R-squared, the r 2 is listed
as 0.2984. If you want Pearson r, just use Microsoft Excel or a calculator to take
square root of 0.2984 to get 0.546.
The r 2 is always positive, because square of any number is always positive. But the
correlation coefficient can be positive or negative, depending on whether the fit-
ted line slopes upward or downward. If the fitted line slopes downward, make
your r value negative.
Why did the program give you r 2 instead of r in the first place? It’s because r 2 is a
useful estimate called the coefficient of determination. It tells you what percent of
the total variability in the Y variable can be explained by the fitted line.
» An r 2 value of 1 means that the points lie exactly on the fitted line, with no
scatter at all.

» An r 2 value of 0 means that your data points are all over the place, with no
tendency at all for the X and Y variables to be associated.
» An r 2 value of 0.3 (as in this example) means that 30 percent of the variance in
the dependent variable is explainable by the independent variable in this
straight-line model.
Note: Figure 18-4 also lists the Adjusted R-squared at the bottom right. We talk
about the adjusted r 2 value in Chapter 17 when we explain multiple regression, so
for now, you can just ignore it.
The F statistic
The last line of the sample output in Figure 17-4 presents the F statistic and asso-
ciated p value (under F-statistic). These estimates address this question: Is the
straight-line model any good at all? In other words, how much better is the
straight-line model, which contains an intercept and a predictor variable, at pre-
dicting the outcome compared to the null model?
The null model is a model that contains only a single parameter representing a
constant term with no predictor variables at all. In this case, the null model would
only include the intercept.
Under α = 0.05, if the p value associated with the F statistic is less than 0.05, then
adding the predictor variable to the model makes it statistically significantly bet-
ter at predicting SBP than the null model.
For this example, the p value of the F statistic is 0.013, which is statistically sig-
nificant. It means using weight as a predictor of SBP is statistically significantly
better than just guessing that everyone in the data set has the mean SBP (which is
how the null model is compared).
Scientific fortune-telling with

the prediction formula
As we describe in Chapter 15, one reason to do regression in biostatistics is to
develop a prediction formula that allows you to make an educated guess about
value of a dependent variable if you know the values of the independent variables.
You are essentially developing a predictive model.
Some statistics programs show the actual equation of the best-fitting straight
line. If yours doesn’t, don’t worry. Just substitute the coefficients of the intercept
and slope for a and b in the straight-line equation: Y a bX .

With the output shown in Figure 16-4, where the intercept (a) is 76.9 and the
slope (b) is 0.487, you can write the equation of the fitted straight line like this:
SBP = 76.9 + 0.487 Weight.
Then you can use this equation to predict someone’s SBP if you know their weight.
So, if a person weighs 100 kilograms, you can estimate that that person’s SBP will
be around 76.9 100 0.487, which is 76.9 48.7 , or about 125.6 mmHg. Your
prediction probably won’t be exactly on the nose, but it should be better than not
using a predictive model and just guessing.
How far off will your prediction be? The residual SE provides a unit of measure-
ment to answer this question. As we explain in the earlier section “Summary sta-
tistics for the residuals,” the residual SE indicates how much the individual points
tend to scatter above and below the fitted line. For the SBP example, this number
is 9.8 , so you can expect your prediction to be within about 10 mmHg most of
the time.
Recognizing What Can Go Wrong with

Fitting a straight line to a set of data is a relatively simple task, but you still have
to be careful. A computer program does whatever you tell it to, even if it’s some-
thing you shouldn’t do.
Those new to straight-line regression may slip up in the following ways:
» Fitting a straight line to curved data: Examining the pattern of residuals in

the residuals versus fitted chart in Figure 16-5 can let you know if you have
this problem.
» Ignoring outliers in the data: Outliers — especially those in the corners of a

scatterplot like the one in Figure 16-3 — can mess up all the classical statistical
analyses, and regression is no exception. One or two data points that are way
off the main trend of the points will drag the fitted line away from the other
points. That’s because the strength with which each point tugs at the fitted
line is proportionate to the square of its distance from the line, and outliers
have a lot of distance, so they have a strong influence.
Always look at a scatter plot of your data to make sure outliers aren’t present.
Examine the residuals to ensure they are distributed normally above and
below the fitted line.

Calculating the Sample Size You Need
To estimate how many data points you need for a regression analysis, you need to
first ask yourself why you’re doing the regression in the first place.
» Do you want to show that the two variables are statistically significantly
associated? If so, you want to calculate the sample size required to achieve a
certain statistical power for the significance test (see Chapter 3 for an introduc-
tion to statistical power).
» Do you want to estimate the value of the slope (or intercept) to within a
certain margin of error? If so, you want to calculate the sample size required
to achieve a certain precision in your estimate.
Testing the statistical significance of a slope is exactly equivalent to testing the

statistical significance of a correlation coefficient, so the sample-size calculations
are also the same for the two types of tests. If you haven’t already, check out
Chapter 15, which contains guidance and formulas to estimate how many partici-
pants you need to test for any specified degree of correlation.
If you’re using regression to estimate the value of a regression coefficient — for

example, the slope of the straight line — then the sample-size calculations
become more complicated. The precision of the slope depends on several factors:
» The number of data points: More data points give you greater precision. SEs
vary inversely with the square root of the sample size. Alternatively, the
required sample size varies inversely with the square of the desired SE. So, if
you quadruple the sample size, you cut the SE in half. This is a very important
and generally applicable principle.
» Tightness of the fit of the observed points to the line: The closer the data
points hug the line, the more precisely you can estimate the regression
coefficients. The effect is directly proportional, in that twice as much Y-scatter
of the points produces twice as large a SE in the coefficients.
» How the data points are distributed across the range of the X variable:
This effect is hard to quantify, but in general, having the data points spread
out evenly over the entire range of X produces more precision than having
most of them clustered near the middle of the range.
Given these factors, how do you strategically design a study and gather data for a
linear regression where you’re mainly interested in estimating a regression coef-
ficient to within a certain precision? One practical approach is to first conduct a
study that is small and underpowered, called a pilot study, to estimate the SE of the

regression coefficient. Imagine you enroll 20 participants and measure them, then
create a regression model. If you’re really lucky, the SE may be as small as you
wanted, or even smaller, so you know if you conduct a larger study, you will have
enough sample.
But the SE from a pilot study usually isn’t small enough (unless you’re a lot luck-
ier that we’ve ever been). That’s when you can reach for the square-root law as a
remedy! Follow these steps to calculate the total sample size you need to get the
precision you want:
1. Divide the SE that you got from your pilot study by the SE you want your
full study to achieve.
2. Take the square of this ratio.
3. Multiply the square of the ratio by the sample size of your pilot study.
Imagine that you want to estimate the slope to a precision or SE of ±5. If a pilot
study of 20 participants gives you a SE of ±8.4 units, then the ratio is 8.4 / 5, which
is 1.68. Squaring this ratio gives you 2.82, which tells you that to get an SE of 5,
you need 2.82 20 , or about 56 participants. And because we assume you took our
advice, we’ll assume you’ve already recruited the first 20 participants for your
pilot study. Now, you only have to recruit only another 36 participants to have a
total of 56.
Admittedly, this estimation is only approximate. But it does give you at least a
ballpark idea of how big your sample size needs to be to achieve the desired
precision.

IN THIS CHAPTER
» Understanding what multiple

regression is
» Preparing your data and interpreting

the output
» Understanding how interactions and

collinearity affect regression analysis
» Estimating the number of

participants you need for a multiple
regression analysis
Chapter 17
More of a Good Thing:
Multiple Regression
C
hapter 15 introduces the general concepts of correlation and regression, two
related techniques for detecting and characterizing the relationship between
two or more variables. Chapter 16 describes the simplest kind of regression —
fitting a straight line to a set of data consisting of one independent variable (the
predictor) and one dependent variable (the outcome). The formula relating the pre-
dictor to the outcome, known as the model, is of the form Y a bX , where Y is the
outcome, X is the predictor, and a and b are parameters (also called regression
coefficients). This kind of regression is usually the only one you encounter in an
introductory statistics course, because it is a relatively simple way to do a regres-
sion. It’s good for beginners to learn!
This chapter extends simple straight-line regression to more than one

predictor — to what’s called the ordinary multiple linear regression model, or more
simply, multiple regression.
CHAPTER 17 More of a Good Thing: Multiple Regression 233

Multiple Regression
In Chapter 16, we outline the derivation of the formulas for determining the
parameters of a straight line so that the line — defined by an intercept at the Y
axis and a slope — comes as close as possible to all the data points (imagine a
scatter plot). The term as close as possible is operationalized as a least-squares line,
meaning we are looking for the line where the sum of the squares (SSQ) of vertical
distances of each point from to the line is the smallest. SSQ for a fitted line is
smallest for the least-squares line than for any other line you could possibly draw.
The same idea can be extended to multiple regression models containing more
than one predictor (which estimates more than two parameters). For two predic-
tor variables, you’re fitting a plane, which is a flat sheet. Imagine fitting a set of
points to this plane in three dimensions (meaning you’d be adding a Z axis to your
X and Y). Now, extend your imagination. For more than two predictors, in regres-
sion, you’re fitting a hyperplane to points in four-or-more-dimensional space.
Hyperplanes in multidimensional space may sound mind-blowing, but luckily for
us, the actual formulas are simple algebraic extensions of the straight-line
formulas.
In the following sections, we define some basic terms related to multiple regres-
sion, and explain when you should use it.
Defining a few important terms

Multiple regression is formally known as the ordinary multiple linear regression
model. What a mouthful! Here’s what the terms mean:
» Ordinary: The outcome variable is a continuous numerical variable whose

random fluctuations are normally distributed (see Chapter 24 for more about
normal distributions).
» Multiple: The model has more than two predictor variables.
» Linear: Each predictor variable is multiplied by a parameter, and these

products are added together to estimate the predicted value of the outcome
variable. You can also have one more parameter thrown in that isn’t multi-
plied by anything — it’s called the constant term or the Intercept. The following
are examples of linear functions used in regression:

• Y a bX (This is the straight-line model from Chapter 16, where X is the
predictor variable, Y is the outcome, and a and b are parameters.)
• Y a bX cX 2 dX 3 (In this multiple regression model, variables can

be squared or cubed. But as long as they’re multiplied by a coefficient —
which is a slope from the model — and the products are added together,
the function is still considered linear in the parameters.)
• Y a bX cZ dXZ (This multiple regression model is special because of

the XZ term, which can be written as X * Z , and is called an interaction. It is
where you multiple two predictors together to create a new interaction
term in the model.)
In textbooks and published articles, you may see regression models written in
various ways:
» A collection of predictor variables may be designated by a subscripted

variable and the corresponding coefficients by another subscripted variable,
like this: Y b0 b1 X 1 b2 X 2.
» In practical research work, the variables are often given meaningful names,
like Age, Gender, Height, Weight, Glucose, and so on.
» Linear models may be represented in a shorthand notation that shows only

the variables, and not the parameters, like this: Y = X + Z + X * Z instead of Y = a
+ bX + cZ + dX * Z or Y = 0 + X + Z + X * Z to specify that the model has no
intercept. And sometimes you’ll see a “~” instead of the “=”. If you do, read the
“~” as “is a function of,” or “is predicted by.”
Being aware of how the calculations work

Fitting a linear multiple regression model essentially involves creating a set of
simultaneous equations, one for each parameter in the model. The equations
involve the parameters from the model and the sums of various products of the
dependent and independent variables. This is also true of the simultaneous
equations for the straight-line regression in Chapter 16, which involve estimating
the slope and intercept of the straight line and the sums of X ,Y , X 2 , and XY. Your
statistical software solves these simultaneous equations to obtain the parameter
values, just as is done in straight-line regression, except now, there are more
equations to solve. In multiple as in straight-line regression, you can also get the
information you need to estimate the standard errors (SEs) of the parameters.

Executing a Multiple Regression
Analysis in Software
Before executing your multiple regression analysis, you may need to do some prep
work on the variables you intend to include in your model. In the following sec-
tions, we explain how to handle the categorical variables you plan to include. We
show you how to examine these variables through making several charts before
you run your analysis. If you need guidance on what variables to consider for your
models, read Chapter 20.
Preparing categorical variables

The predictors in a multiple regression model can be either numerical or categori-
cal (Chapter 8 discusses the different types of data). In a categorical variable, each
category is called a level. If a variable, like Setting, can have only two levels, like
Inpatient or Outpatient, then it’s called a dichotomous or a binary categorical vari-
able. If it can have more than two levels, it is called a multilevel variable.
Figuring out the best way to introduce categorical predictors into a multiple
regression model is always challenging. You have to set up your data the right
way, or you’ll get results that are either wrong, or difficult to interpret properly.
Following are two important factors to consider.
Having enough participants in each level

of each categorical variable
Before using a categorical variable in a multiple regression model, you should
tabulate how many participants (or rows) are included in each level. If you have
any sparse levels — row frequencies in the single digits — you will want to con-
sider collapsing them into others. Usually, the more evenly distributed the num-
ber of rows are across all the levels, and the fewer levels there are, the more
precise and reliable the results. If a level doesn’t contain enough rows, the pro-
gram may ignore that level, halt with a warning message, produce incorrect
results, or crash. Worse, if it produces results, they will be impossible to interpret.
Imagine that you create a one-way frequency table of a Primary Diagnosis vari-
able from a sample of study participant data. Your results are: Hypertension: 73,
Diabetes: 35, Cancer: 1, and Other: 10. To deal with the sparse Cancer variable, you
may want to create another variable in which Cancer is collapsed together with
Other (which would then have 11 rows). Another approach is to create a binary
variable with yes/no levels, such as: Hypertension: 73 and No Hypertension: 46.
But binary variables don’t take into account the other levels. You could also make

a binary Diabetes variable, where 35 were coded as yes and the rest were no, and
so on for Cancer and Other.
Similarly, if your model has two categorical variables with an interaction term (like
Setting + Primary Diagnosis + Setting * Primary Diagnosis), you should prepare a
two-way cross-tabulation of the two variables first (in our example, Setting by
Primary Diagnosis). You will observe that you are limited by having to ensure that
you have enough rows in each cell of the table to run your analysis. See Chapter 12
for details about cross-tabulations.
Choosing the reference level wisely

For each categorical variable in a multiple regression model, the program consid-
ers one of the categories to be the reference level and evaluates how each of the
other levels affects the outcome, relative to that reference level. Statistical soft-
ware lets you specify the reference level for a categorical variable, but you can also
let the software choose it for you. The problem is that the software uses some
arbitrary algorithm to make that choice (such as whatever level sorts alphabeti-
cally as first), and usually chooses one you don’t want. Therefore, it is better if you
instruct the software on the reference level to use for all categorical variables. For
specific advice on choosing an appropriate reference level, read the next section,
“Recoding categorical variables as numerical.”
Recoding categorical variables as numerical

Data may be stored as character variables — meaning the variable for primary
diagnosis (PrimaryDx) may be contain character data, such as Hypertension, Diabe-
tes, Cancer, and Other. Because it is difficult for statistical programs to work with
character data, these variables are usually recoded with a numerical code before
being used in a regression. This means a new variable is created, and is coded as 1
for hypertension, 2 for diabetes, 3 for cancer, and so on.
It is best to code binary variables as 0 for not having the attribute or state, and 1
for having the attribute or state. So a binary variable named Cancer should be
coded as Cancer = 1 if the participant has cancer, and Cancer = 0 if they do not.
For categorical variables with more than two levels, it’s more complicated. Even if
you recode the categorical variable from containing characters to a numeric code,
this code cannot be used in regression unless we want to model the category as an
ordinal variable. Imagine a variable coded as 1 = graduated high school, 2 = gradu-
ated college, and 3 = obtained post-graduate degree. If this variable was entered
as a predictor in regression, it assumes equal steps going from code 1 to code 2,
and from code 2 to code 3. Anyone who has applied to college or gone to graduate
school knows these steps are not equal! To solve this problem, you could select

one level for the reference group (let’s choose 3), and then create two binary indi-
cator variables for the other two levels — meaning one for 1 = graduated high
school and 2 = graduated college. Here’s another example of coding multilevel
categorical variables as a set of indicator variables, where each level is assigned its
own binary variable that is coded 1 if the level applies to the row, and 0 if it does
not (see Table 17-1).
TABLE 17-1 Coding a Multilevel Category into a Set of Binary

Indicator Variables
StudyID PrimaryDx HTN Diab Cancer OtherDx
1 Hypertension 1 0 0 0
2 Diabetes 0 1 0 0
3 Cancer 0 0 1 0
4 Other 0 0 0 1
5 Diabetes 0 1 0 0
Table 17-1 shows theoretical coding for a data set containing the variables StudyID
(for participant ID) and PrimaryDx (for participant primary diagnosis). As shown
in Table 17-1, you take each level and make an indicator variable for it: Hyperten-
sion is HTN, diabetes is Diab, cancer is Cancer, and other is OtherDx. Instead of
including the variable PrimaryDx in the model, you’d include the indicator vari-
ables for all levels of PrimaryDx except the reference level. So, if the reference level
you selected for PrimaryDx was hypertension, you’d include Diab, Cancer, and
OtherDx in the regression, but would not include HTN. To contrast this to the edu-
cation example, in the set of variables in Table 17-1, participants can have a 1 for
one or more indicator variables or just be in the reference group. However, with
the education example, they can only be coded at one level, or be in the reference
group.
Don’t forget to leave the reference-level indicator variable out of the regression,
or your model will break!
Creating scatter charts before you jump

into multiple regression analysis
One common mistake researchers make is immediately running a regression or
another advanced statistical analysis before thoroughly examining their data. As

soon as your data are available in electronic format, you should run error-checks,
and generate summaries and histograms for each variable you plan to use in your
regression. You need to assess the way the values of the variables are distributed
as we describe in Chapter 11. And if you plan to analyze your data using multiple
regression, you need special preparation. Namely, you should chart the relation-
ship between each predictor variable and the outcome variable, and also the rela-
tionships between the predictor variables themselves.
Imagine that you are interested in whether the outcome of systolic blood pressure
(SBP) can be predicted by age, body weight, or both. Table 17-2 shows a small data
file with variables that could address this research question that we use through-
out the remainder of this chapter. It contains the age, weight, and SBP of 16 study
participants from a clinical population.
TABLE 17-2 Sample Age, Weight, and Systolic Blood Pressure Data for a
Multiple Regression Analysis
Participant ID Age (years) Weight (kg) SBP (mmHg)
1 60 58 117
2 61 90 120
3 74 96 145
4 57 72 129
5 63 62 132
6 68 79 130
7 66 69 110
8 77 96 163
9 63 96 136
10 54 54 115
11 63 67 118
12 76 99 132
13 60 74 111
14 61 73 112
15 65 85 147
16 79 80 138

Looking at Table 17-2, let’s assume that your variable names are StudyID for Par-
ticipant ID, Age for age, Weight for weight, and SBP for SBP. Imagine that you’re
planning to run a regression model with this formula (using the shorthand nota-
tion described in the earlier section “Defining a few important terms”): SBP ~ Age
+ Weight. In this case, you should first prepare several scatter charts: one of SBP
(outcome) versus Age (predictor), one of SBP versus Weight (another outcome ver-
sus predictor), and one of Age versus Weight (both predictors). For regression
models involving many predictors, there can be a lot of scatter charts! Fortu-
nately, many statistics programs can automatically prepare a set of small thumb-
nail scatter charts for all possible pairings among a set of variables, arranged in a
matrix as shown in Figure 17-1.
FIGURE 17-1:
A scatter chart
matrix for a set
of variables prior
to multiple
regression.
These charts can give you insight into which variables are associated with each
other, how strongly they’re associated, and their direction of association. They
also show whether your data have outliers. The scatter charts in Figure 17-1
indicate that there are no extreme outliers in the data. Each scatter chart also
shows some degree of positive correlation (as described in Chapter 15). In fact, if
you refer to Figure 17-1, you may guess that the charts in Figure 17-1 correspond
to correlation coefficients between 0.5 and 0.8. In addition to the scatter charts,
you can also have your software calculate correlation coefficients (r values)
between each pair of variables. For this example, here are the results: r 0.654 for
Age versus Weight, r 0.661 for Age versus SBP, and r 0.646 for Weight versus SBP.

Taking a few steps with your software
The exact steps you take to run a multiple regression depend on your software, but
here’s the general approach:
1. Assemble your data into a file with one row per participant and one
column for each variable you want in the model.
2. Tell the software which variable is the outcome and which are the
predictors.
3. Specify whatever optional output you want from the software, which
could include graphs, summaries of the residuals (observed minus
predicted outcome values), and other useful results.
4. Execute the regression (run or submit the code).
Now, you should retrieve the output, and look for the optional output you
requested.
Interpreting the Output of a Multiple

Regression Analysis
The output from a multiple regression run is formatted like the output from the
straight-line regression described in Chapter 16.
Examining typical multiple

regression output
Figure 17-2 shows the output from a multiple regression analysis on the data in
Table 17-2, using R statistical software as described in Chapter 4. Other statistical
software produces similar output, but the results are arranged and formatted
differently.
Here we describe the components of the output in Figure 17-2:
» Code: The first line starting with Call: reflects back the code run to execute the
regression, which contains the linear model using variable names: SBP ~
Age + Weight.

FIGURE 17-2:
Output from
multiple
regression using
the data from
Table 17-2.
» Residual information: As a reminder, the residuals are the observed

outcome values minus predicted values coming from the model. Under
Residuals, the minimum, first quartile, median, third quartile and maximum
are listed (under the headings Min, IQ, Median, 3Q, and Max, respectively).
The maximum and minimum indicate that one observed SBP value was
17.8 mmHg greater than predicted by the model, and one was 15.4 mmHg
smaller than predicted.
» Regression or coefficients table: This is presented under Coefficients:, and

includes a row for each parameter in the model. It also includes columns for
the following:
• Estimate: The estimated value of the parameter, which tells you how much
the outcome variable changes when the corresponding variable increases
by exactly 1.0 unit, holding all the other variables constant. For example, the
model predicts that if all participants have the same weight, every addi-
tional year of age is associated with an increase in SBP of 0.84 mmHg.
• Standard error: The standard error (SE) is the precision of the estimate,
and is in the column labeled Std. Error. The SE for the Age coefficient is
0.52 mmHg per year, indicating the level of uncertainty around the
0.84 mmHg estimate.
• t value: The t value (which is labeled t value) is the value of the parameter
divided by its SE. For Age, the t value is 0.8446 / 0.5163 , or 1.636.
• p value: The p value is designated Pr(>|t|) in this output. The p value

indicates whether the parameter is statistically significantly different from
zero at your chosen α level (let’s assume 0.05). If p 0.05 , then the
predictor variable is statistically significantly associated with the outcome
after controlling for the effects of all the other predictors in the model. In

this example, neither the Age coefficient (p = 0.126) nor the Weight coeffi-
cient (p = 0.167) is statistically significantly different from zero.
» Model fit statistics: These are calculations that describe how well the model
fits your data overall.
• Residual standard error: In this example, the Residual standard error:

(bottom of output) indicates that the observed-minus-predicted residuals
have a standard deviation of 11.23 mmHg.
• Multiple r2: This refers to the square of an overall correlation coefficient

for the multivariate fit of the model, and is listed under Multiple R-squared.
• F statistic: The F statistic and associated p value (on the last line of the
output) indicate whether the model predicts the outcome statistically
significantly better than a null model. A null model contains only the
intercept term and no predictor variables at all. The very low p value
(0.0088) indicates that age and weight together predict SBP statistically
significantly better than the null model.
Checking out optional output to request

Depending on your software, you may also be able to request several other useful
calculations from the regression to be included:
» Predicted values for the dependent variable for each participant. This can be
output either as a listing, or as a new variable placed into your data file.
» Residuals (observed minus predicted value) for each participant. Again, this
can be output either as a listing, or as a new variable placed into your data file.
Deciding whether your data are suitable

for regression analysis
Before drawing conclusions from any statistical analysis, you need to make sure
that your data fulfill assumptions on which that analysis was based. Two assump-
tions of ordinary linear regression include the following:
» The amount of variability in the residuals is fairly constant, and not dependent
on the value of the dependent variable.
» The residuals are approximately normally distributed.

Figure 17-3 shows two graphs you can optionally request that help you determine
or diagnose whether these assumptions are met, so they are called diagnostic graphs
(or plots).
» Figure 17-3a provides an indication of variability of the residuals. To interpret

this plot, visually evaluate whether the points seem to scatter evenly above
and below the line, and whether the amount of scatter seems to be the same
across the left, middle, and right parts of the graph. That seems to be the case
in this figure.
» Figure 17-3b provides an indication of the normality of the residuals. To

interpret this plot, visually evaluate whether the points appear to lie along the
dotted line or are noticeably following a curve. In this figure, the points are
consistent with a straight line except in the very lower-left part of the graph.
FIGURE 17-3:
Diagnostic
graphs from a
regression.
Determining how well the

model fits the data
Several calculations in standard regression output indicate how closely the model
fits your data:
» The residual SE is the average scatter of the observed points from the fitted
model. You want them to be close to the line. As shown in Figure 17-2, the
residual SE is about 11 mmHg.

» The multiple r2 value represents the amount of variability in the dependent
variable explained by the model, so you want it to be high. As shown in
Figure 17-2, it is 0.52 in this example, indicating a moderately good fit.
» A statistically significant F statistic indicates that the model predicts the

outcome significantly better than the null model. As shown in Figure 17-2, the
p value on the F statistic is 0.009, which is statistically significant at α = 0.05.
Figure 17-4 shows another way to judge how well the model predicts the outcome.
It’s a graph of observed and predicted values of the outcome variable, with a
superimposed identity line (Observed Predicted). Your program may offer this
observed versus predicted graph, or you can generate it from the observed and pre-
dicted values of the dependent variable. For a perfect prediction model, the points
would lie exactly on the identity line. The correlation coefficient of these points is
the multiple r value for the regression.
FIGURE 17-4:
Observed versus
predicted
outcomes for the
model SBP ~ Age
+ Weight, for the
data in
Table 17-2.
Watching Out for Special Situations

that Arise in Multiple Regression
Here we describe two topics that come up in multiple regression: interactions
(both synergistic and anti-synergistic), and collinearity. Both relate to how the
simultaneous behavior of two predictors can influence an outcome.

MODEL BUILDING
If you have a big data set with many variables, how do you plan which predictors to try
to include in your multiple regression model? Once you choose the ones you want to
consider, how do you decide which ones to keep and which ones to remove to achieve
the best-fitting model? Biostatisticians approach model building using different meth-
ods, but the goal of all of these is to achieve the best-fitting model that explains the rela-
tionship between the predictors and outcome, and to do it in a transparent way.
Chapter 20 includes a section explaining how to develop a modeling plan if you have a
choice of many potential predictor variables.
Synergy and anti-synergy

Sometimes, two predictor variables exert a synergistic effect on an outcome. That is,
if both predictors were to be associated with an increase in the outcome by one
unit, the outcome would change by more than the sum of the two increases, which
is what you’d expect from changing each value individually by one unit. You can
test for synergy between two predictors with respect to an outcome by fitting a
model that contains an interaction term, which is the product of those two vari-
ables. In this equation, we predict SBP using Age and Weight, and include an
interaction term for Age and Weight:
SBP = Age + Weight + Age * Weight
If the estimate of the slope for the interaction term has a statistically significant
p value, then the null hypothesis of no interaction is rejected, and the two variables
are interpreted to have a significant interaction. If the sign on the interaction term
is positive, it is a synergistic interaction, and if it is negative, it is called an anti-
synergistic or antagonistic interaction.
Introducing interaction terms into a fitted model and interpreting their

significance — both clinically and statistically — must be done contextually.
Interaction terms may not be appropriate for certain models, and may be required
in others.
Collinearity and the mystery of the

disappearing significance
When developing multiple regression models, you are usually considering more
predictors than just two as we used in our example. You develop iterative models,

meaning models with the same outcome variable, but different groups of predic-
tors. You also use some sort of strategy in choosing the order in which you intro-
duce the predictors into the iterative models, which is described in Chapter 20. So
imagine that you used our example data set and — in one iteration — ran a model
to predict SBP with Age and other predictors in it, and the coefficient for Age was
statistically significant. Now, imagine you added Weight to that model, and in the
new model, Age was no longer statistically significant! You’ve just been visited by
the collinearity fairy.
In the example from Table 17-2, there’s a statistically significant positive correla-
tion between each predictor and the outcome. We figured this out when running
the correlations for Figure 17-1, but you could check our work by using the data in
Figure 17-2 in a straight-line regression, as described in Chapter 16. In contrast,
the multiple regression output in Figure 17-2 shows that neither Age nor Weight
are statistically significant in the model, meaning neither has regression coeffi-
cients that are statistically significantly different from zero! Why are they associ-
ated with the outcome in correlation but not multiple regression analysis?
The answer is collinearity. In the regression world, the term collinearity (also
called multicollinearity) refers to a strong correlation between two or more of the
predictor variables. If you run a correlation between Age and Weight (the two pre-
dictors), you’ll find that they’re statistically significantly correlated with each
other. It is this situation that destroys your statistically significant p value seen on
some predictors in iterative models when doing multiple regression.
The problem with collinearity is that you cannot tell which of the two predictor
variables is actually influencing the outcome more, because they are fighting over
explaining the variability in the dependent variable. Although models with col-
linearity are valid, they are hard to interpret if you are looking for cause-and-
effect relationships, meaning you are doing causal inference. Chapter 20 provides
philosophical guidance on dealing with collinearity in modeling.
Calculating How Many Participants

You Need
Studies should target enrolling a large enough sample size to ensure that you get
a statistically significant result for your primary research hypothesis in the case
that the effect you’re testing in that hypothesis is large enough to be of clinical
importance. So if the main hypothesis of your study is going to be tested by a
multiple regression analysis, you should theoretically do a calculation to deter-
mine the sample size you need to support that analysis.

Unfortunately, that is not possible in practice, because the equations would be too
complicated. Instead, considerations are aimed more toward being able to gather
enough data to support a planned regression model. Imagine that you plan to
gather data about a categorical variable where you believe only 5 percent of the
participants will fall in a particular level. If you are concerned about including that
level in your regression analysis, you would want to greatly increase your estimate
for target sample size. Although regression models tend to converge in software if
they include at least 100 rows, that may not be true depending upon the number
and distribution of the values in the predictor variables and the outcome. It is best
to use experience from similar studies to help you develop a target sample size
and analytic plan for a multiple regression analysis.

IN THIS CHAPTER
» Figuring out when to use logistic

regression
» Getting a grip on the basics of logistic

regression
» Running a logistic regression model

and making sense of the output
» Watching for common issues with

logistic regression
» Estimating the sample size you need

for logistic regression
Chapter 18
A Yes-or-No Proposition:
Logistic Regression
Y
ou can use logistic regression to analyze the relationship between one or
more predictor variables (the X variables) and a categorical outcome vari-
able (the Y variable). Typical categorical outcomes include the following
two-level variables (which are also called binary or dichotomous):
» Lived or died by a certain date
» Did or didn’t get diagnosed with Type II diabetes
» Responded or didn’t respond to a treatment
» Did or did not choose a particular health insurance plan
In this chapter, we explain logistic regression. We describe the circumstances

under which to use it, the important related concepts, how to execute it with soft-
ware, and how to interpret the output. We also point out the pitfalls with logistic
regression and show you how to determine the sample sizes you need to execute
such a model.
CHAPTER 18 A Yes-or-No Proposition: Logistic Regression 249

Using Logistic Regression
Following are typical uses of logistic regression analysis:
» To test whether one or more predictors and an outcome are statistically

significantly associated. For example, to test whether age and/or obesity
status are associated with increased likelihood to be diagnosed with
Type II diabetes.
» To overcome the limitations of the 2x2 cross-tab method (described in

Chapter 12), which can analyze only one predictor at a time (and the predictor
has to be binary). With logistic regression, you can analyze multiple predictor
variables at a time. Each predictor can be a numeric variable or a categorical
variable having two or more levels.
» To quantify the extent or magnitude of an association between a particular

predictor and an outcome that have been established to have an association.
In other words, you are seeking to quantify the amount by which a specific
predictor influences the chance of getting the outcome. As an example, you
could quantify the amount obesity plays a role in the likelihood of a person
being diagnosed with Type II diabetes.
» To develop a formula to predict the probability of getting an outcome based

on the values of the predictor variables. For example, you may want to predict
the probability that a person will be diagnosed with Type II diabetes based on
the person’s age, gender, obesity status, exercise status, and medical history.
» To make yes or no predictions about the outcome that take into account the
consequences of false-positive and false-negative predictions. For example,
you can generate a tentative cancer diagnosis from a set of observations and
lab results using a formula that balances the different consequences of a
false-positive versus a false-negative diagnosis.
» To see how one predictor influences the outcome after adjusting for the
influence of other variables. One example is to see how the number of
minutes of exercise per day influences the chance of having a heart attack
after controlling for the for the effects of age, gender, lipid levels, and other
patient characteristics that could influence the outcome.
» To determine the value of a predictor that produces a certain probability of

getting the outcome. For example, you could determine the dose of a drug
that produces a favorable clinical response in 80 percent of the patients
treated with it, which is called the ED 80, or 80 percent effective dose.

Logistic Regression
In this section, we explain the concepts underlying logistic regression using an
example from a fictitious animal study involving data on mortality due to radia-
tion exposure. This example illustrates why straight-line regression wouldn’t
work and why you have to use logistic regression instead.
Gathering and graphing your data

As in the other chapters in Part 5, we present a real-world problem here. This
example examines the lethality of exposure to gamma-ray radiation when given
in acute, large doses. It is already known that gamma-ray radiation is deadly in
large-enough doses, so this animal study is focused only at the short-term lethal-
ity of acute large doses. Table 18-1 presents data on 30 animals in two columns.
TABLE 18-1 Radiation Dose and Survival Data for 30 Animals,

Sorted Ascending by Dose Level
Dose in REMs Outcome 0 Lived; 1 Died Dose in REMS Outcome 0 Lived; 1 Died
0 0 433 0
10 0 457 1
31 0 559 1
82 0 560 1
92 0 604 1
107 0 632 0
142 0 686 1
173 0 691 1
175 0 702 1
232 0 705 1
266 0 774 1
299 0 853 1
303 1 879 1
326 0 915 1
404 1 977 1

In Table 18-1, dose is the radiation exposure expressed in units called Roentgen
Equivalent Man (REM). Because Table 18-1 is sorted ascending by dose, by looking
at the Dose and Outcome columns, you can get a rough sense of how survival
depends on dose. At low levels of radiation, almost all animals live, and at high
doses, almost all animals die.
How can you analyze these data with logistic regression? First, make a scatter plot
(see Chapter 16) with the predictor — the dose — on the X axis, and the outcome
of death on the Y axis, as shown in Figure 18-1a.
FIGURE 18-1:
Dose versus
mortality from
Table 18-1: each
individual’s
data (a) and
grouped (b).
In Figure 18-1a, because the outcome variable is binary, the points are restricted
to two horizontal lines, making the graph difficult to interpret. You can get a bet-
ter picture of the dose-lethality relationship by grouping the doses into intervals.
In Figure 18-1b, we grouped the intervals into 200 REM classes (see Chapter 9),
and plotted the fraction of individuals in each interval who died. Clearly,
Figure 18-1b shows the chance of dying increases with increasing dose.
Fitting a function with an

S shape to your data
Don’t try to fit a straight line if you have a binary outcome variable because the
relationship is almost certainly not a straight line. For one thing, the fraction of
individuals who are positive for the outcome can never be smaller than 0 nor
larger than 1. In contrast, a straight line, a parabola, or any polynomial

distribution would very happily violate those limits at extreme doses, which is
obviously illogical.
If you have a binary outcome, you need to fit a function that has an S shape. The
formula calculating Y must be an expression involving X that — by design — can
never produce a Y value outside of the range from 0 to 1, no matter how large or
small X may become.
Of the many mathematical expressions that produce S-shaped graphs, the logistic
function is ideally suited to this kind of data. In its simplest form, the logistic
function is written like this: Y 1 / 1 e X , where e is the mathematical con-
stant 2.718, known as a natural logarithm (see Chapter 2). We will use e to repre-
sent this number for the rest of the chapter. Figure 18-2a shows the shape of the
logistic function.
The logistic function shown in Figure 18-2 can be made more versatile for repre-
senting observed data by being generalized. The logistic function is generalized by
adding two adjustable parameters named a and b like this: Y 1 / 1 e ( a bX ) .
FIGURE 18-2:
The first graph (a)
shows the shape
of the logistic
function. The
second graph (b)
shows that when
b is 0, the logistic
function becomes
a horizontal
straight line.
Notice that the a bX part looks just like the formula for a straight line (see
Chapter 16). It’s the rest of the logistic function that bends the straight line into
its characteristic S shape. The middle of the S (where Y 0.5 ) always occurs when
X b / a . The steepness of the curve in the middle region is determined by b, as
follows:
» If b is positive, the logistic function is an upward-sloping S-shaped curve, like

the one shown in Figure 18-2a.

» If b is 0, the logistic function is a horizontal straight line whose Y value is equal
to 1 / 1 e a , as shown in Figure 18-2b.
» If b is negative, the curve is flipped upside down, as shown in Figure 18-3a.

Notice that this is a mirror image of Figure 18-2a.
» If b is a very large number, either positive or negative, the logistic curve

becomes so steep that it looks like what mathematicians call a step function, as
shown in Figure 18-3b.
FIGURE 18-3:
The first graph (a)
shows that when
b is negative,
the logistic
function slopes
downward. The
second graph (b)
shows that when
b is very large,
the logistic
function becomes
a “step function.”
Because the logistic curve approaches the limits 0.0 and 1.0 for extreme values of
the predictor(s), you should not use logistic regression in situations where the
fraction of individuals positive for the outcome does not approach these two lim-
its. Logistic regression is appropriate for the radiation example because none of
the individuals died at a radiation exposure of zero REMs, and all of the individu-
als died at doses of 686 REMs and higher. If we imagine a study of patients with a
disease where the outcome is a cure, if taking a drug in very high doses would not
always cause a 100 percent cure, and the disease could resolve on its own without
any drug, the data would not be appropriate. This is because some patients with
high doses would still have an outcome value of 0, and some patients at zero dose
would have an outcome value of 1.
Logistic regression fits the logistic model to your data by finding the values of a
and b that make the logistic curve come as close as possible to all your plotted
points. With this fitted model, you can then predict the probability of the outcome.
See the later section “Predicting probabilities with the fitted logistic formula” for
more details.

GETTING INTO THE NITTY-GRITTY OF
LOGISTIC REGRESSION
You don’t need to know all the theoretical and computation details for logistic regres-
sion because the software will do them for you. However, you should have a general
idea of what it is doing behind the scenes. The calculations are much more complicated
than those for ordinary straight-line or multivariate least-squares regression. In fact, it’s
impossible to write down a set of formulas that give the logistic regression coefficients
in terms of the observed X and Y values. The only way to obtain them is through a com-
plex iterative procedure that would not be practical to do manually.
Logistic regression determines the values of the regression coefficients that are most
consistent with the observed data using what’s called the maximum likelihood criterion.
The likelihood of any statistical model is the probability (based on the model) of obtain-
ing the values you observed in your data. There’s a likelihood value for each row in the
data set, and a total likelihood (L) for the entire data set. The likelihood value for each
data point is the predicted probability of getting the observed outcome result. For indi-
viduals who died (refer to Table 18-1), the likelihood is the probability of dying (Y) pre-
dicted by the logistic formula. For individuals who survived, the likelihood is the
predicted probability of not dying, which is 1 Y . The total likelihood (L) for the whole
set of individuals is the product of all the calculated likelihoods for each individual.
To find the values of the coefficients that maximize L, it is most practical to find the val-
ues that minimize the quantity 2 multiplied by the natural logarithm of L, which also called
the –2 log likelihood and abbreviated –2LL. Statisticians also call –2LL the deviance. The
closer the curve designated by the regression formula comes to the observed points,
the smaller this deviance value will be. The actual value of the deviance for a logistic
regression model doesn’t mean much by itself. It’s the difference in deviance between
two models you might be comparing that is important.
Once deviance is calculated, the final step is to identify the values of the coefficients that
will minimize the deviance of the observed Y values from the fitted logistic curve. This
may sound challenging, but statistical programs employ elegant and efficient ways to
minimize such a complicated function involving several variables, and uses these meth-
ods to obtain the coefficients.
Handling multiple predictors

in your logistic model
The data in Table 18-1 have only one predictor variable, but you may have several
predictors of a binary outcome. If the data in Table 18-1 were about humans, you

would assume the chance of dying from radiation exposure may depend not only
on the radiation dose received, but also on age, gender, weight, general health,
radiation wavelength, and the amount of time over which the person was exposed
to radiation. In Chapter 17, we describe how the straight-line regression model
can be generalized to handle multiple predictors. You can generalize the logistic
formula to handle multiple predictors in the same way.
Suppose that the outcome variable Y is dependent on three predictors called X, V,

and W. Then the multivariate logistic model looks like this:
_
Y 1 / 1 e (a bX CV dW )
Logistic regression finds the best-fitting values of the parameters a, b, c, and d

given your data. That way, for any particular set of values for X, V, and W, you can
use the equation to predict Y, which is the probability of being positive for the
outcome.
Running a Logistic Regression

Model with Software
The theory behind logistic regression is difficult to grasp, and the calculations are
complicated (see the sidebar “Getting into the nitty-gritty of logistic regression”
for details). The good news is that most statistical software (as described in
Chapter 4) can run a logistic regression model, and it is similar to running a
straight-line or multiple linear regression model (see Chapters 16 and 17). Here
are the steps:
1. Make sure your data set has a column for the outcome variable that is
coded as 1 where the individual is positive for the outcome, and 0 when
they are negative.
If you do not have an outcome column coded this way, use the data manage-
ment commands in your software to generate a new variable coded as 0 for
those who do not have the outcome, and 1 for those who have the outcome,
as shown in Table 18-1.
2. Make sure your data set has a column for each predictor variable, and
that these columns are coded the way you want them to be entered
them into the model.

The predictors can be quantitative, such as age or weight. They can also be
categorical, like gender or treatment group. You will need to make decisions
about how to recode these variables to enter them into the regression model.
See Chapter 17, where we describe how to set up categorical predictor
variables.
3. Tell your software which variables are the predictors and which is the
outcome.
Depending on the software, you may do this by typing the variable names,
or by selecting the variables from a menu or list.
4. Request the optional output from the software if available which may
include:
• A summary of information about the variables, goodness-of-fit measures,

and a graph of the fitted logistic curve
• A table of regression coefficients, including odds ratios (ORs) and their

95 percent confidence intervals (CIs)
• Predicted probabilities of getting the outcome for each individual, and a

classification table of observed outcomes versus predicted outcomes
• Measures of prediction accuracy, which include overall accuracy, sensitiv-

ity, and specificity, as well as a Receiver Operator Characteristics (ROC)
curve
5. Execute the model in the software.
Obtain the output you requested, and interpret the resulting model.
Interpreting the Output of Logistic

Regression
Figure 18-4 shows two kinds of statistical software output from a logistic regres-
sion model produced from the data in Table 18-1. The output presented is not from
a particular command in a particular software. Instead, typical output for the dif-
ferent items is presented to enable us to cover many different potential
scenarios.

FIGURE 18-4:
Typical output
from a logistic
regression model.
The output on the
left (a) shows
statistical results
output for the
model, and the
output on
the right (b)
shows predicted
probabilities for
each individual
that can be
output as a
data set.
Seeing summary information

about the variables
At the top of Figure 18-4a, you can see information about the variables under
Descriptives. It can include means and standard deviations of predictors that are
numerical variables, and a count of how many individuals did or did not have the
outcome event. In Figure 18-4a, you can see that 15 of the 30 individuals lived and
15 died.
Assessing the adequacy of the model

In Figure 18-4a, the middle section starting with Deviance and ending with AIC
provides model fit information. These are measures that indicate how well the
fitted function represents the data, which is called goodness-of-fit. Some have test
statistics and an associated p value (see Chapter 3 for a refresher on p values),
while others produce metrics. Although there are different model fit statistics, you
will find that they usually agree on how well a model fits the data.

You may see the following model fit measures, depending on your software:
» A p value associated with the decrease in deviance between the null

model and the final model: This information is shown in Figure 18-4a under
Deviance. Under α = 0.05, if this p value < 0.05, it indicates that adding the
predictor variables to the null model statistically significantly improves its
ability to predict the outcome. In Figure 18-4a, p < 0.0001, which means that
adding radiation dose to the model makes it statistically significantly better at
predicting an individual animal’s chance of dying than the null model.
However, it’s not very hard for a model with any predictors to be better than
the null model, so this is not a very sensitive model fit statistic.
» A p value from the Hosmer-Lemeshow (H-L) test: In Figure 18-4a, this is

listed under Hosmer-Lemeshow Goodness of Fit Test. The null hypothesis for this
test is your data are consistent with the logistic function’s S shape, so if p <
0.05, your data do not qualify for logistic regression. The focus of the test is to
see if the S is getting distorted at very high or very low levels of the predictor
(as shown in Figure 18-4b). In Figure 18-4a, the H-L p value is 0.842, which
means that the data are consistent with the shape of a logistic curve.
» One or more pseudo–r2 values: Pseudo–r2 values indicate how much of the
total variability in the outcome is explainable by the fitted model. They are
analogous to how r2 is interpreted in ordinary least-squares regression, as
described in Chapter 17. In Figure 18-4a, two such values are provided under
the labels Cox/Snell R-square and Nagelkerke R-square. The Cox/Snell r2 is 0.577,
and the Nagelkerke r2 is 0.770, both of which indicate that a majority of the
variability in the outcome is explainable by the logistic model.
» Akaike’s Information Criterion (AIC): AIC is a measure of the final model

deviance adjusted for how many predictor variables are in the model. Like
deviance, the smaller the AIC, the better the fit. The AIC is not very useful on
its own, and is instead used for choosing between different models. When all
the predictors in one model are nested — or included — in another model
with more predictors, the AIC is helpful for comparing these models to see if it
is worth adding the extra predictors.
Checking out the table of regression

coefficients
Your intention when developing a logistic regression model is to obtain estimates
from the table of coefficients, which looks much like the coefficients table from
ordinary straight-line or multivariate least-squares regression (see Chapters 16
and 17). In Figure 18-4a, they are listed under Coefficients and Standard Errors. Observe:
» Every predictor variable appears on a separate row.

» There’s one row for the constant term labeled Intercept.
» The first column usually lists the regression coefficients (under Coeff. in
Figure 18-4a).
» The second column usually lists the standard error (SE) of each coefficient
(under StdErr in Figure 18-4a).
» A p-value column indicates whether the coefficient is statistically significantly

different from 0. This column may be labeled Sig or Signif or Pr(> lzl), but in
Figure 18-4a, it is labeled p-value.
For each predictor variable, the output should also provide the odds ratio (OR) and
its 95 percent confidence interval. These are usually presented in a separate table
as they are in Figure 18-4a under Odds Ratios and 95% Confidence Intervals.
Predicting probabilities with the

fitted logistic formula
The output may include the fitted logistic formula. At the bottom of Figure 18-4a,
the formula is shown as:
Prob Death = 1 / 1 + Exp - -4.828 + 0.01146 * Dose
You can write out the formula manually by inserting the value of the regression
coefficients from the regression table into the logistic formula. The final model
produced by the logistic regression program from the data in Table 18-1 and the
resulting logistic curve are shown in Figure 18-5.
Once you have the fitted logistic formula, you can predict the probability of having
the outcome if you know the value of the predictor variable. For example, if an
individual is exposed to 500 REM of radiation, the probability of the outcome is
given by this formula: Probability of Death 1 / 1 e _( 4 , 828 0.01146 500 ) , which
equals 0.71. An individual exposed to 500 REM of radiation has a predicted proba-
bility of 0.71 — or a 71 percent chance — of dying shortly thereafter. The predicted
probabilities for each individual are shown in the data listed in Figure 18-4b. You
can also calculate some points of special significance on a logistic curve, as you
find out in the following sections.
Be careful with your algebra when evaluating these formulas! The a coefficient in
a logistic regression is often a negative number, and subtracting a negative num-
ber is like adding its absolute value.

FIGURE 18-5:
The logistic curve
that fits the data
from Table 18-1.
Calculating effective doses on a logistic curve

One point of special significance on a logistic curve with a numerical predictor is
a median effective dose. This is a dose (X) that produces a 50 percent response,
meaning where Y 0.5 , and is designated ED50. Similarly, the X value that makes
Y 0.8 is called the 80 percent effective dose and is designated ED80, and so on. You
can calculate these dose levels from the a and b parameters of the fitted logistic
model in the preceding section.
Using your high-school algebra, you can solve the logistic formula
Y 1/ 1 e _( a
bX )
for X as a function of Y. If you don’t remember how to do that,
don’t worry, here’s the answer:
Y
log a
1 Y
X
b
where log stands for natural logarithm. If you substitute 0.5 for Y in the preceding
equation because you want to calculate the ED50, the answer is a / b. Similarly,
1.39 a
substituting 0.8 for Y gives the ED80 as .
b
Imagine a logistic regression model based on a study of participants taking a drug

at different doses where the predictor is level of drug dose, and the outcome is
that it produces a therapeutic response. The model has a 3.45 and b 0.0204
mg/dL. In this case, the ED80 (or 80 percent effective dose) would be equal to
1.39 3.45 / 0.0234 , which works out to about 207 mg/dL.

Calculating lethal doses on a logistic curve
When death is the outcome event, the corresponding terms are median lethal dose
(abbreviated LD50) and 80 percent lethal dose (abbreviated LD80), and so on. To cal-
culated the LD50 using the data in Table 18-1, a 4.83 and b 0.0115, so
a/b 4.83 / 0.0115, which works out to 420 REMs. An LD50 of 420 REMs
dose of radiation means an individual has a 50 percent chance of dying shortly
after being exposed to this level of radiation.
Making yes or no predictions

If you fit a logistic regression model, then learn of the value of predictor variables
for an individual, you can plug them into the equation and calculate the predicted
probability of the individual having the outcome. But sometimes, you are trying to
actually predict the outcome — whether the event will happen or not, yes or
no — to an individual. You can do this by setting a cut value on predicted probabil-
ity. Imagine you select 0.5 as the cut value, and you make a rule that if the
individual’s predicted probability is 0.5 or greater, you’ll predict yes; otherwise,
you’ll predict no.
In the following sections, we talk about yes or no predictions. We explain how they
expose the ability of the logistic model to make predictions, and how you can stra-
tegically select the cut value that gives you the best tradeoff between wrongly
predicting yes and wrongly predicting no.
Measuring accuracy, sensitivity, and specificity

with classification tables
Software output for logistic regression provides several goodness-of-fit measures
(see the earlier section “Assessing the adequacy of the model”). One intuitive
indicator of goodness-of-fit is the extent to which your yes or no predictions from
the logistic model match the actual outcomes. You can cross-tabulate the pre-
dicted and observed outcomes into a fourfold classification table. To do this, you
would ask the software to generate a classification table for you from the data
based on a cut value in the predicted probability. Most software assumes a cut
value of 0.5 unless you tell it to use some other value. Figure 18-6 shows the clas-
sification table of observed versus predicted outcomes from radiation exposure,
using a cut value of 0.5 predicted probability.
From the classification table shown in Figure 18-6, you can calculate several use-
ful measures of the model’s predicting ability for any specified cut value, includ-
ing the following:

FIGURE 18-6:
The classification
table for the
radiation
example.
» Overall accuracy: This refers to the proportion of accurate predictions, as

shown in the concordant cells, which are the upper-left and lower-right. Of the
30 individuals in the data set from Table 18-1, the logistic model predicted
correctly 13 13 / 30 = 0.87, or about 87 percent of the time. This means with
the cut value where you placed it, the model would make a wrong prediction
only about 13 percent of the time.
» Sensitivity: This refers to the proportion of yes outcomes predicted accu-

rately. As seen in the upper-left cell in Figure 18-6, with the cut value where it
was placed, the logistic model predicted 13 of the 15 observed deaths (yes
outcomes). So the sensitivity is 13 / 15 = 0.87, or about 87 percent. This means
the model would have a false-negative rate of 13 percent.
» Specificity: This refers to the proportion of no outcomes predicted accurately.

In the lower-right cell of Figure 18-6, the model predicted survival in 13 of the
15 observed survivors. So, the specificity is 13 / 15 = 0.87, or about 87 percent.
This means the model would have a false-positive rate of 13 percent.
Sensitivity and specificity are especially relevant to screening tests for diseases.
An ideal test would have 100 percent sensitivity and 100 percent specificity, and
therefore, 100 percent overall accuracy. In reality, no test could meet these stan-
dards, and there is a tradeoff between sensitivity and specificity.
By judiciously choosing the cut value for converting a predicted probability into a
yes or no decision, you can often achieve high sensitivity or high specificity, but
it’s hard to maximize both simultaneously. Screening tests are meant to detect
disease, so how you select the cut value depends upon what happens if it produces
a false-positive or false-negative result. This helps you decide whether to priori-
tize sensitivity or specificity.
The sensitivity and specificity of a logistic model depends upon the cut value you
set for the predicted probability. The trick is to select a cut value that gives the
optimal combination of sensitivity and specificity, striking the best balance
between false-positive and false-negative predictions, in light of the different

consequences of the two types of false predictions. A false-positive screening
result from a mammogram may mean the patient is worried until the negative
diagnosis is confirmed by ultrasound, and a false-negative screening results from
a prostate cancer screening may result in a delay in identifying the prostate tumor.
To find this optimal cut value, you need to know precisely how sensitivity and
specificity play against each other — that is, how they simultaneously vary with
different cut values. There’s a neat way to do that which we explain in the follow-
ing section.
Rocking with ROC curves

The graph used to display the sensitivity/specificity tradeoff for any fitted logistic
model is called the Receiver Operator Characteristics (ROC) graph. The name comes
from its original use during World War II to analyze the performance characteris-
tics of people who operated RADAR receivers, but the name has stuck, and now it
is also referred to as an ROC curve.
An ROC graph has a curve that shows you the complete range of sensitivity and
specificity that can be achieved for any fitted logistic model based on the selected
cut value. The software generates an ROC curve by effectively trying all possible
cut values of predicted probability between 0 and 1, calculating the predicted out-
comes, cross-tabbing them against the observed outcomes, calculating sensitivity
and specificity, and then graphing sensitivity versus specificity. Figure 18-7 shows
the ROC curve from the logistic model developed from the data in Figure 18-1
(using R software; see Chapter 4).
FIGURE 18-7:
ROC curve from
dose mortality
data.

As shown in Figure 18-7, the ROC curve always starts in the lower-left corner of
the graph, where 0 percent sensitivity intersects with 100 percent specificity. It
ends in the upper-right corner, where 100 percent sensitivity intersects with
0 percent specificity. Most software also draws a diagonal straight line between
the lower-left and upper-right corners because that represents the formula:
sensitivity 1 – specificity . If your model’s ROC curve were to match that line, it
would indicate the total absence of any predicting ability at all of your model.
Like Figure 18-7, every ROC graph has sensitivity running up the Y axis, which is
displayed either as fractions between 0 and 1 or as percentages between 0 and 100.
The X axis is either presented from left to right as 1 – specificity , or like it is in
Figure 18-7, where specificity is labeled backwards — from right to left — along
the X axis.
Most ROC curves lie in the upper-left part of the graph area. The farther away
from the diagonal line they are, the better the predictive model is. For a nearly
perfect model, the ROC curve runs up along the Y axis from the lower-left corner
to the upper-left corner, then along the top of the graph from the upper-left cor-
ner to the upper-right corner.
Because of how sensitivity and specificity are calculated, the graph appears as a
series of steps. If you have a large data set, your graph will have more and smaller
steps. For clarity, we show the cut values for predicted probability as a scale along
the ROC curve itself in Figure 18-7, but unfortunately, most statistical software
doesn’t do this for you.
Looking at the ROC curve helps you choose a cut value that gives the best tradeoff
between sensitivity and specificity:
» To have very few false positives: Choose a higher cut value to give a high
specificity. Figure 18-7 shows that by setting the cut value to 0.6, you can
simultaneously achieve about 93 percent specificity and 87 percent sensitivity.
» To have very few false negatives: Choose a lower cut value to give higher
sensitivity. Figure 18-7 shows you that if you set the cut value to 0.3, you can
have almost perfect sensitivity because you’ll be at almost 100 percent, but
your specificity will be only about 75 percent, meaning you’ll have a 25 percent
false positive rate.
The software may optionally display the area under the ROC curve (abbreviated
AUC), along with its standard error and a p value. This is another measure of how
good the predictive model is. The diagonal line has an AUC of 0.5, and there is a
statistical test comparing your AUC to the diagonal line. Under α = 0.05, if the p
value < 0.05, it indicates that your model is statistically significantly better than
the diagonal line at accurately predicting your outcome.

Heads Up: Knowing What Can Go Wrong
with Logistic Regression
Logistic regression presents many of the same potential pitfalls as ordinary least-
squares regression (see Chapters 16 and 17), as well as several that are specific to
logistic regression. Watch out for some of the more common pitfalls:
» Don’t fit a logistic function to non-logistic data: Don’t use logistic regres-
sion to fit data that doesn’t behave like the logistic S curve. Plot your grouped
data (as shown earlier in Figure 18-1b), and if it’s clear that the fraction of
positive outcomes isn’t leveling off at Y 0 or Y 1 for very large or very
small X values, then logistic regression is not the correct modeling approach.
The H-L test described earlier under the section “Assessing the adequacy of
the model” provides a statistical test to determine if your data qualify for
logistic regression. Also, in Chapter 19, we describe a more generalized logistic
model that contains other parameters for the upper and lower leveling-
off values.
» Watch out for collinearity and disappearing significance: When you are
doing any kind of regression and two or more predictor variables are strongly
related with each other, you can be plagued with problems of collinearity. We
describe this problem in Chapter 17, and potential modeling solutions in
Chapter 20.
» Check for inadvertent reverse-coding of the outcome variable: The

outcome variable should always be coded as 1 for a yes outcome and 0 for a
no outcome (refer to Table 18-1 for an example). If the variable in the data set
is coded using characters, you should recode an outcome variable using the
0/1 coding. It is important you do the coding yourself, and do not leave it to an
automated function in the program, because it may inadvertently reverse the
coding so that 1 = no and 0 = yes. This error of reversal won’t affect any p
values, but it will cause all your ORs and their CIs to be the reciprocals of what
they would have been, meaning they will refer to the odds of no rather than
the odds of yes.
» Don’t misinterpret odds ratios for categorical predicators: Categorical

predictors should be coded numerically as we describe in Chapter 8. It is
important to ensure that proper indicator variable coding is used, and these
variables are introduced properly in the model, as described in Chapter 17.
Also, be careful not to misinterpret odds ratios for numerical predictors, and be
mindful of the complete separation problem, as described in the following
sections.

Don’t misinterpret odds ratios
for numerical predictors
The OR always represents the factor by which the odds of getting the outcome
event increases when the predictor increases by exactly one unit of measure,
whatever that unit may be. Sometimes you may want to express the OR in
more convenient units than what the data was recorded in. For the example in
Table 18-1, the OR for dose as a predictor of death is 1.0115 per REM. This isn’t too
meaningful because one REM is a very small increment of radiation. By raising
1.0115 to the 100th power, you get the equivalent OR of 3.1375 per 100 REMs, and
you can express this as, “Every additional 100 REMs of radiation more than triples
the odds of dying.”
The value of a regression coefficient depends on the units in which the corre-
sponding predictor variable is expressed. So the coefficient of a height variable
expressed in meters is 100 times larger than the coefficient of height expressed in
centimeters. In logistic regression, ORs are obtained by exponentiating the coef-
ficients, so switching from centimeters to meters corresponds to raising the OR
(and its confidence limits) to the 100th power.
Beware of the complete separation

problem
Imagine your logistic regression model perfectly predicted the outcome, in that
every individual positive for the outcome had a predicted probability of 1.0, and
every individual negative for the outcome had a 0 predicted probability. This is
called perfect separation or complete separation, and the problem is called the perfect
predictor problem. This is a nasty and surprisingly frequent problem that’s unique
to logistic regression, which highlights the sad fact that a logistic regression
model will fail to converge in the software if the model fits perfectly!
If the predictor variable or variables in your model completely separate the yes
outcomes from the no outcomes, the maximum likelihood method will try to make
the coefficient of that variable infinite, which usually causes an error in the soft-
ware. If the coefficient is positive, the OR tries to be infinity, and if it is negative,
it tries to be 0. The SE of the OR tries to be infinite, too. This may cause your CI to
have a lower limit of 0, an upper limit of infinity, or both.
Check out Figure 18-8, which visually describes the problem. The regression is
trying to make the curve come as close as possible to all the data points. Usually it
has to strike a compromise, because there’s a mixture of 1s and 0s, especially in
the middle of the data. But with perfectly separated data, no compromise is neces-
sary. As b becomes infinitely large, the logistic function morphs into a step func-
tion that touches all the data points (observe where b = 5).

While it is relatively easy to identify if there is a perfect predictor in your data set
by looking at frequencies, you may run into the perfect predictor problem as a
result of a combination of predictors in your model. Unfortunately, there aren’t
any great solutions to this problem. One proposed solution called the Firth correc-
tion allows you to add a small number roughly equivalent to half an observation
to the data set that will disrupt the complete separation. If you can do this correc-
tion in your software, it will produce output, but the results will likely be unstable
(very near 0, or very near infinity). The approach of trying to fix the model by
changing the predictors would not make sense, since the model fits perfectly. You
may be forced to abandon your logistic regression plans and instead provide a
descriptive analysis.
FIGURE 18-8:
Visualizing the
complete
separation (or
perfect predictor)
problem in
logistic
regression.
Figuring Out the Sample Size You

Need for Logistic Regression
Estimating the required sample size for a logistic regression can be a pain, even
for a simple one-predictor model. You will have no problem specifying desired
power and α level (see Chapter 3 for more about these items). And, you can state
the effect size of importance as an OR.

Assuming a one-predictor model, the required sample size for logistic regression
also depends on the relative frequency of yes and no outcomes, and how the pre-
dictor variable is distributed. And with multiple predictors in the model, deter-
mining sample size is even more complicated. So for a rigorous sample-size
calculation for a study that will use a logistic regression model with multiple pre-
dictors, you may have no choice but to seek the help of a professional
statistician.
Here are two simple approaches you can use if your logistic model has only one
predictor. In each case, you replace the logistic regression equation with another
equation that is somewhat equivalent, and then do a sample-size calculation
based on that. It’s not an ideal solution, but it can give you an answer that’s close
enough for planning purposes.
» If the predictor is a dichotomous category (a yes/no variable), logistic

regression gives the same p value you get from analyzing a fourfold table.
Therefore, you can use the sample-size calculations we describe in
Chapter 12.
» If the predictor is a continuous numerical quantity (like age), you can

pretend that the outcome variable is the predictor, and age is the outcome.
We realize this flips the cause-and-effect relationship backwards, but if you
allow that conceptual flip, then you can ask whether the two different
outcome groups have different mean values for the predictor. You can test
that question with an unpaired Student t test, so you can use the sample-size
calculations we describe in Chapter 11.

IN THIS CHAPTER
» Using Poisson regression to analyze

counts and event rates
» Getting a grip on nonlinear regression
» Smoothing data without making any

assumptions
Chapter 19
Other Useful Kinds
of Regression
T
his chapter covers regression approaches you’re likely to encounter in bio-
statistical work that are not covered in other chapters. They’re not quite as
common as straight-line regression, multiple regression, and logistic regres-
sion (described in Chapters 16, 17, and 18, respectively), but you should be aware of
them. We don’t go into a lot of detail, but we describe what they are, the circum-
stances under which you may want to use them, how to execute the models and
interpret the output, and special situations you may encounter with these models.
Note: We also don’t cover survival regression in this chapter, even though it’s one
of the most important kinds of regression analysis in biostatistics. Survival analy-
sis is the theme of Part 6 of this book, and is the topic of Chapter 23.
Analyzing Counts and Rates

with Poisson Regression
Statisticians often have to analyze outcomes consisting of the number of occur-
rences of an event over some interval of time, such as the number of fatal highway
accidents in a city in a year. If the occurrences seem to be getting more common
CHAPTER 19 Other Useful Kinds of Regression 271

as time goes on, you may want to perform a regression analysis to see whether the
upward trend is statistically significant (meaning not due to natural random fluc-
tuations). If it is, you may want to create an estimate of the annual rate of increase,
including a standard error (SE) and confidence interval (CI).
Some analysts use ordinary least-squares regression as described in Chapter 16 on

such data, but event counts don’t really meet the least-squares assumptions, so
the approach is not technically correct. Event counts aren’t well-approximated as
continuous, normally-distributed data unless the counts are very large. Also, their
variability is neither constant nor proportional to the counts themselves. So
straight-line or multiple least-squares regression is not the best choice for event
count data.
Because independent random events like highway accidents should follow a Pois-
son distribution (see Chapter 24), they should be analyzed by a kind of regression
designed for Poisson outcomes. And — surprise, surprise — this type of special-
ized regression is called Poisson regression.
Introducing the generalized linear model

Most statistical software packages don’t offer a command or function explicitly
called Poisson regression. Instead, they offer a more general regression technique
called the generalized linear model (GLM).
Don’t confuse the generalized linear model with the very similarly named general
linear model. It’s unfortunate that these two names are almost identical, because
they describe two very different things. Now, the general linear model is usually
abbreviated LM, and the generalized linear model is abbreviated GLM, so we will use
those abbreviations. (However, some old textbooks from the 1970s may use GLM to
mean LM, because the generalized linear model had not been invented yet.)
GLM is similar to LM in that the predictor variables usually appear in the model as
the familiar linear combination:
c0 c1 x 1 c2 x 2 c3 x 3 ...
where the x’s are the predictor variables, and the c’s are the regression coefficients
(with c 0 being called a constant term, or intercept).
But GLM extends the capabilities of LM in two important ways:
» With LM, the outcome is assumed to be a continuous, normally distributed

variable. But with GLM, the outcome can be continuous or an integer. It can

follow one of several different distribution functions, such as normal, expo-
nential, binomial (as in logistic regression), or Poisson.
» With LM, the linear combination becomes the predicted value of the outcome,
but with GLM, you can specify a link function. The link function is a transforma-
tion that turns the linear combination into the predicted value. As we note in
Chapter 18, logistic regression applies exactly this kind of transformation: Let’s
call the linear combination V. In logistic regression, V is sent through the
logistic function 1 / 1 e V to convert it into a predicted probability of
having the outcome event. So if you select the correct link function, you can
use GLM to perform logistic regression.
GLM is the Swiss army knife of regression. If you select the correct link function,
you can use it to do ordinary least-squares regression, logistic regression, Poisson
regression, and a whole lot more. Most statistical software offers a GLM function;
that way, other specialized regressions don’t need to be programmed. If the soft-
ware you are using doesn’t offer logistic or Poisson regression, check to see
whether it offers GLM, and if it does, use that instead. (Flip to Chapter 4 for an
introduction to statistical software.)
Running a Poisson regression

Suppose that you want to study the number of fatal highway accidents per year in
a city. Table 19-1 shows some made-up fatal-accident data over the course of
12 years. Figure 19-1 shows a graph of this data, created using the R statistical
software package.
Running a Poisson regression is similar in many ways to running the other com-
mon kinds of regression, but there are some differences. Here are the steps:
1. As with any regression, prepare your predictor and outcome variables in

your data.
For this example, you have a row of data for each year, so year is the experi-
mental unit. For each row, you have a column containing the outcome values,
which is number of accidents each year (Accidents). Since you have one
predictor — which is year — you have a column for Year.
2. Tell the software which variables are the predictor variables, and which
one is the outcome.

3. Tell the software what kind of regression you want it to carry out by
specifying the family of the dependent variable’s distribution and the link
function.
Step 3 is not obvious, and you may have to consult your software’s help file. In
the R program, as an example, you have to specify both family and link in a
single construction, which looks like this:
glm formula Accidents Year, family poisson link “identity”
This code tells R that the outcome is the variable Accidents, the predictor is the
variable Year, and the outcome variable follows the Poisson family of distribu-
tions. The code link “identity” tells R that you want to fit a model in which
the true event rate rises in a linear fashion, meaning that it increases by a
constant amount each year.
4. Execute the regression and obtain the output.
The next step is to interpret the output.
TABLE 19-1 Yearly Data on Fatal Highway Accidents in One City

Calendar Year Fatal Accidents
2010 10
2011 12
2012 15
2013 8
2014 8
2015 15
2016 4
2017 20
2018 20
2019 17
2020 29
2021 28

FIGURE 19-1:
Yearly data on
fatal highway
accidents
in one city.
Interpreting the Poisson regression output

After you follow the steps for running a Poisson regression in the preceding
section, the program produces output like that shown in Figure 19-2, which is the
Poisson regression output from R’s generalized linear model function (glm).
FIGURE 19-2:
Poisson
regression
output.
This output has the same general structure as the output from other kinds of
regression. The most important parts of it are the following:
» In the coefficients table (labeled Coefficients:), the estimated regression

coefficient for Year is 1.3298, indicating that the annual number of fatal
accidents is increasing by an estimated 1.33 accidents per year.
» The standard error (SE) is labeled Std. Error and is 0.3169, indicating the
precision of the estimated rate increase per year. From the SE, using the rules
given in Chapter 10, the 95 percent confidence interval (CI) around the
estimated annual increase is approximately 1.3298 1.96 x 0.3169, which
gives a 95 percent CI of 0.71 to 1.95 (around the estimate 1.33).

» The column labeled z value contains the value of the regression coefficient
divided by its SE. It’s used to calculate the p value that appears in the last
column of the table.
» The last column, labeled Pr | z | , is the p value for the significance of the
increasing trend estimated at 1.33. The Year variable has a p value of 2.71
e-05, which is scientific notation (see Chapter 2) for 0.0000271. Using α = 0.05,
the apparent increase in rate over the 12 years would be interpreted as highly
statistically significant.
» AIC (Akaike’s Information Criterion) indicates how well this model fits the data.
The value of 81.72 isn’t useful by itself, but it’s very useful when choosing
between two alternative models, as we explain later in this chapter.
R software can also provide the predicted annual event rate for each year, from
which you can add a trend line to the scatter graph, indicating how you think the
true event rate may vary with time (see Figure 19-3).
FIGURE 19-3:
Poisson
regression,
assuming a
constant increase
in accident rate
per year with
trend line.
Discovering other uses for

Poisson regression
The following sections describe other uses of R’s GLM function in performing
Poisson regression.

Examining nonlinear trends
The straight line in Figure 19-3 doesn’t account for the fact that the accident rate
remained low for the first few years and then started to climb rapidly after 2016.
Perhaps the true trend isn’t a straight line, where the rate increases by the same
amount each year. It may instead be an exponential increase, where the rate
increases by a certain percentage each year. You can have R fit an exponential
increase by changing the link option from identity to log in the statement that
invokes the Poisson regression:
glm(formula = Accidents ~ Year, family = poisson(link = “log”))
This produces the output shown in Figure 19-4 and graphed in Figure 19-5.
FIGURE 19-4:
Output from an
exponential trend
Poisson
regression.
Because of the log link used in this regression run, the coefficients are related to
the logarithm of the event rate. Thus, the relative rate of increase per year is
obtained by taking the antilog of the regression coefficient for Year. This is done
by raising e (the mathematical constant 2.718. . .) to the power of the regression
coefficient for Year: e 0.10414 , which is about 1.11. So, according to an exponential
increase model, the annual accident rate increases by a factor of 1.11 each year —
meaning there is an 11 percent increase each year. The dashed-line curve in
Figure 19-4 shows this exponential trend, which appears to accommodate the
steeper rate of increase seen after 2016.
Comparing alternative models

The bottom of Figure 19-4 shows the AIC value for the exponential trend model is
78.476, which is about 3.2 units lower than for the linear trend model in
Figure 19-2 (AIC 81.72 ). Smaller AIC values indicate better fit, so the true trend
is more likely to be exponential rather than linear. But you can’t conclude that the
model with the lower AIC is really better unless the AIC is about six units better.
So in this example, you can’t say for sure whether the trend is linear or exponen-
tial, or potentially another distribution. But the exponential curve does seem to
predict the high accident rates seen in 2020 and 2021 better than the linear trend
model.

FIGURE 19-5:
Linear and
exponential
trends fitted to
accident data.
Working with unequal observation intervals

In this fatal accident example, each of the 12 data points represents the accidents
observed during a one-year interval. But imagine analyzing the frequency of
emergency department visits for patients after being treated for emphysema,
where there is one data point per patient. In that case, the width of the observa-
tion interval may vary from one individual in the data to another. GLM lets you
provide an interval width along with the event count for each individual in the
data. For arcane reasons, many statistical programs refer to this interval-width
variable as the offset.
Accommodating clustered events

The Poisson distribution applies when the observed events are all independent
occurrences. But this assumption isn’t met if events occur in clusters. Suppose
you count individual highway fatalities instead of fatal highway accidents. In that
case, the Poisson distribution doesn’t apply, because one fatal accident may kill
several people. This is what is meant by clustered events.
The standard deviation (SD) of a Poisson distribution is equal to the square root of
the mean of the distribution. But if clustering is present, the SD of the data is
larger than the square root of the mean. This situation is called overdispersion.
GLM in R can correct for overdispersion if you designate the distribution family
quasipoisson rather than poisson, like this:
glm(formula = Accidents ~ Year, family = quasipoisson(link = “log”))

Anything Goes with Nonlinear Regression
Here, we finally present the potentially most challenging type of least-squares
regression, and that’s general nonlinear least-squares regression, or nonlinear
curve-fitting. In the following sections, we explain how nonlinear regression is
different from other kinds of regression. We also describe how to run and inter-
pret a nonlinear regression using an example from drug research, and we show
you some tips involving equivalent functions.
Distinguishing nonlinear regression

from other kinds
In the kinds of regression we describe earlier in this chapter and in Chapters 16, 17,
and 18, the predictor variables and regression coefficients always appear in the
model as a linear combination: c0 c1 x 1 c 2 x 2 c 3 x 3 ... c n x n. But in non-
linear regression, the coefficients no longer have to appear paired up to be multi-
plied by predictor variables (like c 2 x 2). In nonlinear regression, coefficients have a
more independent existence, and can appear on their own anywhere in the
formula. Actually, the term coefficient implies a number that’s multiplied by a
variable’s value. This means that technically, you can’t have a coefficient that
isn’t multiplied by a variable, so when this happens in nonlinear regression,
they’re referred to instead as parameters.
The formula for a nonlinear regression model may be any algebraic expression. It
can involve sums, differences, products, ratios, powers, and roots. These can be
combined together in a formula with logarithmic, exponential, trigonometric, and
other advanced mathematical functions (see Chapter 2 for an introduction to these
items). The formula can contain any number of predictor variables, and any num-
ber of parameters. In fact, nonlinear regression formulas often contain many
more parameters than predictor variables.
Unlike other types of regression covered in this chapter and book, where a regres-
sion command and code are used to generate output, developing a full-blown
nonlinear regression model is more of a do-it-yourself proposition. First, you
have to decide what function you want to fit to your data, making this choice from
the infinite number of possible functions you could select. Sometimes the general
form of the function is determined or suggested by a scientific theory. Using a
theory to guide your development of a nonlinear function means relying on a the-
oretical or mechanistic function, which is more common in the physical sciences
than life sciences. If you choose your nonlinear function based on a function with
a generally similar shape, you are using an empirical function. After choosing the
function, you have to provide starting estimates for the value of each of the
parameters appearing in the function. After that, you can execute the regression.
The software tries to refine your estimates using an iterative process that may or

may not converge to an answer, depending on the complexity of the function
you’re fitting and how close your initial estimates are to the truth. And in addition
to attending to these unique issues, analysts running a nonlinear regression face
all the other complications of multivariate regression, such as collinearity, as
described in Chapter 17.
Checking out an example

from drug research
One common nonlinear regression problem arises in drug development research.
As soon as scientists start testing a promising new compound, they want to deter-
mine some of its basic pharmacokinetic (PK) properties. PK properties describe
how the drug is absorbed, distributed, modified, and eliminated by the body. Typ-
ically, the earliest Phase I clinical trials attempt to obtain basic PK data as a sec-
ondary objective of the trial, while later-phase trials may be designed specifically
to characterize the PKs of the drug accurately and in great detail.
Raw PK data often consist of the concentration level of the drug in the partici-
pant’s blood at various times after a dose of the drug is administered. Consider a
Phase I trial, in which 10,000 micrograms (μg) of a new drug is given as a single
bolus, which is a rapid injection into a vein, in each participant. Blood samples are
drawn at predetermined times after dosing and are analyzed for drug concentra-
tions. Hypothetical data from one participant are shown in Table 19-2 and graphed
in Figure 19-6. The drug concentration in the blood is expressed in units of μg per
deciliter ( g / dL ). Remember, a deciliter is one-tenth of a liter.
Several basic PK parameters, such as maximum concentration, time of maximum

concentration, area under the curve (AUC), are usually calculated directly from the
concentration-versus-time data, without having to fit any curve to the points. But
two important parameters are usually obtained from a regression analysis:
» The volume of distribution (V d): This is the effective volume of fluid or tissue
through which the drug is distributed in the body. This effective volume could
be equal to the blood volume, but could be greater if the drug also spreads
through fatty tissue or other parts of the body. If you know the dose of the drug
infused (Dose), and you know the blood plasma concentration at the moment of
infusion (C 0), you can calculate the volume of distribution as Vd Dose / C 0.
But you can’t directly measure C 0. By the time the drug has distributed evenly
throughout the bloodstream, some of it has already been eliminated from the
body. So C 0 has to be estimated by extrapolating the measured concentrations
backward in time to the moment of infusion (Time 0).
» The elimination half-life (λ): The time it takes for half of the drug in the body
to be eliminated.

TABLE 19-2 Blood Drug Concentration versus Time for One Participant
Time after Dosing (In Hours) Drug Concentration in Blood (μg/dL)
0.25 57.4
0.5 54.0
1 44.8
1.5 52.7
2 43.6
3 40.1
4 27.9
6 20.6
8 15.0
12 10.0
FIGURE 19-6:
The blood
concentration of
an intravenous
drug decreases
over time in one
participant.
PK theory is well-developed and predicts under a set of reasonable assumptions

that the drug concentration (Conc) in the blood following a bolus infusion should
vary with time (Time) according to the equation:
keTime
Conc C oe

where ke is the elimination rate constant. ke is related to the elimination half-life (λ)
according to the formula: 0.693 / ke , where 0.693 is the natural logarithm of 2.
So, if you can fit the preceding equation to your Conc-versus-Time data in
Table 19-2, you can estimate C 0, from which you can calculate Vd . You can also
estimate ke , from which you can calculate λ.
The preceding equation is nonlinear and includes parameters, withke appearing in

the exponent. Before nonlinear regression software became widely available, ana-
lysts would take a shortcut by shoehorning this nonlinear regression problem into
a straight-line regression by working with the logarithms of the concentrations.
But that approach can’t be generalized to handle more complicated equations that
often arise.
Running a nonlinear regression

Nonlinear curve-fitting is supported by many modern statistics packages, like
SPSS, SAS, GraphPad Prism, and R (see Chapter 4). It is possible (though not easy)
to set up calculations in Microsoft Excel. In addition, the web page http://
StatPages.info/nonlin.html can fit any function you can write involving up to
eight independent variables and up to eight parameters. Here are the steps we use
to do nonlinear regression in R:
1. Create a vector of time data, and a vector of concentration data.
In R, you can develop arrays called vectors from data sets, but in this example,
we create each vector manually, naming them Time and Conc, using the data
from Table 19-2:
Time c 0.25, 0.5, 1, 1.5, 2, 3, 4, 6, 8, 12
Conc c 57.4, 54.0, 44.8, 52.7, 43.6, 40.1, 27.9, 20.6, 15.0, 10.0
In the two preceding equations, c is a built-in R function for combine that

creates an array (see Chapter 2) as a vector from the lists of numbers.
2. Specify the equation to be fitted to the data, using the algebraic syntax
your software requires.
We write the equation using R’s algebraic syntax this way: Conc ~ C0 * exp(– ke *
Time), where Conc and Time are your vectors of data, and C0 and ke are
parameters you set.
3. Tell the software that C0 and ke are parameters to be fitted, and provide
initial estimates for these values.
Nonlinear curve-fitting is a complicated task solved by the software through

iteration. This means you give it some rough estimates, and it refines them into

closer estimates to the truth, repeating this process until it arrives at the
best-fitting, least-squares solution.
Coming up with starting estimates for nonlinear regression problems can be

tricky. It’s more of an art than a science. If the parameters have physiological
meaning, you may be able to make a guess based on known physiology or past
experience. Other times, your estimates have to be trial and error. To improve
your estimates, you can graph your observed data in Microsoft Excel, and then
superimpose a curve from values calculated from the function for various
parameter guesses that you type in. That way, you can play around with the
parameters until the curve is at least in the ballpark of the observed data.
In this example, C0 (variable C0) is the concentration you expect at the moment
of dosing (at t 0). From Figure 19-6, it looks like the concentration starts out
around 50, so you can use 50 as an initial guess for C0. The ke parameter (variable
ke) affects how quickly the concentration decreases with time. Figure 19-6
indicates that the concentration seems to decrease by half about every few
hours, so λ should be somewhere around 4 hours. Because 0.693 / ke ,
a little algebra gives the equation ke = 0.693/X. If you plug in 4 hours for X, you
get ke = 0.693/4 = 0.2, so you may try 0.2 as a starting guess for ke. You tell
R the starting guesses by using the syntax: start=list(C0 50 , ke 0.2).
The statement in R for nonlinear regression is nls, which stands for nonlinear
least-squares. The full R statement for executing this nonlinear regression model
and summarizing the output is:
summary nls Conc C0 * exp ke * Time , start list C0 50, ke 0.2
Interpreting the output

As complicated as nonlinear curve-fitting may be, the output is actually quite
simple. It is formatted and interpreted like the output from ordinary linear regres-
sion. Figure 19-7 shows the relevant part of R’s output for this example.
FIGURE 19-7:
Results of
nonlinear
regression in R.

In Figure 19-7, the output first restates the model being fitted. Next, what would
normally be called the coefficients table is presented, only this time, it is labeled
Parameters. It has a row for every adjustable parameter that appears in the func-
tion. Like other regression tables, it shows the fitted value for the parameter under
Estimate, its standard error (SE) under Std. Error, and the p value under Pr(>|t|)
indicating whether that parameter was statistically significantly different from
zero. The output estimatesC 0 at 59.5 2.3 μg/dL and ke at 0.163 0.0164hr 1
because first-order rate constants have units of per time. From these values, you
can calculate the PK parameters you want:
» Volume of distribution: Vd Dose / C 0 10,000 g / 59.5 g / dL 168 dL,

or 16.8 liters. Since this amount is several times larger than the blood volume
of the average human, the results indicate that this drug is going into other
parts of the body besides the blood.
» Elimination half-time: λ = 0.693/ke = 0.693/0.163hr1, or 4.25 hours. This

result means that after 4.25 hours, only 50 percent of the original dose is left
in the body. After twice as long, which is 8.5 hours, only 25 percent of the
original dose remains, and so on.
How precise are these PK parameters? In other words, what is their SE? Unfortu-
nately, uncertainty in any measured quantity will propagate through a mathemat-
ical expression that involves that quantity, and this needs to be taken into account
in calculating the SE. To do this, you can use the online calculator at https://
statpages.info/erpropgt.html. Choose the estimator designed for two vari-
ables, and enter the information from the output into the calculator. You can cal-
culate that the Vd 16.8 0.65 liters, and 4.25 0.43 hours.
R can be asked to generate the predicted value for each data point, from which you
can superimpose the fitted curve onto the observed data points, as in
Figure 19-8.
R also provides the residual standard error (labeled Residual std. err. in Figure 19-7),
which is defined as the standard deviation of the vertical distances of the observed
points from the fitted curve. The value from the output of 3.556 means that the
points scatter about 3.6 μg/dL above and below the fitted curve. Additionally, R
can be asked to provide Akaike’s Information Criterion (AIC), which is useful in
selecting which of several possible models best fits the data.

FIGURE 19-8:
Nonlinear model
fitted to drug
concentration
data.
Using equivalent functions to fit the

parameters you really want
It’s inconvenient, annoying, and error-prone to have to perform manual calcula-
tions on the parameters you obtain from nonlinear regression output. It’s so much
extra work to read the output that contains the estimates you need, likeC 0 and the
ke rate constant, then manually calculate the parameters you want, like Vd and λ.
It’s even more work to obtain the SEs. Wouldn’t it be nice if you could get Vd and
λ and their SEs directly from the nonlinear regression program? Well, in many
cases, you can!
Because nonlinear regression involves algebra, some fancy math footwork can
help you out. Very often, you can re-express the formula in an equivalent form
that directly involves calculating the parameters you actually want to know.
Here’s how it works for the PK example we use in the preceding sections.
Algebra tells you that because Vd Dose / C 0 , then C 0 Dose / Vd . So why not use
Dose / Vd instead of C 0 in the formula you’re fitting? If you do, it becomes
Conc Dose e keTime
. And you can go even further than that. It turns out that
vd
keTime
a first-order exponential-decline formula can be written either as e or as the
Time
t 1/2
algebraically equivalent form 2 .

Applying both of these substitutions, you get the equivalent model:
Dose Time /
C 2 , which produces exactly the same fitted curve as the original
vd
model. But it has the tremendous advantage of giving you exactly the PK param-
eters you want, which are Vd and λ, rather than C 0 and ke which require post-
processing with additional calculations.
From the original description of this example, you already know that Dose = 10,000
μg, so you can substitute this value for Dose in the formula to be fitted. You’ve
already estimated λ (variable tHalf) as 4 hours. Also, you estimated C 0 as about 50
μg/dL from looking at Figure 19-6, as we describe earlier. This means you can
estimate Vd (variable Vd) as 10, 000 / 50, which is 200 dL. With these estimates, the
final R statement is
summary(nls(Conc ~ (10000 / Vd) * 2 ^ ( Time / tHalf),

start list(Vd 200, tHalf 4)))
which produces the output shown in Figure 19-9.
FIGURE 19-9:
Nonlinear
regression that
estimates the
PK parameters
you want.
From Figure 19-9, you can see the direct results for Vd and tHalf. Using the output,
you can estimate that the Vd is 168.2 6.5 dL (or 16.8 0.66 liters), and λ is
4.24 0.43 hours.
Smoothing Nonparametric
Data with LOWESS
Sometimes you want to fit a smooth curve to a set of points that don’t seem to
conform to a common, recognizable distribution, such as normal, exponential,
logistic, and so forth. If you can’t write an equation for the curve you want to fit,
you can’t use linear or nonlinear regression techniques. What you need is essen-
tially a nonparametric regression approach, which would not try to fit any formula/
model to the relationship, but would instead just try to draw a smooth line through
the data points.

Several kinds of nonparametric data-smoothing methods have been developed.
A popular one, called LOWESS, stands for Locally Weighted Scatterplot Smoothing.
Many statistical programs can perform a LOWESS regression. In the following
sections, we explain how to run a LOWESS analysis and adjust the amount of
smoothing (stiffness) of the curve.
Running LOWESS
Suppose that you discover a new hormone called XYZ believed to be produced in
women’s ovaries throughout their lifetimes. Research suggests blood levels of
XYZ should vary with age, in that they are low before going through puberty and
after completing menopause, and high during child-bearing years. You want to
characterize and quantify the relationship between XYZ levels and age as accu-
rately as possible.
Suppose that for your analysis, you are allowed to obtain 200 blood samples drawn
from consenting female participants aged 2 to 90 years for another research proj-
ect, and analyze the specimens for XYZ levels. A graph of XYZ level versus age may
look like Figure 19-10.
FIGURE 19-10:
The relationship
between age and
hormone
concentration
doesn’t conform
to a simple
function.
In Figure 19-10, you can observe a lot of scatter in these points, which makes it
hard to see the more subtle aspects of the XYZ-age relationship. At what age does
the hormone level start to rise? When does it peak? Does it remain fairly constant
throughout child-bearing years? When does it start to decline? Is the rate of
decline after menopause constant, or does it change with advancing age?

It would be easier to answer those questions if you had a curve that represented
the data without all the random fluctuations of the individual points. How would
you go about fitting such a curve to these data? LOWESS to the rescue!
Running LOWESS regression in R is similar to other regression. You need to tell R

which variable represents x and which one represents y, and it does the rest. If
your variables in R are actually named x and y, the R instruction to run a LOWESS
regression is the following: lowess( x , y , f 0.2). (We explain the f 0.2 part in the
following section.)
Unlike other forms of regression, LOWESS doesn’t produce a coefficients table.

The only output is a table of smoothed y values, one for each data point, which you
can save as a data file. Next, using other R commands, you can plot the x and y
points from your data, and add a smoothed line superimposed on the scatter graph
based on the smoothed y values. Figure 19-11 shows this plot.
FIGURE 19-11:
The fitted
LOWESS curve
follows the shape
of the data,
whatever it
may be.
In Figure 19-11, the smoothed curve seems to fit the data quite well across all ages
except the lower ones. The individual data points don’t show any noticeable
upward trend until age 12 or so, but the smoothed curve starts climbing right
from age 3. The curve completes its rise by age 20, and then remains flat until
almost age 50, when it starts declining. The rate of decline seems to be greatest
between ages 50 to 65, after which it declines less rapidly. These subtleties would
be very difficult to spot just by looking at the individual data points without any
smoothed curve.

Adjusting the amount of smoothing
R’s LOWESS program allows you adjust the stiffness of the fitted curve by specify-
ing a smoothing fraction, called f, which is a number between 0 and 1. Figure 19-12
shows what the smoothed curve looks like using three different smoothing
fractions.
FIGURE 19-12:
You can adjust
the smoothness
of the fitted curve
by adjusting the
smoothing
fraction.
Looking at Figure 19-12, you can observe the following:
» Setting f 0.667 produces a rather stiff curve that rises steadily between ages
2 and 40, and then declines steadily after that (see dashed line). The value
0.667 represents 2/3, which is what R uses as the default value of the f
parameter if you don’t specify it. This curve misses important features of the
data, like the low pre-puberty hormone levels, the flat plateau during child-
bearing years, and the slowing down of the yearly decrease above age 65. You
can say that this curve shows excessive bias, systematically departing from
observed values in various places along its length.
» Setting f 0.1, which is at a lower extreme, produces a very jittery curve with
a lot of up-and-down wiggles that can’t possibly relate to actual ages, but
instead reflect random fluctuations in the data (see dark, solid line). You can
say that this curve shows excessive variance, with too many random fluctua-
tions along its length.

» Setting f 0.2 produces a curve that’s stiff enough not to have random
wiggles (see medium, solid line). Yet, the curve is flexible enough to show that
hormone levels are fairly low until age 10, reach their peak at age 20, stay
fairly level until age 50, and then decline, with the rate of decline slowing down
after age 70. This curve appears to strike a good balance, with low bias and
low variance.
Whenever you do LOWESS regression, you have to explore different smoothing

fractions to find the sweet spot that gives the best tradeoff between bias and var-
iance. You are trying to display real features of the curve while smoothing out the
random noise. When used properly, LOWESS regression can be helpful as a way of
gleaning insight from noisy data.

IN THIS CHAPTER
» Choosing potential confounders for

your regression model
» Using a modeling approach to

develop a final model
» Adding interactions to the final

model
» Interpreting the final model for

causal inference
Chapter 20
Getting the Hint from
Epidemiologic Inference
I
n Parts 5 and 6, we describe different types of regression, such as ordinary
least-squares regression, logistic regression, Poisson regression, and survival
regression. In each kind of regression we cover, we describe a situation in which
you are performing multivariable or multivariate regression, which means you are
making a regression model with more than one independent variable. Those
chapters describe the mechanics of fitting these multivariable models, but they
don’t provide much guidance on which independent variables to choose to try to
put in the multivariable model.
The chapters in Parts 5 and 6 also discuss model-fitting, which means the act of
trying to refine your regression model so that it optimally fits your data. When you
have a lot of candidate independent variables (or candidate covariates), part of
model-fitting has to do with deciding which of these variables actually fit in the
model and should stay in, and which ones don’t fit and should be kicked out. Part
of what guides this decision-making process are the mechanics of modeling and
model-fitting. The other main part of what guides these decisions is the hypoth-
esis you are trying to answer with your model, which is the focus of this chapter.
CHAPTER 20 Getting the Hint from Epidemiologic Inference 291

In this chapter, we revisit the concept of confounding from Chapter 7 and explain
how to choose candidate covariates for your regression model. We also discuss
modeling approaches and explain how to add interaction terms to your
final model.
Staying Clearheaded about Confounding

Chapter 7 discusses study design and terminology in epidemiology. As a reminder,
in epidemiology, exposure refers to a factor you hypothesize to cause a disease (or
outcome). In your regression model, the outcome is the dependent variable. The
exposure will be one of the covariates in your model. But what other covariates
belong in the model? How do you decide on a collection of candidate-independent
variables that you would even consider putting in a model with the exposure? The
answer is that you choose them on the basis of their status as a potential
confounder.
A confounder is a factor that meets these three criteria:
» It is associated with the exposure.
» It is associated with the outcome.
» It is not on the causal pathway between the exposure and outcome.
As an example, look at Figure 20-1, which illustrates a study of patients with

Type II diabetes where there is a hypothesized causal relationship between the
exposure of having served in the military and the negative outcome of having an
amputation due to diabetic complications.
As shown Figure 20-1, inability to exercise and low income are both seen as potential
confounders. That is because they are associated with both the exposure of mili-
tary service and the outcome of amputation, and they are not on the causal path-
way between military service and amputation. In other words, what is causing the
outcome of amputation is not also causing the patient’s inability to exercise, nor
is it also causing the patient to have low income. But whatever is causing the
patient’s amputation is also causing the patient’s retinopathy. That’s because
Type II diabetes causes poor circulation, which causes both retinopathy and
amputation. This means that retinopathy and amputation are on the same causal
pathway, and retinopathy cannot be considered a potential confounder.

FIGURE 20-1:
Example of how
confounders are
associated with
exposure and
outcome but are
not on the causal
pathway between
exposure and
outcome.
Avoiding overloading
You may think that choosing what covariates belong in a regression model is easy.
You just put all the confounders and the exposure in as covariates and you’re done,
right? Well, unfortunately, it’s not that simple. Each time you add a covariate to a
regression model, you increase the amount of error in the model by some amount —
no matter what covariate you choose to add. Although there is no official maximum
to the number of covariates in a model, it is possible to add so many covariates that
the software cannot compute the model, causing an error. In a logistic regression
model as discussed in Chapter 18, each time you add a covariate, you increase the
overall likelihood of the model. In Chapter 17, which focuses on ordinary least-
squares regression, adding a covariate increases your sum of squares.
What this means is that you don’t want to add covariates to your model that just
increase error and don’t help with the overall goal of model fit. A good strategy is
to try to find the best collection of covariates that together deal with as much error
as possible. For example, think of it like roommates who share apartment-
cleaning duties. It’s best if they split up the apartment and each clean different
parts of it, rather than insisting on cleaning up the same rooms, which would be
a waste of time. The term parsimony refers to trying to include the fewest
covariates in your regression model that explain the most variation in the depen-
dent variable. The modeling approaches discussed in the next section explain
ways to develop such parsimonious models.

Adjusting for confounders
When designing a regression analysis, you first have to decide: Are you doing an
exploratory analysis, or are you doing a hypothesis-driven analysis? If you are
doing an exploratory analysis, you do not have a pre-supposed hypothesis. Instead,
your aim is to answer the research question, “What group of covariates do I need
to include as independent variables in my regression to predict the outcome and
get the best model fit?” In this case, you need to select a set of candidate covari-
ates and then come up with modeling rules to decide which groups of covariates
produce the best-fitting model. In each chapter on regression in this book, we
provide methods of comparing models using model-fit statistics. You would use
those to choose your final model for your exploratory analysis. Exploratory analy-
ses are considered descriptive studies, and are weak study designs (see
Chapter 7).
But if you collected your data based on a hypothesis, you are doing a hypothesis-
driven analysis. Epidemiologic studies require hypothesis-driven analyses, where
you have already selected your exposure and outcome, and now you have to fit a
regression model predicting the outcome, but including your exposure and con-
founders as covariates. You know you need to include the exposure and the out-
come in every model you run. However, you may not know how to decide on which
confounders stay in the model.
Regardless of whether you are doing exploratory or hypothesis-driven modeling,

you need to make rules before you start modeling that describe how you will make
decisions about your final model and during your modeling process. You may
make a rule that all the covariates in your final model must be associated with a p
value that is statistically significant at α = 0.05. You can make other stipulations
about the final model, or the process of achieving the final model. What is impor-
tant is that you make the modeling rules and write them down before you start
modeling.
You then need to choose a modeling approach, which is the approach you will use
to determine which candidate confounders stay in the model with the exposure
and which ones are removed. There are three common approaches in regression
modeling (although analysts have their customized approaches). These approaches
don’t have official names, but we will use terms that are commonly used. They
are: forward stepwise, backward elimination, and stepwise selection.
» Forward stepwise: This is where one confounder covariate at a time is added

to the model in iterative models. If it does not meet rules to be kept in the
model, it is removed and never considered again in the model. Imagine you
were fitting a regression model with one exposure covariate and eight candi-
date confounders. Suppose that you add the first covariate with the exposure
and it meets modeling rules, so you keep it. But when you add the second

covariate, it does not meet the rules, so you leave it out. You keep doing this
until you run out of variables. Although forward stepwise can work if you have
very few variables, most analysts do not use this approach because it has been
shown to be sensitive to the order you choose in which to enter variables.
» Backward elimination: In this approach, the first model you run contains all
your potential covariates, including all the confounders and the exposure.
Using modeling rules, each time you run the model, you remove or eliminate
the confounder contributing the least to the model. You decide which one
that is based on modeling rules you set (such as which confounder has the
largest p value). Theoretically, after you pare away the confounders that do
not meet the rules, you will have a final model. In practice, this process can
run into problems if you have collinear covariates (see Chapters 17 and 18 for
a discussions of collinearity). Your first model — filled with all your potential
covariates — may error out for this reason, and not converge. Also, it is not
clear whether once you eliminate a covariate you should try it again in the
model. This approach often sounds better on paper than it works in practice.
» Stepwise selection: This approach combines the best of forward stepwise

and backward elimination. Starting with the same set of candidate covariates,
you choose which covariate to introduce first into a model with the exposure.
If this covariate meets modeling rules, it is kept, and if not, it is left out. This
continues along as if you are doing forward stepwise — but then, there’s a
twist. After you are done trying each covariate and you have your forward
stepwise model, you go back and try to add back the covariates you left out
one by one. Each time one seems to fit back in, you keep it and consider it part
of the working model. It is during this phase that collinearity between covariates
can become very apparent. After you try back the covariates you originally left
out and are satisfied that you were able to add back the ones that fit the
modeling rules, you can declare that you have arrived at the final model.
Once you produce your final model, check the p value for the covariate or covari-
ates representing your exposure. If they are not statistically significant, it means
that your hypothesis was incorrect, and after controlling for confounding, your
exposure was not statistically significantly associated with the outcome. However,
if the p value is statistically significant, then you would move on to interpret the
results for your exposure covariates from your regression model. After controlling
for confounding, your exposure was statistically significantly associated with
your outcome. Yay!
Use a spreadsheet to keep track of each model you run and a summary of the
results. Save this in addition to your computer code for running the models. It can
help you communicate with others about why certain covariates were retained and
not retained in your final model.

Computer software may include automated processes you can use for fitting mod-
els. We discourage you from using these in biostatistics because you want to have
a lot of control over how a model is being fitted to make it possible for you to
interpret the results. However, these processes can be used to create comparison
models — or to simulate improved models — which are perfectly reasonable
methods to explore ways to improve your model.
Understanding Interaction
(Effect Modification)
In Chapter 17, we touch on the topic of interaction (also known as effect modifica-
tion). This is where the relationship between an exposure and an outcome is
strongly dependent upon the status of another covariate. Imagine that you con-
ducted a study of laborers who had been exposed to asbestos at work, and you
found that being exposed to asbestos at work was associated with three times the
odds of getting lung cancer compared to not being exposed. In another study, you
found that individuals who smoked cigarettes had twice the odds of getting lung
cancer compared to those who did not smoke.
Knowing this, what would you predict are the odds of getting lung cancer for
asbestos-exposed workers who also smoke cigarettes, compared to workers who
aren’t exposed to asbestos and do not smoke cigarettes? Do you think it would be
additive — meaning three times for asbestos plus two times for smoking equals
five times the odds? Or do you think it would be multiplicative — meaning three
times two equals six times the odds?
Although this is just an example, it turns out that in real life, the effect of being
exposed to both asbestos and cigarette smoking represents a greater than multi-
plicative synergistic interaction (meaning much greater than six) in terms of the
odds for getting lung cancer. In other words, the risk of getting lung cancer for
cigarette smokers is dependent upon their asbestos-exposure status, and the risk
of lung cancer for asbestos workers is dependent upon their cigarette-smoking
status. Because the factors work together to increase the risk, this is a synergistic
interaction (with the opposite being an antagonistic interaction).
How and when do you model an interaction in regression? Typically, you first fit
your final model using a multivariate regression approach (see the earlier section
“Adjusting for confounders in regression” for more on this). Next, once the final
model is fit, you try to interact the exposure covariate or covariates with a con-
founder that you believe is the other part of the interaction. After that, you look at
the p value on the interaction term and decide whether or not to keep the
interaction.

Imagine making a model for the study of asbestos workers, cigarette smoking,
and lung cancer. The variable asbestos is coded 1 for workers exposed to asbestos
and 0 for workers not exposed to asbestos, and the variable smoker is coded 1 for
cigarette smokers and 0 for nonsmokers. The final model would already have
asbestos and smoker in it, so the interaction model would add the additional covari-
ate asbestos × smoker, which is called the higher order interaction term. For indi-
viduals who have a 0 for either asbestos or smoker or both, this term falls out of
their individual predicted probability (because 1 × 0 = 0, and 0 × 0 = 0). Therefore,
if this term is statistically significant, then individuals who qualify to include this
term in their individual predicted probability have a statistically significantly
greater risk of the outcome, and the interaction term should be kept in the model.
Getting Casual about Cause

Chapter 7 explains epidemiologic study designs and presents them in a pyramid
format. The closer to the top of the pyramid, the better the study design is at pro-
viding evidence for causal inference, meaning providing evidence of a causal
association between the exposure with the outcome (or in the case of a clinical
trial of an intervention, the intervention and the outcome). At the top of the
pyramid are systematic review and meta-analysis, where the results of similar
studies are combined and interpreted. Because systematic reviews and meta-
analyses combine results from other high-quality studies, they are at the very
top of the pyramid — meaning they provide the strongest evidence of a causal
association between the exposure or intervention and outcome.
An international organization called the Cochrane Collaboration organizes the

production of systematic reviews and meta-analyses to help guide clinicians.
Their reviews are internationally renowned for being high-quality and are avail-
able at www.cochrane.org.
The study designs on the evidence-based pyramid that could be answered with a
regression model include clinical trial, cohort study, case-control study, and
cross-sectional study. If in your final model your exposure is statistically signifi-
cantly associated with your outcome, you now have to see how much evidence you
have that the exposure caused the outcome. This section provides two methods by
which to evaluate the significant exposure and outcome relationship in your
regression: Rothman’s causal pie and Bradford Hill’s criteria of causality.
Rothman’s causal pie

Kenneth Rothman described how causes of an outcome are not determinate. In
other words, two people can have the same values of covariates and one will get

the outcome, and the other will not. We can’t say for sure what values of covari-
ates will mean that you will for sure get the outcome. But that doesn’t mean you
can’t make causal inferences. Rothman conceptualized cause as an empty pie tin,
and when the pie tin is filled 100 percent with pieces of risk contributed by various
causes, then the individual will experience the outcome. The exposure and con-
founders in your regression model represent these pieces.
For example, cigarette smoking is a very strong cause of lung cancer, as is occu-
pational exposure to asbestos. There are other causes, but for each individual,
these other causes would fill up small pieces of the causal pie for lung cancer.
Some may have a higher genetic risk factor for cancer. However, if they do not
smoke and stay away from asbestos, they will not fill up much of their pie tin, and
may have necessary but insufficient cause for lung cancer. However, if they include
both asbestos exposure and smoking in their tin, they are risking filling it up and
getting the outcome.
Bradford Hill’s criteria of causality

» Sir Bradford Hill was a British epidemiologist who put forth criteria for
causality that can be useful to consider when thinking of statistically signifi-
cant exposure–outcome relationships from final regression models. Although
there are more than the criteria we list here, we find the following criteria to
be the most useful when evaluating potential exposure–outcome causal
relationships in final models:
» First, consider if the data you are analyzing are from a clinical trial or cohort
study. If they are, then you will have met the criterion of temporality, which
means the exposure or intervention preceded the outcome and is especially
strong evidence for causation.
» If the estimate for the exposure in your regression model is large, you can say
you have a strong magnitude of association, and this is evidence of causation.
This is especially true if your estimate is larger than those of the confounders
in the model as well as similar estimates from the scientific literature.
» If your exposure shows a dose-response relationship with the outcome, it is

evidence of causation. In other words, if your regression model shows that
the more individuals smoke, the higher their risk for lung cancer, this is
evidence of causation (see Chapter 18 for more on dose-response
relationships).
» If the estimate is consistent in size and direction with other analyses, including
previous studies you’ve done and studies in the scientific literature, there is
more evidence for causation.

6
Analyzing
Survival Data
IN THIS PART . . .
Learn the definition of survival data and how to

handle censored observations (where you do not
observe the outcome as part of your study).
Prepare survival curves using the life-table and

Kaplan-Meier methods.
Estimate median survival times and survival rates at

specified times.
Compare survival curves between two or more groups

using the log-rank test.
Analyze multiple predictors of survival using Cox

proportional hazards regression.
IN THIS CHAPTER
» Beginning with the basics of

survival data
» Generating life tables and trying the

Kaplan-Meier method
» Applying some handy guidelines for

survival analysis
» Using survival data for even more

calculations
Chapter 21
Summarizing and
Graphing Survival Data
T
his chapter describes statistical techniques that deal with a special kind of
numerical data called survival data or time-to-event data. These data reflect
the interval from a particular starting point in time, such the date a patient
receives a certain diagnosis or undergoes a certain procedure, to the first or only
occurrence of a particular kind of event that represents an endpoint. Because these
techniques are often applied to situations where the endpoint event is death, we
usually call the use of these techniques survival analysis, even when the endpoint is
something less drastic (or final) than death. Survival data could include time from
resolution of a chronic illness symptom to its relapse, but it can also be a desirable
endpoint, such as time to remission of cancer, or time to recovery from an acute
condition. Throughout this chapter, we use terms and examples that imply that
the endpoint is death, such as saying survival time instead of time to event. However,
everything we say also applies to other kinds of endpoints.
You may wonder why you need a special kind of analysis for survival data in the
first place. Why not just treat survival times as ordinary numerical variables? Why
not summarize them as means, medians, standard deviations, and so on, and
graph them as histograms and box-and-whiskers charts? Why not compare sur-
vival times between groups with t tests and ANOVAs? Why not use ordinary least-
squares regression to explore how various factors influence survival time?
CHAPTER 21 Summarizing and Graphing Survival Data 301

In this chapter, we explain how survival data aren’t like ordinary numerical data
and why you need to use specific techniques to analyze them properly. We describe
two ways to construct survival curves: the life-table and the Kaplan-Meier meth-
ods. We guide you in preparing and interpreting survival curves and show you how
to glean useful information from these curves, such as median survival time and
five-year survival rates.
Understanding the Basics of Survival Data

To understand survival analysis, you first have to understand survival data. Sur-
vival times are intervals between a designated starting time point and the time
point an event occurs. These intervals have can have a specific type of missing
data due to a phenomenon called censoring. Because survival data usually include
censored data, they must be analyzed in a very specific way to avoid generating
biased estimates that lead to incorrect conclusions.
Examining how survival times are intervals

The techniques described in this chapter for summarizing, graphing, and com-
paring survival data deal with the time interval from a defined starting point to
the first occurrence of an endpoint event. The event can be designated as death or
a relapse of a particular condition, such as a recurrence of cancer. Or you could
designate the event to be surgical removal (called an explant) of a failed mechani-
cal component, such as an artificial heart valve. If a patient’s heart valve was
implanted on January 10 (beginning of time interval), but their body rejected it
and the explant took place on January 30 (time of event), then the time interval
from implant to explant is 30 – 10, or 20 days.
A person can die only once, so survival analysis can obviously be used for one-
time events. But other endpoints can occur multiple times, such as having a stroke
or having cancer go into remission. The techniques we describe in this chapter
only analyze time to the first occurrence of the event. More advanced survival
analysis methods are needed for models that can handle multiple occurrences of
an event, and these are beyond the scope of this book.
The starting point of the time interval is somewhat arbitrary, so it must be defined
explicitly every time you do a survival analysis. Imagine that you’re studying the
progression of chronic obstructive pulmonary disease (COPD) in a group of
patients. If you want to study the natural history of the disease, the starting point
can be the diagnosis date. But if you’re instead interested in evaluating the
efficacy of a treatment, the starting point can be defined as the date the
treatment began.
302 PART 6 Analyzing Survival Data

Recognizing that survival times
aren’t normally distributed
Even though survival times are numerical quantities, they’re almost never
normally distributed. Because of this, it’s generally not a good idea to use the
following:
» Means and standard deviations to describe survival times
» T tests and ANOVAs to compare survival times between groups
» Least-squares regression to investigate how survival time is influenced by

other factors
If non-normality were the only problem with survival data, you’d be able to sum-
marize survival times as medians and centiles instead of means and standard
deviations. Also, you could compare survival between groups with nonparametric
Mann-Whitney and Kruskal-Wallis tests instead of t tests and ANOVAs. But time-
to-event data are susceptible to a specific type of missingness called censoring.
Typical parametric and nonparametric regression methods are not equipped to
deal with censoring, so we present survival analysis techniques in this chapter.
Considering censoring
Survival data are defined as the time interval between a selected starting point and
an endpoint that represents an event. But unfortunately, the time the event takes
place can be missing in survival data. This can happen in two general ways:
» You may not be able to observe all the participants in the data until they
have the event. Because of time constraints, at some point, you have to end
the study and analyze your data. If your endpoint is death, hopefully at the
end of your study, some of the participants are still alive! At that point, you
would not know how much longer these participants will ultimately live. You
only know that they were still alive up to the last date they were measured in
the study as part of data collection, or the last date study staff communicated
with them in some way (such as through a follow-up phone call). This date is
called the date of last contact or the last-seen date, and would be the date that
these participants would be censored in your data.
» You may lose track of some participants during the study. Participants
who enroll in a study may be lost to follow-up (LFU), meaning that it is no
longer possible for study staff to locate them and continue to collect data for
the study. These participants are also censored at their date of last contact,
but in the case of LFU, this date is typically well before the observation
period ends.

You can describe these two situations in one general way. You know that every
participant in the study either died on a certain date (in which case they have the
event), or was alive up to some last-seen date when they stopped being observed,
in which case they are censored.
Figure 21-1 shows the results of a small study of survival in cancer patients after
a surgical procedure to remove a tumor. Ten patients were recruited to participate
in the study and were enrolled at the time of their surgery. The recruitment period
went from Jan. 1, 2010, to the end of Dec. 31, 2011 (meaning a two-year enrollment
period). All participants were then followed until they died, or until the conclusion
of the study, on Dec. 31, 2016, which added five years of additional observation
time after the last enrollment. Each participant has a horizontal timeline that
starts on the date of surgery and ends with either the date of death or the
censoring date.
FIGURE 21-1:
Survival of ten
study participants
following surgery
for cancer.
In Figure 21-1, observe that each line ends with a code, and there’s a legend at the
bottom. Six of the ten participants (#’s 1, 2, 4, 6, 9, and 10, labeled X) died during
the course of the follow-up study. Two participants (#5 and #7, labeled L) were
LFU at some point during the study, and two participants (#3 and #8, labeled E)
were still alive at the end of the study. So this study has four participants — the
Ls and the Es — with censored survival times.

So, how do you analyze survival data containing censoring? The following sections
explain the correct ways to proceed as well as mistakes to avoid.
Analyzing censored data properly

Statisticians have developed techniques to utilize the partial information con-
tained in censored observations. We describe two of the most popular techniques
later in this chapter, which are the life-table method and the Kaplan-Meier (K-M)
method. To understand these methods, you need to first understand two funda-
mental concepts — hazard and survival:
» The hazard rate is the probability of the participant dying in the next small
interval of time, assuming the participant is alive right now.
» The survival rate is the probability of the participant living for a certain
amount of time after some starting time point.
The first task when analyzing survival data is usually to describe how the hazard
and survival rates vary with time. In this chapter, we show you how to estimate
the hazard and survival rates, summarize them as tables, and display them as
graphs. Most of the larger statistical packages (such as those described in
Chapter 4) allow you to do the calculations we describe automatically, so you may
never have to do them manually. But without first understanding how these
methods work, it’s almost impossible to understand any other aspect of survival
analysis, so we provide a demonstration for instructional purposes.
Making mistakes with censored data

Here are two mistakes you need to avoid when working with survival data:
» You shouldn’t exclude participants with a censored survival time from any
survival analysis!
» You shouldn’t substitute the censored date with some other value, which is
called imputing. When you impute numerical data to replace a missing value, it
is common to use the last observed value for that participant (called last
observation carried forward, or LOCF, imputation). However, you should not
impute dates in survival analysis.
Exclusion and imputation don’t work to fix the missingness in censored data. You
can see why in Figure 21-2, where we’ve slid the timelines for all the participants
over to the left as if they all had their surgery on the same date. The time scale
shows survival time in years after surgery instead of chronological time.

FIGURE 21-2:
Survival times
from the date of
surgery.
If you exclude all participants who were censored in your analysis, you may be left
with analyzable data on too few participants. In this example, there are only six
uncensored participants, and removing them would weaken the power of the
analysis. Worse, it would also bias the results in subtle and unpredictable ways.
Using the last-seen date in place of the death date for a censored observation may
seem like a legitimate use of LOCF imputation, but because the participant did not
die during the observation period, it is not acceptable. It’s equivalent to assuming
that all censored participants died immediately after the last-contact date. But
this assumption isn’t reasonable, because it would not be unusual for them to live
on many years. This assumption would also bias your results toward artificially
shorter survival times.
In your analytic data set, only include one variable to represent time observed
(such as Time in days, months, or years), and one variable to represent event
status (such as Event or Death), coded as 1 if they are have the event during the
observation period, and 0 if they are censored. Calculate these variables from raw
date variables stored in other parts of the data (such as date of death, date of visit,
and so on).

Looking at the Life-Table Method
To estimate survival and hazard rates in a population from a set of observed sur-
vival times, some of which are censored, you must combine the information from
censored and uncensored observations properly. How is this done? Well, it’s not
done by dividing the number of participants alive at a certain time point in the
study by the total number of participants in the study, because this fails to account
for censored observations.
Instead, think of the observation period in a study as a series of slices of time.

Think about how each time a participant survives a slice of time and encounters
the next one, they have a certain probability of surviving to the end of that slice
and continuing on to encounter the next. The cumulative survival probability can
then be obtained by successively multiplying all these individual time-slice sur-
vival probabilities together. For example, to survive three years, first the partici-
pant has to survive the first slice (Year 1), then survive the second slice (Year 2),
and then survive the third slice (Year 3). The probability of surviving all three
years is the product of the probabilities of surviving through Year 1, Year 2,
and Year 3.
These calculations can be laid out systematically in a life table, which is also called
an actuarial life table because of its early use by insurance companies. The calcula-
tions only involve addition, subtraction, multiplication, and division, so they can
be done manually. They are easy to set up in a spreadsheet format, and there are
many life-table templates available for Microsoft Excel and other spreadsheet
programs that you can use.
Making a life table

To create a life table from your survival data, you should first break the entire
range of survival times into convenient time slices. These can be months, quar-
ters, or years, depending on the time scale of the event you’re studying. Also, you
have to consider the time increments in which you want to report your results.
You should arrange to have at least five slices or else your survival and hazard
estimates will be too coarse to show any useful features. Having many skinny
slices doesn’t disturb the calculations, but the life table will have many rows and
may become unwieldy. For the survival times shown in Figure 21-2, a natural
choice would be to use seven 1-year time slices.

Next, count how many participants experienced the event during each slice, and
how many were censored, meaning they were last observed during this time slice
and had not experienced the event. From Figure 21-2, you see that
» During the first year after surgery, one participant died (#1), and one partici-
pant was censored (#5, who was LFU).
» During the second year, no participants died or were censored.
» During the third year, two participants died (#4 and #9), and none were
censored.
Continue tabulating deaths and censored times for the fourth through seventh
years, and enter these counts into the appropriate cells of a spreadsheet like the
one shown in Figure 21-3.
FIGURE 21-3:
A partially
completed life
table to analyze
the survival times
shown in
Figure 21-2.
To fill in the table shown in Figure 21-3:
» Put the description of the time interval that defines each slice into Column A.
» Enter the total number of participants alive at the start into Column B in the
0–1 yr row.
» Enter the counts of participants who died within each time slice into Column C
(labeled Died).
» Enter the counts of participants who were censored during each time slice
into Column D (labeled Last Seen Alive).
After you’ve entered all the counts, the spreadsheet will look like Figure 21-3.
Then you perform the calculations shown in the Formula row at the top of the
figure to generate the numbers in all the other cells of the table. (To see what it
looks like when the table is completely filled in, take a sneak peek at
Figure 21-4.)

Columns B, C, and D
Column B includes the number of participants known to be alive at the start of
each year after surgery. This is equal to the number of participants alive at the
start of the preceding year minus the number who died (Column C) or were
censored (Column D) during the preceding year. Here’s the formula, written in
terms of the column letters: B for any year = B – C – D from the preceding year.
Here’s how this process plays out in Figure 21-3:
» Out of the ten participants alive at the start, one died and one was last seen
alive during the first year. This means eight participants (10 – 1 – 1) are
known to still be alive at the start of the second year. The missing partici-
pant is #5, who was LFU during the first year. They are censored and not
counted in any subsequent years.
» Zero participants died or were last seen alive during the second year. So, the
same eight participants are still known to be alive at the start of the third year.
» Calculations continue the same way for the remaining years.
Column E
Column E shows the number of participants at risk for dying during each year. You
may guess that this is the number of participants alive at the start of the interval,
but there’s one minor correction. If any were censored during that year, then
they weren’t technically able to be observed for the entire year. Though they may
die that year, if they are censored before then, the study will miss it. What if you
don’t know exactly when during that year they became censored? If you don’t
have the exact date, you can consider them being observed for half the time
period (in this case, 0.5 years). So the number at risk can be estimated as the
number alive at the start of the year, minus one-half of the number who became
censored during that year, as indicated by the formula for Column E: E = B – D/2.
(Note: To simplify the example, we are using years, but you could use months
instead if you have exact censoring and death dates in your data to improve the
accuracy of your analysis.)
Here’s how this formula works in Figure 21-3:
» Ten participants were alive at the start of Year 1, and one participant was
censored during Year 1. To correct for censoring, divide 1 by 2, which is
0.5. Next, subtract 0.5 from 10 to get 9.5. After correcting for censoring, only
9.5 participants are at risk of dying during Year 1.

» Eight participants were alive at the start of Year 2, and zero were censored
during Year 2. So all eight participants continued to be at risk during Year 2.
» Calculations continue in the same way for the remaining years.
Column F
Column F shows the Probability of Dying during each interval, assuming the partic-
ipant has survived up to the start of that interval. To calculate this, divide the Died
column by the At Risk column. This represents the fraction of those who were at
risk of dying at the beginning of the interval who actually died during the interval.
Formula for Column F: F = C/E.
Here’s how this formula works in Figure 22-3:
» For Year 1, the probability of dying is calculated by dividing the one death by
the 9.5 participants at risk: 1/9.5, or 0.105 (10.5 percent).
» Zero participants died in Year 2. So, for participants surviving Year 1 and alive
at the beginning of Year 2, the probability of dying during Year 2 is 0. Woo-hoo!
Column G
Column G shows the Probability of Surviving during each interval for participants
who have survived up to the start of that interval. Since surviving means not
dying, the equation for this column is 1 – Probability of Dying, as indicated by the
formula for Column G: G = 1 – F.
Here’s how this formula works out in Figure 22-3:
» The probability of dying in Year 1 is 0.105, so the probability of surviving in

Year 1 is 1 – 0.105, or 0.895.
» The probability of dying in Year 2 is 0.000, so the probability of surviving in

Year 2 is 1 – 0.000, or 1.000.
Column H
Column H shows the cumulative probability of surviving from the time of surgery
all the way through the end of this time slice. To survive from the start time
through the end of any given year (year N), the participant must survive each of

the years from Year 1 through Year N. Because surviving each year is an indepen-
dent accomplishment, the probability of surviving all N of the years is the product
of the individual years’ probabilities. So Column H is a running product of Column G.
In other words, the value of Column H for Year N is the product of the first N
values in Column G.
Here’s to fill in Figure 22-3 (with the results shown in Figure 22-4):
» For Year 1, H is the same as G: a 0.895 probability of surviving one year.
» For Year 2, H is the product of G for Year 1 times G for Year 2; that is, 0.895 ×
1.000, or 0.895.
» For Year 3, H is the product of the Gs for Years 1, 2, and 3; that is, 0.895 ×
1.000 × 0.750, or 0.671.
Putting everything together

Figure 21-4 shows the spreadsheet with the results of all the preceding
calculations.
FIGURE 21-4:
Completed life
table to analyze
the survival
times shown in
Figure 22-2.
Interpreting a life table

Figure 21-4 contains the hazard rates (in Column F) and the cumulative survival
probabilities (in Column H) for each year following surgery, based on your sample
of ten participants. Keep in mind these life-table features:
» The hazard and survival values obtained from this life table are estimates
from a sample of the true population hazard and survival functions (in this
case, using one-year intervals).

» The intervals are often the same size for all the rows in a life table like in the
example, but they don’t have to be. You may choose these on the basis of the
probabilities in your data.
» The hazard rate obtained from a life table for each time slice is equal to the
Probability of Dying (Column F) divided by the width of the slice. Therefore,
the hazard rate for the first year would be expressed as 0.105 per year, or
10.5 percent per year.
» The Cumulative Survival probability, in Column H, is the probability of surviving

from the start date through to the end of the interval. It has no units, and it
can be expressed as a fraction or as a percentage. The value for any time slice
applies to the moment in time at the end of the interval.
» The cumulative survival probability is always 1.0 (or 100 percent) at time 0,
whenever you designate that time 0 is (in the example, date of surgery), but
it’s not included in the table.
» The cumulative survival function decreases only at the end of an interval that
has at least one observed death, because censored observations don’t cause
a decrease in the estimated survival. Censored observations however
influence the size of the decreases when subsequent events occur. This is
because censoring reduces the number at risk, which is used in the denomi-
nator in the calculation of the death and survival probabilities.
» If an interval contains no events and no censoring, like in the 1–2 years row in
the table in Figure 21-4, it has no impact on the calculations. Notice how all
subsequent values for Column B and for Columns E through H would remain
identical if that row were removed.
Graphing hazard rates and survival

probabilities from a life table
Graphs of hazard rates and cumulative survival probabilities (Columns F and H
from Figure 21-4, respectively) can be prepared from life-table results using
Microsoft Excel or another spreadsheet or statistical program with graphing
capabilities. Figure 21-5 illustrates the way these results are typically presented.
» Figure 21-5a is a graph of hazard rates. Hazard rates are often graphed as
bar charts, because each time slice has its own hazard rate in a life table.
» Figure 21-5b is a graph of cumulative survival probabilities, also known as

the survival function. Survival values are usually graphed as stepped line charts,

where a horizontal line represents the cumulative survival probability during
each time slice. The cumulative survival for the Year 0 to 1 time slice is 1.0
(100 percent), so the horizontal line stays at y = 1.0. But between Year 1 and
Year 3, the cumulative survival probability drops to 0.895, so a vertical line is
dropped from 1.0 to y = 0.895 at the time the Year 1 to Year 2 interval starts.
It goes across both that interval and the next one because there are no deaths
in these intervals. This stepped line continues downstairs and finally ends at
the end of the last interval where the cumulative survival probability is 0.128.
FIGURE 21-5:
Hazard function
(a) and survival
function (b)
results from
life-table
calculations.
Digging Deeper with the

Kaplan-Meier Method
Using very narrow time slices doesn’t hurt life-table calculations. In fact, you can
define slices so narrow that each participant’s survival time falls within its own
private little slice. Imagine you had N participants. Your life table would have N
rows with data from one participant each. You could theoretically add all rest of
the rows to fill out the rest of the time slices. These would not have any data in
them, and since empty rows don’t affect the life-table calculations, you could just
stick with your life table where each row has one participant’s data. And if you
happen to have two or more participants with exactly the same survival or censor-
ing time, it’s okay to put each one in their own row.
The life-table calculations work fine with only one participant per row and pro-
duce what’s called Kaplan-Meier (K-M) survival estimates. You can think of the K-M
method as a very fine-grained life table. Or, you can see a life table as a grouped
K-M calculation.
A K-M worksheet for the survival times is shown in Figure 21-6. It is based on the
one-participant-per-row idea and is laid out much like the usual life-table

worksheet shown in Figure 22-4, but with a few differences in the raw data cells
and minor differences in the calculations:
» Instead of a column identifying the time slices, there are two columns
identifying the individual participant (Column A) and their survival or censor-
ing time (Column B). The table is ordered from the shortest time to
the longest.
» Instead of two columns containing the number who died and were censored
in each interval, you need only one column indicating whether or not the
participant in that row died (Column C). If they died during the observation
period, use code 1, and if not and they were censored, use code 0.
» These changes mean that Column D labeled Alive at Start now decreases by 1
for each subsequent row.
» The At Risk column in Figure 21-4 isn’t needed, because it can be calculated
from the Alive at Start column. That’s because if the participant is censored,
the probability of dying is calculated as 0, regardless of the value of the
denominator.
» To calculate Column E, the Probability of Dying, divide the Died indicator by the
number of participants alive for that time period in Column D, Alive at Start.
Formula: E = C/D.
» The probability of surviving (Column F) and the cumulative survival (Column G)

are calculated the same way as in the life-table method.
FIGURE 21-6:
Kaplan-Meier
calculations.
Figure 21-7 shows graphs of the K-M hazard and survival estimates from
Figure 21-6. These charts were created using the R statistical software. Most soft-
ware that performs survival analysis can create graphs similar to this. The K-M

survival curve in Figure 21-7b has smaller steps than the life-table survival curve
in Figure 21-5b, so it’s more fine-grained. This is because the step curve now
decreases at every time point at which a participant died. You can tell from the
figures where participant #1 died at 0.74 years, #9 died at 2.27 years, #4 died at
2.34 years, and so on.
FIGURE 21-7:
Kaplan-Meier
estimates of the
hazard (a) and
survival (b)
functions.
While the K-M survival curve tends to be smoother than the life table survival
curve, just the opposite is true for the hazard curve. In Figure 21-7a, each partic-
ipant has their own very thin bar, and the resulting chart isn’t easy to interpret.
Heeding a Few Guidelines for Life-Tables

and the Kaplan-Meier Method
Most of the larger statistical packages (see Chapter 4) can perform life-table and
Kaplan-Meier calculations for you and directly generate survival curves. You have
to identify two variables for the software: one with the survival time for each par-
ticipant, and a binary variable coded 1 if the survival time represents time to death
or the event, and 0 if it represents censored time. It sounds simple, but it’s sur-
prisingly easy to mess up. Here are some pointers for setting up your data and
interpreting the results properly.
Recording survival times correctly

It is important to draw a distinction between data collection and data analysis.
When recording the raw data, it’s best to collect all the relevant dates for the
study. Before the study starts, the dates of interest for data collection should be
specified, which could include date of diagnosis, start of therapy, end of therapy,

start of improvement or remission, date of relapse, or others. For events, you
should record date of each event if it recurs, and even if death is not the event of
interest, date of death should be recorded if available. For censoring purposes,
ensure that you are collecting dates of contact so you can identify a last-seen date
if needed. If you collect your data properly, you will later be able to calculate any
time interval needed, as well as create an event status indicator needed.
Dates and times should be recorded to suitable precision. If your study timeline
is years, it’s best to keep track of dates to the day. In a Phase I clinical trial
(see Chapter 5), participants may be studied for events that happen in a span
of a few days. In those cases, it’s important to record dates and times to the near-
est hour or minute. You can even envision laboratory studies of intracellular
events where time would have to be recorded with millisecond — or even
microsecond — precision!
Dates and times can be stored in different ways in different statistical software (as
well as Microsoft Excel). Designating columns as being in date format or time for-
mat can allow you to perform calendar arithmetic, allowing you to obtain time
intervals by subtracting one date from another.
Miscoding censoring information

It can be surprisingly easy to miscode the event status indicator. If the name of the
variable is Death, and is coded as 1 if the participant died during the observation
period and 0 if they were censored, this seems intuitive. But analysts may want to
identify all the censored observations in their data, so they may create a censored
indicator named Censored, and code it as 1 if the participant is censored, and 0 if
they are not. Because data may be used for different types of survival analyses,
there could be other event indicators included in the data as well also coded
as 1 and 0.
The problem is that if you accidentally use your censored indicator instead of your
event indicator when running your survival analysis, you will unknowingly flip
your analysis, and you won’t get any warning or error message from the program.
You’ll only get incorrect results. Worse, depending on how many censored and
uncensored observations you have, the survival curve may also not hint at any
errors. It may look like a perfectly reasonable survival curve for your data, even
though it’s completely wrong.
You have to read your software’s documentation carefully to make sure you code
your event variable correctly. Also, you should always check the program’s output
for the number of censored and uncensored observations and compare them to the
known count of censored and uncensored participants in your data file.

IN THIS CHAPTER
» Using the log-rank test to compare

two groups
» Thinking about more complicated

ways to compare the survival
experience
» Calculating the necessary sample size

to compare survival times
Chapter 22
Comparing Survival
Times
T
he life table and Kaplan-Meier survival curves described in Chapter 21 are
ideal for summarizing and describing the time to the first or only occur-
rence of a particular event based on times observed in a sample of individu-
als. They correctly incorporate data that reflect when an individual is observed
during the study but does not experience the event, which is called censored data.
Animal and human studies involving endpoints that occur on a short time-scale,
like measurements taking during an experimental surgical procedure, may yield
totally uncensored data. However, the more common situation is that during the
observation period of studies, not all individuals experience the event, so you usu-
ally have censored data on your hands.
In biological research and especially in clinical trials (discussed in Chapter 5), you
often want to compare survival times between two or more groups of individuals.
In humans, this may have to do with survival after cancer surgery. In animals, it
may have to do with testing the toxicity of a potential therapeutic. This chapter
describes an important method for comparing survival curves between two groups
called the log-rank test, and explains how to calculate the sample size you need to
have sufficient statistical power for this test (see Chapter 3). The log-rank test can
be extended to handle three or more groups, but this discussion is beyond the
scope of this book.
CHAPTER 22 Comparing Survival Times 317

In this chapter, as in Chapters 21 and 23, we use the term survival in reference to
the outcome of death. However, all the calculations pertain to any type of outcome
event being studied, including good ones, such as cancer going into remission.
There is some ambiguity associated with the name log-rank test. It has also been
called different names (such as the Mantel-Cox test), and has been extended into
variants such as the Gehan-Breslow test. You may also observe that different
software may calculates the log-rank test slightly differently. In this chapter, we
describe the most commonly used form of the log-rank test.
If have no censored observations in your data, you can skip most of this chapter.
This may happen if, for example, death is your outcome and at the end of your
study period no individuals are alive anymore — they all have died in your study.
As you may guess, this situation is much more common in animal studies than
human studies. But if you have followed all the individuals in your data until they
all experienced the outcome, and you have two or more groups of numbers indi-
cating survival times that you want to compare, you can use approaches described
in Chapter 11. One option is to use an unpaired Student t test to test whether one
group has a statistically significantly longer mean survival time than the other. If
you have three or more groups, you would use an ANOVA instead. But because
survival times are very likely to be non-normally distributed, you may prefer to
use a nonparametric test, such as the Wilcoxon Sum-of-Ranks test or Mann-
Whitney U test, to compare the median survival time between two groups. With
more than two groups, you would use the nonparametric Kruskal-Wallis test.
Suppose that you conduct a toxicity study with laboratory animals of a potential
cancer drug. You obtain 90 experimental mice. The mice are randomly placed in
groups such that 60 receive the drug in their food, and 30 are given control food
with no drug. A laboratory worker observes them and records their vital status
every day after the experiment starts, taking note of when each animal dies or is
censored, meaning they are taken out of the study for another reason (such as not
eating). You perform a life-table analysis on each group of mice — the drug com-
pared to control — as described in Chapter 21, and graph the results. The graph
displays the survival curves shown in Figure 22-1. As a bonus, the two life tables
generated to support this display also provide the summary information needed
the log-rank test.
The two survival curves in Figure 22-1 look different. The drug group seems to be
showing better survival than the control group. But is this apparent difference
real, or could it be the result of random fluctuations only? The log-rank test
answers this question.

FIGURE 22-1:
Survival curves
for two groups of
laboratory
animals.
Comparing Survival between Two Groups

with the Log-Rank Test
The log-rank test can be performed using individual-level data, or on data that
has been summarized into a life-table format. In this section, we describe how to
run a log-rank test with statistical software, which is how it is usually done. Next,
to help you understand the underlying calculations, we describe the log-rank test
calculations in detail using the life-table as you might carry them out manually
using spreadsheet software such as Microsoft Excel.
Understanding what the log-rank

test is doing
A two-group log-rank test asks whether events — which are deaths in our
example — are split between the two groups in the same proportion as the num-
ber of at-risk individuals in the two groups. The computer selects a group and
sums the difference between the observed and expected number of deaths in each
time slice over all the time slices to get the total excess deaths for that group. The
excess death sum is then scaled down, meaning it is divided by an estimate of its
standard deviation. (Later in this chapter we describe how to calculate that stan-
dard deviation estimate.) The scaled-down excess deaths sum is a number whose
random sampling fluctuations should follow a normal distribution, and from
which a p value can be easily calculated. The null hypothesis of the log-rank test
is that there is no difference in survival between the two groups, so a p value less
than your selected α (usually 0.05) indicates a statistically significant difference.

Don’t worry if the preceding paragraph makes your head spin. It is only meant to
give you a general sense behind the calculations in the log-rank test.
Running the log-rank test on software

Most commercial statistical software packages (like those described in Chapter 4)
can perform a log-rank test. You first organize your data into a table that has one
row per individual, and these three columns:
» Group: The group variable contains a code indicating the individual’s group. In
this example, we could use the code Drug = 1 and Control = 2.
» Time: A numerical variable containing the individual’s survival time. For

individuals experiencing the event during the study, it represents time to
event. For censored individuals, it is time to the end of observation.
» Event status: A variable that indicates the individual’s status at the end of
observation. If they got the event, it is usually coded as 1, and if not or they
are censored, it is coded as 0.
To run the log-rank test, you tell your computer program which variable repre-
sents the group variable, which one means time, and which one contains the
event status. The program should produce a p value for the log-rank test. If you
set α = 0.05 and the p value is less than that, you reject the null and conclude that
the two groups have statistically significantly different survival curves.
In addition to the p value, the program may output median survival time for each
group along with confidence intervals, and difference in median times between
groups. If possible, you will also want to request graphs that show whether your
data are consistent with the hazard proportionality assumption that we describe
later in “Assessing the assumptions.”
Looking at the calculations

The log-rank test should not be done manually because it is an error-prone task.
But we believe you’ll have a better appreciation of the log-rank test if you under-
stand how it works, so we describe how the calculations could theoretically be
carried out using a Microsoft Excel spreadsheet.
The log-rank test utilizes information from the life tables needed to produce the
graph shown earlier in Figure 22-1. Figure 22-2 shows a portion of the life tables
that produced the curves shown in Figure 22-1, with the data for the two groups
displayed side by side.

FIGURE 22-2:
A portion of the
life-table
calculations for
two groups of
laboratory
animals.
In Figure 22-2, the Drug group’s results are in columns B through E, and the Con-
trol group’s results are in columns F through I. The only measurements needed
from Figure 22-2 for the log-rank test are At risk (columns E and I), meaning
number at risk in each time slice for each group, and Died (columns C and G),
meaning the number of observed deaths in that time slice for each group. The log-
rank test calculations are in a second spreadsheet (shown in Figure 22-3).
FIGURE 22-3:
Basic log-rank
calculations done
manually (but
please use
software
instead!).
The spreadsheet shown in Figure 22-3 has the following columns:
» Column A identifies the time slices, consistent with Figure 22-2.
» Columns B and C pertain to the Drug group, and reprint the At risk and Died
columns from Figure 22-2 for that group. Columns D and E pertain to the
Control group and reprint the At risk and Died columns from Figure 22-2 for
that group.
» Columns F and G show the combined total number of individuals at risk and
the total number of individuals who died, which is obtained by combining the
corresponding columns for the two groups.
» Column H, labeled % At Risk, shows Group 1’s percentage of the total number
of at-risk individuals per time slice.

» Column I, labeled Expected Deaths, shows the number of deaths you’d expect
to see in Group 1 based on apportioning the total number of deaths (in both
groups) by Group 1’s percentage of total individuals at-risk. For the 0–1 day
row, Group 1 had about 2 / 3 of the 89 individuals at risk, so you’d expect it
to have about 2 / 3 of the nine deaths.
» Column J, labeled Excess Deaths, shows the excess number of actual deaths
compared to the expected number for Group 1.
» Column K shows the variance (equal to the square of the standard deviation)
of the excess deaths. It’s obtained from this complicated formula that’s based
on the properties of the binomial distribution (see Chapter 24):
V DT N 1 /N T N 2 /N T NT DT / N T 1
For the first time slice (0–1 day), this becomes:

V 9 59.5/ 89 29.5/ 89 89 9 / 89 1 , which equals approximately 1.813.
N refers to the number of individuals at risk, D refers to deaths, the subscripts

1 and 2 refer to groups 1 and 2, and T refers to the total of both groups
combined.
Next, you add up the excess deaths in all the time slices to get the total number of
excess deaths for Group 1 compared to what you would have expected if the deaths
had been distributed between the two groups in the same ratio as the number of
at-risk individuals.
Then you add up all the variances. You are allowed to do that, because the sum of
the variances of the individual numbers is equal to the variance of the sum of a set
of numbers.
Finally, you divide the total excess deaths by the square root of the total variance
to get a test statistic called Z:
Z ExcessDeaths / Variances
The Z value is approximately normally distributed, so you can obtain a p value

from a table of the normal distribution or from an online calculator. For the data
in Figure 23-3, z 5.65/ 6.64 , which is 2.19. This z value corresponds to a p
value of 0.028, so the null hypothesis is rejected, and you can conclude that the
two groups have a statistically significantly different survival curve.
Note: By the way, it doesn’t matter which group you assign as Group 1 in these
calculations. The final results come out the same either way.

Assessing the assumptions
Like all statistical tests, the log-rank test assumes that you studied an unbiased
sample from the population about which you’re trying to draw conclusions. It also
assumes that any censoring that occurred was due to circumstances unrelated to
the treatment being tested (for example, individuals didn’t drop out of the study
because the drug made them sick).
Also, the log-rank test looks for differences in overall survival time. In other
words, it’s not good at detecting differences in shape between two survival curves
with similar overall survival time, like the two curves shown in Figure 22-4. These
two curves actually have the same median survival time, but the survival experi-
ence is different, as shown in the graph. When two survival curves cross over each
other, as shown in Figure 22-4b, the excess deaths are positive for some time
slices and negative for others. This leads them to cancel out when they’re added
up, producing a smaller z value as a test statistic z value, which translates to
larger, non-statistically significant p value.
FIGURE 22-4:
Proportional (a)
and nonpropor-
tional (b) hazards
relationships
between two
survival curves.
Therefore, one very important assumption of the log-rank test is that the two
groups have proportional hazards, which means the two groups must have gener-
ally similar survival shapes, as shown in Figure 22-4a. Flip to Chapter 21 for more
about survival curves, and read about hazards in more detail in Chapter 23.

Considering More Complicated
Comparisons
The log-rank test is good for comparing survival between two or more groups. But
it doesn’t extend well to more complicated situations. What if you want to do one
of the following?
» Test whether survival depends on age or some other continuous variable
» Test the simultaneous effect of several variables, or their interactions, on

survival
» Correct for the presence of confounding variables or other covariates
In other areas of statistical testing, such situations are handled by regression

techniques. Survival analysis regression uses survival outcomes with censored
observations, and can accommodate these analyses. We describe survival regres-
sion in Chapter 23.
Estimating the Sample Size Needed for

Survival Comparisons
We introduce power and sample size in Chapter 3. Calculating the sample size for
survival comparisons is complicated by several factors:
» The need to specify an alternative hypothesis: This hypothesis can take the
form of a hazard ratio, described in Chapter 23, where the null hypothesis is
that the hazard ratio = 1. Or, you can hypothesize the difference between two
median survival times.
» The impact of censoring: How censoring impacts sample size needed

depends on the accrual rate, dropout rate, and the length of follow-up.
» The shape of the survival curves: For sample-size calculations, it is often

assumed that the survival curve is exponential, but that may not be realistic.
In Chapter 4, we recommend using free software G*Power for your sample-size

calculations. However, because G*Power does not offer a survival sample-size
estimator, for this, we recommend you use another free software package called
PS (Power and Sample Size Calculation), which is available from Vanderbilt

University Medical Center (https://biostat.app.vumc.org/wiki/Main/
PowerSampleSize).
After opening the PS program, choose the Survival tab, fill in the form, and click
Calculate. The median survival times for the two groups are labeled m1 and m2,
the accrual interval is labeled A, the post-accrual follow-up period is labeled F,
and the group allocation proportion is labeled m. Note that the time variables must
always be entered in the same units (days, in this example). You will also need to
enter your chosen α and power.
Here is an example. Suppose that you’re planning a study to compare an experi-

mental drug for keeping cancer remission to placebo in two equal-sized groups of
cancer patients whose cancer is in remission. You expect to observe participants
for a total of three years to see whether their cancer returns (which is the out-
come). From existing studies, you expect the median placebo time to be 20 months,
and you think the drug should extend this to 30 months. If it truly does extend
survival (time to remission) that much, you want to be able to detect this. You set
α = 0.05 and power at 80 percent so that you have an 80 percent chance of getting
a p value of less than 0.05 when you compare drug to placebo using the log-rank
test. If you fill in the PS form with these estimates and select Calculate, under
Sample Size the software will say you need 170 participants in each group (a total
of 340 participants).

IN THIS CHAPTER
» Knowing when to use survival

regression
» Grasping the concepts behind

survival regression
» Running and interpreting the

outcome of survival regression
» Peeking at prognosis curves
» Estimating sample size for survival

regression
Chapter 23
Survival Regression
S
urvival regression is one of the most commonly used techniques in
biostatistics. It overcomes the limitations of the log-rank test (see
Chapter 22) and allows you to analyze how survival time is influenced by one
or more predictors (the X variables), which can be categorical or numerical. In this
chapter, we introduce survival regression. We specify when to use it, describe its
basic concepts, and show you how to run survival regressions in statistical
software and interpret the output. We also explain how to build prognosis curves
and estimate the sample size you need to support a survival regression.
Note: Because time-to-event data so often describe actual survival, when the
event we are talking about is death, we use the terms death and survival time. But
everything we say about death applies to the first occurrence of any event, like
pre-diabetes patients restoring their blood sugar to normal levels, or cancer sur-
vivors suffering a recurrence of cancer.
CHAPTER 23 Survival Regression 327

Knowing When to Use Survival Regression
In Chapter 21, we examine the special problems that come up when the researcher
can’t continue to collect data during follow-up on a participant long enough to
observe whether or not they ever experience the event being studied. To recap, in
this situation, you should censor the data. This means you should acknowledge the
participant was only observed for a limited amount of time, and then was lost to
follow-up. In that chapter, we also explain how to summarize survival data using
life tables and the Kaplan-Meier method, and how to graph time-to-event data as
survival curves. In Chapter 22, we describe the log-rank test, which you can use to
compare survival among a small number of groups — for example, participants
taking drug versus placebo, or participants initially diagnosed at four different
stages of the same cancer.
But the log-rank test has limitations:
» The log-rank test doesn’t handle numerical predictors well. Because this
test compares survival among a small number of categories, it does not work
well for a numerical variable like age. To compare survival among different
age groups with the log-rank test, you would first have to categorize the
participants into age ranges. The age ranges you choose for your groups
should be based on your research question. Because doing this loses the
granularity of the data, this test may be less efficient at detecting gradual
trends across the whole age range.
» The log-rank test doesn’t let you analyze the simultaneous effect of
different predictors. If you try to create subgroups of participants for each
distinct combination of categories for more than one predictor (such as three
treatment groups and three diagnostic groups), you will quickly see that you
have too many groups and not enough participants in each group to support
the test. In this example — with three different treatment groups and three
diagnostic groups — you would have 3 × 3 groups, which is nine, and is already
too many for a log-rank test to be useful. Even if you have 100 participants in
your study, dividing them into nine categories greatly reduces the number of
participants in each category, making the subgroup estimate unstable.
Use survival regression when the outcome (the Y variable) is a time-to-event

variable, like survival time. Survival regression lets you do all of the following,
either in separate models or simultaneously:
» Determine whether there is a statistically significant association between

survival and one or more other predictor variables

» Quantify the extent to which a predictor variable influences survival, including
testing whether survival is statistically significantly different between groups
» Adjust for the effects of confounding variables that also influence survival
» Generate a predicted survival curve called a prognosis curve that is customized

for any particular set of values of the predictor variables
Grasping the Concepts behind

Survival Regression
Note: Our explanation of survival regression has a little math in it, but nothing
beyond high school algebra. In laying out these concepts, we focus on multiple
survival regression, which is survival regression with more than one predictor. But
everything we say is also true when you have only one predictor variable.
Most kinds of regression require you to write a formula to fit to your data. The
formula is easiest to understand and work with when the predictors appear in the
function as a linear combination in which each predictor variable is multiplied by a
coefficient, and these terms are all added together (perhaps with another coeffi-
cient, called an intercept, thrown in). Here is an example of a typical regression
formula: y c0 c1 x 1 c 2 x 2 c 3 x 3 . Linear combinations (such as c2x2 from the
example formula) can also have terms with higher powers — like squares or
cubes — attached to the predictor variables. Linear combinations can also have
interaction terms, which are products of two or more predictors, or the same pre-
dictor with itself.
Survival regression takes the linear combination and uses it to predict survival.
But survival data presents some special challenges:
» Censoring: Censoring happens when the event doesn’t occur during the
observation time of the study (which, in human studies, means during
follow-up). Before considering using survival regression on your data, you
need to evaluate the impact censoring may have on the results. You can do
this using life tables, the Kaplan-Meier method, and the log-rank test, as
described in Chapters 21 and 22.
» Survival curve shapes: Some business disciplines develop models for

estimating time to failure of mechanical or electronic devices. They estimate
the times to certain kinds of events, like a computer’s motherboard wearing
out or the transmission of a car going kaput, and find that they follow

remarkably predictable shapes or distributions (the most common being the
Weibull distribution, covered in Chapter 24). Because of this, these disciplines
often use a parametric form of survival regression, which assumes that you
can represent the survival curves by algebraic formulas. Unfortunately for
biostatisticians, biological data tends to produce nonparametric survival curves
whose distributions can’t be represented by these parametric distributions.
As described earlier, nonparametric survival analyses using life tables, Kaplan-

Meier plots, and log-rank tests are limiting. But as biostatisticians, we could not
rely on using parametric distributions in our models; we wanted to use a hybrid,
semi-parametric kind of survival regression. We wanted one that was partly non-
parametric, meaning it didn’t assume any mathematical formula for the shape of
the overall survival curve, and partly parametric, meaning we could use some
parameter (or predicted survival distribution shape) to guide our formulas the
way other industries used the Weibull distribution. In 1972, a statistician named
David Cox developed a workable method for doing this. The procedure is now
called Cox proportional hazards regression, which we call PH regression for the rest of
this chapter for brevity. In the following sections, we outline the steps of per-
forming a PH regression.
Since 1972, many issues have been identified when using survival regression for
biological data, especially with respect to its appropriateness for the type of data.
One way to examine this is by running a logistic regression model (see Chapter 18)
with the same predictors and outcome as your survival regression model without
including the time variable, and seeing if the interpretation changes.
The steps to perform a PH regression

You can understand PH regression in terms of several conceptual steps, although
when using statistical software like is described in Chapter 4, it may appear that
these steps take place simultaneously. That is because the output created is
designed for you — the biostatistician — to walk through the following steps in
your mind and make decisions. You must use the output to:
1. Determine the shape of the overall survival curve produced from the
Kaplan-Meier method.
2. Estimate how your hypothesized predictor variables may impact the

bends in this curve — in other words, in what ways your predictors may
affect survival.

3. Create a PH regression model that fits your data the best it can, and
interpret the values of the regression coefficients so you can calculate
predicted survival times.
Determining the baseline in time-to-event

analyses
Your software may define the baseline survival function in one of two ways:
» The survival curve of an average participant: This curve is calculated as if

the value of each predictor is equal to the group average value for that
variable. The average-participant baseline is like the overall survival curve you
get from a Kaplan-Meier calculation by using all the available participants.
» The survival curve of a hypothetical zero participant: This curve is calcu-

lated assuming the value of each predictor is equal to 0. Some mathematicians
prefer to use the zero-participant baseline because can make formulas simpler,
but biostatisticians don’t like it because it corresponds to a hypothetical
participant who can’t possibly exist in the real world. No actual person has an
age equal to 0 years, or weighs 0 kilograms, and so on. The survival curve for
this impossible person doesn’t look like a regular survival curve, so as biostatis-
ticians, we can’t really use the zero-participant baseline survival function.
Luckily, the way your software defines its baseline function doesn’t affect any of
the calculated measures on your output, so you don’t have to worry about it. But
you should be aware of these definitions if you plan to generate prognosis curves,
because the formulas to generate these are slightly different depending upon the
way the computer calculates the baseline survival function.
Bending the baseline

Now for the tricky part. How do you bend or flex this baseline curve to express how
survival may increase or decrease for different predictor values? Survival curves
always start at 1.0 at time 0, meaning 100 percent of the sample do not have the
event at time 0. The bending process must preserve that time starts at 0, and
maximum survival is 1.0. If you raise 0 or 1 to any power, you will find that they
stay the same — 0 stays 0, and 1 stays 1. But, exponentiating any number between
0 and 1 smoothly raises or lowers all the values between 0 and 1.
We will demonstrate what we mean by imagining our baseline function was a

straight line (even though no actual biological survival curve would ever be exactly
a straight line). Look at Figure 23-1a, which is a graph of the equation y 1 x .

FIGURE 23-1:
Bending a
straight line into
different shapes
by raising each
point on the line
to some power: h.
To understand the flex, look at what happens when you raise this straight line to
various powers, which we refer to as h and illustrate in Figure 23-1b:
» Squaring: If you set h = 2, the y value for every point on the line always comes
out smaller, because they are always less than 1. For example, 0.8 2 is 0.64.
» Taking the square root: If we set h = 0.05, the y value of every point on the
line becomes larger. For example, the square root of 0.25 is 0.5.
Notice in Figure 23-1b, both 12 and 10.5 remain 1, and 0 2 and 0 0.5 both remain 0, so
those two ends of the line don’t change.
Does the same trick work for a survival curve that doesn’t follow any particular
algebraic formula? Yes, it does! Look at Figure 23-2.
FIGURE 23-2:
Raising to a
power works
for survival
curves, too.

Here are some important points about Figure 23-2:
» Figure 23-2a shows a typical survival curve. It’s not defined by any algebraic
formula. It just graphs the table of values obtained by a life-table or Kaplan-
Meier calculation.
» Figure 23-2b shows how the baseline survival curve is flexed by raising every
baseline survival value to a power. You get the lower curve by setting h = 2
and squaring every baseline survival value. You get the upper curve by setting
h = 0.05 and taking the square root of every baseline survival value. Notice
that the two flexed curves keep all the distinctive zigs and zags of the baseline
curve, in that every step occurs at the same time value as it occurs in the
baseline curve.
• The lower curve represents a group of participants who had a worse

survival outcome than those making up the baseline group. This means
that at any instant in time, they were somewhat more likely to die than a
baseline participant at that same moment. Another way of saying this is
that the participants in the lower curve have a higher hazard rate than the
baseline participants.
• The upper curve represents participants who had better survival than
a baseline person at any given moment — meaning they had a lower
hazard rate.
Obviously, there is a mathematical relationship between the chance of dying at

any instant in time, which is called hazard, and the chance of surviving up to some
point in time, which we call survival. It turns out that raising the survival curve to
the h power is exactly equivalent to multiplying the hazard curve by the natural
logarithm of h. Because every point in the hazard curve is being multiplied by the
same amount — by Log(h) — raising a survival curve to a power is referred to as
a proportional hazards transformation.
But what should the value of h be? The h value varies from one individual to
another. Keep in mind that the baseline curve describes the survival of a perfectly
average participant, but no individual is completely average. You can think of
every participant in the data as having her very own personalized survival curve,
based on her very own h value, that provides the best estimate of that partici-
pant’s chance of survival over time.
Seeing how predictor variables influence h

The final piece of the PH regression puzzle is to figure out how the predictor vari-
ables influence h, which influences survival. As you likely know, all regression
procedures estimate the values of the coefficients that make the predicted values
agree as much as possible with the observed values. For PH regression, the

software estimates the coefficients of the predictor variables that make the pre-
dicted survival curves agree as much as possible with the observed survival times
of each participant.
How does PH regression determine these regression coefficients? The short

answer is, “You’ll be sorry you asked!” The longer answer is that, like all other
kinds of regression, PH regression is based on maximum likelihood estimation.
The software uses the data to build a long, complicated expression for the proba-
bility of one particular individual in the data dying at any point in time. This
expression involves that individual’s predictor values and the regression coeffi-
cients. Next, the software constructs a longer expression that includes the likeli-
hood of getting exactly the observed survival times for all the participants in the
data set. And if this isn’t already complicated enough, the expression has to deal
with the issue of censored data. At this point, the software seeks to find the values
of the regression coefficients that maximize this very long likelihood expression
(similar to the way maximum likelihood is described with logistic regression in
Chapter 18).
Hazard ratios
Hazard ratios (HRs) are the estimates of relative risk obtained from PH regression.
HRs in survival regression play a similar role that odds ratios play in logistic
regression. They’re also calculated the same way from regression output — by
exponentiating the regression coefficients:
» In logistic regression: Odds ratio e Regression Coefflclent
» In PH regression: Hazard ratio e Regression Coefflclent
Keep in mind that hazard is the chance of dying in any small period of time. For
each predictor variable in a PH regression model, a coefficient is produced that —
when exponentiated — equals the HR. The HR tells you how much the hazard rate
increases for the participants positive for the predictor compared to the compari-
son group when you increase the variable’s value by exactly 1.0 unit. Therefore, a
HR’s numerical value depends on the units in which the variable is expressed in
your data. And for categorical predictors, interpreting the HR depends on how you
code the categories.
For example, if a survival regression model in a study of emphysema patients

includes number of cigarettes smoked per day as a predictor of survival, and if the
HR for this variable comes out equal to 1.05, then a participant’s chances of dying
at any instant increase by a factor of 1.05 (5 percent) for every additional cigarette
smoked per day. A 5 percent increase may not seem like much, but it’s applied for
every additional cigarette per day. A person who smokes one pack (20 cigarettes)

per day has that 1.05 multiplication applied 20 times, which is like multiplying by
1.05 20, which equals 2.65. One pack contains 20 cigarettes, so if you change the
units in which you record smoking levels from cigarettes per day to packs per day,
you would use units that are 20 times larger. In that case, the corresponding
regression coefficient is 20 times larger, and the HR is raised to the 20th power
(2.65 instead of 1.05 in this example).
And a two-pack-per-day smoker’s hazard increases by a factor of 2.65 over a

one-pack-per-day smoker. This translates to a 2.65 2 increase (approximately
sevenfold) in the chances of dying at any instant for the smoker compared to a
nonsmoker.
Executing a Survival Regression

As with all statistical methods dealing with time-to-event data, your dependent
variable is actually a pair of variables:
» Event status: The event status variable is coded this way:
• Equal to 1 if the event was known to occur during the observation period
(uncensored)
• Equal to 0 if the event didn’t occur during the observation period (censored)
» Time-to-event: In participants who experienced the event during the

observation period, this is the time from the start of observation to the
occurrence of the event. In participants who did not experience the event
during the observation period, this is the time from the start of observation to
the last time the participant was observed. We describe time-to-event data in
more detail in Chapter 21.
And as with all regression methods, you designate one or more variables as the
predictors. The rules for representing the predictor variables are the same as
described in Chapter 18:
» For continuous numerical variables, choose units of a convenient magnitude.
» For categorical predictors, carefully consider how you recode the data,
especially in terms of selecting a reference group. Consider a five-level age
group variable. Would you want to model it as an ordinal categorical variable,
assuming a linear relationship with the outcome? Or would you prefer using
indicator variables, allowing each level to have its own slope relative to the
reference level? Flip to Chapter 8 for more on recoding categorical variables.

You may not be sure which variables in your data to include as predictors in the
regression. We provide advice on model-building in Chapter 17.
After you assemble and properly code the data, you execute the regression in sta-
tistical software using a similar approach as you use when doing ordinary least-
squares or logistic regression. You need to specify the variables in the regression
model:
1. Specify the two outcome variables.
• The event status variable
• The time-to-event variable
2. Specify the predictor variables.
Make sure you are careful when you include categorical predictors, especially
indicator variables. All the predictors you introduce should make sense
together in the model.
Most software also lets you specify calculations you want to see on the output. You
should always request at least the following:
» Coefficients table, including HRs and their confidence intervals
» Tests of whether the hazard proportionality assumption is valid
You may also want to request the following output:
» Summary descriptive statistics about the data. These can include number of
censored and uncensored observations, median survival time, and mean and
standard deviation for each predictor variable in the model
» One or more measures of goodness-of-fit for the model
» Baseline survival function, which outputs as a table of values and a survival

curve
» Baseline hazard function values, which output as a table and graph
After you specify all the input to the program, execute the code, retrieve the out-
put, and interpret the results.

Interpreting the Output of
a Survival Regression
Suppose that you have conducted a long-term survival study of 200 Stage 4 cancer
patients who were enrolled from four clinical centers (A, B, C, and D) and were
randomized to receive either chemotherapy or radiation therapy. Participants
were followed for up to ten years, after which the survival data were summarized
by treatment (see Figures 23-3a and 23-3b for the two treatments) and by clinical
center.
It would appear, from Figure 23-3, that radiation (compared to chemotherapy)

and clinical centers A and B (compared to C and D) are associated with better sur-
vival. But are these apparent effects statistically significant? PH regression can
answer these and other questions.
FIGURE 23-3:
Kaplan-Meier
survival curves by
treatment and
clinical center.
To run a PH regression on the data from this example, you must indicate the fol-
lowing to the software in your code:
» The time-to-event variable. We named this variable Time, and it was coded
in years. For participants who died during the observation period, it was
coded as the number of years from observation beginning until death. For
participants who did not die during the observation period, it contains
number of years they were observed.
» The event status variable. We named this variable Status, and coded it as
1 if the participant was known to have died during the observation period,
and 0 if they did not die.

» The treatment group variable. In this case, we created the variable
Radiation, and coded it as 1 if the participant was in the radiation group,
and 0 if they were in the chemotherapy group. That way, the coefficient
produced in the output will indicate the increase or decrease in proportional
hazard associated with being in the radiation group compared to the chemo-
therapy group.
» The clinical center variable. In this case, we choose to create an indicator

variable called CenterCD, which is 1 if the participant is from Center C or
Center D, and is 0 if they are from A or B. Alternatively, you could choose to
create one indicator variable for each center, as described in Chapter 18.
If you use a numerical variable such as age as a predictor and enter it into the
model, the resulting coefficient will apply to increasing this variable by one unit
(such as for one year of age).
Using the R statistical software, the PH regression can be invoked with a single
command:
coxph(formula = Surv(Time, Status) ~ CenterCD + Radiation)
Figure 23-4 shows R’s output, using the data that we graph in Figure 23-3. The
output from other statistical programs won’t look exactly like Figure 23-4, but
you should be able to find the main components described in the following
sections.
FIGURE 23-4:
Output of a PH
regression
from R.

Testing the validity of the assumptions
When you’re analyzing data using PH regression, you’re assuming that your data
are consistent with the idea of flexing a baseline survival curve by raising all the
points in the entire curve to the same power (shown as h in Figures 23-1b and
23-2b). You’re not allowed to twist the curve so that it goes higher than the base-
line curve (h 1) for small time values and lower than baseline (h 1) for large
time values. That would be a non-PH flexing of the curve.
One quick check to see whether a predictor is affecting your data in a non-PH way
is to take the following steps:
1. Split your data into two groups, based on the predictor.
2. Plot the Kaplan-Meier survival curve for each group (see Chapter 22).
If the two survival curves for a particular predictor display the slanted figure-
eight pattern shown in Figure 23-5, either don’t use PH regression on those
data, or don’t use that predictor in your PH regression model. That’s because it
violates the assumption of proportional hazards underlying PH regression.
FIGURE 23-5:
Don’t try PH
regression on this
kind of data
because it
violates the PH
assumption.
Your statistical software may offer several options to test the hazard-
proportionality assumption. Check your software’s documentation to see what
it offers and how to interpret the output. It may offer the following:
» Graphs of the hazard functions versus time, which let you see the extent to
which the hazards are proportional.
» A statistical test for significant hazard non-proportionality. R provides a

function called cox.zph for this purpose, and other packages may offer
a comparable option.

Checking out the table of regression
coefficients
A regression coefficients table in a survival regression looks very much like the
tables produced by almost all kinds of regression: ordinary least-squares, logistic,
Poisson, and so on. The survival regression table has a row for every predictor
variable, usually containing the following items:
» The value of the regression coefficient. This says how much the log of the HR
increases when the predictor variable increases by exactly 1.0 unit. It’s hard to
interpret unless you exponentiate it into a HR. In Figure 23-4, the coefficient
for CenterCD is 0.4522, indicating that every increase of 1 in CenterCD (which
literally means comparing everyone at Centers A and B to those at Centers C
and D), there is an increase the logarithm of the hazard by 0.4522. When
exponentiated, this translates into a HR of 1.57 (listed on the output under
exp(coef)). As predicted from looking at Figure 24-3, this indicates that those at
Centers C and D together are associated with a higher hazard compared with
those at Centers A and B together. For indicator variables, there will be a row
in the table for each non-reference level, so in this case, you see a row for
Radiation. The coefficient for Radiation is –0.4323, which when exponentiated,
translates to an HR of 0.65 (again listed under exp(coef)). The negative sign
indicates that in this study, radiation treatment is associated with less hazard
and better survival than the comparison treatment, which is chemotherapy.
Interpreting the HRs and their confidence intervals is described in the next
section “Homing in on hazard ratios and their confidence intervals.”
» The coefficient’s standard error (SE), which is a measure of the precision of

the regression coefficient. The SE of the CenterCD coefficient is 0.1013, so you
would express the CenterCD coefficient as 0.45 ± 0.10.
» The coefficient divided by its SE often labeled t or Wald, but designated as z

in Figure 23-4.
» The p value. Under the assumption that α = 0.05, if the p value is less than
0.05, it indicates that the coefficient is statistically significantly different from 0
after adjusting for the effects of all the other variables that may appear the
model. In other words, a p value of less than 0.05 means that the correspond-
ing predictor variable is statistically significantly associated with survival. The p
value for CenterCD is shown as 8.09e–06, which is scientific notation for
0.000008, indicating that CenterCD is very significantly associated with survival.
» The HR and its confidence limits, which we describe in the next section.
You may be surprised that no intercept (or constant) row is in the coefficient table
in the output shown in Figure 23-4. PH regression doesn’t include an intercept in
the linear part of the model because the intercept is absorbed into the baseline
survival function.

Homing in on hazard ratios and their
confidence intervals
HRs from survival and other time-to-event data are used extensively as safety
and efficacy outcomes of clinical trials, as well as in large-scale epidemiological
studies. Depending on how the output is formatted, it may show the HR for each
predictor in a separate column in the regression table, or it may create a separate
table just for the HRs and their confidence intervals (CIs).
If the software doesn’t output HRs or their CIs, you can calculate them from the
regression coefficients and standard errors (SEs) as follows:
» Hazard ratio e Coef
» Lower 95 percent confidence limit e Coef 1.96 SE
» Upper 95 percent confidence limit e Coef 1.96 SE
In Figure 23-4, the coefficients are listed under coef, and the SEs are listed under
se(coef). HRs are useful and meaningful measures of the extent to which a variable
influences survival.
» A HR of 1 corresponds to a regression coefficient of 0, and indicates that the

variable has no effect on survival.
» The CI around the HR estimated from your sample indicates the range in
which the true HR of the population from which your sample was drawn
probably lies.
In Figure 23-4, the HR for CenterCD is e 0.4522 1.57, with a 95 percent CI of 1.29 to
1.92. This means that an increase of 1 in CenterCD (meaning being a participant at
Centers A or B compared to being one at Centers C or D) is statistically signifi-
cantly associated with a 57 percent increase in hazard. This is because multiplying
by 1.57 is equivalent to a 57 percent increase. Similarly, the HR for Radiation (rela-
tive to the comparison, which is chemotherapy) is 0.649, with a 95 percent CI of
0.43 to 0.98. This means that those undergoing radiation had only 65 percent the
hazard of those undergoing chemotherapy, and the relationship is statistically
significant.
Risk factors, or predictors associated with increased risk of the outcome, have HRs
greater than 1. Protective factors, or predictors associated with decreased risk of the
outcome, have HRs less than 1. In the example, CenterCD is a risk factor, and Radi-
ation is a protective factor.

Assessing goodness-of-fit and predictive
ability of the model
There are several measures of how well a regression model fits the survival data.
These measures can be useful when you’re choosing among several different
models:
» Should you include a possible predictor variable (like age) in the model?
» Should you include the squares or cubes of predictor variables in the model
(meaning including age2 or age3 in addition to age)?
» Should you include a term for the interaction between two predictors?
Your software may offer one or more of the following goodness-of-fit measures:
» A measure of agreement between the observed and predicted outcomes

called concordance (see the bottom of Figure 23-4). Concordance indicates the
extent to which participants with higher predicted hazard values had shorter
observed survival times, which is what you’d expect. Figure 23-4 shows a
concordance of 0.642 for this regression.
» An r (or r2) value that’s interpreted like a correlation coefficient in ordinary

regression, meaning the larger the r2 value, the better the model fits the data.
In Figure 23-4, r2 (labeled Rsquare) is 0.116.
» A likelihood ratio test and associated p value that compares the full model,
which includes all the parameters, to a model consisting of just the overall
baseline function. In Figure 23-4, the likelihood ratio p value is shown as
4.46e 06 , which is scientific notation for p 0.00000446 , indicating a model
that includes the CenterCD and Radiation variables can predict survival
statistically significantly better than just the overall (baseline) survival curve.
» Akaike’s Information Criterion (AIC) is especially useful for comparing alternative

models but is not included in Figure 23-4.
Focusing on baseline survival

and hazard functions
The baseline survival function is represented as a table with two columns — time
and predicted survival — and a row for each distinct time at which one or more
events were observed.

The baseline survival function’s table may have hundreds of rows for large data
sets, so instead of printing it, you should save the table as a data file. Then, you
can use it to generate a customized prognosis curve (described in the next section)
for any specific set of values for the predictor variables.
The software may also offer a graph of the baseline survival function. If your soft-
ware is using an average-participant baseline (see the earlier section, “The steps
to perform a PH regression”), this graph is useful as an indicator of the entire
group’s overall survival. But if your software uses a zero-participant baseline, the
curve is not helpful.
How Long Have I Got, Doc? Constructing

Prognosis Curves
A primary reason to use regression analysis is to predict outcomes from any par-
ticular set of predictor values. For survival analysis, you can use the regression
coefficients from a PH regression along with the baseline survival curve to con-
struct an expected survival (prognosis) curve for any set of predictor values.
Suppose that you’re survival time (from diagnosis to death) for a group of cancer
patients in which the predictors are age, tumor stage, and tumor grade at the time
of diagnosis. You’d run a PH regression on your data and have the program gen-
erate the baseline survival curve as a table of times and survival probabilities.
After that, whenever a patient is newly diagnosed with cancer, you can take that
person’s age, stage, and grade, and generate an expected survival curve tailored
for that particular patient. (The patient may not want to see it, but at least it could
be done.)
You’ll probably have to do these calculations outside of the software that you use
for the survival regression, but the calculations aren’t difficult and can be done in
a Microsoft Excel spreadsheet. The example in the following sections uses the
small set of sample data that’s preloaded into the online calculator for PH regres-
sion at https://statpages.info/prophaz.html. This particular example has
only one predictor, but the basic idea extends to multiple predictors.
Obtaining the necessary output

Figure 23-6 shows the output from the built-in example (omitting the Iteration
History and Overall Model Fit sections). Pretend that this model represents sur-
vival, in years, as a function of age for patients just diagnosed with some partic-
ular disease. In the output, the age variable is called Variable 1.

FIGURE 23-6:
Output of PH
regression for
generating
prognostic
curves.
Looking at Figure 23-6, first consider the table in the Baseline Survivor Function
section, which has two columns: time in years, and predicted survival expressed
as a fraction. It also has four rows — one for each time point in which one or more
deaths was actually observed. The baseline survival curve for the example data
starts at 1.0 (100 percent survival) at time 0, as survival curves always do, but this
row isn’t shown in the output. The survival curve remains flat at 100 percent until
year two, when it suddenly drops down to 99.79 percent, where it stays until year
seven, when it drops down to 98.20 percent, and so on.
In the Descriptive Stats section near the start of the output in Figure 23-6, the
average age of the 11 patients in the example data set is 51.1818 years, so the base-
line survival curve shows the predicted survival for a patient who is exactly 51.1818
years old. But suppose that you want to generate a survival curve that’s custom-
ized for a patient who is a different age — like 55 years old. According to the PH
model, you need to raise the entire baseline curve to some power h. This means
you have to exponentiate the four tabulated points by h.
In general, h depends on two factors:
» The value of the predictor variable for that patient. In this example, the value
of age is 55.
» The values of the corresponding regression coefficients. In this example, in

Figure 23-6, you can see 0.3770 labeled as Coeff. in the regression table.

Finding h
To calculate the h value, do the following for each predictor:
1. Subtract the average value from the patient’s value.
In this example, you subtract the average age, which is 51.18, from the patient’s
age, which is 55, giving a difference of +3.82.
2. Multiply the difference by the regression coefficient and call the

product v.
In this example, you multiply 3.82 from Step 1 by the regression coefficient for
age, which is 0.377, giving a product of 1.44 for v.
3. Calculate the v value for each predictor in the model.
4. Add all the v values, and call the sum of the individual v values V.
This example has only one predictor variable, which is age, so V equals the v
value you calculate for age in Step 2, which is 1.44.
5. Calculate e V .
This is the value of h. In this example, e 1.44 gives the value 4.221, which is the h
value for a 55-year-old patient.
6. Raise each of the baseline survival values to the power of h to get the
survival values for the prognosis curve.
In this example, you have the following prognosis:
• For year-zero survival 1.000 4.221 1.000, or 100 percent
• For two-year survival: 0.9979 4.221

0.9912, or 99.12 percent
• For seven-year survival 0.9820 4.221

0.9262, or 92.62 percent
• For nine-year survival 0.9525 4.221

0.8143 , or 81.43 percent
• For ten-year survival 0.8310 4.221

0.4578 , or 45.78 percent
You then graph these calculated survival values to give a customized survival
curve for this particular patient. And that’s all there is to it!
Here’s a short version of the procedure:
1. V = sum of [(patient value – average value) * coefficient] summed over all

the predictors
2. h eV
3. Customized survival baseline survival h

Some points to keep in mind:
» If your software outputs a zero-based baseline survival function, you don’t

subtract the average value from the patient’s value. Instead, calculate the v
term as the product of the patient’s predictor value multiplied by the regres-
sion coefficient.
» If a predictor is a categorical variable, you have to code the levels as numbers.

If you have a dichotomous variable like pregnancy status, you could code not
pregnant = 0 and pregnant = 1. Then, if in a sample only including women,
47.2 percent of the sample is pregnant, the average pregnancy status is 0.472. If
the patient is not pregnant, the subtraction in Step 1 is 0 – 0.472, giving –0.472.
If the patient is pregnant, you would use the equation 1 – 0.472, giving 0.528.
Then you carry out all the other steps exactly as described.
» It’s even a little trickier for multivalued categories (such as different clinical
centers) because you have to code each of these variables as a set of indicator
variables.
Estimating the Required Sample Size

for a Survival Regression
Note: Elsewhere in this chapter, we use the word power in its algebraic sense, such
as in x 2 is x to the power of 2. But in this section, we use power in its statistical
sense to mean the probability of getting a statistically significant result when
performing a statistical test.
Except for straight-line regression discussed in Chapter 16, sample-size calcula-

tions for regression analysis tend not to be straightforward. If you find software
that will calculate sample-size estimates for survival regression, it often asks for
inputs you don’t have.
Very often, sample-size estimates for studies that use regression methods are
based on simpler analytical methods. We recommend that when you’re planning
a study that will be analyzed using PH regression, you base your sample-size esti-
mate on the simpler log-rank test, described in Chapter 22. The free PS program
handles these calculations very well.

You still have to specify the following:
» Desired α level: We recommend 0.05
» Desired power: We usually use 80 percent
» Effect size of importance: This is typically expressed as a HR or as the

difference in median survival time between groups
You also need estimates of the following:
» Anticipated enrollment rate: How many participants you hope to enroll per
time period
» Planned duration of follow-up: How long you plan to continue following all
the participants after the last participant has been enrolled before ending the
study and analyzing your data
If you are uncomfortable with estimating sample size for a large study that will be
evaluated with a regression model, consult a statistician with experience in devel-
oping sample-size estimates for similarly-designed studies. They will be able to
guide you in the tips and tricks they use to arrive at an adequate sample-size cal-
culation given your research question and context.

7
The Part of Tens
IN THIS PART . . .
Find common statistical distributions that may

describe the shape of your data, and common
distribution functions that arise in statistical
significance testing.
Follow simple rules for getting quick estimates of the

number of participants you need for a properly
designed study.
IN THIS CHAPTER
» Delving into distributions that may

describe your data
» Digging into distributions that arise

during statistical significance testing
Chapter 24
Ten Distributions Worth
Knowing
T
his chapter describes ten statistical distribution functions you’ll probably
encounter in biological research. For each one, we provide a graph of what
that distribution looks like, as well as some useful or interesting facts and
formulas. You find two general types of distributions here:
» Distributions that describe random fluctuations in observed data: Your

study data will often conform to one of the first seven common distributions.
In general, these distributions have one or two adjustable parameters that
allow them to fit the fluctuations in your observed data.
» Common test statistic distributions: The last three distributions don’t describe
your observed data. Instead, they describe how a test statistic that is calculated
as part of a statistical significance test will fluctuate if the null hypothesis is true.
The Student t, chi-square, and Fisher F distributions allow you to calculate test
statistics to help you decide if observed differences between groups, associations
between variables, and other effects you want to test should be interpreted as
due to random fluctuations or not. If the apparent effects in your data are due
only to random fluctuations, then you will fail to reject the null hypothesis. These
distributions are used with the test statistics to obtain p values, which indicate
the statistical significance of the apparent effects. (See Chapter 3 for more
information on significance testing and p values.)
CHAPTER 24 Ten Distributions Worth Knowing 351

This chapter provides a very short table of critical values for the t, chi-square, and
F distributions. A critical value is the value that your calculated test statistic must
exceed in order for you to declare statistical significance at the α = 0.05 level. For
example, the critical value for the normal distribution is 1.96 at α = 0.05.
The Uniform Distribution

The uniform distribution is the simplest distribution. It’s a continuous number
between 0 and 1. To generalize, it is a continuous number between a and b, with
all values within that range equally likely (see Figure 24-1). The uniform distribu-
tion has a mean value of ( b a ) / 2 and a standard deviation of b a / 12 . The
uniform distribution arises in the following contexts:
» Round-off errors are uniformly distributed. For example, a weight recorded

as 85 kilograms (kg) can be thought of as a uniformly distributed random
variable between a = 84.5 kg and b = 85.5 kg. This causes the mean to be
(84.5 + 85.5)/2 = 85 kg, with a standard error of (84.4 – 84.5)/√12, which is
1/3.46 = 0.29 kg.
» In the case the null hypothesis is true, the p value from any exact significance
test is uniformly distributed between 0 and 1.
FIGURE 24-1:
The uniform
distribution.
The Microsoft Excel formula RAND() generates a random number drawn from
the standard uniform distribution.
352 PART 7 The Part of Tens

The Normal Distribution
The most popular and widely-used distribution is the normal distribution (also
called the Gaussian distribution and the probability bell curve). It describes variables
whose fluctuations are the combined result of many independent causes.
Figure 24-2 shows the shape of the normal distribution for various values of the
mean and standard deviation. Many other distributions (binomial, Poisson,
Student t, chi-square, Fisher F) become nearly normal-shaped for large samples.
The Microsoft Excel statement NORMSINV(RAND()) generates a normally

distributed random number, with mean 0 and SD 1.
FIGURE 24-2:
The normal
distribution at
various means
and standard
deviations.
The Log-Normal Distribution

This distribution is also called skewed. If a set of numbers is log-normally distrib-
uted, then the logarithms of those numbers will be normally distributed (see the
preceding section “The normal distribution”). Laboratory values such as enzyme
and antibody concentrations are often log-normally distributed. Hospital lengths
of stay, charges, and costs are also approximately log-normal.
You should suspect log-normality if the standard deviation of a set of numbers is

so big it’s in the ballpark of the size of the mean. Figure 24-3 shows the relation-
ship between the normal and log-normal distributions.
If a set of log-normal numbers has a mean A and standard deviation D, then the
natural logarithms of those numbers will have a standard deviation
s Log 1 D / A 2 , and a mean m Log A s 2 / 2.

FIGURE 24-3:
The log-normal
distribution.
The Binomial Distribution

The binomial distribution helps you estimate the probability of getting x successes
out of N independent tries when the probability of success on one try is p. (See
Chapter 3 for an introduction to probability.) A common example of the binomial
distribution is the probability of getting x heads out of N flips of a coin. If the coin
is fair, p = 0.5, but if it is lopsided, p could be greater than or less than 0.5 (such
as p = 0.7). Figure 24-4 shows the frequency distributions of three binomial dis-
tributions, all having p 0.7 but having different N values.
FIGURE 24-4:
The binomial
distribution.
The formula for the probability of getting x successes in N tries when the proba-
bility of success on one try is p is Pr x , N , p ) p x 1 p ) N x N! / [x!( N x ! .
Looking across Figure 24-4, you might have guessed that as N gets larger, the
binomial distribution’s shape approaches that of a normal distribution with mean
Np and standard deviation Np(1 p ) .

The arc-sine of the square root of a set of proportions is approximately normally
distributed, with a standard deviation of 1 / 4 N . Using this transformation, you
can analyze data consisting of observed proportions with t tests, ANOVAs, regres-
sion models, and other methods designed for normally distributed data. For
example, using this transformation, you could use these methods to statistically
compare proportions of participants who responded to treatment in two different
treatment groups in a study. However, whenever you transform your data, it can
be challenging to back-transform the results and interpret them.
The Poisson Distribution

The Poisson distribution gives the probability of observing exactly N independent
random events in some interval of time or region of space if the mean event rate
is m. The Poisson distribution describes fluctuations of random event occurrences
seen in biology, such as the number of nuclear decay counts per minute, or the
number of pollen grains per square centimeter on a microscope slide. Figure 24-5
shows the Poisson distribution for three different values of m.
FIGURE 24-5:
The Poisson
distribution.
The formula to estimate probabilities on the Poisson distribution

is Pr( N , m ) m N e m / N .
Looking across Figure 24-5, you might have guessed that as m gets larger, the
Poisson distribution’s shape approaches that of a normal distribution, with
mean m and standard deviation m.
The square roots of a set of Poisson-distributed numbers are approximately nor-

mally distributed, with a standard deviation of 0.5.

The Exponential Distribution
If a set of events follows the Poisson distribution, the time intervals between con-
secutive events follow the exponential distribution, and vice versa. Figure 24-6
shows the shape of two different exponential distributions.
The Microsoft Excel statement LN(RAND()) makes exponentially distributed

random numbers with mean 1.
FIGURE 24-6:
The exponential
distribution.
The Weibull Distribution

This distribution describes failure times for devices (such as light bulbs), where
the failure rate can be constant, or can change over time depending on the shape
parameter, k. It is also used in human survival analysis, where failure is an out-
come (such as death). In the Weibull distribution, the failure rate is proportional
to time raised to the k – 1 power, as shown in Figure 24-7a.
» If k 1, the failure rate has a lot of early failures, but these are reduced
over time.
» If k 1, the failure rate is constant over time, following an exponential

distribution.
» If k 1, the failure rate increases over time as items wear out.
Figure 24-7b shows the corresponding cumulative survival curves.
The Weibull distribution shown in Figure 24-7 leads to survival curves of the form
k
Survival l e Time , which are widely used in industrial statistics. But survival
methods that don’t assume a distribution for the survival curve are more common
in biostatistics (we cover examples in Chapters 21, 22, and 23).

FIGURE 24-7:
The Weibull
distribution.
The Student t Distribution

This family of distributions is most often used when comparing means between
two groups, or between two paired measurements. Figure 24-8 shows the shape
of the Student t distribution for various degrees of freedom. (See Chapter 11 for
more info about t tests and degrees of freedom.)
FIGURE 24-8:
The Student t
distribution.
In Figure 24-8, as the degrees of freedom increase, the shape of the Student t
distribution approaches that of the normal distribution.
Table 24-1 shows the critical t value for various degrees of freedom at α = 0.05.
Under α = 0.05, random fluctuations cause the t statistic to exceed the critical t
value only 5 percent of the time. This 5 percent includes exceeding t on either the
positive or negative side. From the table, if you determine your critical t is 2.01 at
50 df, and your test statistic is 2.45, it exceeds the critical t, and is statistically
significant at α = 0.05. But this would also be true if your test statistic was –2.45,
because the table only presents absolute values of critical t.

TABLE 24-1 Critical Values of Student t for α = 0.05
Degrees of Freedom t crit
1 12.71
2 4.30
3 3.18
4 2.78
5 2.57
6 2.45
8 2.31
10 2.23
20 2.09
50 2.01
∞ 1.96
For other α and df values, the Microsoft Excel formula =T.INV.2T(α, df) gives the
critical Student t value.
The Chi-Square Distribution

This family of distributions is used most commonly for two purposes: testing
goodness-of-fit between observed and expected event counts, and for testing for
association between categorical variables. Figure 24-9 shows the shape of the
chi-square distribution for various degrees of freedom.
As you look across Figure 24-9, you may notice that as the degrees of freedom
increase, the shape of the chi-square distribution approaches that of the normal
distribution. Table 24-2 shows the critical chi-square value for various degrees of
freedom at α = 0.05.
Under α = 0.05, random fluctuations cause the chi-square statistic to exceed the
critical chi-square value only 5 percent of the time. If the chi-square value from
your test exceeds the critical value, the test is statistically significant at α = 0.05.
For other α and df values, the Microsoft Excel formula = CHIINV(α, df) gives the
critical 2 value.

FIGURE 24-9:
The chi-square
distribution.
TABLE 24-2 Critical Values of Chi-Square for α = 0.05

Degrees of Freedom 2
Crit
1 3.84
2 5.99
3 7.81
4 9.49
5 11.07
6 12.59
7 14.07
8 15.51
9 16.92
10 18.31
The Fisher F Distribution

This family of distributions is frequently used to obtain p values from an analysis
of variance (ANOVA). Figure 24-10 shows the shape of the Fisher F distribution for
various degrees of freedom.

FIGURE 24-10:
The Fisher F
distribution.
Random fluctuations cause F to exceed the critical F value only 5 percent of the
time. If the F value from your ANOVA exceeds this value, the test is statistically
significant at α = 0.05. For other values of α, df1, and df2 , the Microsoft Excel
formula = FINV(α, df1, df2) will give the critical F value.

IN THIS CHAPTER
» Quickly estimating sample size for

several basic statistical tests
» Adjusting for different levels of

power and α
» Adjusting for unequal group sizes and

for attrition during the study
Chapter 25
Ten Easy Ways to
Estimate How Many
Participants You Need
S
ample-size calculations (also called power calculations) tend to frighten
researchers and send them running to the nearest statistician. But if you you
need a ballpark idea of how many participants are needed for a new research
project, you can use these ten quick and dirty rules of thumb.
Before you begin, take a look at Chapter 3 — especially the sections on hypothesis
testing and the power of a test. That way, you’ll refresh your memory about what
power and sample-size calculations are all about. For your study, you will need to
select the effect size of importance that you want to detect. An effect size could be
the difference of at least 10 mmHg in mean systolic blood-pressure lowering
between groups on two different hypertension drugs, or it could be having the
degree of correlation between two laboratory values of at least 0.7. Once you select
your effect size and compatible statistical test, look in this chapter for the rule for
the statistical test you selected to calculate the sample size.
The first six sections tell you how many participants you need to provide complete
data for you to analyze in order to have an 80 percent chance of getting a p value
CHAPTER 25 Ten Easy Ways to Estimate How Many Participants You Need 361
that’s less than 0.05 when you run the test if a true difference of your effect size
does indeed exist. In other words, we are setting the parameters 80 percent power
at α = 0.05, because they are widely used in biological research. The remaining
four sections tell you how to modify your estimate for other power or α values, and
how to adjust your estimate for unequal group size and dropouts from the study.
Comparing Means between Two Groups

» Applies to: Unpaired Student t test, Mann-Whitney U test, and Wilcoxon
Sum-of-Ranks test.
» Effect size (E): The difference between the means of two groups divided by
the standard deviation (SD) of the values within a group.
» Rule: You need 16 /E 2 participants in each group, or 32 /E 2 participants

altogether.
For example, say you’re comparing two hypertension drugs — Drug A and
Drug B — on lowering systolic blood pressure (SBP). You might set the effect size
of 10 mmHg. You also know from prior studies that the SD of the SBP change is
known to be 20 mmHg. Then the equation is E 10 / 20, or 0.5, and you need
16 / 0.5 2, or 64 participants in each group (128 total).
Comparing Means among Three,

Four, or Five Groups
» Applies to: One-way Analysis of Variance (ANOVA) or Kruskal-Wallis test.
» Effect size (E): The difference between the largest and smallest means among
the groups divided by the within-group SD.
» Rule: You need 20 /E 2 participants in each group.
Continuing the example from the preceding section, if you’re comparing three
hypertension drugs — Drug A, Drug B, and Drug C — and if any mean difference
of 10 mmHg in SBP between any pair of drug groups is important, then E is still
10 / 20, or 0.5, but you now need 20/ 0.5 2 , or 80 participants in each group (240
total).

Comparing Paired Values
» Applies to: Paired Student t test or Wilcoxon Signed-Ranks test.
» Effect size (E): The average of the paired differences divided by the SD of the
paired differences.
» Rule: You need 8 /E 2 participants (pairs of values).
Imagine that you’re studying test scores in struggling students before and after
tutoring. You determine a six-point improvement in grade points is the effect size
of importance, and the SD of the changes is ten points. Then E 6 / 10 , or 0.6, and
2
you need 8 / 0.6 , or about 22 students, each of whom provides a before score and
an after score.
Comparing Proportions between

Two Groups
» Applies to: Chi-square test of association or Fisher Exact test.
» Effect size (D): The difference between the two proportions (P1 and P2 ) that
you’re comparing. You also have to calculate the average of the two propor-
tions: P P1 P2 / 2.
» Rule: You need 16 P 1 P / D 2 participants in each group.
For example, if a disease has a 60 percent mortality rate, but you think your drug
can cut this rate in half to 30 percent, then P 0.6 0.3 / 2 , or 0.45, and
D 0.6 0.3 , or 0.3. You need 16 0.45 1 0.45 / 0.3 2, or 44 participants
in each group (88 total).
Testing for a Significant Correlation

» Applies to: Pearson correlation test. It is also a good approximation for the
non-parametric Spearman correlation test.
» Effect size: The correlation coefficient (r) you want to be able to detect.
» Rule: You need 8 /r 2 participants (pairs of values).
Imagine that you’re studying the association between weight and blood pressure,
and you want the correlation test to come out statistically significant if these two
variables have a true correlation coefficient of at least 0.2. Then you need to study
8 / 0.2 2 , or 200 participants.
Comparing Survival between Two Groups

» Applies to: Log-rank test or Cox proportional-hazard regression.
» Effect size: The hazard ratio (HR) you want to be able to detect.
» Rule: The required total number of observed deaths/

2
events 32 / natural log of HR .
Here’s how the formula works out for several values of HR greater than 1:
Hazard Ratio Total Number of Events
1.1 3,523
1.2 963
1.3 465
1.4 283
1.5 195
1.75 102
2.0 67
2.5 38
3.0 27
Your enrollment must be large enough and your follow-up must be long enough
to ensure that the required number of events take place during the observation
period. This may be difficult to estimate beforehand as it involves considering
recruitment rates, censoring rates, the shape of the survival curve, and other fac-
tors difficult to forecast. Some research protocols provide only a tentative esti-
mate of the expected enrollment for planning, budgeting, and ethical purposes.
Many state that enrollment and/or follow-up will continue until the required
number of events has been observed. Even with ambiguity, it is important to fol-
low conventions described in this book when designing to avoid criticism for
departing from good general principles.

Scaling from 80 Percent to
Some Other Power
Here’s how you take a sample-size estimate that provides 80 percent power from
one of the preceding rules and scale it up or down to provide some other power:
» For 50 percent power: Use only half as many participants — multiply the
estimate by 0.5.
» For 90 percent power: Increase the sample size by 33 percent — multiply the
estimate by 1.33.
» For 95 percent power: Increase the sample size by 66 percent — multiply the
estimate by 1.66.
For example, if you know from doing a prior sample size calculation that a study
with 70 participants provides 80 percent power to test its primary objective,
then a study that has 1.33 70, or 93 participants will have about 90 percent
power to test the same objective. The reason to consider power of levels other than
80 percent is because of limited sample. If you know that 70 participants provides
80 percent power, but you will only have access to 40, you can estimate maximum
power you are able to achieve.
Scaling from 0.05 to Some

Other Alpha Level
Here’s how you take a sample-size estimate that was based on testing at the
α = 0.05 level, and scale it up or down to correspond to testing at some other
α level:
» For α = 0.10: Decrease the sample size by 20 percent — multiply the

estimate by 0.8.
» For α = 0.025: Increase the sample size by 20 percent — multiply

the estimate by 1.2.
» For α = 0.01: Increase the sample size by 50 percent — multiply the

estimate by 1.5.
For example, imagine that you’ve calculated you need a sample size of 100 partici-
pants using α = 0.05 as your criterion for significance. Then your boss says you
have to apply a two-fold Bonferroni correction (see Chapter 11) and use α = 0.025
as your criterion instead. You need to increase your sample size to 100 x 1.2, or 120
participants, to have the same power at the new α level.
Adjusting for Unequal Group Sizes

When comparing means or proportions between two groups, you usually get the
best power for a given sample size — meaning it’s more efficient — if both groups
are the same size. If you don’t mind having unbalanced groups, you will need
more participants overall in order to preserve statistical power. Here’s how to
adjust the size of the two groups to keep the same statistical power:
» If you want one group twice as large as the other: Increase one group by
50 percent, and reduce the other group by 25 percent. This increases the total
sample size by about 13 percent.
» If you want one group three times as large as the other: Reduce one
group by a third, and double the size of the other group. This increases the
total sample size by about 33 percent.
» If you want one group four times as large as the other: Reduce one group
by 38 percent and increase the other group by 250 percent. This increases the
total sample size by about 56 percent.
Suppose that you’re comparing two equal-sized groups, Drug A and Drug B. You’ve
calculated that you need two groups of 32, for a total of 64 participants. Now, you
decide to randomize group assignment using a 2:1 ratio for A:B. To keep the same
power, you’ll need 32 1.5 , or 48 for Drug A, an increase of 50 percent. For B,
you’ll want 32 0.75, or 24, a decrease of 25 percent, for an overall new total 72
participants in the study.
Allowing for Attrition

Sample size estimates apply to the number of participants who give you complete,
analyzable data. In reality, you have to increase this estimate to account for those
who will drop out of the study, or provide incomplete data for other reasons (called

attrition). Here’s how to scale up your sample size estimate to develop an enroll-
ment target that compensates for attrition, remembering longer duration studies
may have higher attrition:
Enrollment = Number Providing Complete Data × 100/(100 – %Attrition)
Here are the enrollment scale-ups for several attrition rates:
Expected Attrition Increase the Enrollment by
10% 11%
20% 25%
25% 33%
33% 50%
50% 100%
If your sample size estimate says you need a total of 60 participants with complete
data, and you expect a 25 percent attrition rate, you need to enroll 60 1.33 , or 80
participants. That way, you’ll have complete data on 60 participants after a quar-
ter of the original 80 are removed from analysis.
Index
Symbols
alternative hypothesis, 40, 43, 144, 150–151, 324
Alzheimer’s disease, 91, 146
α (alpha)
amputation, 292–293
Bonferroni, 153–154
analysis of variance (ANOVA)
definition of, 41
assessing, 152–157
level, 206
introduction to, 11, 47–49
relation to sample sizes, 365–366
using, 143–145, 158
setting, 43
analytic dataset, 76
* (asterisk), 155
analytic research, 88–90
β (beta), 41
analytic study designs, 91
λ (half-life), 280–282
analytic suite, 57–58
κ (kappa), 189–190
analyzing data, 7, 9–10, 74–76
μg (micrograms), 280–282
and rule, 31
Π (pi), 27–28
animal research, 1
√ (radical sign), 21
ANOVA (analysis of variance)
Σ (sigma), 27–28
assessing, 152–157
γ (skewness coefficient), 121
introduction to, 11, 47–49
using, 143–145, 158
A anticipated enrollment rate, 347
absolute values, 23 antilogarithm, 22, 118–119
accuracy, 37, 38, 262–264 anti-synergy, 245–247
active group, 187 area under the ROC curve (AUC), 265, 280
actuarial life tables, 307, 311–316, 320–321 arguments, 23
addition, 18–19 arithmetic mean, 115–116
additive, 296 arrays, 25–27
administrative measurements, 63 asbestos, 296–298
adverse events, 70 asymptomatic confidence limits, 134
agriculture, 1 attrition, 366–367
Akaike’s Information Criterion (AIC), 259, 276– average values, 11, 39–40, 141–158
277, 342
B
alcohol intake, 94–98
alpha (α)
background information, 69
Bonferroni, 153–154
backward elimination approach, 295
definition of, 41
bad fit line, 215–216
level, 206
balanced groups, 154
relation to sample sizes, 365–366
bar charts, 113, 126
setting, 43
Index 369
base-2 logarithms, 22 case report forms (CRFs), 74
base-10 logarithms, 22 case reports, 91
base-e logarithms, 22 case series, 91
baseline survival function, 342–344 case studies, 91
baseline values, 96 case-control study, 90, 93–96
beta (β), 41 CAT (computerized tomography) scans, 188
bimodal (two-peaked) distribution, 114, 117 categorical data, 112–114
binary logarithms, 22 categorical variables, 236–238
binary variables, 173–174, 177, 236–238 causal inference, 11–12, 87, 90–95, 247
binomial distribution, 36, 354–355 CBD (cannabidiol), 159–166
biology, 1 censoring, 302–306, 329, 335
biopsy specimens, 188 census, 33–34
biostatisticians, 60, 141, 161, 168–169 center, 115
biostatistics, definition of, 1, 7 centiles, 120
bivariate analysis, 11, 128, 145, 160 central tendency, 115
blinding, 66, 70, 171 central-limit theorem (CLT), 134
blocked randomization, 67 charts and charting. See also graphing
blood pressure bar and pie, 113, 126
of study participants, 116–117 box-and-whiskers, 127–128
variable name, 18 categorical data, 112–114
body mass index (BMI), 177 correlation coefficients, 202–203
bolus, 280 hazard rates and survival probabilities, 312–315
Bonferroni adjustment, 153–156 multiple regression, 243–254
box-and-whiskers charts, 127–128 numerical data, 124–128
Bradford Hill’s criteria of causality, 297–298 Poisson regression, 273–276
Receiver Operator Characteristics (ROC),
264–265
C residuals, 222–223
calculated values, 37 scatter plots, 214, 219, 221, 238–240
calculations, 8 software for, 57–58
cancer s-shaped data, 252–256
as a categorical variable, 236–238 student t test, 45–47
liver, 93–97 CHD (coronary heart disease), 92–93
lung, 296–298 chemotherapy, 337
relation to weight, 11 chi-square distribution, 165–166, 358–359
remission, 318, 325 chi-square test
stages, 102 pros and cons, 167–169
survival data, 208, 301–306, 337, 343 sample size, 171–172
candidate covariates, 291 tables, 13
cannabidiol (CBD), 159–166 using, 11, 161–167, 174
car accidents, 274–278

chronic obstructive pulmonary disease confidence level (CL), 39
(COPD), 302 confidence limits (CL), 130, 134
CI (confidence interval) confidentiality, 71
calculating, 226 confounding
definition of, 10, 39, 130–131 adjusting for, 145, 294–296
using, 131–138, 183, 194–197 criteria for, 292
CIR (cumulative incidence rate), 178–179 definition of, 66, 94
CL (confidence level), 39 residual, 68
CL (confidence limits), 130, 134 constants, 17
classification tables, 262–264 contingency tables, 173–178, 184, 187–188
ClinCalc, 172 control group, 98, 154, 187
clinical center variable, 338 convenience sampling, 84–85
clinical trials, 9–10, 61–74 COPD (chronic obstructive pulmonary
cloud-based storage, 60 disease), 302
CLT (central-limit theorem), 134 coronary heart disease (CHD), 92–93
cluster sampling, 84 correlation, 40, 201–202, 220
Cochrane Collaboration, 297 correlation coefficient
code, 241 analyzing, 203–207
code-based methods, 60 description of, 202–203
coding categories, 105–107 example of, 10, 40
coefficient of determination, 227 straight-line regression, 227–228
coefficient of variation (CV), 120 table, 242
Cohen’s Kappa, 189 correlational studies, 91–93
cohort study, 90, 93–96 COVID-19, 183–184
collecting data Cox, David (biostatistician), 330
introduction to, 1–2, 9–10 Cox proportional hazards regression, 330
manually and digitally, 74, 101–110 Cox/Snell R-square, 259
collinearity, 246–247, 266 CRAN (Comprehensive R Archive Network), 58
commercial software, 54–58 CRFs (case report forms), 74
common logarithms, 22 crossover design, 65
comparing averages, 142 cross-sectional studies, 87–96, 176
complete separation, 267–268 cross-tabulated data
complicated formulas, 24 analyzing, 160–167, 171–172
Comprehensive R Archive Network (CRAN), 58 example of, 112–113, 166
computer science, 18 introduction to, 11
computerized tomography (CAT) scans, 188 relation to logistic regression, 250
concordance, 342 tables, 173–174
confidence interval (CI) cumulative incidence rate (CIR), 178–179
calculating, 226 cumulative survival probability, 311–314
definition of, 10, 39, 130–131 curved-line relationships, 214
using, 131–138, 183, 194–197 CV (coefficient of variation), 120
Index 371
D difference, 147–148, 163
difference table, 163
data. See also cross-tabulated data
disease, 191, 193–194
analyzing and collecting, 1, 7, 9–10, 74–76,
101–110 dispersion, 115, 119
categorical, 112–114 distribution center, 115
free-text, 103 distributions
interval and ordinal, 102 bimodal (two-peaked), 114–117
ratio, 102, 107–108 binomial, 36, 354–355
skewed and unskewed, 11, 114, 121, 353 chi-square, 165–166, 358–359
survival, 208, 301–306, 337, 343 exponential, 356
time, 108–110 Fisher F, 152, 359–360
data close-out, 76 frequency, 47–48
data dictionary, 110 leptokurtic and platykurtic, 122
data safety monitoring board (DSMB), 73, 76 normal, 13, 36, 114, 353
data safety monitoring committee (DSMC), 73 probability, 35–37
data snapshot, 76 sampling, 38
date data, 108–110 statistical, 13
date of last contact, 303 student t, 357–358
DBP (diastolic blood pressure), 116–117 weibull, 330, 356–357
deciliter (dL), 280 District of Columbia, 34–35
decision theory, 10, 39–40 division, 20
degrees of freedom (df) dL (deciliter), 280
calculating, 147–148, 152–153 dose-response relationship, 298
for chi-square tests, 166–167, 358–359 double-blinding, 66, 97
dementia, 91, 146, 187–188 double-precision numbers, 107
denominator, 41–42 drug description, 70
dependent variable, 208, 213, 235–245 drug development research, 280–282
descriptive research, 88–90 DSMB (data safety monitoring board), 73, 76
descriptive study designs, 91 DSMC (data safety monitoring committee), 73
desired power, 347 Dupont, William D. (biostatistician), 60
desired α level, 347
determinants, 191
E
deviation, 119, 258–259
ECG (electrocardiogram), 188
df (degrees of freedom)
ecologic fallacy, 93
calculating, 147–148
ecologic studies, 91–93
for chi-square test, 166–167, 358–359
effect modification, 296–297
numerator and denominator, 152
effect size
diabetes, 135–136, 236–240.See also Type II
compared to power and sample size, 45–47
diabetes
definitions of, 362–364
diagnostic procedures, 183–188
example of, 39
diastolic blood pressure (DBP), 116–117, 123–124
of importance, 158, 206, 361
dichotomous variables, 173–174, 249, 269

efficacy objectives, 62–63
electrocardiogram (ECG), 188
F
F ratio, 152
elements, 24, 27–28
F statistic, 228. 242
elimination constant rate, 282
F value (value of F statistic), 155
engineering, 18
factorials, 22
enzyme levels, 125–128
failure times, 356–357
Epi Info, 59
fasting glucose values, 25–26, 153
epidemiological research, 1, 9–12
fat intake, 92–93
epidemiology, 191, 291–298
FDA (Food and Drug Administration), 71
Epidemiology for Dummies (Mitra), 92
first quartile, 222
equal variance, 143
Fisher, Ronald Aylmer (biostatistician), 196
equations, 15, 24–25, 35–36
Fisher Exact test, 11, 13, 169–172
error, 78
Fisher F distribution, 152, 359–360
estimate value, 242
Fisher z transformation, 204–205
estimation theory, 10
fleas, 7
event status variable, 337
floating point numbers, 107
evidence levels, 91
Food and Drug Administration (FDA), 71
Excel (Microsoft)
formulas
for data collection, 103, 105, 107–110
building blocks of, 17
functions of, 57
creating, 11–12
for log-rank tests, 319–320
definition of, 15
for randomization, 67
introduction to, 1–2, 7–8
for straight-line regression, 217
types of, 16, 24
for survival regression, 343
forward stepwise approach, 294–295
exclusion, 74
fourfold tables, 11, 173–178, 184, 187–188
exclusion criteria, 64
fractional numbers, 107
expected count, 164
free software, 58–60
expected survival, 329, 331, 343–346
free-text data, 103
experimental research, 61, 88–90, 97–98
frequency bar charts, 113
experiments, 1, 9–10, 91
frequency distribution, 47–48
expert opinion, 91
functions, 23
explicit constants, 17
exploratory analysis, 294
exploratory efficacy objective, 63 G
exploratory objectives, 62–63 gamma-ray radiation, 251–256, 260–263, 267,
exponential distribution, 356 337–338
exponential increase, 277 Gaussian distribution, 353
exponentiating, 21 generalized linear model (GLM), 272–278
exposure, 83, 175, 178, 193, 292 genetics, 1
extrapolation, 108 geometric mean (GM), 118–119
Index 373
GLM (generalized linear model), 272–278 historical control, 142
glucose values, 25–26, 153 H-L test, 259
gold standard test, 183–184 homogeneity of variances, 155
good fit line, 215–216 hormone concentration, 287–290
goodness of fit, 227–228, 258–259 Hosmer-Lemeshow Goodness of Fit test, 259
G*Power HR (hazard ratios), 334–335, 340
description of, 59–60, 68 HTN (hypertension), 90, 94–98, 177–178
using, 158, 198, 207, 324–325 human health research, 88
graphing. See also charts and charting human subjects protection certification, 73
categorical data, 112–114 hyperplane, 234
correlation coefficients, 202–203 hypertension (HTN), 90, 94–98, 177–178
hazard rates and survival probabilities, 312–315 hypothesis, 40–47, 63, 94, 247.See also null
multiple regression, 243–245 hypothesis
numerical data, 124–128 hypothesis-driven analysis, 294
Poisson regression, 273–276 hypothesized cause, 83, 175, 178, 193, 292
Receiver Operator Characteristics (ROC),
264–265
residuals, 222–223
I
ICF (Informed Consent Form), 72–73
software for, 57–58
ICH (International Conference on
s-shaped data, 252–256 Harmonization), 72
student t test, 45–47 icons explained, 3
GraphPad, 67 identification (ID) numbers, 104
Greek letters, 17 identity line, 245
GUI (guided user interface), 56, 58 imputation, 75
incidence, 191–198
H incidence rate, 192–198
h value, 344–346 inclusion criteria, 64
half-life (λ), 280–282 independent t test, 148
hazard rate independent variable, 208–210, 213, 215, 291–294
definition of, 305 indicator variables, 237–238
from life tables, 311–315 indices, 174
relation to survival rate, 333 individual-level data, 160
hazard ratios (HR), 334–335, 340, 364 inferential statistics, 34, 77
health insurance, 112–114 inferring, 10
healthcare, 9–10 infinity, 33
highway accidents, 274–278 influenza, 192
Hill, Bradford (epidemiologist) Informed Consent Form (ICF), 72–73
Bradford Hills’ criteria of causality, 297–298 inner mean, 118
histogram, 34–35, 124–125 integers, 107

interaction, 296–297 Kruskal, William (statistician), 141
interaction terms, 237, 329 Kruskal-Wallis test, 47–49, 144, 157
intercept, 215, 221, 224–225, 272, 329 kurtosis, 121–122
intercept row, 224
interim analysis, 76
International Conference on Harmonization
L
(ICH), 72 Last Observation Carried Forward (LOCF), 75
International Review Board (IRB), 72 last-seen date, 303
Internet sources. See also G*Power; Microsoft LazStats, 59
Excel least-squares line, 234
ClinCalc, 14 left skewed data, 121
Comprehensive R Archive Network (CRAN), 58 leptokurtic distribution, 122
Epi Info, 59 lethal dose, 262
GraphPad, 67 levels of measurement, 102–103
International Business Machines, 57 LFU (lost to follow-up), 303–304, 308–309
National Health and Nutrition Examination life sciences, 1
Survey (NHANES), 82, 86, 93, 148–157 life-table method, 307, 311–316, 320–321
National Institutes of Health (NIH), 72–73 likert scale, 102, 106
OpenStat and LazStats, 59 line of best fit, 215–216
Power and Sample Size Calculation (PS), 60, linear combination, 329
171–172, 324–325
linear function, 210–211
SAS OnDemand for Academics (ODA), 55–56
linear model, 272–273
Statista, 34
linear regression, 210
StatPages, 158, 190, 282
link function, 273–274
interpolation, 208
liver cancer, 93–97
inter-quartile range (IQR), 120
locally weighted scatterplot smoothing (LOWESS)
inter-rater reliability, 188–190 curve-fitting, 12, 286–290
interval data, 102 LOCF (Last Observation Carried Forward), 75
intervention-related measurements, 63 logarithms, 21–22, 118–119
interventions, 61–62 logistic regression
intra-rater reliability, 188–190 basics of, 251–254
IQR (inter-quartile range), 120 definition of, 12, 210
IRB (International Review Board), 72 disadvantages of, 266–268
iterative models, 246–247 evaluating, 257–265
sample size for, 268–269
K using, 249–250, 255–257
Kaplan-Meier method, 313–316, 328, 330 log-normal distribution, 36, 125, 353–354
Kaplan-Meier (K-M) survival estimate, 313, 317 log-rank test, 317–325, 328–330
kappa (κ), 189–190 longitudinal research, 90
kilograms (kg), 218–220, 225–226, 239 lost to follow-up (LFU), 303–304, 308–309
Index 375
LOWESS (locally weighted scatterplot smoothing) Mitra, Amal K. (author)
curve-fitting, 12, 286–290 Epidemiology for Dummies, 92
lung cancer, 296–298 mmHg (millimeters of mercury), 218–226,
229, 239
M mode, 117
model building, 246
Mann, Henry (professor), 141
model fit statistics, 242
Mann-Whitney U test, 47–49, 143, 362
models
Mantel-Cox test. See log-rank test
generalized linear (GLM), 272–278
Mantel-Haenszel chi-square test, 168
linear, 272–273
margin of error (ME), 134
null, 228, 242, 259
marginal totals, 160
parsimonious, 293
masking, 66, 70, 171
predictive, 228–229
mathematical expressions, 15
regression, 68, 208–209
mathematical operations, 18–25
molecular biology, 7
matrix, 26
multicollinearity, 246–247
matrix algebra, 26
multi-dimensional arrays, 26–27
maximum value, 222
multilevel variable, 236
ME (margin of error), 134
multiple regression
mean
basics of, 234–235
arithmetic, 115–116
introduction to, 26
compared to other values, 142–157, 362
confidence limits, 134–135
special considerations, 245–247
mean square (mean Sq), 155
using, 236–245
measurements, 63–64, 102–103
multiple R-squared, 242
mechanical function, 279
multiplication, 18–20
median, 116–117, 123, 222
multiplicative, 296
meta-analyses, 97–98
multiplicity, 75–76
metadata, 110
multi-site study, 104
mice, 1, 318
multi-stage sampling, 85–86
micrograms (μg), 280–282
multivariable regression, 291
Microsoft Excel
multivariate analysis, 128, 145, 291
for data collection, 103, 105, 107–110
functions of, 57
for log-rank tests, 319–320 N
for randomization, 67 Nagelkerke R-square, 259
for straight-line regression, 217 National Health and Nutrition Examination Survey
for survival regressions, 343 (NHANES), 82, 86, 93, 148–157
millimeters of mercury (mmHg), 218–226, 229, 239 National Institutes of Health (NIH), 72–73
minimum value, 222 natural logarithms, 22
missing data, 74–75 negative predicted value (NPV), 187

negatively skewed data, 121 observed versus predicted graph, 245
NHANES (National Health and Nutrition ODA (SAS OnDemand for Academics), 55–56
Examination Survey), 82, 86, 93, 148–157 odds, 32–33, 181
NIH (National Institutes of Health), 72–73 odds ratio (OR), 94–96, 181–183, 266–267
nominal variables, 102 Office for Human Research Protections (OHRP), 72
non-code-based methods, 60 one-dimensional arrays, 25
nonlinear function, 211 one-group t test, 148
nonlinear least-squares regression, 12 one-sided confidence interval, 133
nonlinear regression, 279–286 one-way ANOVA, 144
nonlinear trends, 277 open-source software, 58–59
nonparametric regression, 286–290 OpenStat, 59
nonparametric tests, 48–49, 157 OR (odds ratio), 94–96, 181–183, 266–267
non-proportional hazards, 323 or rule, 31
non-sampling error, 78 order of operation, 24
non-steroidal anti-inflammatory drugs (NSAIDS), ordinal data, 102
159–166
ordinary multiple linear regression model. See
normal distribution, 13, 36, 114, 353 multiple regression
normal Q-Q graph, 222–223 ordinary regression, 210
normal-based confidence intervals, 134 outcome, 208
normal-based confidence limits, 134 outcome-related measurements, 64
normality assumption, 143 outliers, 229
not rule, 31 overall accuracy, 185
notches, 128
NPV (negative predicted value), 187
NSAIDs (non-steroidal anti-inflammatory drugs), P
159–166 p value
nuisance variables, 145 definition of, 41, 242
null hypothesis determining, 165–167
definition of, 40 evaluating, 42, 144, 147–157, 226–227, 340
evaluating, 42–43, 150, 152–153, 322, 351–352 from the H-L test, 259
example of, 161–165, 319 paired t test, 148
null model, 228, 242, 259 paired values, 363
numerator, 41–42 parabolic relationship, 214, 252–253
numerical data, 107–109, 114–123 parallel design, 65
parameters, 33–34, 77, 208, 279
O parametric tests, 48–49

parsimonious models, 293
obesity, 177–178, 182–183
parsimony, 293
observational research, 88–90
participant identification (ID), 240
observed count, 164
participant study identifier, 104
Index 377
participants. See also sample size; samples PPV (positive predictive value), 187
enrolling, 68 precision, 37, 38, 147
protection for, 71–73 predicted values, 242
selecting, 64–65, 236–237 predictive model, 228–229
PatSat, 106 predictive value negative, 187
PCR (polymerase chain reaction), 184 predictive value positive, 187
Pearson, Karl (biostatistician), 161 predictors
Pearson Correlation test, 47–49, 227 introduction to, 208, 233
Pearson kurtosis index, 122 in iterative models, 246–247
percentile, 120 in logistic models, 255–256
perfect predictor problem, 267–268 in regression models, 273–274, 279
perfect separation, 267–268 relation to the outcome, 242, 245–246, 250
periodicity, 83 types of, 209, 235–236
PH (proportional hazards regression), 330–331, pregnancy, 171, 185–187
333–334 prevalence, 179, 186, 191–194
pharmacokinetic (PK) properties, 280–282 prevalence ratio, 179
pi (Π), 27–28 primary diagnosis (PrimaryDx), 236–238
pie charts, 113 primary efficacy objective, 62
pilot study, 230 primary objectives, 62
placebo, 66–67, 171, 187–188 primary sampling units (PSU), 86
placebo effect, 66, 187–188 privacy, 71
plain text format, 16, 24 probability, 30–33
platykurtic distribution, 122 probability bell curve, 353
Plummer, Walton D. (biostatistician), 60 probability distributions, 35–37
pointy-topped distribution, 114 probability of independence, 166
poisson distribution, 36, 355 procedural descriptions, 70
poisson regression product, of an array, 27
definition of, 12, 210 prognosis curves, 329, 331, 343–346
using, 271–278 proportional hazards (PH) regression, 330–331,
polymerase chain reaction (PCR), 184 333–334
populations, 33–37, 175 proportions, 11, 135–136, 363
positive predictive value (PPV), 187 protective factor, 178
positively skewed data, 121 protractor, 113
post-hoc tests, 143, 152–157 PS (Power and Sample Size Calculation)
potential confounding variables, 64 for chi-square and Fisher exact tests,
Power and Sample Size Calculation (PS) 171–172
for chi-square and Fisher exact tests, 171–172 definition of, 60
definition of, 60 for survival comparisons, 324–325
for survival comparisons, 324–325 pseudo-r-squared values, 259
power calculations, 47, 171–172, 361, 365 PSU (primary sampling units), 86
powers, 20–21, 41, 44, 45–47, 206 Python, 58

R straight-line
basics of, 215–216
R (software)
disadvantages of, 229–231
description of, 58
evaluating, 220–224
nonlinear regression, 282–286
using, 216–220
odds ratio calculation, 183
when to use, 213–215
risk ratio calculation, 180–181
survival
straight-line regression, 221
concepts of, 329–335
r value, 203–206
definition of, 210
radiation exposure, 251–256, 260–263, 267,
337–338 evaluating, 337–343
radical sign (√), 21 sample size for, 346–347
random number generator (RNG), 80–81 using, 335–336
random shuffling, 67 when to use, 328–329, 343
random variability, 158 univariate, 209
randomization, 97, 171 regression analysis, 12, 207–208
randomized controlled trials (RCTs), 65–67, 98 regression models, 68, 208–209
randomness, 33 relative frequency, 30
range, 120 relative risk, 95, 178–181
ranks, 49 REM (Roentgen Equivalent Man), 251–254,
260–262, 267
rate ratio (RR), 195–196
research
ratio data, 102, 107–108
analytic, descriptive and observational, 88–90
rationale, 69
animal, 1
RCTs (randomized controlled trials), 65–67, 98
epidemiological, 1, 9–12
Receiver Operator Characteristics (ROC), 257,
264–265 experimental, 61, 88–90, 97–98
reference level, 237 human health, 88
regression longitudinal, 90
logistic research studies, 1
basics of, 251–254 residual information, 242
definition of, 12, 210 residual standard error, 222, 242
disadvantages of, 266–268 residuals, 222–224, 242–245
evaluating, 257–265 residuals versus fitted graph, 222–223
sample size for, 268–269 retinopathy, 292–293
using, 249–250, 255–257 right skewed data, 121
multiple risk ratio, 96, 178–181
basics of, 234–235 RMS (root-mean-square), 222
introduction to, 26 RNG (random number generator), 80–81
sample size for, 247–248 ROC (Receiver Operator Characteristics), 257,
264–265
special considerations, 245–247
Roentgen Equivalent Man (REM), 251–254,
using, 236–245 260–262, 267
multivariable, 291 Roman letters, 17
ordinary, 210
Index 379
root-mean-square (RMS), 222 SAS OnDemand for Academics (ODA), 55–56
roots, 21 SBP (systolic blood pressure)
Rothman’s causal pie, 297–298 comparing, 39
round-off error, 352 effect of drugs on, 62, 123–125
RR (rate ratio), 195–196 relation between weight and, 218–229
RStudio, 58 variable name of, 18
Rumsey, Deborah J. (author) scatter plots
Statistics For Dummies, 2, 29 creating, 219, 238–240
Statistics II for Dummies, 29 example of, 221
types of, 214
S Scheffe’s test, 154–156

scientific notation, 28
safety considerations, 70
screening tests, 184–187
safety objectives, 62–63
SD (standard deviation), 119, 123, 143, 222
safety study, 62
SE (standard error)
sample size
calculating, 147–148, 179–180
for chi-square and Fisher exact tests, 171–172
of coefficients, 225–226, 340
for cohort studies, 95–96
compared to confidence intervals, 130–131
compared to power and effect size, 44–47
description of, 38, 129, 163, 242
for comparing averages, 158
in a fraction, 41
for correlation tests, 206–207
secondary efficacy objective, 62
estimating, 198, 361–367
secondary objectives, 62
introduction to, 14
sensitivity, 185–186, 262–264
for logistic regression, 268–269
sigma (Σ), 27–28
for multiple regression, 247–248
significance, 40, 42–43, 174
relation to confidence intervals, 130–131
significance tests. See statistical tests; specific test
for straight-line regression, 230–231 names
for survival comparisons, 324–325 significant association, 11
for survival regression, 346–347 significant correlation, 363–364
sample statistic, 175 simple formulas, 24
samples. See also sample size simple random samples (SRS), 80–81
framing, 78–79 simple randomization, 66
introduction to, 9–10, 14 simple regression, 209
selecting, 33–37, 68, 78–80 simulation, 79
types of, 80–86 single-blinding, 66
sampling clusters, 83–84 single-precision numbers, 107
sampling distribution, 38 single-site study, 104
sampling error, 34, 78 skewed data, 11, 114, 121, 353
sampling frame, 78 skewness, 121
sampling strategies, 175–176 skewness coefficient (γ), 121
SAP (Statistical Analysis Plan), 70 slope row, 224
SAS (Statistical Analysis System), 54–56, 58, 110 slopes, 215–217, 221, 224–231

smoking, 296–298, 334–335 Statistics II For Dummies (Rumsey), 29
smoothing fraction, 289–290 StatPages, 158, 190, 282
Social Science Statistics, 170 stepped line charts, 312–313
software stepwise selection, 195
for data collection, 105–110, 241 storage modes, 107
evolution of, 54 straight-line regression
introduction to, 8 basics of, 215–216
for logistic regression, 256–257 disadvantages of, 229–231
for power calculations, 47 evaluating, 220–224
for straight-line regression, 217 using, 216–220
types of, 54–60 when to use, 213–215
variables in, 18 stratified samples, 81–82
Spearman Rank Correlation test, 47–49 strong linear relationship, 214
specificity, 185–186, 262–264 Structured Query Language (SQL), 110
spot-checking, 109 student t distribution, 357–358
SPSS (Statistical Package for the Social Sciences), student t test, 41–42, 47–49, 142–152, 362–363
54–58 student t value, 226
SQL (Structured Query Language), 110 study design, 2, 7, 14, 88–90
square root, 332 study protocol, 68–71
square root law, 131 study rationale, 69
squaring, 332 study title, 69
SRS (simple random samples), 80–81 subtraction, 18–19
s-shaped relationship, 214, 252–256 sum, 27
SSQ (sum of squares), 155, 215–216, 234 sum of squares (SSQ), 155, 215–216, 234
standard deviation (SD), 119, 123, 143, 222 summarizing data
standard error (SE) categorical, 112–114
calculating, 147–148, 179–180 numerical, 114–123
of coefficients, 225–226, 340 survival, 302–316
compared to confidence intervals, 130–131 summary statistics, 111
description of, 38, 129, 163, 242 surveillance, 89–90
in a fraction, 41 survival analysis, 13, 68, 301
statistic, definition of, 40 survival curve shapes, 329–330
Statistical Analysis Plan (SAP), 70 survival data, 208, 301–306, 337, 343
Statistical Analysis System (SAS), 54–56, 58, 110 survival rate, 305, 364
statistical distributions, 13 survival regression
statistical estimation theory, 10, 37–39 concepts of, 329–335
statistical inference, 10, 37 definition of, 210
Statistical Package for the Social Science (SPSS), evaluating, 337–343
54–58
statistical tests, 8, 40–41, 44, 47–49.See also tests
using, 335–336
statistically rare, 93
when to use, 328–329, 343
Statistics For Dummies (Rumsey), 2, 29
survival time, 301–306, 317–325
Index 381
symbolic constants, 17 treatment periods, 65
symmetry, 115 treatments, 187–188
synergy, 245–247 trees, 1
systematic error, 37 trend line, 276
systematic sampling, 82–83 trimmed mean. See inner mean
systemic reviews, 97–98 true value, 37
systolic blood pressure (SBP) Tukey-Kramer test, 154–156
comparing, 39 Tukey’s HSD (“honestly” significant difference test),
effect of drugs on, 62, 123–125 154–156
relation between weight and, 218–229 two-dimensional arrays, 26
variable name of, 18 two-peaked (bimodal) distribution, 114, 117
type I error, 41, 42–44, 75–76, 152
Type II diabetes, 10–12, 192–197, 250, 292
T type II error, 41–44
t tests, 41–42, 47–49, 142–152, 362–363 typeset format, 16, 24–25
t value, 242 typographic effects, 16
Tableau, 57, 60
terminal elimination rate constant, 12
test statistic, 37, 40, 41–42, 147 U
tests ultrasound, 188
chi-square, 11, 13, 161–169, 171–172, 174 unbalanced groups, 154
Fisher exact, 11, 13, 169–172 under-coverage, 78
H-L, 259 unequal-variance t test, 143
log-rank, 317–325, 328–330 uniform distribution, 352
Mann-Whitney U, 47–49, 143, 362 United States
nonparametric, 48–49, 157 airports, 34–35
post-hoc, 143, 152–157 census, 84
Scheffe’s, 154–156 International Review Board, 72
Spearman Rank Correlation, 47–49 surveillance study, 86
student t, 41–42, 47–49, 142–152, 362–363 univariate analysis, 128
Tukey-Kramer, 154–156 univariate regression, 209
unequal-variance t, 143 Universität Düsseldorf, 59
Wilcoxon Signed-Ranks (WSR), 47–49, 142, 146 unskewed data, 121
Wilcoxon Sum-of-Ranks, 47–49, 143, 157, 362
theoretical function, 279 V
third quartile, 222 value of F statistic (F value), 155
three-way ANOVA, 144–145 values. See also p value
tied values, 49 absolute, 23
time data, 108–110 average, 11, 141–158
time-to-event variable, 337 calculated, 37
treatment bias, 66 estimate, predicted and t values, 242

F value, 155 Comprehensive R Archive Network (CRAN), 58
h value, 345–346 Epi Info, 59
paired, 363 Graphpad, 67
positive predictive, 187 International Business Machines, 57
pseudo-r-squared, 259 National Health and Nutrition Examination
r value, 203–206 Survey (NHANES), 82, 86, 93, 148–157
tied, 49 National Institutes of Health (NIH), 72–73
Vanderbilt University, 60 OpenStat and LazStats, 59
Vanderbilt University Medical Center, 324–325 Power and Sample Size Calculation (PS), 60,
171–172, 324–325
variable names, 110
SAS OnDemand for Academics (ODA), 55–56, 57
variable width, 128
Social Science Statistics, 170
variables
Statista, 34
binary, 173–174, 177, 236–238
StatPages, 158, 190, 282
categorical, 236–238
weibull distribution, 330, 356–357
clinical center, 338
weight, 218–220, 225–226, 239–240, 246
dependent, 208, 213, 235–245
Welch, Bernard Lewis (statistician), 141
dichotomous, 173–174, 249, 269
Welch test, 143, 150–151
event status, 337
Whitney, Donald Ransom (statistician), 141
independent, 208–210, 213, 215, 291–294
whole numbers, 107
indicator, 237–238
Wilcoxon, Frank (statistician), 141
multilevel, 236
Wilcoxon Signed-Ranks (WSR) test, 47–49, 142, 146
nominal, 102
Wilcoxon Sum-of-Ranks test, 47–49, 143, 157, 362
nuisance, 145
withdrawal criteria, 64
potential confounding, 64
World War II, 71, 264
variance, 119, 143, 155
variance table, 155
viruses, 1 X
Viya, 56, 60 X variable, 213–215, 225–228
volume of distribution (Vd), 280–282 x-rays, 188
W Y
Wallis, Wilson Allen (statistician), 141 Y variable, 213–216, 224–228
washout intervals, 65 Yates, Frank (statistician), 168–169
waves, 96, 176 Yates continuity correction, 168–169
weak linear relationship, 214
websites. See also G*Power; Microsoft Excel
ClinCalc, 172
Cochrane, 297
Index 383
About the Authors
Monika M. Wahi, MPH, CPH, is a well-published data scientist with more than
20 years of experience, and president of the public health informatics and educa-
tion firm DethWench Professional Services (DPS) (www.dethwench.com). She is
the author of Mastering SAS Programming for Data Warehousing and has coauthored
over 35 peer-reviewed scientific articles. After obtaining her master of public
health degree in epidemiology from the University of Minnesota School of Public
Health, she has served many roles at the intersection of study design, biostatis-
tics, informatics, and research in the public and private sectors, including at
Hennepin County Department of Corrections in Minneapolis, the Byrd
Alzheimer’s Institute in Tampa, and the U.S. Army. After founding DPS, she
worked as an adjunct lecturer at Laboure College in the Boston area for several
years, teaching about the U.S. healthcare system and biostatistics in their bachelor
of nursing program. At DPS, she helps organizations upgrade their analytics
pipelines to take advantage of new research approaches, including open source.
She also coaches professionals moving into data science from healthcare and
other fields on research methods, applied statistics, data governance, informatics,
and management.
John C. Pezzullo, PhD, spent more than half a century working in the physical,
biological, and social sciences. For more than 25 years, he led a dual life at Rhode
Island Hospital as an information technology programmer/analyst (and later
director) while also providing statistical and other technical support to biological
and clinical researchers at the hospital. He then joined the faculty at Georgetown
University as informatics director of the National Institute of Child Health and
Human Development’s Perinatology Research Branch. He created the StatPages
website (https://statpages.info), which provides online statistical calculating
capability and other statistics-related resources.
Dedication
This book is dedicated to my favorite “dummy,” my dear old dad, Bhupinder Nath
“Ben” Wahi. He is actually a calculus whiz. He used to point to the For Dummies
books when we’d see them at the bookstore and say, “Am I a dummy?” Of course,
I won’t answer that! — Monika
Author’s Acknowledgments
First and foremost, I want to acknowledge and honor the late Dr. John Pezzullo,
the original author of this work. It has been a pleasure to revise his thorough and
interesting writing, delivered with the enthusiasm of a true educator. I am also
extremely grateful for Matt Wagner of Fresh Books, who opened the doors
necessary for my coauthorship of this book. Further, I am indebted to editor
Katharine Dvorak for her constant support, guidance, and helpfulness through the
writing process. Additionally, I want to extend a special “thank you” to my amaz-
ing colleague, Sunil Gupta, who provided his typically excellent technical review.
Finally, I want to express my appreciation for all the other members of the For
Dummies team at Wiley who helped me along the way to make this book a success.
Thank you for helping me be the best writer I can be! — Monika
Publisher’s Acknowledgments
Executive Editor: Lindsay Berg Production Editor: Pradesh Kumar

Managing Editor: Murari Mukundan Cover Image: © ArtemisDiana/Shutterstock
Project Editor: Katharine Dvorak
Technical Editor: Sunil K. Gupta, Founder of
R-Guru.com and SASSavvy.com
WILEY END USER LICENSE AGREEMENT
Go to www.wiley.com/go/eula to access Wiley’s ebook EULA.

Biostatistics For Dummies, 2nd Edition (True PDF

Uploaded by

Copyright:

Available Formats

Biostatistics For Dummies, 2nd Edition (True PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Biostatistics For Dummies, 2nd Edition (True PDF

Uploaded by

Copyright:

Available Formats

Biostatistics

by Monika Wahi, MPH and

Published simultaneously in Canada

LIMIT OF LIABILITY/DISCLAIMER OF WARRANTY: THE PUBLISHER AND THE AUTHOR MAKE NO

Library of Congress Control Number: 2024939111

ISBN 978-1-394-25146-9 (pbk); ISBN 978-1-394-25147-6 (ebk); ISBN 978-1-394-25148-3 (ebk)

Part 1: Getting Started with Biostatistics . . . . . . . . . . . . . . . . . . . . . . 5

Part 2: Examining Tools and Processes . . . . . . . . . . . . . . . . . . . . . . . . 51

Part 3: Getting Down and Dirty with Data . . . . . . . . . . . . . . . . . . . . 99

Part 4: Comparing Groups. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139

CHAPTER 13: Taking a Closer Look at Fourfold Tables. . . . . . . . . . . . . . . . . . . . . . . . . . 173

Part 5: Looking for Relationships with Correlation

CHAPTER 23: Survival Regression. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327

Part 7: The Part of Tens. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349

PART 1: GETTING STARTED WITH BIOSTATISTICS. . . . . . . . . . . . 5

CHAPTER 1: Biostatistics 101. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

CHAPTER 2: Overcoming Mathophobia: Reading and

PART 2: EXAMINING TOOLS AND PROCESSES. . . . . . . . . . . . . . . 51

CHAPTER 4: Counting on Statistical Software . . . . . . . . . . . . . . . . . . . . . . 53

CHAPTER 5: Conducting Clinical Research. . . . . . . . . . . . . . . . . . . . . . . . . . . 61

vi Biostatistics For Dummies

CHAPTER 6: Taking All Kinds of Samples. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

CHAPTER 7: Having Designs on Study Design. . . . . . . . . . . . . . . . . . . . . . . 87

PART 3: GETTING DOWN AND DIRTY WITH DATA . . . . . . . . . . 99

CHAPTER 8: Getting Your Data into the Computer . . . . . . . . . . . . . . 101

Table of Contents vii

CHAPTER 9: Summarizing and Graphing Your Data . . . . . . . . . . . . . 111

CHAPTER 10: Having Confidence in Your Results. . . . . . . . . . . . . . . . . . 129

PART 4: COMPARING GROUPS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139

CHAPTER 11: Comparing Average Values between Groups. . . . . . 141

viii Biostatistics For Dummies

CHAPTER 12: Comparing Proportions and Analyzing

CHAPTER 13: Taking a Closer Look at Fourfold Tables. . . . . . . . . . . . 173

CHAPTER 14: Analyzing Incidence and Prevalence Rates in

CHAPTER 15: Introducing Correlation and Regression . . . . . . . . . . . 201

CHAPTER 16: Getting Straight Talk on Straight-Line

CHAPTER 17: More of a Good Thing: Multiple Regression. . . . . . . . 233

x Biostatistics For Dummies

CHAPTER 18: A Yes-or-No Proposition: Logistic Regression. . . . . . 249

CHAPTER 19: Other Useful Kinds of Regression. . . . . . . . . . . . . . . . . . . . 271

CHAPTER 20: Getting the Hint from Epidemiologic Inference. . . 291

PART 6: ANALYZING SURVIVAL DATA. . . . . . . . . . . . . . . . . . . . . . . 299

CHAPTER 21: Summarizing and Graphing Survival Data. . . . . . . . . 301

CHAPTER 22: Comparing Survival Times . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317

CHAPTER 23: Survival Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327

xii Biostatistics For Dummies

PART 7: THE PART OF TENS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349

CHAPTER 24: Ten Distributions Worth Knowing. . . . . . . . . . . . . . . . . . . 351

CHAPTER 25: Ten Easy Ways to Estimate How

Table of Contents xiii

About This Book

» Students at the undergraduate or graduate level who are taking a course in

» Professionals who have had no formal biostatistical training, and possibly no

If you’re interested in biostatistics, then you’re no dummy! But perhaps you

2 Biostatistics For Dummies

Beyond the Book

Where to Go from Here

» If you’re uncomfortable with mathematical notation, then Chapter 2 is the

» You can get an introduction to human research and clinical trials in

» Using in plain text formulas, such as 5 3