Data Analysis with
Stata
A Comprehensive Guide for Data Analysis and Interpretation of Outputs

First Edition
Stata Version 13

Mohammad Tajul Islam

MBBS, DTM&H, MSc (CTM), MPH
Professor (Adjunct), Department of Public Health
North South University and State University of Bangladesh

Russell Kabir
BDS, MPH, MSc, PhD
Senior Lecturer (Research Methods), Course Leader (MSc Public Health and
Community Wellbeing and MPH Global Public Health)
Anglia Ruskin University, UK

Monjura Nisha
BDS, MPH, PhD
Postdoctoral Research Fellow
The Daffodil Centre, The University of Sydney,
A joint venture with Cancer Council NSW
NSW, Australia

ASA Publications
Dhaka, Bangladesh
ISBN 978-984-35-3165-0 [First Edition]
Copyright © 2022 by Authors
First published 2022
This book is copyrighted by the authors. However, the e-book is free for everyone.
Anyone can download it from the links, print it out for personal use, and share it with
others, but it is strictly prohibited to use it for any kind of profit-making venture with-
out the written permission of the first author. Its contents may be used and incorporated
into other materials with proper acknowledgements and citations. The datasets provid-
ed in the links and used in this book are hypothetical and can be used for practice.

Publisher
ASA Publications
Dhaka, Bangladesh
Email: asapublications7@gmail.com

Distributor
Altaf Medical Book Center
Shop: 121 & 128; Lane: 3; Islamia Market, Nilkhet, Dhaka 1205
Cell: +880-1711-985991; +880-1611-985991; +880-1511-985991

Price: BDT 500 (GBP 15; US$ 20)

Production credits
Editor: Mohammad Tajul Islam
Proofreading: Russell Kabir and Md. Golam Kibria
Composition and Cover Design: Zariath Al-Mamun Badhon
Production Director: Md Altaf Hossain
Printing: ASA Publications
Binding: Rahim Bindings

Suggested Citation
Islam MT, Kabir R, Nisha M, eds. Data Analysis with Stata: A Comprehensive Guide
for Data Analysis and Interpretation of Outputs. Dhaka, Bangladesh: Altaf Publica-
tions; December 2022.

Printed in Bangladesh
To all our family members, students, and promising researchers in
health and social sciences
Preface
Stata is a popular and versatile data analysis software. Many books and guidelines on
Stata are available online or in the market. Stata books focusing on the analysis of
health-related data are scarce. Most of the books used data from sources that are related
to business or psychology, with which health researchers are mostly unfamiliar. All
these factors, and the inspiration of students, encouraged us to write this book. In this
book, we have used simple variables that are commonly used in health and social
science-related research as examples to make it easy and understandable for users.
This book is intended for students (MPH, FCPS, MD, MS, MPhil, PhD, and others),
teachers, and young researchers in health and social sciences. It is written in very
simple language. This book answers three basic questions about data analysis. These
are: a) what to do (what statistics to use for data analysis in order to achieve the objec-
tives); b) how to do (how to analyze data using Stata); and c) what do the outputs mean
(how to interpret the outputs). All these questions are answered in an understandable
manner with examples.
This book covers more than the basic statistical methods of data analysis that are com-
monly used in health and social sciences research. It is the gateway to learning statis-
tics and Stata together, and will help the users go further. This book covers data man-
agement, descriptive statistics, hypothesis testing using bivariate and multivariable
analysis, and others. It is easier to learn through exploration than only reading. Users
are encouraged to explore further once the basics are known. From our understanding,
using the statistics covered in this book, students and researchers will be able to
analyze most of their data from epidemiological studies and publish them in interna-
tional peer-reviewed journals.
We are optimistic that students and researchers will find this book useful while analyz-
ing their data and interpreting the outputs. If you have any comments or suggestions
about this book, feel free to write to the e-mail address below.

M. Tajul Islam
abc.taj@gmail.com

i
Foreword
The book titled "Data Analysis with Stata", written by Dr. M. Tajul Islam, Dr. Russell
Kabir, and Dr. Monjura Nisha, is an immense contribution to teaching, training, and
research in the field of medical sciences and relevant disciplines. The authors of the
book have provided a step-by-step approach from data management to data analysis
using Stata in this book.
The book is a useful resource for data analysis using Stata for undergraduate, postgrad-
uate, and doctoral students of medical sciences and other health-related disciplines,
such as Dental, Pharmacy, Nursing, Occupational Health, Physical Medicine, Biomed-
ical and Social Sciences.
Each chapter is written in a simple manner, focusing on the needs of students. Contents
are systematically presented and described so that students can easily grasp the process
of data analysis and interpret the outputs.
This book will serve as a helpful training tool for university lecturers, dissertation
supervisors, and data analysis instructors.
Readers will benefit from using this book, and I wish for the extensive circulation of
the book.

Professor Hafiz T.A. Khan

PhD (Edin Napier), PostDoc (Oxon), FRSPH, CStat (UK)
Professor of Public Health and Statistics
College of Nursing, Midwifery and Healthcare
University of West London
Middlesex TW8 9GB
United Kingdom
Email: hafiz.khan@uwl.ac.uk

ii
Acknowledgements
We are particularly indebted and grateful to Dr. Shakil Ahmed and Mr. Md. Golam
Kibria, who reviewed all the sections, provided their critical views and constructive
suggestions, and took the lead role in publishing this book. We are particularly thankful
to Professor Hafiz T. A. Khan of the University of West London for his careful review
and writing the foreword for this book.
We would like to say thanks to Dr. Ali Davod Parsa, Dr. Richard Hayhoe, Dr. Tim
Hayes, Dr. Rachael Cubberley, Dr. Oonagh Corrigan, and Dr. Shannon Doherty of
Anglia Ruskin University, UK; Mrs. Madhini Sivasubramanian and Dr. Divya Vinna-
kota of University of Sunderland, London; Dr. Julia Morgan of Greenwich University,
UK; Dr. Md. Anwarul Azim Majumder of University of West Indies; Dr. Brijesh Sathi-
an of Hamad Medical Corporation, Qatar; Dr. Sujita Kumar Kar of King George’s
Medical University, Lucknow, India; Dr. Shah Jalal Sarker of UCL, UK; and Dr. S. M.
Yasir Arafat of Enam Medical College and Hospital, Bangladesh.
Many of our students, who continually pushed and encouraged us to write this book,
deserve a share of the credit, including Dr. SM Anwar Sadat and Dr. Md. Naymul
Hasan. We, as authors, accept full responsibility for any deficiencies or errors that the
book may have. Finally, we express our sincerest thanks for the assistance that we have
received from ASA Publications in publishing this book.

Links for the e-book and datasets

Users may download the e-book and datasets used in this book from the links below.
• https://drive.google.com/drive/folders/1tdTEoBuMwObThyuFjehfb
MkHlyUcoIp_?usp=sharing
• https://drive.google.com/drive/folders/1D2LSmaz7eYJQwh
py5vZ0TTx7p_nGcTOD?usp=sharing

To the users
This e-book is free for all. However, if you can afford, please donate BDT 100 (US$ 2
for users outside Bangladesh) to a charity or a needy person. This little amount is suffi-
cient to offer a meal for an orphan in developing countries.

iii
Table of Contents
Chapter 1 Introduction 1
1.1 Version dilemma 1
1.2 Stata interface 2
1.3 Steps of data analysis 2

Chapter 2 Generating Data Files 7

2.1 Generating data files 7
2.1.1 Generating a data file by typing data in the data editor 7
2.1.2 Generating a data file using copy and paste commands 15
2.1.3 Importing a data file from other programs 15
2.1.4 Deleting and inserting variables 16
2.1.4.1 Deleting a variable 16
2.1.4.2 Inserting a new variable 16
2.1.4.3 Copy a variable into the same data file 17

Chapter 3 Basics of Stata 19

3.1 Stata files 19
3.1.1 Data file 19
3.1.1.1 Opening (retrieving) an existing data file 19
3.1.2 Output file or log-file 20
3.1.2.1 Saving outputs in a log-file 20
3.1.2.2 Opening an existing log-file 21
3.1.2.3 Browsing Stata outputs in Results window 22
3.1.2.4 Copy tables from Stata outputs to MS Word 22
3.1.2.5 Transformation of a log-file from smcl to ASCII format 22
3.1.3 Do-file or Stata commands file 23
3.1.3.1 Generating a Do-file 23
3.1.3.2 Saving commands in a Do-file 24
3.1.3.3 Executing the commands in a Do-file 24
3.2 Basic command syntax 25
3.3 Knowing the dataset 27
3.3.1 Generating brief description of a dataset 27
3.3.2 Codebook 27
3.4 Sorting of data 29

iv
 3.5 Stata operator symbols 30
3.6 Getting help in Stata 30

Chapter 4 Data Cleaning and Data Screening 33

4.1 Checking for out-of-range errors 33
4.2 Checking for outliers 35
4.3 Assessing normality of data 37

Chapter 5 Data Management 39

5.1 Converting string variables to numeric variables 39
5.1.1 Converting a string variable with non-numeric values into a
numeric variable 40
5.1.2 Converting a string variable with numeric values into a
numeric variable 40
5.2 Recoding of data 41
5.2.1 Recoding a string variable into the same variable 41
5.2.2 Recoding a string variable into a different string variable 41
5.2.3 Recoding a numeric variable into the same variable 41
5.2.4 Recoding a numeric variable into a different variable 41
5.3 Making class intervals 43
5.4 Combining data into a new variable 45
5.5 Data transformation 47
5.6 Calculation of total score 48
5.7 Extraction of duration from dates 49
5.8 Relocating variables in the dataset 50
5.9 Selecting a sub-group for analysis 51
5.10 The “egen” command 52
5.11 Creating dummy variables 53
5.12 Transformation of data formats 53

Chapter 6 Data Analysis: Descriptive Statistics 57

6.1 Frequency distribution 57
6.2 Central tendency and dispersion 59
6.2.1 Interpretation 63
6.3 Descriptive statistics disaggregated by a categorical
variable 66

v
 Chapter 7 Generating Graphs 69
7.1 Histogram 69
7.1.1 Saving graphs 70
7.3 Box and plot chart 73
7.4 Bar graph 76
7.4.1 Bar graph for the mean of a quantitative variable across a
categorical variable 76
7.4.2 Bar graph for the frequencies of a categorical variable 76
7.5 Line graph 77
7.6 Pie chart 78

Chapter 8 Checking Data for Normality 81

8.1 Assessing normality of data 81
8.1.1 Interpretation 83

Chapter 9 Testing of Hypothesis 85

Chapter 10 Student’s t-test for Hypothesis Testing 91

10.1 One-sample t-test 91
10.1.1 Interpretation 92
10.2 Independent samples t-test 93
10.2.1 Test for equality of variances: Levene’s test 93
10.2.2 Commands for independent samples t-test 95
10.2.3 Interpretation 95
10.3 Paired t-test 97
10.3.1 Commands 97
10.3.2 Interpretation 98

Chapter 11 Analysis of Variance (ANOVA) 99

11.1 One-way ANOVA 99
11.1.1 Commands 100
11.1.2 Interpretation 101
11.1.3 Post hoc test 101
11.1.4 Interpretation of post hoc test results 102
11.1.5 One-way ANOVA for unequal variances 103
11.2 Two-way ANOVA 103

vi
 11.2.1 Commands for two-way ANOVA 105
11.2.2 Interpretation 106
11.2.3 Post hoc test for two-way ANOVA 106

Chapter 12 Repeated Measures ANOVA 109

12.1 One-way repeated measures ANOVA 109
12.1.1 Commands 110
12.1.2 Interpretation 113

Chapter 13 Association Between Two Categorical

Variables: Chi-square Test of Independence 117
13.1 Chi-square test of independence 117
13.1.1 Commands 118
13.1.2 Interpretation 120
13.2 Relative risk and odds ratio 121
13.2.1 Interpretation 125
13.3 Stratified analysis 126
13.3.1 Interpretation 127

Chapter 14 Hypothesis Test of Proportions 131

14.1 One-sample test of proportion 131
14.2 Two-sample test of proportions 133

Chapter 15 Association Between Two Continuous

Variables: Correlation 135
15.1 Pearson’s correlation 135
15.1.1 Scatter plot 136
15.1.2 Commands for Pearson’s correlation 136
15.1.3 Interpretation 138
15.2 Spearman and Kendall’s tau-b correlations 139
15.2.1 Interpretation 140
15.3 Partial correlation 141
15.3.1 Interpretation 141

Chapter 16 Linear Regression 143

16.1 Simple linear regression 144

vii
 16.1.1 Commands for simple linear regression 145
16.1.2 Interpretation 146
16.2 Multiple linear regression 149
16.2.1 Sample size for multiple regression 150
16.2.2 Commands for multiple linear regression analysis 151
16.2.3 Interpretation 153
16.2.4 Regression diagnostics 154
16.2.4.1 Checking for multicollinearity 155
16.2.4.2 Checking for linearity 156
16.2.4.3 Checking for normality of residuals 158
16.2.4.4 Checking for homoscedasticity 161
16.2.4.5 Checking for outliers 163
16.2.4.6 Test for independence 164
16.2.5 Variable selection for a model 166
16.2.5.1 Backward selection method 167
16.2.5.2 Forward selection method 168
16.2.5.3 Interpretation 169

Chapter 17 Logistic Regression 171

17.1 Mathematical concept of logistic regression model 172
17.2 Binary logistic regression 173
17.2.1 Unconditional binary logistic regression 173
17.2.1.1 Interpretation 176
17.2.2 Logistic regression diagnostics 178
17.2.2.1 Checking for multicollinearity 179
17.2.2.2 Checking for model fit 180
17.2.2.3 ROC curve 182
17.2.2.4 Other postestimation commands 184
17.2.3 Variable selection for a model 185
17.2.4 Incorporating interaction terms in the model 186
17.2.5 Sample size for logistic regression 188
17.2.6 Conditional logistic regression 188
17.2.6.1 Interpretation 190
17.3 Analysis of cross-sectional data: Estimation of prevalence
ratio 191
17.3.1 Poisson regression 193

viii
 17.3.2 Cox regression with constant time 193
17.3.3 Generalized linear model 194

Chapter 18 Multinominal Logistic Regression 197

18.1 Interpretation 198
18.2 Post-estimation commands 200

Chapter 19 Survival Analysis 203

19.1 Survival analysis: Kaplan-Meier method 204
19.1.1 Preparing data for analysis 205
19.1.2 Commands for Kaplan-Meyer method 206
19.1.3 Interpretation 209
19.2 Cox regression 213
19.2.1 Commands 213
19.2.2 Interpretation 215
19.2.3 Checking for assumptions 217

Chapter 20 Nonparametric Methods 221

20.1 Mann-Whitney U test 221
20.2 Median test 223
20.3 Wilcoxon signed ranks test 223
20.4 Kruskal-Wallis test 224
20.5 Friedman test 225

Chapter 21 Analysis of Covariance (ANCOVA) 229

21.1 One-way ANCOVA 229
21.1.1 Commands 231
21.1.2 Interpretation: One-way ANCOVA 233
21.2 Two-way ANCOVA 235
21.2.1 Commands 236
21.2.2 Interpretation: Two-way ANCOVA 238

Chapter 22 Miscellaneous 241

22.1 Reliability of scales: Cronbach’s alpha 241
22.1.1 Interpretation 242
22.2 Constructing wealth quintiles 244

References 249

ix
 1
Introduction

Stata is a data analysis software that allows us to perform a wide range of statistical
analyses. Stata can be operated by using its drop-down menu as well as by writing the
commands directly in the command window. The commands of Stata are straightfor-
ward and simple. Therefore, data analysis by writing commands is more popular
among users. This book is written focusing on the use of Stata’s commands for data
analysis.
This book is for students, young researchers, and teachers. It provides more than a
basic understanding and techniques for statistical analysis of data related to health and
social sciences research. This book is based on Stata version 13. There are upper as
well as lower versions of Stata. Stata commands are sufficiently similar, and users can
use this book for both upper and lower versions. The newer versions are mainly for the
more advanced and recently developed techniques that the users most likely do not
require.
This book focuses on statistical decision-making, data analysis, and interpretation of
the outputs. It covers commonly used data analysis techniques in health and social
sciences research. The topics covered in this book include data management, descrip-
tive statistics, and bivariate and multivariable analysis for hypothesis testing, including
nonparametric methods and others.

1.1 Version dilemma

Stata is continuously evolving over time. That means that the commands, options,
language elements, and others may change in future versions. However, Stata ensures
2 Introduction

that the higher versions execute the commands regardless of the version of Stata in
which they are written. It is, therefore, expected that all the commands (syntax) used in
this book will run in higher (or lower) versions.
This book is based on Stata version 13. If you are using a different version (e.g.,
version 17), you can still use the commands. If you find any problem in executing a
command provided in this book in version 17 (or other versions), type the following
command at the top of every Do-file (Chapter 3) that you create:
version 13
This simple step will ensure that your Do-file or program will continue to run not only
in version 17 but also in all future versions of Stata, even if that future version has
changes in the syntax of some of the commands or programming constructs.
You can also use the above command as a prefix while writing a command in the com-
mand window. For example, if you want to execute an ANOVA syntax in version 17,
which is written in version 13, use the command:
version 13: anova …..
This command will set Stata’s version to 13, run the anova command, and then reset
Stata’s version to whatever it was before the command was executed. For further infor-
mation, visit: https://www.stata.com/manuals/pversion.pdf.

1.2 Stata interface

Once you open the Stata (double-clicking the Stata icon), it looks like Figure 1.1. It has
six windows, as shown in Figure 1.1 by the numbers 1 to 6. The purposes of the
windows are described in Table 1.1. You will also find some useful icons on the Stata
toolbar (Fig 1.2). The functions of the icons are provided in Table 1.2. We will discuss
Stata in further detail in the subsequent chapters.

1.3 Steps of data analysis

We collect data for our studies using various tools and methods. The most commonly
used tools for data collection are questionnaires and record sheets, while the common-
ly used data collection methods are interviews (face-to-face, telephonic or online),
observations, physical examinations, and lab tests. Sometimes we use available data
(secondary data) for our research studies, such as hospital records or data from other
 Introduction 3

studies (e.g., Bangladesh Demographic and Health Survey data). Once data is collect-
ed, the steps of data analysis are:
• Data coding, if a pre-coded questionnaire or record sheet is not used
• Development of a data file and data entry
• Data cleaning (checking for errors during data entry)
• Data screening (checking assumptions for statistical tests)
• Data analysis
• Interpretation of results

Figure 1.1 Stata interface and windows

Figure 1.2 Stata toolbar icons

4 Introduction

Table 1.1 Stata windows and purposes

Window Purpose
Command window This is for writing the commands. We write commands in this
[1] window. The commands are exec uted when the “Enter” key is
pressed. You can use the “Page Up” and “Page Down” keys on
the keyboard to recall commands from the “Review” window
(window 6).
Results or outputs This window displays the outputs along with the commands
window [2] executed. It also shows error messages if there is any
problem with the commands. The Results window keeps about
500 lines of the outputs. When this limit is exceeded, Stata
deletes the earlier outputs. If you want to save the outputs, you
must generate a log-file to store the outputs. You can browse the
results using the mouse or <Shift+Page Up/Page Down or Arrow>
buttons.
Variables window [3] Displays the variable names of the dataset currently active in
memory (i.e., currently being used). If you double-click on a
variable in this window, the variable will appear in the
"Command" window.
Properties window: Under the “Properties” window, there is a “Variables” window
Variables window [4] (4) and a “Data” window (5). The “Variables” window shows
the variable properties, such as variable name, variable label,
etc.
Data window [5] This window shows the file name, path, number of variables in
the dataset, and total observations.
Review window [6] This window displays the commands already executed during an
analysis session. If you click on a command in this window, the
command will appear in the “Command” window and can be
executed by pressing the “Enter” key. If you double -click on a
command in the “Review” window, the command will be
directly executed, and the outputs can be seen in the “Results”
window.
 Introduction 5

Table 1.2 Stata toolbar icons and functions

Icon Function
To open a data file from the desired location.

To save the data file.

To save the log-file (results or outputs file). You can begin, close, suspend,
or resume the log-file using this icon.
This is the new Do-file Editor icon. The Do-file is for writing, editing, and
saving the commands. You can generate a Do-file (command file) and edit
it using this icon.
This icon is for going to the Data Editor (Edit mode). In this mode, you can
edit or change data in the data file.
This icon is for going to the Data Editor (Browse mode). You can only
browse and see data in this mode, but you cannot change them.
This icon is for going to the Variables Manager, where you can edit
(change) the variable names, variable labels, and value labels.
6
 2
Generating Data Files

Like other data analysis programs, Stata has to read a data file to analyze the data. It is,
therefore, necessary to develop a data file or import it from other programs for use by
Stata. Data files can be generated by Stata, but it is not a popular practice. Researchers
mostly convert or import data files generated in other programs for use by Stata. Data
files generated in other programs can be easily transformed into Stata format for analy-
sis. In this chapter, we will discuss how to generate a data file in Stata.

2.1 Generating data files

2.1.1 Generating a data file by typing data in the data editor

For generating a data file, the first and basic step is to decide on a name for each of the
variables included in the questionnaire or record sheet. To name a variable, we need to
follow certain rules. They are:
• The variable names must be unique (i.e., all the variables should have different
names)
• A variable name must be between 1 and 32 characters long. But try to keep it
as short as possible.
• Variable names must begin with a letter (small or capital) or an underscore.
Variable names cannot begin with a number. Although underscore can be used
to begin a variable name, it is strongly discouraged because such variable
names are used to indicate temporary variables in Stata.
• Variables cannot include full stop (.), space, or symbols like, ?, *, µ, λ, ~, !, -,
@, and #.
8 Generating Data Files

• Stata is case-sensitive. For example, "Gender", "gender", and "GENDER" will

not be considered as the same variable by Stata. During analysis, you need to
type the variable names correctly for execution. We recommend using the
variable names with all lowercase letters (e.g., gender).
Once the variable names are decided, the next step is to generate a data file. In Stata,
for generating a data file, we start by entering data before inserting the variable names.
Suppose that we have collected data using a pre-coded questionnaire (codes are shown
in parenthesis) with the following variables:

Categorical variables:
• Sex (m= male; f= female)
• Religion (1= Islam/Muslim; 2= Hindu; 3= Others)
• Occupation (1= Business; 2= Government job; 3= Private job; 4= Others)
• Marital status (1= Married; 2= Unmarried; 3= Others)
• Have diabetes mellitus (1= Yes; 2= No)

Quantitative variables (numerical variables):

• ID number
• Age of the respondent
• Monthly family income
• Systolic blood pressure (BP)
• Diastolic BP
Assume that we have decided to use "age" as the variable name for age, "sex" for sex,
and "religion" for religion. Instead of age, sex, and religion, you can use any other
names for the variables, such as v1, v2, and v3. It is always convenient to develop a
codebook in MS Word or MS Excel where the Stata variable names, actual variable
names (variable labels), and variable codes (value labels) are recorded (Table 2.1). The
codebook is helpful during data analysis.
It is convenient to use numeric variables instead of string (character) variables for a
data file. The numeric variables have numeric codes (e.g., 1= male; 2= female). The
string variables may or may not be coded. If a string variable is coded, it is coded with
letters (e.g., m= male; f= female). When a string variable is not coded, the data is
directly entered into the data file. For example, the data of gender (male/female),
religion (Islam/Hindu/Others), and occupation (business/job holder/others) may be
entered directly into the data file. Note that Stata does not allow value labels for the
 Generating Data Files 9

coded string variables (e.g., m= male; f= female).

Table 2.2 shows some data (as an example) that has been collected using a question-
naire (Table 2.1). We will use this data to generate a data file in Stata.

Table 2.1 Questionnaire codebook

Stata variable Actual variable name/ variable Variable code/ value
name label labels
idno Identification number Actual value
age Age in years Actual value
sex Sex* m= male
f= female
religion Religion 1= Islam/Muslim
2= Hindu
3= Others
occu Occupation 1= Business
2= Government job
3= Private job
4= Others
income Monthly family income in Tk. Actual value
marital Marital status 1= Married
2= Unmarried
3= Others
diabetes Have diabetes mellitus 1= Yes
2= No
sbp Systolic blood pressure in mmHg Actual value
dbp Diastolic blood pressure in mmHg Actual value
*For practical purposes, it is better to consider a numeric variable (e.g., 1= male; 2=
female) rather than a string (character) variable for sex (e.g., m= male; f= female) as well
as for other variables.
10 Generating Data Files

Table 2.2 Data collected from the study subjects (only a portion is shown)
idno age sex religion occu income marital
1 26 m 1 2 25000 1
2 28 f 2 2 35000 1
3 29 f 1 1 60000 1
4 34 m 1 3 20000 2

Open the Stata program by double-clicking the Stata icon. You will see the Stata inter-
face (Stata/SE 13.0) as shown in Figure 1.1 (Chapter 1). The simplest way to generate
a data file is through the Data Editor. To go to the Data Editor, select from the menu
bar:
Window > Data Editor
Or,
Data > Data Editor > Data Editor (Edit)
Or,
Click the icon on the toolbar.
You will see the “Data Editor (Edit) – Untitled” as shown in Figure 2.1. This is the
window for defining the variables as well as for data entry. Use the following steps to
generate a data file:

Step 1: Our first variable is "idno" (Table 2.2). When the cursor is placed in the first
column of the first row, it will show "var1[1]" in the box above. Type the first value of
the variable "idno" as shown in Table 2.2 (which is 1, i.e., just type 1 and press the
"Enter" button). You will notice that "var1" appears at the top of the first column (Fig
2.2).
Now, type the value (which is 26) of the second variable "age" in the first box of the
second column and press "Enter". You will see that "var2" appears at the top of the
second column. Like this, enter the values of other variables into the Stata Data Editor
spreadsheet.
If the first value entered for a variable is a number, Stata will consider it a numeric
variable and will permit only numbers as its values subsequently. The numeric values
may begin with a plus or minus sign, and include decimal points. However, the num-
bers should not have any commas (,), such as 10,000 or 1,000,000.
 Generating Data Files 11

If the first value entered for a variable is a non-numeric character (such as m, f, or any
other letter), Stata will consider it a string (text) variable. A string variable may have
values up to 244 characters long and may have any combination of letters, numbers,
symbols, and spaces.

Figure 2.1 Stata data editor spreadsheet

Figure 2.2 Stata data editor with variable names

12 Generating Data Files

In the Data Editor or Data Browser, the string variable values appear as red, numeric
variable values appear as black, and labeled (coded) numeric variable values appear as
blue.

Step 2: In this step, we will rename (replace) the variable names that have been gener-
ated automatically by Stata (like var1, var2, and var3) with the variable names as
shown in the codebook (Table 2.1). For example, we need to rename the first variable
"var1" as "idno", the second variable "var2" as "age", and so on.
You can see that there are three windows on the right side of the Data Editor (Fig 2.1).
These are the "Variables", "Properties", and "Data" windows. In the "Variables"
window, all the variable names have appeared, like var1, var2, and var3 (Fig 2.2).
Click on "var1" in the "Variables" window. Now look at the "Properties" window. It
shows "var1" against "Name" in the window. Double-click on "Name" in the "Proper-
ties" window, delete "var1" and write "idno". This will replace "var1" with "idno". You
can see the new variable name (idno) in the spreadsheet as well as in the "Variables"
window. Like this, rename all the variables generated automatically by Stata with the
variable names of your choice. You can also use the following command to change the
variable name:
rename var1 idno
This command will rename the variable “var1” to “idno”.

Step 3: We will now label the variables one by one. To write the variable label for
“idno”, select (click on) the variable “idno” in the “Variables” window (Fig 2.2). Dou-
ble-click on “Label” in the “Properties” window and write “serial no”. This will be the
variable label for “idno” as shown in the “Variables” window for the variable “idno”
against “Label”. In this way, complete the variable labels of all variables as shown in
the codebook. An alternative to labeling a variable is by using the following command
in the command window:
label var idno “serial no”
This will label the variable “idno” as “serial no”.

Step 4: Step 4 is assigning the value labels. Since the variables "idno" and "age" are
not categorical variables (i.e., these variables are not coded), they do not need to have
value labels. Value labels are only needed for the categorical variables that are coded
numerically. Stata does not allow value labels for string variables. Therefore, we need
 Generating Data Files 13

to assign value labels only for the numerically coded variables, such as "religion",
"occupation", and others.
In our example, the variable "religion" is a numerically coded variable, and the code
numbers are 1= Muslim; 2= Hindu; 3= Others. We need to assign value labels to this
variable. To assign value labels, either use the following command or the following
steps:
label define religion 1”Muslim” 2”Hindu” 3”Others”
Or simply,
la de religion 1”Muslim” 2”Hindu” 3”Others”
Or, use the following steps:
• Select the variable “religion’ in the “Variables” window (in Data Editor) (Fig
2.2)
• Click on “Value Label” in “Properties” window (under “variables”)
• You will see a dropdown arrow and a small box with 3 dots
• Click on the 3 dot box
• Click on “Create Label”
• Write “religion” in the box “Label name”
• Write 1 in the “Value” box; write “Muslim” in the “Label” box; and click Add
• Write 2 in the “Value” box; write “Hindu” in the “Label” box; and click Add
• Write 3 in the “Value” box; write “Others” in the “Label” box; and click Add
• Click “OK” and then “Close”
• Click on “Value Label” in “Properties” window and using the dropdown
arrow, select “religion”
The above steps will insert value labels for the variable "religion". In this way (or using
the command), provide value labels to all the coded variables in the data file and enter
the data one by one. If there is any missing value, just keep the cell blank (or type the
assigned missing value code). Stata will consider it (when the cell is kept blank) a
missing value and will indicate it with a dot (.).
Closing the Data Editor window at any time during data entry will keep the data in
memory (data will not be lost) unless you exit the Stata program. Once data entry is
completed (or partially completed), we need to save the data file (Stata data files have
the extension ".dta"). To save a data file, select:
File (from the menu bar) > Save As… > Select the folder where you want to save
14 Generating Data Files

the file > Give a file name (.dta will appear by default) > Save
Or,
You can save the data file using the command “save”
For example, if you want to save the data file by the name "practice" on your desktop,
use the following command (you need to specify the path):
save "C:\Users\HP\Desktop\practice.dta"
When you are in use of a data file for analysis and want to save the data file with the
same name, use the command:
save, replace
This will save the data file with the same name and in the same location. The replace
option in the command replaces the old file with the same name.
So far, we have discussed how to rename and label a variable and how to assign value
labels using the Stata Data Editor as well as by using the commands. However, it is
convenient to rename and label variables and assign value labels by writing commands
in the command window of Stata. The following is the summary of how to use Stata
commands for the above functions:
• If you want to change (rename) the variable name “var1” to “idno”, use the
following command:
rename var1 idno
• If you want to label the variable “idno” as “serial no.”, use the following com
mand:
label var idno “serial no.”
• To assign the value labels (1= Muslim, 2= Hindu, 3= Others) to the variable
“religion”, use the following command:
label define religion 1”Muslim” 2”Hindu” 3”Others”
After executing the above command (label define), you need to select “religion” for the
“Value Label” in the “Properties” window using the dropdown arrow, or simply use the
following command:
label values religion religion
• If you want to change (rename) the variable label of a variable (e.g., variable
 Generating Data Files 15

label of “age” by “age in years”), use the following command:

label var age “age in years”
• If you want to change the value labels, for example, you want to change the
value label of religion as 3= Christian (instead of Others), use the following
command:
label define religion 1”Muslim” 2”Hindu” 3”Christian”, replace
• By default, the data file displays the value labels of the variables in the Data
Editor (Edit or Browse mode). If you want to see the code numbers (values) of
the variables, use the following steps (in data editor mode) or command:
Tools > Value Labels > Hide All Value Labels
Or,
Browse, nolabel

2.1.2 Generating a data file using copy and paste commands

The easiest way to generate a data file in Stata is to copy and paste data from another
data file, e.g., from Excel, dBase, SPSS, or others. For example, suppose that you have
a data file generated in SPSS, Excel, or other program. You can bring all the data from
these files to Stata. To do this, open the "Data Editor" window in Stata (Window > Data
Editor). Go to the data file from where you want to copy data into Stata. Select and
copy (Control-C) data you want to import and just paste (Control-V) them in the
spreadsheet of Stata's "Data Editor". Stata will provide the variable names automatical-
ly (by default) as var1, var2, and var3, etc. Rename the variables (var1, var2, var3 and
others) with the variable names of your choice. Also provide the variable labels and
value labels as discussed in section 2.1.1. Finally, save the data file at your desired
location (File > Save as > …). This method is suitable for a small dataset with few
variables.

2.1.3 Importing a data file from other programs

Stata version 16 and above has the option to directly import SPSS data into Stata by
using the dropdown menu [File > Import > SPSS data (*.sav)]. The lower versions do
not have this option. However, the best way to get a SPSS data file for use in Stata is
to save the data file into Stata format in SPSS.
For example, if you want to convert the SPSS data file “wealth.sav” into “wealth.dta”
in Stata format, first open the data file in SPSS. Then use the following steps:
16 Generating Data Files

File > Save as… > Select Stata Version 13 SE (*.dta) for the “Save as type” box
using the dropdown arrow > Write “wealth” in the file name box > Save
This will convert and save the SPSS data file “wealth” into Stata format. You can also
use a suitable data conversion program to import data from SPSS or other formats into
Stata format.

2.1.4 Deleting and inserting variables

2.1.4.1 Deleting a variable

You can delete unwanted variables from a data file. To delete a variable, use the follow-
ing steps:
• Select a variable in the “Variables” window of Stata
• Right-click the mouse and select “Drop selected variables”
You can also use the command “drop” to delete a variable. For example, if you want to
delete the variable “sex” from the data file, use the following command:
drop sex

2.1.4.2 Inserting a new variable

You cannot insert a new variable in the data file without any value or missing value.
Therefore, you need to select a value (or missing value) for the variable. For example,
you want to insert a new variable “gender”, all the values of which will be 1. To insert
the new variable (gender), use the following command:
generate gender=1
Or,
gen gender=1
If you want to insert the variable “gender” with the missing values, use the following
command:
gen gender=.
The new variable will appear as the last variable in the data file. You also need to
provide the variable label and value labels for the new variable as discussed earlier
(Section 2.1.1).
 Generating Data Files 17

2.1.4.3 Copy a variable into the same data file

You may be interested in copying a variable in the same data file that you are using.
Suppose that you want to have a copy of the variable "religion" already present in the
dataset. If you want this, you need to generate a new variable since the data file cannot
take the same variable name as its copy. Let us name the new variable "religion2". Use
the following command to get an exact copy (clone) of the variable "religion":
clonevar religion2=religion
You can also use the following command:
gen religion2=religion
The advantages of using the "clonevar" command are that it keeps the variable label
and value labels the same, including the missing value code in the same way as they
were in the old variable. On the other hand, if the command "gen" is used, it will not
keep the variable label and value labels in the new variable. You need to provide them
separately.
18
 3
Basics of Stata

In this chapter, we will discuss some of the basic functions related to the use of Stata
program, such as Stata files, command syntax, and others. Before we proceed to data
analysis, it is important to know all these basic things.

3.1 Stata files

Stata generally involves/generates three types of files. They are:
• Data file: The data file contains data for the analysis by Stata and has the
extension “.dta”;
• Output file: The output file contains the results of data analysis and the
commands. It is also called a “log” file and has an extension of either “.smcl”
or “.log”;
• Command file (Do-file): The command file is called a “Do-file” in Stata and
has the extension “.do”. This file is for storing a collection of commands for
data analysis. The commands saved in a Do-file can be reused or executed at
any time as required. The commands can be written directly and/or edited in a
Do-file for analysis.

3.1.1 Data file

3.1.1.1 Opening (retrieving) an existing data file

In the previous chapter (Chapter 2), we discussed how to generate a data file in Stata. If
you already have a Stata data file saved on your computer, you can open it in different
20 Basics of Stata

ways, such as using the drop-down menu, writing the command, or using the icon.
Suppose that you have a Stata data file named "Data_3.dta" on your computer and its
location is C:\Users\HP\Desktop. If you want to open the data file, use the following
steps or command:
File (on the menu bar) > Open > Go to Desktop and select the data file “Data_3”
> Open
Or,
use C:\Users\HP\Desktop\Data_3, clear
Or,
Click on the icon > Go to Desktop > Select “Data_3” > Open
Or,
Double click on the data file that you want to open

3.1.2 Output file or log-file

When you analyze data, the outputs (results) are shown in the results window of Stata.
The outputs will not be saved automatically. You need to save the outputs in a file. In
Stata, the output files are called “log” files. The log-files will have the outputs and
commands, but not the graphs. Graphs generated in Stata need to be saved separately,
as discussed in Section 7.1.1.

3.1.2.1 Saving outputs in a log-file

A log-file can be saved in two different formats.
• Stata format (also called smcl format): In Stata format, the log-file will have
the extension “.smcl”. The Stata format log-files preserve the formats that we
see in the results window and are the defaults.
• ASCII format: In ASCII (ordinary text) format, the file will have the exten-
sion “.log”.
We generally use the Stata format (smcl format) to save the outputs. However, the
“smcl” format can be converted to “ASCII” format whenever necessary.
Suppose that you want to analyze your data and save the outputs in a file named
“Results_3”. To save the outputs, create a log-file at the beginning of the analysis (you
can also create the log-file in between an analysis) using the menu bar options or com-
mand as described below:
 Basics of Stata 21

File (on the menu bar) > Log > Begin… > Select the location where you want to
save the file > Write “Results_3” in the “File name” box > Select the format .smcl
(usually the default) > Save
Or,
log using C:\Users\HP\Desktop\Results_3
The above command will open/generate (begin) a log-file with the name
“Results_3.smcl” on the desktop.
Once a log-file is opened, you can temporarily stop (suspend) saving the outputs at any
time during analysis by using the following steps or command:
File > Log > Suspend
Or,
log off
You can restart (resume) saving the outputs at any point in your analysis session by
using the following steps or command:
File > Log > Resume
Or,
log on
The log-file is saved and closed automatically at the end of an analysis session when
you exit Stata. You can, however, save and close the log-file anytime during analysis
by using the following steps or command:
File > Log > Close
Or,
log close

3.1.2.2 Opening an existing log-file

To open an existing log-file already saved on your computer (e.g., you want to open the
Results_3.smcl file, which is saved on the desktop), use the following steps:
File > Log > View > Browse > Select the file “Results_3” from Desktop > Open
> Ok
If you want to add (append) the results of a new analysis to a previously saved log-file
(say, Results_3.smcl), use the following steps or command:
File > Log > Begin > Select the file “Results_3” from Desktop (or from the
location where it is saved) > Save > Select “Append to existing file” > Ok
22 Basics of Stata

Or,
log using C:\Users\HP\Desktop\Results_3, append
You need to specify the file’s path, otherwise the command will not be executed.
If you want to overwrite (replace the contents of a log-file with the outputs of a new
analysis) the outputs in a log-file (e.g., Results_3.smcl), use the following command:
log using C:\Users\HP\Desktop\Results_3, replace
This command will delete all the contents of the log-file previously saved and will save
the outputs of the new (subsequent) analysis session.

3.1.2.3 Browsing Stata outputs in Results window

You can browse the outputs of analyses in the results window by using the mouse or
keyboard buttons (<Shift+Page Up> or <Shift+Arrow button>).
Stata normally (by default) pauses each time the results window is full of information
while executing a command unless you press any key on the keyboard. We can ask
Stata to continue scrolling (i.e., providing the outputs without pausing) till the comple-
tion of an output by using the following command:
set more off
To go back to the pause mode, use the following command:
set more on

3.1.2.4 Copy tables from Stata outputs to MS Word

You can copy a table (or commands or other information) from the Stata outputs
(results) window to MS Word. To do this:
Select the table from the output window by dragging down the mouse > Click the
right button of the mouse > Select copy > Go to MS Word file where you want to
paste the table > Right click the mouse > Paste
Note: Use the Courier New or Consolas font to preserve the table alignments. We have
used the Consolas font in this book for better resolution.

3.1.2.5 Transformation of a log-file from smcl to ASCII format

You can convert a Stata log-file from smcl (.smcl) format to ASCII (.log) format (and
vice versa). Suppose that you want to transform the log-file “Results_3.smcl” located
on your desktop to “Results_3.log”. Use the following command or steps:
 Basics of Stata 23

translate C:\Users\HP\Desktop\Results_3.smcl C:\Users\HP\Desktop\Results_3.log

Or,
File > Log > Translate > Select the file “Results_3.smcl” from Desktop clicking
the “Browse” tab of “Input File” > Open > Click the “Browse” tab of “Output
File” > Give the file name “Results_3.log” > Save > Translate

3.1.3 Do-file or Stata commands file

3.1.3.1 Generating a Do-file

One can generate a Do-file and write the commands in the Do-file Editor for subse-
quent use. To open the Do-file Editor (a new Do-file), use the following steps or com-
mand:
Window > Do-file Editor > New Do-file Editor
Or,
doedit
Or,
Click on the icon
The above command is for opening a new Do-file. You can save the Do-file like we
save a file in the MS word:
File (in Do-file editor) > Save as > Select location and file name > Save
To open a saved Do-file (e.g., to open the Do-file “Test.do”), use either of the follow-
ing commands. You need to specify the path of the file and file name to open the
Do-file, otherwise the command will not work.
doedit Test.do
doedit C:\Users\HP\Desktop\Test.do
This command will open the Do-file “Test.do” saved on the desktop. The alternative
way is:
Click on the icon , then
File (in Do-file editor) > Open > File… ctrl+O > Go to the file location and select
the file > Open
All the commands must be written in lowercase letters in the Do-file. While writing the
commands, Stata considers the end of a command line as the end of that command. If
your command exceeds a single line, use three back slashes (///) at the end of the line
24 Basics of Stata

before continuing to the next line. Then Stata will consider that the command is contin-
ued to the next line.
Once a new Do-file is generated, you can copy the commands from Stata’s Review
window and paste them into the Do-file.

3.1.3.2 Saving commands in a Do-file

Suppose that you have used some commands to analyze your data. You can see those
commands pooled in the Review window. If you want to save the commands in the
Review window into a Do-file, use the following steps:
• Select the command(s) you want to copy
• Right click the mouse button
• Select “Send to Do-file Editor” (this will automatically open a Do-file with the
selected commands in it)
• Save the file (File > Save as…)
You will see that all the selected commands are in a separate window (Do-file Editor).
You can edit the commands in this file and save the file for future execution. Com-
mands can also be copied and pasted into a Do-file from the Review window, Results
window, or log-file using the “Copy” and “Paste” options.

3.1.3.3 Executing the commands in a Do-file

It is simple to execute the commands saved in a Do-file. First, open the Do-file in the
Do-file Editor (Section 3.1.3.1) containing the commands that you want to execute. To
execute a single command or several connected commands at a time written in the
Do-file, use the following steps:
Select the command(s) that you want to execute by using the mouse
Tools > Execute (Do), or click on the icon in the Do-file Editor
This will execute the commands selected in the Do-file. If you use the following steps
without selecting any command in the Do-file (after opening the Do-file in the Do-file
Editor), Stata will execute all the commands stored in the Do-file.
Tools > Execute (do)
To execute all the commands in a Do-file (say, the Do-file “Test.do” saved on the desk-
top), use the command:
do C:\Users\HP\Desktop\Test.do
 Basics of Stata 25

This will execute all the commands saved in the Do-file “Test.do” without opening the
Do-file in the Do-file Editor.

3.2 Basic command syntax

Most researchers use Stata commands for the analysis of data because of their simplici-
ty and ease of use. A typical form of Stata’s command syntax is like:
command [varlist] [if exp] [in] [weight] [, options]
Or,
[prefix:] command [varlist] [if exp] [in] [weight] [, options]

command: Indicates Stata’s command for the analysis of data. It tells us what Stata is
supposed to analyze. Stata commands are case sensitive. All the commands must be
written in lowercase format, otherwise they will not work.

varlist: “varlist” stands for “variable list”. It indicates the list of variables needed for a
command to execute. The variable list is optional in many commands. If “varlist” is not
specified, the command runs on all the variables in the dataset. For example, if you use
the command:
summarize age
Stata will provide the summary statistics of the variable “age”. If you use only the com-
mand “summarize” without any variable name, Stata will provide the summary statis-
tics of all the variables in the dataset. Instead of writing the command “summarize”,
you can use only the first three initial letters, such as “sum” to get the summary statis-
tics.

if exp: “if exp” means “if expression”. It specifies the conditions to be considered
during an analysis. It is optional. For example, if you want to get the summary statistics
of age for males only (supposing that males are coded as 1 for the variable “sex”), use
the following command:
sum age if sex==1

in: “in” indicates the range restrictions in terms of observation numbers. It is optional.
For example, if you want to list the first (or last) 10 values of the variable “age” in the
dataset, use the following command:
26 Basics of Stata

list age in 1/10

list age in -10/-1
The first command will list the first 10 values (1/10 indicates 1 to 10), while the second
command will display the last 10 values of the variable “age”.

[ ]: All the syntax in [ ] is optional. You may not need to select anything. For example,
you can use the following command (without using anything for “if”, “in”, and
“weight”) to get the summary statistics for age.
sum age

weight: “weight” indicates the “weight variable”. If there is any weight variable
(frequency or sampling weight) that you want to include in an analysis, put the variable
after “in”. For example,
sum age [fweight = v2]
Here, “fweight” indicates frequency weight (“pweight” indicates sampling weight)
and “v2” is the weight variable that you want to consider.

options: “, options” indicate an optional instruction for data analysis. Note that there
is a comma (,) before the options, which must be used. For example,
sum age, detail
Here, we have used the option “detail”. Once we use this option (detail), Stata will give
the detailed summary statistics (mean, SD, skewness, kurtosis, percentile, and others)
of the variable.

prefix: The “prefix” is not mandatory for an analysis. The prefix is used to get the
results by sub-groups, such as by sex, occupation, or other variables. For example, if
you want to get the summary statistics of age by sex (i.e., disaggregated by males and
females), the prefix is needed. Then the commands would be like:
sort sex
by sex: sum age
Or,
bysort sex: sum age
Or,
By sex, sort: sum age
 Basics of Stata 27

Here, “by sex”, “bysort sex”, and “by sex, sort” are the prefixes, while “sum” is the
main command to get the summary statistics of age. We have used the command “sort
sex” to sort (in ascending order) the variable “sex”. Stata needs sorting (in ascending
order) of the prefix variable (in this example, sex) before executing the main command
“sum”. That’s why we used the first command, “sort sex”. You can, however, use a
single command like “bysort sex” that would first sort the “by variable” (here it is sex)
before executing the main command.

3.3 Knowing the dataset

3.3.1 Generating brief description of a dataset

Before data analysis, we need to know about the data, such as the number of observa-
tions and variables in the dataset, the nature of the variables, variable labels, and value
labels. Once the data file is loaded in Stata, you can get a brief description of the
dataset (variables) by using the following command:
describe
Or,
Data > Describe data > Describe data in memory > Ok
The above command (instead of “describe”, you can simply use “des”) will display the
characteristics of all the variables in the dataset, since we did not specify any variable
in the command (Table 3.1). Table 3.1 shows that there are 210 observations and 17
variables in the dataset. There is also other information. We can, however, get a
description of the variables specified with the command, such as:
des age sex occupation religion
This command will display a description of the variables “age”, sex”, “occupation”,
and “religion”.

3.3.2 Codebook
Before we start data analysis, it is important to know the variables in the dataset, espe-
cially how they are coded and how the missing values are identified. A good practice is
to either develop a codebook for the data file (as discussed in Chapter 2) or to look at
the data before analysis, so that you understand the structure of the information. This
can be done by using the command “codebook”.
28 Basics of Stata

Table 3.1 Description of a data file

. describe

Contains data from

C:\Users\HP\Documents\Taj_1_desktop_Home\STATA_Book_Taj\Data\Data_3.dta
obs: 210
vars: 17
size: 25,410
------------------------------------------------------------------------------------
storage display value
variable name type format label variable label
------------------------------------------------------------------------------------
ID_no double %10.0g Id number
age double %10.0g Age
sex str1 %1s Sex: string
sex_1 double %10.0g sex_1 Sex: numeric
religion double %10.0g religion Religion
religion_2 double %10.0g religion_2 Religion 2
occupation double %10.0g occupation Occupation
income double %10.0g Monthly income
sbp double %10.0g Systolic BP
dbp double %10.0g Diastolic BP
f_history double %10.0g f_history Family history of diabetes
pepticulcer double %10.0g pepticulcer Have peptic ulcer
diabetes double %10.0g diabetes Have diabetes mellitus
post_test double %10.0g Post test score
pre_test double %10.0g Pre test score
date_ad long %dD_m_Y Date of admission
date_dis long %dD_m_Y Date of discharge

codebook
codebook age sex_1 diabetes
codebook age sex_1 diabetes, compact
Or,
label list
label list religion occupation diabetes
The first command listed above (codebook), where the variables are not specified,
will provide the codebook of all the variables in the dataset. The second command
will provide the codebook of the variables included in the command (age, sex_1, and
diabetes) (Table 3.2). The “label list” command considers only the numerically coded
variables.
 Basics of Stata 29

Table 3.2 Outputs of codebook command

. codebook age sex_1 diabetes

------------------------------------------------------------------------------
age Age
------------------------------------------------------------------------------
type: numeric (double)

range: [6,45] units: 1

unique values: 35 missing .: 0/210

mean: 26.5143
std. dev: 7.49049

percentiles: 10% 25% 50% 75% 90%

16.5 21 27 32 36.5

-----------------------------------------------------------------------------
sex_1 Sex: numeric
-----------------------------------------------------------------------------
type: numeric (double)
label: sex_1

range: [0,1] units: 1

unique values: 2 missing .: 0/210

tabulation: Freq. Numeric Label

133 0 Female
77 1 Male

----------------------------------------------------------------------------
diabetes Have diabetes mellitus
----------------------------------------------------------------------------
type: numeric (double)
label: diabetes

range: [0,1] units: 1

unique values: 2 missing .: 4/210

tabulation: Freq. Numeric Label

162 0 No
44 1 Yes
4 .

3.4 Sorting of data

Sorting of data is sometimes required for browsing or editing. Sorting is also needed
before executing some commands (Section 3.2). Stata has the option to sort variables
both in ascending (default) or descending order. If you want to sort the variable “age” in
ascending order (lowest to highest value), use the first of the following commands. If you
want to sort it in descending order (highest to lowest value), use the second command:
30 Basics of Stata

sort age
gsort –age

3.5 Stata operator symbols

The arithmetic operator symbols and logical expressions are sometimes used in data
analysis. Logical expressions are used mostly with the “if” qualifier. Table 3.3 shows
the commonly used arithmetic operators and logical symbols. We will see their appli-
cations in the subsequent chapters.

3.6 Getting help in Stata

The basic commands for getting help in Stata are "help" and "search". The primary use
of the “help” command is to learn about a command or function whose name you
already know. For example, if you want to get the help files for the command “gener-
ate”, use:
help generate
On the other hand, the primary use of the command "search" is to learn about a subject.
It is used, especially when you do not know the command for a function. For example,
if you want to know about the ANOVA test and its commands, use:
search anova
This command will provide a search list of several official Stata entries. There are also
unofficial websites that include a wealth of commands. The "search" command can
locate information outside the official Stata files. Such an important site is the Statisti-
cal Software Components (SSC). To get into this site, use:
help ssc
If the “help” command does not find a command or function, it will automatically
continue with the “search” command. Say, you are looking for a command that calcu-
lates the sensitivity and specificity. Use the keywords with the command “search” to
get help, such as:
search sensitivity specificity
For more details about Stata’s help, users may use other resources such as Stata’s
YouTube channel (help youtube), Stata list (help statalist), Stata technical support
 Basics of Stata 31

(search technical support) and others.

Table 3.3 Key operator symbols and logical expressions used in Stata
Operator symbols:
+ Addition
- Subtraction
* Multiplication
/ Division
^ Power
Logical expressions:
< Less than
<= Less than or equal to
== Equal to
> Greater than
>= Greater than or equal to
!= Not equal to (~= can also be used; ! or ~ indicates not)
& And
| Or
32
 4
Data Cleaning and Data Screening

Once data is entered into Stata or imported from other programs, we need to be sure
that the data is free from errors. Data cleaning is commonly done by generating
frequency distribution tables of all the variables to find the out-of-range values and by
cross-tabulations (or by other means) to check the conditional values. If errors are
identified, they need to be corrected.
Simultaneously, we need to check the data if it fulfils the assumptions of the desired
statistical test (data screening). For example, is the data of a quantitative variable
normally distributed to do a t-test? There are other assumptions that need to be checked
before using statistical techniques, especially for hypothesis testing. In this chapter, we
will primarily discuss data cleaning. For the time being, users may skip this chapter
and proceed to Chapter 5. Once the users develop some skills in data analysis, they can
come back to this chapter. Use the data file <Data_3.dta> for practice. The codebook
for this data file can be seen in Chapter 6 (Table 6.1).

4.1 Checking for out-of-range errors

We can check the out-of-range errors by generating a frequency distribution table of a
variable or by checking the minimum and maximum values. For instance, suppose that
you want to check if there are any out-of-range errors in the variable “religion” (the
variable “religion” has 3 levels/values: 1= Muslim; 2= Hindu; 3= Others). To check the
out-of-range errors, we will find the minimum and maximum values of the variable
“religion” in the dataset by using the following command (outputs are shown in Table
4.1):
34 Data Cleaning and Data Screening

tabstat religion, stat(min max)

Or,
sum religion
Table 4.1 shows that the value labels of the variable "religion" range from 1 to 3 (mini-
mum 1 and maximum 3), which is within the range of our code numbers. Therefore,
there is no out-of-range error in this variable. If the maximum value was more than 3
or the minimum value was less than 1, this would indicate that there were errors in data
entry. In such a situation, identify the subjects (by ID no.) and correct the errors by
checking the questionnaire. You can also check the same for age and other variables.

Table 4.1 Minimum and maximum values of religion

. tabstat religion, stats(min max)

variable | min max

-------------+--------------------
religion | 1 3
----------------------------------

. sum religion

Variable | Obs Mean Std. Dev. Min Max

-------------+--------------------------------------------------------
religion | 210 1.52381 .7067039 1 3

If there is any value that is outside of the valid range of a variable, we can identify that
value against the identification (ID) number (variable name of the identification num-
ber in our dataset is "ID_no"). Let us generate a frequency distribution table of "diabe-
tes" to check the values by using the following command:
tab diabetes
This command will produce Table 4.2. The table shows that there is data with a value
of "2" for diabetes, which is outside the data range (the valid codes for diabetes are “0”
and “1”). Now, to identify the subject (ID_no) who has this value, use the following
command (Table 4.2):
list ID_no diabetes if diabetes==2
This command will provide the identification number (ID_no) of the subject for whom
the value of diabetes is “2” (Table 4.2). The table shows that the subject with the serial
number (ID_no) 9 has a value of “2”. Correct this value in the data file in the Data
Editor (Edit) mode and save the file.
 Data Cleaning and Data Screening 35

Table 4.2 Frequency distribution of diabetes

. tab diabetes

Have |
diabetes |
mellitus | Freq. Percent Cum.
------------+-----------------------------------
No | 164 78.10 78.10
Yes | 45 21.43 99.52
2 | 1 0.48 100.00
------------+-----------------------------------
Total | 210 100.00

. list ID_no diabetes if diabetes==2

+-----------------+
| ID_no diabetes |
|-----------------|
9. | 9 2 |
+-----------------+

4.2 Checking for outliers

Outliers are extreme values that deviate from other observations. Outliers may appear
in a dataset because of measurement errors, variability in the measurements, errors in
data entry, or other reasons. A commonly used rule says that a data point is an outlier if
it is more than 1.5×IQR (inter-quartile range) above the 3rd quartile (i.e., Q3 +
1.5×IQR) or below the 1st quartile (i.e., Q1 – 1.5×IQR). There are several ways to iden-
tify outliers in a dataset. The most commonly used method is the visualization of data.
The visualization methods that can be used are the box-plot chart, scatter plot, and
histogram. Statistical methods like Z-score and other methods are also available.
We will check the potential outliers by constructing a box and plot chart. Let us check
if there are any outliers in the variable systolic blood pressure (BP) [variable name is
“sbp”]. To get the box and plot chart for systolic BP, use the following command (Figs
4.1 and 4.2):
graph box spb
Or,
graph box sbp, marker (1, mlabel (ID_no))
Both these commands will generate box and plot charts for systolic BP (Figs 4.1 and
4.2). Figure 4.1 generated by the first command, shows that there are three dots above
the upper whisker without showing the case numbers (ID_no). The second command
36 Data Cleaning and Data Screening

has produced Figure 4.2, which shows the case numbers (ID_no) of the dots. These
dots indicate that there are 3 potential outliers in systolic BP and their case numbers
(ID numbers) are 20, 54, and 193.

200
180
160
Systolic BP
140
120
100

Figure 4.1 Box and plot chart of systolic BP with outliers

200

193
54
20
180
160
Systolic BP
140
120
100

Figure 4.2 Box and plot chart of systolic BP with id no. of outliers
 Data Cleaning and Data Screening 37

4.3 Assessing normality of data

One of the major assumptions for using parametric tests is that the dependent continu-
ous variable is normally distributed. Whether the data has come from a normally
distributed population or not, can be checked in different ways. The commonest meth-
ods of checking the normality of a dataset are through:
• Histogram
• Q-Q plot
• Formal statistical tests [Skewness-Kurtosis test, Shapiro Wilk test, or Kolmogorov
Smirnov (K-S) test]
This topic is further discussed in Chapter 8.
38
 5
Data Management

Data analysis may require data manipulations, such as making class intervals, classify-
ing a group of people with a specific characteristic using a cut-off value (e.g., you may
want to classify people who have hypertension using a cut-off value of either systolic
or diastolic BP), and recoding of data for summarization and other purposes. In this
chapter, we will discuss data manipulations that are commonly needed during data
analysis, like:
• Converting string variables to numeric variables
• Recoding of data
• Making class intervals
• Combine data to generate an additional variable
• Data transformation
• Calculation of total score
• Extraction of duration from dates
• Relocation of variables
• Selection of a subgroup for data analysis
• Transformation of data from wide format to long format
Use the data file <Data_3.dta> for practice.

5.1 Converting string variables to numeric variables

Stata data files may have both string and numeric variables. It is always preferable to
have numeric variables in the dataset. Numeric variables are easier to manipulate and
can be used in various statistical analyses. A string variable may be coded or have
40 Data Management

a direct response (e.g., names of districts or provinces). The codes of a string variable
may be a character (e.g., m= male; f= female) or a number (e.g., 1= male; 2= female).
Even though the codes are made up of numbers, they are actually characters. A detailed
overview of how to convert a string variable into a numeric variable is discussed
below.

5.1.1 Converting a string variable with non-numeric values into a numeric vari-
able
In our dataset, the variable “sex” is a string variable with non-numeric codes (m= male;
f= female). We can convert the string variable “sex” into a numeric variable (say, sex1)
by using the command “encode”.
encode sex, generate(sex1)
The above command (you can use "gen" instead of "generate") will generate a new
numeric variable "sex1" with the codes/values of 1 for female and 2 for male. Stata
provides the values in alphabetical order, beginning with 1 (i.e., 1 would be given to
the alphabetically first value of the original variable).

5.1.2 Converting a string variable with numeric values into a numeric variable
A string variable may have string numeric values (such as 1, 2, 3, etc.). The following
command can be used to change a string variable with numeric string values into a
numeric variable, where the numeric values will be retained as numeric.
For example, suppose that the variable "sex" is a string variable with values/codes of
0= female and 1= male. We want to convert "sex" into a numeric variable (say, sex1),
retaining the values of the string variable "sex". Use the following command:
gen sex1= real(sex)
This command will generate a new numeric variable “sex1” while keeping the values
of the string variable same (0= female; 1= male).
You can also convert a string variable with numeric values without generating a new
variable, i.e., into the same variable. To do this, use the following command:
destring sex, replace
This will convert the string variable "sex" into a numeric variable, keeping the name of
the variable and its values the same as before.
 Data Management 41

5.2 Recoding of data

Suppose that you have the string variable “sex” coded as “m” and “f” in the dataset.
You want to replace the existing code “m” by 1 and “f” by 2. There are two options for
the recoding of data:
a) Recoding into the same variable, and
b) Recoding into a different variable

5.2.1 Recoding a string variable into the same variable

Suppose that we have a string variable “sex” in our dataset, which is coded as m/f. To
recode “m” with 1 and “f” with 2 of the string variable into the same variable, use the
commands:
replace sex=”1” if sex==”m”
replace sex=”2” if sex==”f”

5.2.2 Recoding a string variable into a different string variable

First, we need to generate a new string variable (say, sex1) before we recode the exist-
ing string variable “sex”. Then, we will change the codes. Use the following com-
mands:
gen sex1=” “
replace sex1=”1” if sex==”m”
replace sex1=”2” if sex==”f”
This will generate a new string variable “sex1” with the codes 1= male and 2= female.

5.2.3 Recoding a numeric variable into the same variable

Suppose that you have a numeric variable "diabetes" in your dataset, which is coded as
1= don’t have diabetes and 2= have diabetes. You want to replace the codes "1" with
"0", and "2" with "1". Use the following command:
recode diabetes (1=0) (2=1)

5.2.4 Recoding a numeric variable into a different variable

In this case, first, we will generate a new variable that is identical to the original
variable using the command “generate or gen”. Then we will recode the values of the
42 Data Management

new variable using the command “recode”. Suppose that you have the numeric
variable “diabetes”, which is coded as “1= no diabetes” and “2= have diabetes”. You
want to recode the variable into a new variable “diabetes1”, where “1” will be coded
as “0 (no diabetes)” and “2” will be coded as “1 (have diabetes)”. Use the following
commands:
gen diabetes1=diabetes
recode diabetes1 (1=0) (2=1)
Or,
gen diabetes1=0
replace diabetes1=1 if diabetes==2
The first command (gen) will generate a new variable “diabetes1” equivalent to the
original variable “diabetes” (i.e., the values of diabetes1 and diabetes will be the same).
The second command will change the values of the new variable “diabetes1” from “1”
to “0” and from “2” to “1”.
The alternative approach is to generate a new variable “diabetes1” with all the values
equal to “0” by using the command “gen” (third command). Then replace the value “0”
with “1” if the value of the variable “diabetes” is “2” by using the command “replace”
(fourth command).
You can check the coding scheme of the new variable by using the commands “tab” or
“codebook”, such as:
tab diabetes1
codebook diabetes1
Note that the new variable will always appear as the last variable in the dataset that you
can check in browser mode. When you change the codes, you need to give the variable
label (if coded into a new variable) and value labels of the new codes by using the
following commands (also discussed in Chapter 2):
label var diabetes1 "have diabetes"
la de diabetes1 0”no diabetes” 1”have diabetes”
label values diabetes1 diabetes1
The outputs of the above commands are displayed in Table 5.1. Other examples of
using the “recode” command are provided in Table 5.2.
 Data Management 43

Table 5.1 Recoding of diabetes into a new variable

. gen diabetes1=diabetes
. tab diabetes1

diabetes1 | Freq. Percent Cum.

------------+-----------------------------------
1 | 165 78.57 78.57
2 | 45 21.43 100.00
------------+-----------------------------------
Total | 210 100.00

. recode diabetes1 (1=0) (2=1)

(diabetes1: 210 changes made)

. tab diabetes1

diabetes1 | Freq. Percent Cum.

------------+-----------------------------------
0 | 165 78.57 78.57
1 | 45 21.43 100.00
------------+-----------------------------------
Total | 210 100.00

. label var diabetes1 "have diabetes"

. la de diabetes1 0"no diabetes" 1"have diabetes"
. label values diabetes1 diabetes1
. tab diabetes1

have diabetes | Freq. Percent Cum.

--------------+-----------------------------------
no diabetes | 165 78.57 78.57
have diabetes | 45 21.43 100.00
--------------+-----------------------------------
Total | 210 100.00

5.3 Making class intervals

The central tendency (such as mean and median) and dispersion (such as SD) of quan-
titative data provide meaningful information. Further useful summarization may be
achieved by grouping the data into class intervals or categories. Suppose that you want
to categorize the variable "age" into the following categories or class intervals:
• ≤20 years (to be coded as 1),
• 21-30 years (to be coded as 2),
• 31-40 years (to be coded as 3), and
• >40 years (to be coded as 4)
44 Data Management

For this exercise, we will recode the variable "age" into a new variable "age2". We
recommend that users always recode into a new variable. If you recode into the same
variable, the original data will be lost, and it cannot be recovered once the data file is
saved. To recode into a different variable (age2), use the following commands:
gen age2=age
recode age2 (0/20=1) (21/30=2) (31/40=3) (*=4)
label var age2 "age group"
label define age2 1"<=20 yrs" 2"21-30 yrs" 3"31-40 yrs" 4">40 yrs"
label values age2 age2
tab age2
The above commands will generate a new variable "age2" with four categories of age
as mentioned above (Table 5.3). The "*" in (*=4) indicates all other values.

Table 5.2 Some examples of the use of recode command

Command Outputs
recode var1 (0=1) Will change values of the variable “var1”, 0 to 1
recode var1 (1=0) (2=1) Will change values of the variable “var1”, 1 to 0 and 2
to 1
recode var1 (0=1) (*=2) Will change values of the variable “var1”, 0 to 1 and
all other values to 2
recode var1 2/4=2 Will change values of the variable “var1”, (2 to 4) to 2
recode var1 (1 3 5 = 1) Will change values 1, 3 & 5 of the variable “var1” to 1
recode var1 (.=9) Will change the missing values to 9
recode var1 (9=.) Will change the value “9” as missing value

Using the command "recode", you can also classify people who have hypertension and
those who do not have hypertension (for example). To do this, you need to use a cut-off
value to define hypertension. For example, we have collected data on diastolic BP
(variable name is "dbp"). We want to classify those as "hypertensive", if the diastolic
BP is >90 mmHg. Now, generate a new variable (say, htn) equivalent to the variable
"dbp". Recode the variable "htn" as ≤90= 0 (normal BP) and >90=1 (hypertensive). We
hope you can do it now. If you cannot, use the following commands:
gen htn=dbp
recode htn (0/90=0) (*=1)
 Data Management 45

Table 5.3 Age categories

. gen age2=age
. recode age2 (0/20=1) (21/30=2) (31/40=3) (*=4)
. label var age2"age group"
. label define age2 1"<=20 yrs" 2"21-30 yrs" 3"31-40 yrs" 4">40 yrs"
. label values age2 age2
. tab age2

age group | Freq. Percent Cum.

------------+-----------------------------------
<=20 yrs | 50 23.81 23.81
21-30 yrs | 97 46.19 70.00
31-40 yrs | 58 27.62 97.62
>40 yrs | 5 2.38 100.00
------------+-----------------------------------
Total | 210 100.00

The above two commands will generate a new variable "htn" with values/codes "0"
and "1" (the last variable in the data file). The code "0" indicates people who do not
have hypertension (diastolic BP ≤ 90) and "1" indicates people who have hypertension
(diastolic BP >90).
To give the variable label and value labels, use the following commands:
label var htn "hypertension"
la de htn 0"no hypertension" 1"have hypertension"
label values htn htn
To check whether the commands have been executed properly, make a frequency
distribution table of the new variable "htn". This will also provide you with the propor-
tion of subjects with diastolic hypertension. Use the following command to get the
frequency distribution table for “htn” (Table 5.4):
tab htn

5.4 Combining data into a new variable

Sometimes, the cut-off point of a measurement (e.g., hemoglobin and blood pressure)
for defining a condition (e.g., anemia and hypertension) may vary according to gender
or other characteristics. In such a situation, a single cut-off point for defining a condi-
tion may not be appropriate.
For example, we have collected data on diastolic BP (variable name is “dbp”) both for
46 Data Management

Table 5.4 Frequency distribution of hypertension

. gen htn=dbp
. recode htn (0/90=0) (*=1)
(htn: 210 changes made)

. lab var htn "hypertension"

. la de htn 0"no hypertension" 1"have hypertension"
. la values htn htn
. tab htn

hypertension | Freq. Percent Cum.

------------------+-----------------------------------
no hypertension | 162 77.14 77.14
have hypertension | 48 22.86 100.00
------------------+-----------------------------------
Total | 210 100.00

males and females (variable name is “sex_1”). We have defined hypertension as a

diastolic BP greater than 85 mmHg if it is a female, and a diastolic BP greater than 90
mmHg if it is a male. Now, how to classify those who have hypertension considering
gender and cut-off values?
To do this, first we will generate a new variable, say “htn1”, for which all the values
will be “0” (zero) by using the command “gen”.
gen htn1=0
This will generate a new variable "htn1" with all the values of “0”. Now, use the
following commands to recode the new variable:
replace htn1=1 if dbp>85 & sex_1==0
replace htn1=1 if dbp>90 & sex_1==1
The above commands will replace some of the values of the variable "htn1" from “0”
to “1” based on the conditions provided in the commands. In this example, code “1”
indicates individuals with hypertension. Note that the variable "sex_1" used in the
commands is a numeric variable with codes 0= female and 1= male.
Provide the variable label and value labels as done before and make a frequency distri-
bution table for the variable “htn1” to check whether the commands have worked prop-
erly. You will also get the proportion of subjects with hypertension from the table
(Table 5.5). Data shows that the prevalence of diastolic hypertension is 30.0% in the
sample.
 Data Management 47

Table 5.5 Data combined into a new variable

. gen htn1=0

. replace htn1=1 if dbp>85 & sex_1==0

(55 real changes made)

. replace htn1=1 if dbp>90 & sex_1==1

(8 real changes made)

. lab var htn1 "hypertension"

. la de htn1 0"no hypertension" 1"have hypertension"

. la values htn1 htn1

. tab htn1

hypertension | Freq. Percent Cum.

------------------+-----------------------------------
no hypertension | 147 70.00 70.00
have hypertension | 63 30.00 100.00
------------------+-----------------------------------
Total | 210 100.00

5.5 Data transformation

In many situations, the data that has been collected on a quantitative variable for a
study is not normally distributed. Since the parametric methods (in general) for testing
hypotheses are more effecient than the nonparametric methods, data transformations
are occasionally needed to make the distribution normal and meet the assumptions for
a parametric test. Depending on the shape of the data distribution, there are several
options for data transformation. The following table (Table 5.6) shows some of the
options for data transformation.
The commonly used method of data transformation is log transformation. Let us see
how to transform data into a log value. Suppose that you want to transform diastolic BP
(variable name is “dbp”) into a log of diastolic BP. Use the command "gen", as shown
in Table 5.7 (the first and second commands), to transform the data into a new variable
"dbp1" with the log values of diastolic BP. The table also shows the commands for
other data transformation options.
48 Data Management

Table 5.6 Data transformation options

Method Good for Bad for
Log Right skewed data Zero values and negative values
Square root Right skewed data Negative values
Square Left skewed data Negative values
Reciprocal Making small values bigger and Zero values and negative values
big values smaller

Table 5.7 Data transformation commands

Command Function
Will generate a new variable:
gen dbp1=log(dbp) “dbp1” with the values of “log of dbp”
gen dbp2=ln(dbp) “dbp2” with the values of “natural log of dbp”
gen dbp3=dbp^2 “dbp3” with the values of “square of dbp”
gen dbp4=sqrt(dbp) “dbp4” with the values of “square root of dbp”
gen dbp5=(1/dbp) “dbp5” with the values of “reciprocal of dbp”

5.6 Calculation of total score

A study was conducted by a researcher to assess the knowledge of secondary school
students about how HIV is transmitted. To assess their knowledge, the researcher
collected data on the following questions (Table 5.8; data file: Data_HIV.dta).
When data is coded like 1= correct and 2= incorrect, first recode the data of the knowl-
edge variables as 2 = 0 (i.e., 0 indicates incorrect). In such a coding scheme (0/1), the
total score would range from 0–4, since there are four knowledge questions. Then,
calculate the total score by generating a new variable "tknow". Use the following com-
mands to do this:
recode k1-k4 (2=0)
gen tknow=(k1+k2+k3+k4)
Or,
egen tknow= rowtotal(k1 k2 k3 k4)
 Data Management 49

Table 5.8 Questions for assessing the knowledge on HIV transmission

HIV is transmitted through: Codes
1. Sexual contact (variable name: k1) 1. Yes 2. No
2. Transfusion of unscreened blood (variable name: k2) 1. Yes 2. No
3. Sharing of injection needle (variable name: k3) 1. Yes 2. No
4. Accidental needle stick injury (variable name: k4) 1. Yes 2. No

The first command will recode the knowledge variables k1 to k4 from “2” to “0”, while
the second (or third) command will calculate the total correct answers in the new
variable “tknow” (total knowledge on HIV). You can check the data by generating the
descriptive statistics and a frequency distribution table of the variable “tknow” by
using the following commands (Table 5.9):
sum tknow
tab tknow
Table 5.9 displays the descriptive statistics (mean, standard deviation, and others) and
frequency distribution of the students' overall knowledge on HIV transmission. The
table shows that the mean of total knowledge is 2.18 (SD 0.63), while the minimum
value is 0 and the maximum value is 4. The table also indicates that there are 2 (1.0%)
students who do not have any knowledge on HIV transmission (since the score is 0,
i.e., they could not answer any questions correctly). One hundred and twenty-five
(63.8%) students know two ways of HIV transmission, while only 1.5% of the students
know all the ways of HIV transmission. You can also classify the students as having
"good" or "poor" knowledge by using a cut-off value based on the total knowledge
score.

5.7 Extraction of duration from dates

Stata can extract the difference between two dates. Suppose that you have a dataset
with variables for the date of admission (variable name is date_ad) and date of
discharge (date_dis) of patients admitted to a hospital (data file Data_3.dta). If you
want to calculate the duration of hospital stay (date of discharge minus date of admis-
sion), use the following command to extract the difference in days:
gen dura = date_dis – date_ad
50 Data Management

Table 5.9 Total knowledge on HIV

. sum tknow

Variable | Obs Mean Std. Dev. Min Max

-------------+--------------------------------------------------------
tknow | 196 2.183673 .638057 0 4

. tab tknow

tknow | Freq. Percent Cum.

------------+-----------------------------------
0 | 2 1.02 1.02
1 | 16 8.16 9.18
2 | 125 63.78 72.96
3 | 50 25.51 98.47
4 | 3 1.53 100.00
------------+-----------------------------------
Total | 196 100.00

This command will generate a new variable “dura” (the last variable in the dataset) that
contains the duration of hospital stay in days. You can check the data by getting the
summary statistics of the variable “dura” using the command “sum” (Table 5.10).
sum dura, detail

5.8 Relocating variables in the dataset

You can order the variables according to alphabetical order in the variables window.
Suppose that you want to arrange the variables according to ascending order. Use the
following command:
order _all, first alphabetic
You can also move a variable to your desired position. In Stata, when a new variable is
generated, it appears (by default) as the last variable in the dataset. Suppose that you
have generated a new variable "age1" and you want it to be after the existing variable
"age". Use the following command:
order age1, after(age)
If you want to send a variable (say, sex) at the end of the dataset, use the first command
below. If you want it to be the first variable, use the second command:
order sex, last
order sex, first
 Data Management 51

Table 5.10 Summary statistics of duration of hospital stay

. sum dura, detail
dura
-------------------------------------------------------------
Percentiles Smallest
1% 1 1
5% 3 3
10% 3 3 Obs 21
25% 4 4 Sum of Wgt. 21

50% 6 Mean 5.761905

Largest Std. Dev. 2.343177
75% 7 8
90% 8 8 Variance 5.490476
95% 9 9 Skewness .1741077
99% 11 11 Kurtosis 2.813007

5.9 Selecting a sub-group for analysis

You may have a large dataset with many variables. But, you may not need all the data
or all the variables for your research. Stata has the option to select a small dataset (a
subset) for analysis. Remember that the bigger the dataset, the harder it is for Stata to
manage the data. It is, therefore, good to have a dataset as small as possible while keep-
ing all the relevant information.
The data file <Data_3.dta> has several variables, including the variable "diabetes1".
In the dataset, the variable "diabetes1" is coded as "1= Yes (have diabetes)" and "0= No
(do not have diabetes)". Suppose that you want to analyze the data only for those who
are more than 30 years old and have diabetes. We can select the data with this criteria
using the command either "drop" or "keep". Make sure that you have saved your origi-
nal dataset before using these commands. To select the group who are more than 30
years old and have diabetes (i.e., diabetes1= 1), use the following commands:
drop if diabetes1==0
drop if age<=30
Or,
drop if diabetes1==0 | age<=30
The first two commands or the third command (alone) will exclude subjects who do
not have diabetes and are aged ≤30 years from the subsequent analyses. Instead of the
command "drop", you can use the command "keep" with the same effects as follows:
52 Data Management

keep if diabetes1==1
keep if age>30
Or,
keep if diabetes1==1 & age>30
This will limit your analyses to subjects with diabetes who are over the age of 30. You
can check the data by making a frequency distribution table (using the command "tab
diabetes1") of "diabetes1" and the summary statistics (using the command "sum age,
detail") of "age".
To delete a variable from the dataset, use the command “drop”. For example, if you
want to delete the variables “income” and “sex”, use the following command:
drop income sex
Note that to have analyses for all the data after you use the commands “drop” or
“keep”, you need to read the data file again, either by using the command “use” or by
using the drop-down menu.
You can also select a subset of data or variables for analysis while opening the data file
by using the following commands:
use C:\Users\HP\Desktop\Data_3.dta if diabetes1==1 & age<=30
use age sex diabetes1 using C:\Users\HP\Desktop\Data_3.dta
The first command will select a subset of subjects who are less than or equal to 30
years old and have diabetes. The second command will select only the variables “age”,
“sex”, and “diabetes1” while opening the data file.

5.10 The “egen” command

The command "egen" (which indicates extended generate) is an extended version of
the command "generate" to generate new variables. The "egen" command is used to
generate new variables that require some additional functions, such as mean, median,
z-score, or others. For example, if you want to generate a new variable (say, zage) with
the z-scores of age, use the following command:
egen zage=std(age)
Other examples of the use of the "egen" command are:
egen v5= rowmean(v1 v2 v3 v4)
egen v10= rowtotal(v1 v2 v3 v4)
 Data Management 53

The first command will generate a new variable "v5" that will have the row-mean of
the variables v1 to v4. The second command will generate a new variable "v10", which
will have the row-total of the variables v1 to v4.

5.11 Creating dummy variables

A dummy variable is a dichotomous variable that takes values of “0” and “1”, where
the values indicate the presence or absence of a characteristic (e.g., “0” may indicate a
non-Muslim and “1” may indicate a Muslim). Dummy variables are also known as
indicator variables. Where a categorical variable has more than two categories, it can
be represented by a set of dummy variables, with one variable for each category.
Sometimes it is necessary to use dummy variables during analysis, e.g., in multiple
linear regression analysis (Section 16.2). In Stata, you can generate the dummy
variables easily. Suppose that you want to generate dummy variables for religion. The
variable "religion" has three levels (categories), 1= Muslim, 2= Hindu, and 3= Chris-
tian. To generate dummy variables for religion, use the following command:
tab religion, gen(reli)
This command will generate three dummy variables – “reli1”, “reli2”, and “reli3” in
the data file. All these variables will be coded as 0/1, where for “reli1”, 1 is Muslim and
0 is Others (Hindu and Christian); for “reli2”, 1 is Hindu and 0 is Others (Muslim and
Christian); and for “reli3”, 1 is Christian and 0 is Others (Muslim and Hindu) (Table
5.11). Provide the variable and value labels to the dummy variables as discussed earlier.
You can check the dummy variables by making frequency distribution tables using the
following command (Table 5.11):
tab1 reli1 reli2 reli3

5.12 Transformation of data formats

Repeated measures data may come in two different formats – wide format or long
format. For example, we have measured the blood sugar levels of five study subjects at
three time points, such as at the baseline and at 7 and 14 hours after giving a drug. The
variables generated for the blood sugar levels at each time point are "time0", "time7",
and "time14", respectively. If the data (blood sugar levels) of the individuals are plot-
ted under each time variable, as shown in Table 5.12, it is called a wide data format.
54 Data Management

Table 5.11 Dummy variables of religion

. tab religion, gen(reli)

Religion | Freq. Percent Cum.

------------+-----------------------------------
MUSLIM | 126 60.00 60.00
HINDU | 58 27.62 87.62
Christian | 26 12.38 100.00
------------+-----------------------------------
Total | 210 100.00

. tab1 reli1 reli2 reli3

-> tabulation of reli1

religion==M |
USLIM | Freq. Percent Cum.
------------+-----------------------------------
0 | 84 40.00 40.00
1 | 126 60.00 100.00
------------+-----------------------------------
Total | 210 100.00

-> tabulation of reli2

religion==H |
INDU | Freq. Percent Cum.
------------+-----------------------------------
0 | 152 72.38 72.38
1 | 58 27.62 100.00
------------+-----------------------------------
Total | 210 100.00

-> tabulation of reli3

religion==C |
hristian | Freq. Percent Cum.
------------+-----------------------------------
0 | 184 87.62 87.62
1 | 26 12.38 100.00
------------+-----------------------------------
Total | 210 100.00

Here, the data for each subject is plotted once under each time variable.
In long format, there is a time variable (say, time) with three or more levels/ categories
(in our example, we need three levels, such as 0= time0, 7= time7, and 14= time14) and
there is a separate variable for blood sugar level (say, sugar), as shown in Table 5.13.
In this format, the blood sugar levels of the study subjects are plotted under the variable
"sugar" against the categories of the time variable.
 Data Management 55

Table 5.12 Wide data format Table 5.13 Long data format
Subject Time0 Time7 Time14 Subject Time Sugar
1 110 108 107 1 0 110
2 115 112 110 2 0 115
3 112 110 115 3 0 112
4 106 105 104 4 0 106
5 109 108 107 5 0 109
1 7 108
2 7 112
3 7 110
4 7 105
5 7 108
1 14 107
2 14 110
3 14 115
4 14 104
5 14 107

Occasionally, it may be necessary to transform the data from a wide format to a long
format. For example, Stata cannot use wide data format for the analysis of repeated
measures ANOVA. We need to change the wide data format to a long format, which
can be done by Stata. For this exercise, let us use the data file <Data_repeat_2.dta>.
In the data file, there are seven variables — “sl (serial number)”, “subject (study
subject)”, “treatment (treatment group)”, “sugar0 (baseline blood sugar level)”, “sug-
ar7 (blood sugar level at 7 hours after treatment)”, “sugar14 (blood sugar level at 14
hours after treatment)”, and “sugar24 (blood sugar level at 24 hours after treatment)”.
This data file is in wide format for the data of blood sugar levels. We can transform this
data file into a long data format (for blood sugar) by using the following command:
reshape long sugar, i(subject) j(time)
This command will provide Table 5.14. The table shows that with this command, Stata
has generated a new variable "sugar", which contains the blood sugar levels of all the
subjects at different time points. The Stata also generated another new variable "time"
with four levels – 0 (baseline blood sugar level), 7 (blood sugar level at 7 hours after
treatment), 14 (blood sugar level at 14 hours after treatment) and 24 (blood sugar level
at 24 hours after treatment) [You can check it by using the command “tab time”]. Since
there are 15 study subjects and blood sugar levels are measured four times on each
56 Data Management

subject, the total number of observations is 60. The number of variables in the dataset
has also been reduced from 7 to 5.

Table 5.14 Transformation of data from wide to long format

. reshape long sugar, i(subject) j(time)
(note: j = 0 7 14 24)

Data wide -> long

-----------------------------------------------------------------------------
Number of obs. 15 -> 60
Number of variables 7 -> 5
j variable (4 values) -> time
xij variables:
sugar0 sugar7 ... sugar24 -> sugar
-----------------------------------------------------------------------------

After the transformation of the dataset, provide the variable label and value labels of
the new variables “time” and “sugar” by using the following commands:
la var time “time of blood sugar measurement”
la de time 0”baseline” 7”sugar at 7 hrs” 14”sugar at 14 hrs” 24”sugar at 24 hrs”
la values time time
la var sugar “blood sugar level”
You can also transform a long data format into a wide data format by using the follow-
ing command:
reshape wide sugar, i(subject) j(time)
 6
Data Analysis: Descriptive Statistics

Descriptive statistics are always used at the beginning of data analysis. The objectives
of using descriptive statistics are to organize and summarize data. Commonly used
descriptive statistics are frequency distribution, measures of central tendency (mean,
median, and mode), and measures of dispersion (range, standard deviation, and
variance). Measures of central tendency convey information about the average value of
a dataset, while the measures of dispersion provide information about the amount of
variation present in the dataset. Other descriptive statistics that are used during data
analysis include quartile and percentile. In this chapter, we will discuss how to analyze
data for descriptive statistics. Use the data file <Data_3.dta> for practice. The code-
book for the dataset is given in Table 6.1.

6.1 Frequency distribution

Frequency distribution of variables, especially for categorical variables, is commonly
done during data analysis. The command for the frequency distribution table is "tabu-
late" or simply "tab". If you want to find the frequency distribution of the variable
"sex", use the following command:
tabulate sex
Or,
tab sex
This command will give you the frequency distribution table for sex (Table 6.2). If you
want the frequency distribution tables of two or more variables by a single command,
the command “tab” will not produce separate tables for the variables. The use of two
58 Data Analysis: Descriptive Statistics

variables after the command “tab” will produce a cross-table for the variables includ-
ed in the command. For example, if you use the variables sex and religion with the

Table 6.1 Codebook for the data file “Data_3.dta”

Stata variable name Actual variable name Variable code
ID_no Identification number Actual value
age Age in years Actual value
sex Sex: string m= Male
f= Female
sex_1 Sex: numeric 0= Female
1= Male
religion Religion 1= Islam
2= Hindu
3= Others
religion_2 Religion 2 1= Islam
2= Hindu
3= Christian
4= Buddha
occupation Occupation 1= Government job
2= Private job
3= Business
4= Others
income Monthly family income in Tk. Actual value
sbp Systolic blood pressure in Actual value
mmHg
dbp Diastolic blood pressure in Actual value
mmHg
f_history Family history of diabetes 1= Yes
2= No
peptic Have peptic ulcer 1= Yes
2= No
diabetes Have diabetes mellitus 1= No
2= Yes
diabetes1 Have diabetes mellitus 0= No
1= Yes
posttest Post-test score Actual value
pretest Pre-test score Actual value
datead Date of hospital admission Actual date
datedis Date of discharge Actual date
 Data Analysis: Descriptive Statistics 59

command “tab” (i.e., tab sex religion), Stata will produce a cross-table of sex by
religion. However, if you want to get the frequency distribution of several variables at
a time (e.g., sex and religion), use the command “tab1”. To get the frequency distribu-
tions of sex and religion, use the following command:
tab1 sex religion
This command will produce two separate tables, one each for sex and religion, as
shown in Table 6.2. The table shows that there are a total of 210 subjects in the dataset,
out of which 133 (63.33%) are female and 77 (36.67%) are male. The frequency distri-
bution of religion is provided after the frequency distribution table of sex.
If there is any missing data, the output of the analysis will not show it (by default). If
there are missing values in the data, the total number of subjects will be less than the
number of data collected. For example, if we look at Table 6.2 (frequency distribution
of religion), there are 210 subjects, while Table 6.3 shows that there are 205 subjects in
the frequency distribution table of sex. This indicates that there are five missing values
in the data for "sex". To get the frequency distribution with missing cases (Table 6.3),
use the following command:
tab sex, miss
To get the identification numbers (variable name is "ID_no") of the missing subjects
(for example, for sex), use the following command.
list ID_no if missing(sex)
This will provide the id numbers of the missing cases for the variable “sex” (Table 6.3).

6.2 Central tendency and dispersion

We calculate the central tendency and dispersion for the quantitative variables.
Suppose that you want to find the mean, standard deviation (SD), minimum, and maxi-
mum values of the variable “age”. The command to get these measures is “summarize”
or simply “sum”. Use the following command:
sum age
The above command will provide the mean, SD, minimum (min) and maximum (max)
values of the variable "age" (Table 6.4). However, to get more detailed information,
use the following command:
sum age, detail
60 Data Analysis: Descriptive Statistics

Table 6.2 Frequency distribution of sex and religion

. tab1 sex religion

-> tabulation of sex

Sex: string | Freq. Percent Cum.

------------+-----------------------------------
f | 133 63.33 63.33
m | 77 36.67 100.00
------------+-----------------------------------
Total | 210 100.00

-> tabulation of religion

Religion | Freq. Percent Cum.

------------+-----------------------------------
MUSLIM | 126 60.00 60.00
HINDU | 58 27.62 87.62
Christian | 26 12.38 100.00
------------+-----------------------------------
Total | 210 100.00

Table 6.3 Frequency distribution of sex with missing cases and their id numbers
. tab sex

Sex: string | Freq. Percent Cum.

------------+-----------------------------------
f | 130 63.41 63.41
m | 75 36.59 100.00
------------+-----------------------------------
Total | 205 100.00

. tab sex, miss

Sex: string | Freq. Percent Cum.

------------+-----------------------------------
| 5 2.38 2.38
f | 130 61.90 64.29
m | 75 35.71 100.00
------------+-----------------------------------
Total | 210 100.00

. list ID_no if missing(sex)

+------+
| ID_no |
|------|
6. | 6 |
9. | 9 |
15. | 15 |
20. | 20 |
26. | 26 |
+------+
 Data Analysis: Descriptive Statistics 61

This command will provide the percentiles, mean, SD, variance, skewness, kurtosis,
and four lower and four upper values of the variable “age” (Table 6.4).
You can also get the specific statistics that you may want to have for a variable. For
example, if you want to have the mean, median, SD, and 10th percentile of age (you can
also use multiple variables), use the following command:
tabstat age, stat(mean median sd p10)
tabstat age sbp dbp, stat(n mean sd median p10) col(stat) long
You can also get all those statistics at each level of another categorical variable (e.g.,
sex). Use the following commands (Table 6.5):
tabstat age, stat(n mean sd median) by(sex)
tabstat age sbp, stat(n mean sd p50) by(sex) long format
tabstat age sbp, stat(n mean sd p50) by(sex) long col(stat)
tabstat age sbp, stat(n mean sd q) by(sex) long nototal
Sometimes we need to know the confidence interval (CI) for the mean. For instance, if
you want to get the 95% CI for the mean of age, use the following command:
ci age
ci age, level(99)
The first command will provide the 95% CI (default), while the second command will
provide the 99% CI (Table 6.4) for the mean age. You can use multiple variables with
the command to get the CIs. For example, if you want to get the 95% CIs for the means
of age, income, and systolic BP (variable name is sbp), use the following command
(Table 6.4).
ci age income sbp
When a variable is coded as 0/1, Stata considers it a dichotomous categorical variable.
For example, the variable "diabetes1" in the dataset is coded as 0/1 (Table 6.1), where
code "0" indicates the subjects who do not have diabetes and code "1" indicates the
subjects who have diabetes. With this coding scheme of a variable, if we use the com-
mand "sum", the mean actually indicates the proportion of the subjects coded as “1”,
i.e., the prevalence of diabetes in the sampled population when it is a cross-sectional
data (Table 6.6). We can also get the prevalence (proportion) of diabetes using the com-
mand “tab”.
sum diabetes1
tab diabetes1
62 Data Analysis: Descriptive Statistics

Table 6.4 Summary statistics of age including the confidence intervals

. sum age

Variable | Obs Mean Std. Dev. Min Max

-------------+--------------------------------------------------------
age | 210 26.51429 7.490491 6 45

. sum age, detail

Age
-------------------------------------------------------------
Percentiles Smallest
1% 10 6
5% 14 6
10% 16.5 10 Obs 210
25% 21 11 Sum of Wgt. 210

50% 27 Mean 26.51429

Largest Std. Dev. 7.490491
75% 32 41
90% 36.5 43 Variance 56.10745
95% 38 43 Skewness -.0917613
99% 43 45 Kurtosis 2.690733

. ci age

Variable | Obs Mean Std. Err. [95% Conf. Interval]

-------------+---------------------------------------------------------------
age | 210 26.51429 .516893 25.49529 27.53328

. ci age, level(99)

Variable | Obs Mean Std. Err. [99% Conf. Interval]

-------------+---------------------------------------------------------------
age | 210 26.51429 .516893 25.17059 27.85798

. ci age income sbp

Variable | Obs Mean Std. Err. [95% Conf. Interval]

-------------+---------------------------------------------------------------
age | 210 26.51429 .516893 25.49529 27.53328
income | 210 85194.49 1223.074 82783.34 87605.63
sbp | 210 127.7333 1.384129 125.0047 130.462

The first command (sum) will provide the mean, while the command "tab" will provide
a frequency distribution table of the variable (Table 6.6). To get the CI for a proportion,
use the command “proportion”. For example, to get the 95% CI for the prevalence of
diabetes, use the following command (you can use multiple variables together with this
command) (Table 6.6):
proportion diabetes1
Table 6.6 shows the 95% CI for the proportion (%) of people who have diabetes, which
 Data Analysis: Descriptive Statistics 63

is 0.163 (16.3%) to 0.275 (27.5%). The 95% CI indicates we are 95% confident that the
prevalence of diabetes in the population is between 16.3% and 27.5%. You can use
other options with this command, such as:
proportion diabetes1, level(99)
proportion diabetes1, over(religion) level(99)
The second command will display the 99% CI of diabetes in different religious groups.

Table 6.5 Outputs of “tabstat” command

. tabstat age sbp dbp, stat(n mean sd median p5) col(stat)

variable | N mean sd p50 p5

-------------+--------------------------------------------------
age | 210 26.51429 7.490491 27 14
sbp | 210 127.7333 20.05794 123 101
dbp | 210 82.76667 11.74929 82 66

. tabstat age sbp, stat(n mean sd p50) by(sex) long col(stat)

sex variable | N mean sd p50

--------------------+----------------------------------------
f age | 133 26.88722 6.802021 27
sbp | 133 129.5714 21.37695 124
--------------------+----------------------------------------
m age | 77 25.87013 8.559929 26
sbp | 77 124.5584 17.22108 122
--------------------+----------------------------------------
Total age | 210 26.51429 7.490491 27
sbp | 210 127.7333 20.05794 123
-------------------------------------------------------------

6.2.1 Interpretation
In Table 6.4, we can see all the descriptive statistics (central tendency and dispersion)
of the variable "age", including the statistics for Skewness and Kurtosis.
We believe you understand the meanings of mean (average), SD (average difference of
individual observations from the mean) and variance (square of SD). The analysis did
not give the median directly. As indicated in Table 6.4, the 50th percentile (P50) is the
median (middle value of the data set).
As presented in Table 6.4, the mean age is 26.5 years and the SD is 7.49 years. Let us
discuss the other statistics provided in Table 6.4, especially the skewness, kurtosis,
percentile, and quartile (not provided directly in the analysis).
64 Data Analysis: Descriptive Statistics

Table 6.6 Prevalence and confidence interval (CI) of diabetes

. sum diabetes1

Variable | Obs Mean Std. Dev. Min Max

-------------+--------------------------------------------------------
diabetes | 210 .2142857 .4113064 0 1

. tab diabetes1

Have |
diabetes |
mellitus | Freq. Percent Cum.
------------+-----------------------------------
No | 165 78.57 78.57
Yes | 45 21.43 100.00
------------+-----------------------------------
Total | 210 100.00

. proportion diabetes1

Proportion estimation Number of obs = 210

--------------------------------------------------------------
| Proportion Std. Err. [95% Conf. Interval]
-------------+------------------------------------------------
diabetes1 |
no | .7857143 .0283828 .724512 .8363903
yes | .2142857 .0283828 .1636097 .275488
--------------------------------------------------------------

Skewness and Kurtosis: These two statistics are used to evaluate whether the data has
come from a normally distributed population or not. In Table 6.4, we can see the statis-
tics for skewness (-0.091) and kurtosis (2.69). Skewness indicates the spreadness of
distribution (a measure of symmetry). Skewness "greater than 0" indicates data is
skewed to the right; skewness "less than 0" indicates data is skewed to the left, while
skewness "near to 0" indicates data is symmetrical (normally distributed).
However, the normality of data should not be evaluated based on skewness alone. We
also need to consider the statistics for kurtosis. Kurtosis indicates the "peakness" or
"flatness" of the distribution. Kurtosis is a measure to understand whether the data is
heavy-tailed or light-tailed relative to the normal distribution. Data with high kurtosis
tends to have heavy tails. The heavy-tailed distributions usually have outliers. Data
with low kurtosis tends to have light tails, or a lack of outliers.
The value of kurtosis for a normal distribution is 3. The data for "age" has a skewness
of -0.091 and a kurtosis of 2.69. Since the value of skewness is close to zero and the
value of kurtosis is close to 3, we may consider that the variable "age" has come from
 Data Analysis: Descriptive Statistics 65

a normally distributed population. We can also check the normality of a dataset by

other methods, which are discussed in Chapter 8.

Percentile and Quartile: Stata has provided the percentiles (1%, 5%, 10%, 25%, etc.)
for the variable "age" (Table 6.4). It did not show the quartiles directly. When a dataset
is divided into four equal parts after being arranged in ascending order, each part is
called a quartile. It is expressed as Q1 (first quartile or 25th percentile), Q2 (second
quartile or median or 50th percentile), and Q3 (third quartile or 75th percentile).
On the other hand, when the data is divided into 100 equal parts (after the ordered array
– from lowest to highest), each part is called a percentile (P). We can see in Table 6.4
that the 10th percentile (P10) is 16.5, the 25th percentile or P25 (Q1) is 21.0, the P50 (medi-
an or Q2) is 27.0, and the P75 (Q3) is 32.0 years. The Q1 (the first quartile) is 21 years,
indicating that 25% of the study subjects’ age is less than or equal to 21 years. On the
other hand, the 75th percentile (P75 or Q3) which is 32 years, indicates that 75% of the
study subjects’ age is less than or equal to 32 years. There is another measurement,
called interquartile range (IQR), which is not shown in the analysis. IQR is the differ-
ence between Q3 and Q1 (Q3 – Q1). In this example, the IQR is 11 years (32 – 21).

Sum of weights (wgt): When the "detail" option is used with the command "sum",
Stata provides the statistics "sum of wgt.", indicating the sum of weights (Table 6.4).
The table shows that the sum of weights for age is 210. Since we did not use any weight
variable in the analysis, by default, each subject is given a weight of 1. When the
option "detail" is used, the sum of the weights will therefore be equal to the number of
observations, as shown in our example. We do not need this information to interpret the
descriptive statistics.

95% confidence interval (CI): Table 6.4 shows the 95% CIs for age and other
variables (income and systolic BP). The 95% CI for mean age is 25.4 to 27.5 years.This
means that we are 95% confident that the mean age of the population from which the
sample is drawn is between 25.4 and 27.5 years.
Table 6.6 shows the mean and 95% CI of the dichotomous categorical variable "diabe-
tes1", which is coded as 0/1 (0= don’t have diabetes; 1= have diabetes). Here, the mean
value is 0.2142. This indicates that 21.42% (mean value multiplied by 100) of the
subjects have diabetes. If it is a cross-sectional data, we can also say that the preva-
lence of diabetes is 21.42%. The 95% CI of the prevalence is 16.36% – 27.54%. Here,
the 95% CI indicates that the prevalence of diabetes in the population would be
between 16.36% and 27.54%, and we are 95% confident about it.
66 Data Analysis: Descriptive Statistics

6.3 Descriptive statistics disaggregated by a categorical variable

You can get the descriptive statistics and other measures disaggregated (separately) by
categorical variables. For example, if you want to get the frequency distribution of
religion by sex (i.e., by males and females, separately), use any of the following com-
mands (we need to sort the variable "sex" first by using the command "sort"). Outputs
are provided in Table 6.7.
sort sex
by sex: tab religion
Or,
bysort sex: tab religion
Or,
by sex, sort: tab religion
If you want to get the measures of central tendency and dispersion of systolic BP (vari-
able name is "sbp") by diabetes, use the following command. Stata will generate sepa-
rate outputs for those with and without diabetes (Table 6.8). Other alternative com-
mands are provided in Section 6.2.
bysort diabetes: sum sbp
bysort diabetes: sum sbp, detail
 Data Analysis: Descriptive Statistics 67

Table 6.7 Frequency distribution of religion by sex

. bysort sex: tab religion

-----------------------------------------------------------------------------------
-> sex = f

Religion | Freq. Percent Cum.

------------+-----------------------------------
MUSLIM | 76 57.14 57.14
HINDU | 35 26.32 83.46
Christian | 22 16.54 100.00
------------+-----------------------------------
Total | 133 100.00

-----------------------------------------------------------------------------------
-> sex = m

Religion | Freq. Percent Cum.

------------+-----------------------------------
MUSLIM | 50 64.94 64.94
HINDU | 23 29.87 94.81
Christian | 4 5.19 100.00
------------+-----------------------------------
Total | 77 100.00

Table 6.8 Descriptive statistics of systolic BP by diabetes

. bysort diabetes: sum sbp

---------------------------------------------------------------------------------
-> diabetes = No

Variable | Obs Mean Std. Dev. Min Max

-------------+--------------------------------------------------------
sbp | 165 127.6061 20.87994 91 195

---------------------------------------------------------------------------------
-> diabetes = Yes

Variable | Obs Mean Std. Dev. Min Max

-------------+--------------------------------------------------------
sbp | 45 128.2 16.90428 100 176
68
 7
Generating Graphs

The information derived from data analysis needs to be presented in an effective and
understandable manner. Data and information can be presented in textual, tabular, or
graphical forms. Tables and graphs are powerful communication tools for the presenta-
tion of information. They can make an article easy to understand for the readers. While
a table is suitable for presenting quantitative and qualitative information, a graph is an
effective visual method for data presentation. A graph displays data at a glance, facili-
tates comparison, and can reveal trends and relationships.
The researchers need to carefully decide which type of graph or chart will be the best
way to present the information. The type of graph or chart to be used depends on the
data type, analysis outputs, and the objective of communicating the information. Inap-
propriate use of graphs may fail to convey the right information and may sometimes
confuse the readers, leading to misinterpretation of data. Therefore, the graphs for data
presentation must be chosen carefully.
The graphs commonly used for the presentation of data include histograms, scatter
plots, box and plot charts, bar graphs, line graphs, and pie charts. In this chapter, we
will discuss how to generate these graphs using Stata. Use the data file <Data_3.dta>.

7.1 Histogram
We usually generate a histogram to assess the distribution of a continuous variable. The
histogram provides information about: a) the distribution of a dataset (whether
symmetrical or not); b) the concentration of values; and c) the range of values. To
generate a histogram (say, for the variable "age"), use the following command:
70 Generating Graphs

histogram age
This will display a histogram for age (Fig 7.1 A). You can specify the Y-axis scale in
the command, such as frequency or percentage, by adding the following options:
histogram age, frequency
histogram age, percent
histogram age, norm percent
The first command will display a histogram where the Y-axis value is the number
(frequency; Fig 7.1 B), while the second command will display the Y-axis value in
percentage. The third command will produce a histogram with the overlying normal
curve (Fig 7.1 C).
You can specify the interval width to construct a histogram. For example, if you want
to have the histogram of age at a class interval of 3, use the following command:
histogram age, width(3) frequency
Or,
histogram age, width(3) start(2) frequency
You can also get the histograms dissertated by a categorical variable (say, by sex). To
do this, use the following command:
histogram age, by(sex) frequency
This command will display histograms of age for females and males, separately (Fig
7.1 D). You can use the graph edit options to give a title, subtitle, and change the font,
background, and color schemes of the histogram. These options can also be specified
through commands.
Looking at the histogram (Fig 7.1), it seems that the data is more or less symmetrical.
This indicates that age may be normally (approximately) distributed in the population.

7.1.1 Saving graphs

The graphs generated by Stata cannot be saved in the output files. They need to be
saved separately. The graphs can be saved in various formats, such as Stata Graph
(.gph), Enhanced Metafile (.emf), or in other formats. The disadvantage of a .gph
format is that it can only be read by Stata. On the other hand, the .emf format is usually
the best for use in Microsoft Word documents. Suppose that you have generated a
histogram for age. To save a histogram (or other graphs) generated by Stata with the
name “graph1”, use the following steps in the graph window:
 Generating Graphs 71

File > Save As… > Select location of the file to be saved > Give a file name in the
“File name” box (e.g., Graph1) > Select a format in the “Save as type” box (e.g.,
*.gph) > Save

A B
.08

40
.06

30
Frequency
Density

.04

20
.02

10
0

0
10 20 30 40 50 10 20 30 40 50
Age Age

C D
20

f m
30
15

20
Frequency
Percent

10

10
5

0 50 0 50
0

10 20 30 40 50 Age
Age Graphs by Sex: string

Figure 7.1 Histograms of age

You can also use the following commands to save the graph as “Graph1” on your desk-
top (or other locations):
graph save C:\Users\HP\Desktop\Graph1
graph save C:\Users\HP\Desktop\Graph1.emf, replace
The first command will save the graph on the desktop in .gph format, while the second
command will save the graph on the desktop in .emf format.
72 Generating Graphs

7.2 Scatter plot

A scatter diagram provides useful information about the relationship between two
continuous variables. The scatter diagram provides information/ideas about:
• Whether there is any correlation between the variables;
• Whether the relationship (if there is any) is linear or non-linear;
• The direction of the relationship, i.e., whether it is positive (if the value of one
variable increases with the increase of the other variable) or negative (if the
value of one variable decreases with the increase of the other variable);
• The presence of potential outliers in the dataset, i.e., the values that differ
significantly from other observations.
Stata can generate scatter plots with varieties of options, such as with a fit line (regres-
sion line), 95% CI for the fit line, disaggregated by a categorical variable, and others.
Several commands can be used to generate scatter plots. Use the following commands
to generate a scatter plot for systolic (variable name is “sbp”) and diastolic (variable
name is “dbp”) BP:
twoway scatter sbp dbp
twoway lfit sbp dbp || scatter sbp dbp
twoway lfitci sbp dbp || scatter sbp dbp
The first command will display the basic scatter plot of systolic BP against diastolic BP
(Fig 7.2 A). The first variable after the command is considered for the Y-axis. The
second command will display the regression line (fit line; Fig 7.2 B) for systolic BP on
diastolic BP, while the third command will provide the 95% CI of the regression line
(Fig 7.2 C).
The other commands and options that can be used to generate a scatter plot are:
twoway scatter sbp dbp, mlabel(ID_no)
twoway scatter sbp dbp, by(diabetes)
graph matrix age sbp dbp
The first command will display a scatter plot with the data points labelled by the case
numbers (ID numbers) (Fig 7.3 A), while the second command will display a scatter
plot by the variable “diabetes” (Fig 7.3 B). The last command will display a scatter plot
matrix for the variables listed (age, systolic BP, and diastolic BP) with the command
(Fig 7.4).
 Generating Graphs 73

200 A B

200
180
180

160
Systolic BP
160
Systolic BP

140
140

120
120

100
100

60 80 100 120
Diastolic BP
60 80 100 120 Fitted values Systolic BP
Diastolic BP

C
200
180
160
140
120
100

60 80 100 120
Diastolic BP

95% CI Fitted values

Systolic BP

Figure 7.2 Scatter plots of systolic and diastolic BP

7.3 Box and plot chart

We have shown how to generate the basic box and plot charts in Chapter 4. Here, we
will discuss the other options for generating the box and plot charts. The commands
that may be used for generating a box and plot chart are as follows:
graph box sbp
graph box dbp, over(religion)
graph box age sbp dbp
The first command is the basic command for generating a box and plot chart. The first
command will generate a box and plot chart of systolic BP as shown in Figure 7.5. The
second command will display box and plot charts of diastolic BP for different catego-
74 Generating Graphs

ries of religion (Fig 7.6). You can also generate box and plot charts for multiple
variables simultaneously by using the last command. Horizontal box and plot charts
can also be generated by using the following command:
graph hbox sbp

A B
200

No Yes
193

200
54
20
180

207
159
15 27
122 23
47
197
169 37
160

3 82
196

Systolic BP
166 79 53
210 190 191 61185 68
130
186 16
31

150
204
147 132 7 73 39
86
131
14 103 90
56 144195189
89
114
181
4452 48
140

30 145 6512055
10
138 42
81
100
8574 66
63102
87
129
22175 171
50 161
118 176 36
148 60 142
109
127
96
139 162 26 21 45 12657
78
136 182
133208
71
140
83
205 2420277
116
150
58 6
153 119
84 203
120

164111 163 137

32 17951 43
17714911
10592
12113 9910425
173 124
141134
17488 154
98 112
156 35
34
200 184
157
18
11076 108 168
198
28 188 194
146
70 33
160 29 167 165 101 125
106 46
93 135
121 172178158
91 13 19 62
128 206
38 192 1 183
17 41
72 20159 49

100
95 69 2 97
151 4 64 67 80 115
100

117107 180 2095 155123

9 199
75 143 187
8 40
152
94
170
60 80 100 120 60 80 100 120
60 80 100 120 Diastolic BP
Diastolic BP Graphs by Have diabetes mellitus

Figure 7.3 Scatter plots of systolic BP and diastolic BP with ID numbers (A) and by
diabetes (B)

Figure 7.4 Scatter plots matrix for age, systolic BP and diastolic BP

The box and plot chart provides information about the distribution of a dataset. It also
provides summary statistics of a variable, like Q1 (first quartile or 25th percentile),
median (second quartile or Q2) and Q3 (third quartile or 75th percentile) as well as
information about outliers or extreme values. The lower boundary of the box indicates
 Generating Graphs 75

the value for Q1, while the upper boundary indicates the value for Q3. The median is
represented by the horizontal line within the box. The minimum and maximum values
are indicated by the horizontal lines of the whiskers (Fig 7.5).
In the box and plot chart, the presence of outliers is indicated by the dots (Fig 7.5). The
outliers are the values greater than 1.5 box length distance (i.e., interquartile range or
IQR) from the edge (upper or lower) of the box [i.e., greater than (1.5×IQR + Q3) or
less than (Q1 – 1.5×IQR)]. You can see that there are 3 outliers in the data of systolic
BP (Fig 7.5). If you want to get the case numbers of the outliers, use the following
command ("ID_no" is the variable name for identification number in our dataset):
graph box sbp, marker (1, mlabel (ID_no))

Figure 7.5 Box and plot chart of systolic BP

120
100
Diastolic BP

80
60

MUSLIM HINDU Christian

Figure 7.6 Box and plot chart of systolic BP by religion

76 Generating Graphs

7.4 Bar graph

Bar graphs are very powerful tools for presenting summary statistics, which make it
easier for the readers to understand the relationship between the various values. Here,
we will discuss how to generate simple bar graphs, such as: a) the mean of a quantita-
tive variable across a categorical variable; and b) the frequencies of a categorical
variable.

7.4.1 Bar graph for the mean of a quantitative variable across a categorical vari-
able
You may be interested in presenting the mean income of the respondents by religion.
To do this, use the following command:
graph bar income, over(religion)
graph hbar income, over(religion)
The first command will display a bar graph showing mean income across religious
groups (Fig 7.7A). The second command will display a horizontal bar with the same
information. If you want to get the median, instead of the mean income by religion, use
the following command:
graph bar (median) income, over(religion)
Use the following commands if you want to get the value labels on each bar:
graph bar age, over(religion) blabel(bar)
graph bar age, over(religion) blabel(bar, format (%3.1f))
The first command will provide the value labels (mean age) on the bars. The addition
of the option "format (%3.1f)" in the command will provide the mean up to 3 digits
with one decimal point (Fig 7.7B).

7.4.2 Bar graph for the frequencies of a categorical variable

There is no specific and direct command for generating a bar chart to graphically repre-
sent a frequency table. The “fbar” command can be used to make a bar graph. Howev-
er, before using the “fbar” command, we need to install a module using the following
command:
ssc install fbar
Once the module is installed, use any of the following commands to make a bar graph:
 Generating Graphs 77

fbar religion
fbar religion, percent
fbar religion, by(sex)
The first command will generate a bar graph of religion with a frequency of occurrence
and the second one with a percentage. The last command will generate a frequency bar
graph of religion by sex. The bar graphs generated by the "fbar" command cannot be
edited in graph edits.
An alternative way to generate a bar graph is to use the modified command for histo-
gram. To generate a bar graph for religion with the modified command of histogram,
use the following command:
histogram religion, discrete freq gap(20) xlabel(1 2 3, valuelabel)
histogram religion, discrete percent gap(20) xlabel(1 2 3, valuelabel)
The first command will generate a bar graph with the frequency of the bars, while the
second one will provide the bar with a percentage.
30
80,000

27.0
26.3
24.9
mean of income

mean of age
60,000

20
40,000

10
20,000
0

MUSLIM HINDU Christian MUSLIM HINDU Christian

A B
Fig 7.7 Bar graph of mean income (A) and mean age (B) by religion

7.5 Line graph

A line graph is particularly suitable to demonstrate the trend, i.e., changes over time.
We can check the trend for two or more variables at a time. The basic command for
generating a line graph is "graph twoway line" or simply "line". Suppose that you want
to generate a line graph to demonstrate the change in TFR (total fertility rate; variable
78 Generating Graphs

name is "tfr") over the period of 1975 to 2018 (variable name is "year"). To do this, use
the following command (use the data file <Line.dta>):
line tfr year
twoway connect tfr year
The first command will produce the line graph A without the markers of the scatter
points, while the second command will produce the line graph B with the markers as
shown in Figure 7.8 (graphs have been edited for better resolution). The command
"connect" actually combines the features of scatter with a line, i.e., connects the scatter
points with a line segment. The Y-axis variable (dependent variable) must be written
first after the basic command.
You can also include two dependent variables to generate a line graph. Suppose that
you want to demonstrate the trend of TFR as well as CPR (contraceptive prevalence
rate; variable name is “cpr”) over the years. Use the following command to get the line
graph (Fig 7.8 C).
twoway connect tfr cpr year
Though you can give a title, Y-axis and X-axis labels, and others by using the Stata
commands, it is easier to write them using the graph edits. Users can easily explore the
graph editing options in Stata.
If you have a dataset in wide format (use the data file “Repeated Anova_2”, which is
in wide format), you can make a line graph by using the following command:
profileplot sugar_0 sugar_7 sugar_14 sugar_24, by(treatment)
Or,
profileplot sugar_0 - sugar_24, by(treatment)
Either of the above commands will generate a line graph of mean blood sugar levels at
4 different time points for three treatment groups (“sugar_0 - sugar_24” indicates the
variables sugar_0 to sugar_24).

7.6 Pie chart

Pie charts are also commonly used to present data. Stata can generate pie charts. The
"graph pie" command with the "over" option generates a pie chart representing the
frequency of each group. The "plabel" option places the labels (names or values) of the
categories inside each slice of the pie chart. Note that Stata does not allow a string
variable for making a pie chart. The following commands can be used to make a pie
 Generating Graphs 79

6
6

total fertility rate

5
5
total fertility rate

4
4

3
3

2
2

1970 1980 1990 2000 2010 2020 1970 1980 1990 2000 2010 2020
year year

A B
60
40
20
0

1970 1980 1990 2000 2010 2020

year

total fertility rate contraceptive prev. rate

C
Fig 7.8 Line graphs of TFR (A & B) and trends of TFR and CPR (C) by year

chart for the variable "religion".

graph pie, over(religion)
graph pie, over(religion) plabel(_all name)
graph pie, over(religion) plabel(_all percent)
graph pie, over(religion) pie(1, explode) plabel(_all percent)
The first command will produce a simple pie chart for the variable religion (Fig 7.9 A).
The second command will display a pie chart with category names (Fig 7.9 B), while
the third command will display category percentages (Fig 7.9 C). To explode a slice of
a pie graph, use the last command (Fig 7.7 D).
80 Generating Graphs

Christian

HINDU

MUSLIM

MUSLIM HINDU MUSLIM HINDU

Christian Christian

A B

12.38% 12.38%

27.62% 27.62%

60% 60%

MUSLIM HINDU MUSLIM HINDU

Christian Christian

C D
Fig 7.9 Pie charts of religion
 8
Checking Data for Normality

It is important to know the nature of the distribution of a continuous random variable

before using the statistical tests for hypotheses testing. To use the parametric methods
for hypothesis testing (such as t-test, ANOVA, correlations, and others), one of the
important assumptions is that the data of a continuous dependent variable is normally
distributed. It is, therefore, necessary to check whether the data has come from a
normally distributed population or not before we use the parametric methods. Use the
data file <Data_3.dta> for practice.

8.1 Assessing normality of data

The normality of data is an important assumption for using a parametric method of
testing a hypothesis. Whether the data is from a normally distributed population or not,
can be checked in several different ways. The most commonly used methods are:
a) Graphs, such as histograms and Q-Q (quantile-quantile) plots,
b) Descriptive statistics, such as skewness and kurtosis, and
c) Formal statistical tests, such as the Shapiro Wilk test (commonly used) and the
Skewness-Kurtosis (S-K) test.
In this section, we will discuss how to get a histogram and a Q-Q plot and how to
perform the formal statistical tests (Shapiro Wilk test and S-K test) to check the
normality of data.
We want to assess whether or not the variable "systolic BP" in our dataset is normally
distributed in the population. For this, we will first construct a histogram and a Q-Q
plot for systolic BP (variable name is “sbp”). To construct a histogram for systolic BP,
82 Checking Data for Normality

use the following command (also see Chapter 7):

histogram sbp, norm
histogram sbp, by(diabetes) norm
The first command will generate a histogram of systolic BP with an overlying normal
curve (Fig 8.1). The second command will display the same by diabetes.
.025
.02
.015
Density
.01
.005
0

100 120 140 160 180 200

Systolic BP

Figure 8.1 Histogram of systolic BP

To construct the Q-Q plot for systolic BP, use the first command (Fig 8.2). The second
command will generate a Q-Q plot for males.
qnorm sbp
qnorm sbp if sex==”m”
Formal statistical tests to assess the normality of data can also be used. The statistical
tests for checking the normality of data are the Shapiro Wilk test (commonly used) and
the Skewness-Kurtosis test. To do these tests for systolic BP, use the following com-
mands:
swilk sbp
sktest sbp
swilk sbp if diabetes==1
sktest sbp if diabetes==1
The outputs of first two commands are provided in Table 8.1. You can also use these
tests for multiple variables at a time.
 Checking Data for Normality 83

8.1.1 Interpretation
With Stata commands, we have generated the histogram and Q-Q plot (Figs 8.1 and
8.2) for systolic BP to visually check the distribution of data. We have also used the
formal statistical tests (Shapiro Wilk test and Skewness-Kurtosis test) to assess the
normality of data (Table 8.1).
200
150
Systolic BP
100
50

80 100 120 140 160 180

Inverse Normal

Figure 8.2 Q-Q plot of systolic BP

Table 8.1 Normality tests for systolic BP

. swilk sbp
Shapiro-Wilk W test for normal data

Variable | Obs W V z Prob>z

-------------+--------------------------------------------------
sbp | 210 0.95618 6.821 4.429 0.00000

. sktest sbp
Skewness/Kurtosis tests for Normality
------- joint ------
Variable | Obs Pr(Skewness) Pr(Kurtosis) adj chi2(2) Prob>chi2
-------------+---------------------------------------------------------------
sbp | 210 0.0001 0.2925 14.72 0.0006

The histogram (Fig 8.1) provides an impression about the distribution of the data (whether
the distribution is symmetrical or not). If we look at the histogram of systolic BP, it seems
that the data is slightly skewed to the right (i.e., the distribution is not symmetrical).
84 Checking Data for Normality

The Q-Q plot (Fig 8.2) also provides information on whether the data has come from a
normally distributed population or not. The Q-Q plot compares the distribution of data
with the standardized theoretical distribution from a specified family of distribution (in
this case, from a normal distribution). If the data is normally distributed, all the points
(dots) lie on the diagonal straight line. Our interest is in the central portion of the line
as well as in the tails. Deviation from the central portion of the line means non-normal-
ity. Deviations at the ends of the plot indicate the existence of outliers. We can see (Fig
8.2) that there is a slight deviation of the dots at the central portion as well as at the two
ends. This may indicate that the data may not have come from a normally distributed
population.
The specific tests (objective tests) that we have used to assess if the data has come from
a normally distributed population are the Shapiro Wilk test and the S-K test. The results
of these two tests are provided in Table 8.1.
Look at the "Prob > z" (for the Shapiro Wild test) and "Prob>chi2" (for the S-K test)
columns of Table 8.1. These columns indicate the p-values of the tests. A p-value of
<0.05 indicates that the data has not come from a normally distributed population. In
our example, the p-value is 0.000 for both the tests. This indicates that the data for
systolic BP has not come from a normally distributed population. Here, the null
hypothesis is "data has come from a normally distributed population" and the alterna-
tive hypothesis is "data has not come from a normally distributed population". We will
reject the null hypothesis since the p-values of the tests are <0.05.
The formal tests are very sensitive to sample size. These tests may be significant for
slight deviations in large sample data (n>100). Similarly, the likelihood of getting a
p-value <0.05 for a small sample of data (n<20, for example) is low. Therefore, the
rules of thumb for normality checking are:
1) For a sample size of <30: Assume non-normal;
2) For a moderate sample size (30-100): If the formal statistical test is significant
(p<0.05), consider a non-normal distribution; otherwise, check the normality
using other methods, such as histograms and Q-Q plots;
3) For a large sample size (n>100): If the formal statistical test is not significant
(p>0.05), accept normality; otherwise, check with other methods.
However, for practical purposes, just look at the histogram. If it seems that the distri-
bution is approximately symmetrical, consider that the data has come from a normally
distributed population and use a parametric test. If the sample size is less than 30, use
a nonparametric test.
 9
Testing of Hypothesis

The present and the following chapters provide basic information on how to select
statistical tests for testing hypotheses, perform the statistical tests with Stata, and inter-
pret the results of common problems related to health and social sciences research.
Before we proceed, let us discuss some of the basic concepts of hypothesis testing.
A hypothesis is a statement about one or more populations. A hypothesis is concerned
with the parameter of a population about which the statement is made. A hospital man-
ager may hypothesize that the average length of stay at his hospital is seven days, or a
researcher may hypothesize that the rate of recovery with drug A is better than that of
drug B. By means of hypothesis testing, one determines whether or not such statements
are compatible with the available data.
There are two types of statistical hypotheses: null (H0) and alternative (HA) hypotheses.
The null hypothesis is the hypothesis of equality or no difference. The null hypothesis
always says that two or more quantities (parameters) are equal. We always test the null
hypothesis, not the alternative hypothesis. Using a statistical test, we either reject or do
not reject the null hypothesis. If we can reject the null hypothesis, then we can only
accept the alternative hypothesis. It is, therefore, necessary to have a very clear under-
standing of the null hypothesis.
Suppose that we are interested in determining the association between coffee drinking
and stomach cancer. In this situation, the null hypothesis is "there is no association
between coffee drinking and stomach cancer (or, coffee drinking and stomach cancer
are independent)", while the alternative hypothesis is "there is an association between
coffee drinking and stomach cancer (or, coffee drinking and stomach cancer are not
independent) ". If we can reject the null hypothesis with a statistical test (i.e., if the test
86 Testing of Hypothesis

is significant; p-value <0.05), then we can only say that there is an association between
coffee drinking and stomach cancer.
Various statistical tests are available to test hypotheses. Selecting an appropriate statis-
tical test is the key to analyzing the data. What statistic is to be used to test a hypothesis
depends on the study design, data type, distribution of data, and objective of the study.
It is, therefore, important to understand the nature of the variable (categorical or quan-
titative), measurement type (nominal, ordinal, interval, or ratio scale), as well as the
study design. The following tables (Tables 9.1 to 9.4) provide basic guidelines about
the selection of statistical tests depending on the type of data and situation.

Table 9.1 Association between quantitative and categorical or quantitative variables

Situation for hypothesis testing Data normally Data non-
distributed normal
1. Comparison with a single population mean 1-sample t-test Sign test/
(with a fixed value) Wilcoxon
Signed
Example: You have taken a random sample
Rank test
from a population of diabetic patients to assess
the mean age. Now, you want to test the null
hypothesis that the mean age of the diabetic
patients in the population is 55 years.
2. Comparison of the means of two related Paired t-test Sign test/
samples (pre- and post-test comparison) Wilcoxon
Signed
Example: You want to test the hypothesis of
Rank test
whether the drug "Inderal" is effective in
reducing blood pressure (BP) or not. To test the
hypothesis, you selected a group of subjects
and measured their BP before giving the drug
(measurements before treatment or pre-test).
Then, you administered the drug to all of them
and measured their BP after one hour
(measurements after treatment or post-test).
Now, you want to compare if the mean BP
before (pre-test) and after (post-test)
 Testing of Hypothesis 87

Table 9.1 Association between quantitative and categorical or quantitative variables

Situation for hypothesis testing Data normally Data non-
distributed normal
administration of the drug is the same or not.
3. Comparison between two independent Independent Mann-
sample means [association between a samples t-test Whitney U
quantitative and a categorical variable with test (also
two levels] called the
Wilcoxon
Example: You have taken a random sample of
Rank Sum
students from a university. Now, you want to
test)
test the hypothesis if the mean systolic BP of
male and female students is the same or not.
4. Comparison of more than two independent One-way Kruskal
sample means [association between a ANOVA Wallis test
quantitative and a categorical variable with
more than two levels]
Example: You have taken a random sample
from a population. You want to test the
hypothesis if the mean income of different
religious groups (e.g., Muslims, Hindus, and
Christians) is the same or not.
Another example, you have three drugs, A, B,
and C. You want to investigate whether all
these three drugs are equally effective in
reducing blood pressure or not.
5. Association between two quantitative Pearson’s Spearman’s
variables correlation correlation
(Can also
Example: You want to test the hypothesis if
be used for
there is a correlation between systolic BP and
ordinal
age.
variables)
88 Testing of Hypothesis

Table 9.1 Association between quantitative and categorical or quantitative variables

Situation for hypothesis testing Test Statistics
6. Association between a quantitative and an Spearman’s correlation, if the
ordinal variable ordinal variable has 5 or more
levels.
Example: You want to test the hypothesis if
there is a correlation between systolic BP and Otherwise, use the Kendall’s
severity of anemia. Tau-B statistics
7. Association between two ordinal variables Spearman’s correlation, if
both the ordinal variables
Example: You want to test the hypothesis if
have 5 or more levels.
there is a correlation between severity of pain
and stage of cancer. Otherwise, use the Kendall’s
Tau-B statistics

Table 9.2 Association between two categorical variables

Situation for hypothesis testing Test statistics
1. Association between two categorical variables Chi-square test/
(independent samples) Fisher’s Exact
test
Example: You have taken a random sample from a
population and want to test the hypothesis that there is an
association between sex and asthma. Another example, you
want to assess the association between smoking and stomach
cancer.
2. Association between two categorical variables of related McNemar test
samples, such as data from a matched case-control study
Example: You want to test the hypothesis if there is an
association between diabetes mellitus and heart disease when
the data is matched for smoking or other variables (a matched
case-control study design).
 Testing of Hypothesis 89

Table 9.3 Multivariable analysis

Type of outcome/dependent variable Type of multivariable
analysis
1. The outcome variable (also called the dependent Multiple linear regression;
variable) is on an interval or ratio scale – e.g.,
Analysis of variance
blood pressure, birth weight, and blood sugar.
(ANOVA);
Analysis of covariance
(ANCOVA)
2. The dependent variable is a dichotomous Binary logistic regression
categorical variable (i.e., a nominal categorical
variable with two levels) – e.g., disease (present
or absent); ANC (taken or not taken); and
outcome (cured or not cured).
3. The dependent variable is a nominal categorical Multi-nominal logistic
variable with more than two levels – e.g., regression
treatment seeking behavior (such as treatment
not received; received homeopathic treatment;
received allopathic treatment); and cause of
death (cancer, heart disease, pneumonia).
4. The dependent variable is an ordinal categorical Proportional odds regression
variable – e.g., severity of anemia (no anemia, (Ordinal regression)
mild to moderate anemia, severe anemia); stage
of cancer (stage 1, stage 2, stage 3); severity of
pain (mild, moderate, severe), etc.
5. The dependent variable is time-to- Proportional hazards analysis
outcome/event, such as time-to-death, time-to- (Cox regression)
recurrence, and time-to-cure.
6. The dependent variable is a count – e.g., the Poisson regression
number of post-operative infections; the number
of patients admitted with heart disease to a
hospital; and the number of road traffic accident
cases treated in the emergency department.
90 Testing of Hypothesis

Table 9.3 Multivariable analysis

Type of outcome/dependent variable Type of multivariable
analysis
7. Incidence rates, such as incidence rates of Poisson regression
tuberculosis; incidence rates of pneumonia;
incidence rates of car accidents, etc.

Table 9.4 Agreement analysis

Situation for hypothesis testing Test statistics
1. Agreement between two quantitative variables Bland Altman test/plots
Example: You want to test the hypothesis that the
two methods of blood sugar measurement agree with
each other.
2. Agreement between two categorical variables Kappa estimates
Example: You want to test the hypothesis that the
diagnosis of cataract in patients is agreed upon by
two physicians.
 10
Student’s t-test for Hypothesis Testing

The student’s t-test is commonly known as the t-test. It is a frequently used parametric
statistical method to test a hypothesis. There are several types of t-tests used in differ-
ent situations (Table 9.1), such as: a) One-sample t-test; b) Independent samples t-test;
and c) Paired t-test. In this chapter, we will discuss all these t-tests and the interpreta-
tion of their outputs. Use the data file <Data_3.dta> for practice.

10.1 One-sample t-test

The one-sample t-test is done to compare the mean with a hypothetical value. For
example, we have collected data on diastolic BP (variable name is “dbp”) of students
of the State University of Bangladesh by taking a random sample. We are interested in
knowing if the mean diastolic BP of the students is different from 80 mmHg.

Hypothesis
Null hypothesis (H0): The mean diastolic BP of students is equal to 80 mmHg in
the population (study population is the students of the State University of Bangla-
desh).
Alternative hypothesis (HA): The mean diastolic BP of students is different from
(not equal to) 80 mmHg in the population.

Assumptions
1. The distribution of diastolic BP in the population is normal;
2. The sample is a random sample from the population.
92 Student’s t-test for Hypothesis Testing

The first job, before testing the hypothesis, is to check whether or not the distribution
of diastolic BP is normal in the population (assumption 1). To do this, check the histo-
gram and/or Q-Q plot of diastolic BP or do the formal statistical test of normality (Sha-
piro Wilk test or S-K test) as discussed in Chapter 8. If the assumption is met (diastolic
BP is at least approximately normal), do the 1-sample t-test; otherwise, use the
nonparametric method (Wilcoxon Signed Rank test, as discussed in Chapter 20).
Suppose that the diastolic BP is normally distributed in the population. Use the follow-
ing command to do the 1-sample t-test:
ttest dbp=80
With this command, Stata will generate Table 10.1.

Table 10.1 One-sample t-test

. ttest dbp=80

One-sample t test
------------------------------------------------------------------------------
Variable | Obs Mean Std. Err. Std. Dev. [95% Conf. Interval]
---------+--------------------------------------------------------------------
dbp | 210 82.76667 .8107779 11.74929 81.16832 84.36502
------------------------------------------------------------------------------
mean = mean(dbp) t = 3.4124
Ho: mean = 80 degrees of freedom = 209

Ha: mean < 80 Ha: mean != 80 Ha: mean > 80

Pr(T < t) = 0.9996 Pr(|T| > |t|) = 0.0008 Pr(T > t) = 0.0004

10.1.1 Interpretation
In this example, we have tested the null hypothesis "the mean diastolic BP of the
students is equal to 80 mmHg in the population". Data shows that the mean diastolic
BP of the sample is 82.77 mmHg (95% CI: 81.16 - 84.36) with a SD of 11.75 mmHg
(Table 10.1). The results of the one-sample t-test show that the calculated value of "t"
is 3.412 and the p-value [(Pr(|T| > |t|); 2-tailed] is 0.0008. Since the p-value is <0.05,
we will reject the null hypothesis at the 95% confidence level. This means that the
mean diastolic BP of the students (in the population) from where the sample is drawn
is different from 80 mmHg (p<0.001).
 Student’s t-test for Hypothesis Testing 93

10.2 Independent samples t-test

The independent samples t-test involves one quantitative variable (dependent variable)
and a categorical variable with two levels (categories). This test is done to compare the
mean of a dependent variable between two categories of the categorical variable.
For example, we are interested in knowing if the mean age of diabetic and non-diabetic
patients in the population is the same or not. Here, the test-variable (dependent
variable) is age (a quantitative variable) and the categorical variable is diabetes, which
has two levels/categories (have diabetes and do not have diabetes). Before doing this
test, we need to check assumption 1 [i.e., age is normally distributed at each level of
diabetes], as discussed earlier.

Hypothesis
H0: The mean age of diabetic and non-diabetic patients is the same in the popula-
tion.
HA: The mean age of diabetic and non-diabetic patients is different (not the same)
in the population.

Assumptions
1. The dependent variable (age) is normally distributed at each level of the inde-
pendent variable (diabetes);
2. The variances of the dependent variable (age) at each level of the independent
variable (diabetes) are the same/equal;
3. Subjects represent random samples from the population.

10.2.1 Test for equality of variances: Levene’s test

Before doing the independent samples t-test, we need to check if the variances of the
dependent variable (age) at each level of the independent variable (diabetes) are the
same or not (assumption 2). Levene’s test is the most commonly used statistical test to
determine whether two or more groups have equal variances. To do the Levene’s test
to determine if the variances of age are equal among the diabetic and non-diabetic
patients, use the following command (Table 10.2):
robvar age, by(diabetes)
You can also use the following command to check the equality of variances. It will
provide the results of variance ratio statistics (Table 10.2).
94 Student’s t-test for Hypothesis Testing

sdtest age, by(diabetes)

Table 10.2 Test results for equality of variances

. robvar age, by(diabetes)

Have |
diabetes | Summary of Age
mellitus | Mean Std. Dev. Freq.
------------+------------------------------------
Yes | 27.911111 8.4633494 45
No | 26.133333 7.1835969 165
------------+------------------------------------
Total | 26.514286 7.4904907 210

W0 = 3.2178863 df(1, 208) Pr > F = 0.0742901

W50 = 3.0099114 df(1, 208) Pr > F = 0.08423809
W10 = 3.0595865 df(1, 208) Pr > F = 0.08173708

. sdtest age, by(diabetes)

Variance ratio test

------------------------------------------------------------------------------
Group | Obs Mean Std. Err. Std. Dev. [95% Conf. Interval]
---------+--------------------------------------------------------------------
yes | 45 27.91111 1.261642 8.463349 25.36844 30.45378
no | 165 26.13333 .5592423 7.183597 25.02909 27.23758
---------+--------------------------------------------------------------------
combined | 210 26.51429 .516893 7.490491 25.49529 27.53328
------------------------------------------------------------------------------
ratio = sd(yes) / sd(no) f = 1.3880
Ho: ratio = 1 degrees of freedom = 44, 164

Ha: ratio < 1 Ha: ratio != 1 Ha: ratio > 1

Pr(F < f) = 0.9265 2*Pr(F > f) = 0.1471 Pr(F > f) = 0.0735

The results of Levene’s test (using the command "robvar") are provided in Table 10.2.
The table shows the mean and standard deviation of age as well as the total observa-
tions (Freq) for both diabetic (Yes) and non-diabetic (No) patents. We can see that the
standard deviation of age is higher among diabetic patients (8.46) compared to non-di-
abetic (7.18) patients, but the Levene’s test will tell us whether or not this difference is
statistically significant. The results of the "sdtest" command also provide the same
information as shown after Levene’s test results in Table 10.2.
Stata has provided three options for Levene’s test results:
 Student’s t-test for Hypothesis Testing 95

• W0, which is the test statistic of Levene’s test centered at the mean;
• W50, which is the test statistic centered at the median; and
• W10, which is the test statistic centered at a 10% trimmed mean (i.e., the top
5% and bottom 5% of the values are trimmed out).
We will consider the test statistics centered at the median (W50) and the p-value for
this is 0.084 (>0.05). This means that the variance of age for diabetic and non-diabetic
patients is not different (equal). Similarly, the p-value (0.147) of the "sdtest" indicates
the same, i.e., the variances for diabetic and non-diabetic patients with regard to age
are not different.

10.2.2 Commands for independent samples t-test

The independent samples t-test can be done in two situations:
• When the variances of the dependent variable (e.g., age) are equal at each
level of the categorical variable (e.g., diabetes); and
• When the variances of the dependent variable are unequal at each level of the
categorical variable.
The commands for the independent samples t-test are:
ttest age, by(diabetes)
ttest age, by(diabetes) unequal
The first command is for the t-test when the variances are equal, while the second com-
mand is for the t-test when the variances are unequal. The outputs of the t-tests are
provided in Tables 10.3 (with equal variances) and 10.4 (with unequal variances).

10.2.3 Interpretation
Table 10.3 shows the descriptive measures of age by diabetes. We can see that there are
45 subjects with diabetes and 165 subjects without diabetes. The mean age of the
diabetic patients is 27.9 (SD 8.46) and that of the non-diabetic patients is 26.1 (SD
7.18) years.
The t-test results are provided at the bottom of the descriptive statistics. The calculated
t-value is -1.41 (with the degrees of freedom 208). Since we are interested in under-
standing if the mean age is the same for both diabetic and nondiabetic patients, we will
consider the 2-tailed test. Look at the p-value, provided under "Ha: diff !=0", which is
96 Student’s t-test for Hypothesis Testing

0.158 (>0.05). We cannot, therefore, reject the null hypothesis. This means that the
mean age of diabetic and non-diabetic patients in the population from where samples
are drawn is not different (p=0.158).
The results of the t-test for unequal variances are provided in Table 10.4. The interpre-
tation of the results is the same as above.

Table 10.3 Independent samples t-test with equal variances

. ttest age, by(diabetes)

Two-sample t test with equal variances

------------------------------------------------------------------------------
Group | Obs Mean Std. Err. Std. Dev. [95% Conf. Interval]
---------+--------------------------------------------------------------------
No | 165 26.13333 .5592423 7.183597 25.02909 27.23758
Yes | 45 27.91111 1.261642 8.463349 25.36844 30.45378
---------+--------------------------------------------------------------------
combined | 210 26.51429 .516893 7.490491 25.49529 27.53328
---------+--------------------------------------------------------------------
diff | -1.777778 1.256707 -4.255293 .6997375
------------------------------------------------------------------------------
diff = mean(No) - mean(Yes) t = -1.4146
Ho: diff = 0 degrees of freedom = 208

Ha: diff < 0 Ha: diff != 0 Ha: diff > 0

Pr(T < t) = 0.0793 Pr(|T| > |t|) = 0.1587 Pr(T > t) = 0.9207

Table 10.4 Independent samples t-test with unequal variances

. ttest age, by(diabetes) unequal

Two-sample t test with unequal variances

------------------------------------------------------------------------------
Group | Obs Mean Std. Err. Std. Dev. [95% Conf. Interval]
---------+--------------------------------------------------------------------
No | 165 26.13333 .5592423 7.183597 25.02909 27.23758
Yes | 45 27.91111 1.261642 8.463349 25.36844 30.45378
---------+--------------------------------------------------------------------
combined | 210 26.51429 .516893 7.490491 25.49529 27.53328
---------+--------------------------------------------------------------------
diff | -1.777778 1.380033 -4.536122 .9805668
------------------------------------------------------------------------------
diff = mean(No) - mean(Yes) t = -1.2882
Ho: diff = 0 Satterthwaite's degrees of freedom = 62.3436

Ha: diff < 0 Ha: diff != 0 Ha: diff > 0

Pr(T < t) = 0.1012 Pr(|T| > |t|) = 0.2024 Pr(T > t) = 0.8988
 Student’s t-test for Hypothesis Testing 97

10.3 Paired t-test

The paired t-test is done to compare the difference between the two means of related
samples. Related samples indicate the measurements taken from the same subjects at
two or more different times or situations. For example, you have organized a training
session for 32 staff members of your organization. To evaluate the effectiveness of the
training, you have taken a pre-test before the training to assess the current status of
knowledge of the participants. At the end of the training, you have again taken a test
(post-test). Now, you want to compare if the training has significantly increased the
knowledge or not.
Another example is, suppose that you want to determine the effectiveness of a drug
(e.g., Inderal) in reducing the systolic blood pressure (BP). To do this, you have select-
ed a random sample from a population. You have measured the systolic BP of all the
individuals before giving the drug (pre-test or baseline measurement). You have again
measured their BP one-hour after giving the drug (post-test or endline measurement)".
The paired t-test is the appropriate test to compare the means in both situations.

Hypothesis
H0: There is no difference in mean scores before and after the training (for exam-
ple 1).
HA: The mean scores are different before and after the training.

Assumptions
1. The difference between two measurements (pre- and post-test) of the depen-
dent variable (examination scores) is normally distributed;
2. Subjects represent a random sample from the population.

10.3.1 Commands
We will use the first example (to compare the pre- and post-test scores) to do the paired
t-test. Use the following command (the variable names are: pre_test and post_test) for
the paired t-test (Table 10.5):
ttest post_test = pre_test
98 Student’s t-test for Hypothesis Testing

Table 10.5 Paired t-test results

. ttest post_test= pre_test

Paired t test
------------------------------------------------------------------------------
Variable | Obs Mean Std. Err. Std. Dev. [95% Conf. Interval]
---------+--------------------------------------------------------------------
post_t~t | 32 90.98438 1.492164 8.440957 87.94109 94.02766
pre_test | 32 53.57813 2.727373 15.42835 48.01561 59.14064
---------+--------------------------------------------------------------------
diff | 32 37.40625 2.47848 14.0204 32.35136 42.46114
------------------------------------------------------------------------------
mean(diff) = mean(post_test - pre_test) t = 15.0924
Ho: mean(diff) = 0 degrees of freedom = 31

Ha: mean(diff) < 0 Ha: mean(diff) != 0 Ha: mean(diff) > 0

Pr(T < t) = 1.0000 Pr(|T| > |t|) = 0.0000 Pr(T > t) = 0.0000

10.3.2 Interpretation
Table 10.5 shows the descriptive statistics of both the pre- and post-test results. Look-
ing at the mean scores, we can get an impression of whether the training has increased
the mean score or not. We can see that the post-test mean is 90.9, while the pre-test
mean is 53.5, and the difference is large (37.4). To understand if the difference between
post-test mean and pre-test mean is significant or not, we need to check the paired t-test
results given at the bottom of the descriptive statistics. The results show that the calcu-
lated t-value is 15.09 (degrees of freedom: 31).
Three different p-values for the paired t-test are displayed at the bottom of the table.
Since we are interested in understanding if there is any difference in the mean scores
before and after the training (a two-tailed test), we will consider the p-value provided
at the middle [Ha: mean(diff) != 0], which is 0.000. Since this value is smaller than
0.05 (our level of significance), we will reject the null hypothesis. We can, therefore,
conclude that there is sufficient evidence to say that there is a significant difference in
the mean scores before and after the training; i.e., the mean knowledge score has
increased significantly after the training (p<0.001).
 11
Analysis of Variance (ANOVA)

Analysis of variance (ANOVA) is a commonly used statistical method for testing

hypothesis. An ANOVA test is done to compare the means when the categorical inde-
pendent variable has more than two levels (categories). There are several types of
ANOVA tests, such as one-way ANOVA, two-way ANOVA, repeated-measures ANO-
VA, and others. In this chapter, one-way and two-way ANOVA are discussed. The
repeated measures ANOVA is discussed in Chapter 12. Use the data file <Data_3.dta>
for practice.

11.1 One-way ANOVA

One-way ANOVA is done to compare the means of more than two groups, while the
t-test compares the means of two groups. One-way ANOVA involves two variables:
one categorical variable with more than two levels or categories (for example, the
variable "religion_2", which has 4 categories – Muslim, Hindu, Christian, and
Buddhist); and a quantitative variable (e.g., income, age, and blood pressure). Suppose
that you want to assess if the mean income (variable name is "income") of all the
religious groups (variable name is "religion_2") is the same or not in the population.
One-way ANOVA is the appropriate test for this, if the assumptions are met.

Hypothesis
H0: The mean income of all the religious groups is equal.
HA: Not all means (of income) among the religious groups are equal.
100 Analysis of Variance (ANOVA)

Assumptions
1. The dependent variable (income) is normally distributed at each level of the
independent variable (religion);
2. The variances of the dependent variable (income) at each level of the indepen-
dent variable (religion) are the same (homogeneity of variances); and
3. Subjects represent random samples from the population.
If the variances of the dependent variable in all the categories are not equal (violation
of assumption 2), but the sample size in all the groups is large and similar, ANOVA can
be used.

11.1.1 Commands
All the following commands will provide the ANOVA test results. We recommend
using the first command, which provides the descriptive statistics along with the ANO-
VA test results (Table 11.1).
oneway income religion_2, tabulate
oneway income religion_2
anova income religion_2

Table 11.1 One-way ANOVA test results of income and religion

. oneway income religion_2, tabulate

| Summary of Monthly income

Religion 2 | Mean Std. Dev. Freq.
------------+------------------------------------
MUSLIM | 88180.905 17207.614 126
HINDU | 79166.028 17804.631 36
Christian | 79405.615 19857.021 26
BUDDHISM | 84796.591 14447.348 22
------------+------------------------------------
Total | 85194.486 17724.033 210

Analysis of Variance
Source SS df MS F Prob > F
------------------------------------------------------------------------
Between groups 3.3068e+09 3 1.1023e+09 3.64 0.0136
Within groups 6.2349e+10 206 302663565
------------------------------------------------------------------------
Total 6.5656e+10 209 314141354

Bartlett's test for equal variances: chi2(3) = 2.2804 Prob>chi2 = 0.516

 Analysis of Variance (ANOVA) 101

11.1.2 Interpretation
The outputs of the one-way ANOVA test are provided in Table 11.1. In this example,
we have used "income" as the dependent variable and "religion" as the independent
variable. The independent variable (religion) has 4 categories (levels) – Muslim,
Hindu, Christian, and Buddhist. The results first provided the descriptive measures
(mean, SD, etc.) of income by religion. For example, the mean income of Muslims is
BDT 88,180.9, with a SD of 17,207.6.
The ANOVA test (F-test) results are provided at the bottom of the descriptive statistics.
The value of the F-statistic is 3.64 and the corresponding p-value (Prob > F) is 0.0136.
Since the p-value is <0.05, we can reject the null hypothesis. This means that not all
group means of income are equal.
However, before interpreting the ANOVA test results, we need to check the Bartlett's
test for homogeneity of variances (equal variances) provided at the bottom of the table.
This test is done to assess if all the group-variances in income are equal (assumption
2). The p-value (Prob>chi2) of the Bartlett’s test is 0.516. Since the p-value is >0.05,
the variances of income at all levels of the religious group are equal (i.e., assumption 2
is not violated). The assumption would have been violated if the p-value was <0.05.

11.1.3 Post hoc test

If the ANOVA test is significant, it indicates that not all group means are equal. But it
does not provide information about which group-means are significantly different. To
identify which group means are significantly different, we need to use a post hoc multi-
ple comparison test, such as Bonferroni, Tukey, or Scheffe’s test. Use any of the
following commands (preferably the first one) to get the post hoc test results (Tables
11.2 and 11.3). If the ANOVA test (F-test) is not significant (i.e., p-value is >0.05), we
do not need the post-hoc test.
pwmean income, over(religion_2) mcompare(bonferroni) effects
oneway income religion_2, bonferroni tabulate
You can generate the box and plot charts to display the distribution of medians/means
of the dependent variable across the groups (at each level of the independent variable,
i.e., in different religious groups). To get the box and plot charts (Fig 11.1) of income
for different religious groups, use the following command:
graph box income, over(religion_2)
102 Analysis of Variance (ANOVA)

Table 11.2 Multiple comparisons of mean income by religion

. pwmean income, over(religion_2) mcompare(bonferroni) effects

Pairwise comparisons of means with equal variances

over : religion_2
---------------------------
| Number of
| Comparisons
-------------+-------------
religion_2 | 6
---------------------------
----------------------------------------------------------------------------------------
| Bonferroni Bonferroni
income | Contrast Std. Err. t P>|t| [95% Conf. Interval]
-----------------------+----------------------------------------------------------------
religion_2 |
HINDU vs MUSLIM | -9014.877 3287.767 -2.74 0.040 -17773.41 -256.3402
Christian vs MUSLIM | -8775.289 3747.399 -2.34 0.121 -18758.27 1207.696
BUDDHISM vs MUSLIM | -3384.314 4019.891 -0.84 1.000 -14093.21 7324.585
Christian vs HINDU | 239.5876 4477.525 0.05 1.000 -11688.44 12167.61
BUDDHISM vs HINDU | 5630.563 4707.946 1.20 1.000 -6911.298 18172.42
BUDDHISM vs Christian | 5390.976 5039.677 1.07 1.000 -8034.608 18816.56
----------------------------------------------------------------------------------------

You can also get the error-bar chart of mean income for different religious groups. Use
the following three commands consecutively to get the error-bar (Fig 11.2). If you use
the command "oneway" instead of “anova” for the ANOVA test, the commands “mar-
gins” and “marginsplot” for error-bar will not work.
anova income religion_2
margins religion_2
marginsplot

11.1.4 Interpretation of post hoc test results

Both the commands for multiple comparisons will provide the Bonferroni test results,
as shown in Tables 11.2 and 11.3. Each row of Table 11.2 represents a comparison
between two specific religious groups. For example, the first row compares the mean
income between Hindus and Muslims. We can see that the mean difference in income
between Hindus and Muslims is -9,014.87 and the corresponding p-value is 0.040
(<0.05). This indicates that the mean income of Muslims and Hindus is different in the
population (Hindus have a significantly lower income than Muslims). The differences
in mean income of other religious groups are not significant as the p-values are >0.05.
The table has also provided the 95% CIs of mean differences. Table 11.3 has provided
the same information except for the 95% CIs of mean differences.
 Analysis of Variance (ANOVA) 103

Table 11.3 Multiple comparisons of mean income by religion

. oneway income religion_2, bonferroni tabulate

| Summary of Monthly income

Religion 2 | Mean Std. Dev. Freq.
------------+------------------------------------
MUSLIM | 88180.905 17207.614 126
HINDU | 79166.028 17804.631 36
Christian | 79405.615 19857.021 26
BUDDHISM | 84796.591 14447.348 22
------------+------------------------------------
Total | 85194.486 17724.033 210

Analysis of Variance
Source SS df MS F Prob > F
------------------------------------------------------------------------
Between groups 3.3068e+09 3 1.1023e+09 3.64 0.0136
Within groups 6.2349e+10 206 302663565
------------------------------------------------------------------------
Total 6.5656e+10 209 314141354

Bartlett's test for equal variances: chi2(3) = 2.2804 Prob>chi2 = 0.516

Comparison of Monthly income by Religion 2

(Bonferroni)
Row Mean-|
Col Mean | MUSLIM HINDU Christia
---------+---------------------------------
HINDU | -9014.88
| 0.040
|
Christia | -8775.29 239.588
| 0.121 1.000
|
BUDDHISM | -3384.31 5630.56 5390.98
| 1.000 1.000 1.000

11.1.5 One-way ANOVA for unequal variances

When the group variances are not homogeneous (i.e., when Bartlett’s test p-value is
<0.05), we cannot use the F-test for comparison of group means. Instead, we need to
use the Welch test. Similarly, for the comparison of individual group means (post hoc
test), instead of Bonferroni’s (or Tukey’s) test, we need to use the Games-Howell test.
There is no straight-forward way to get these tests done in Stata. The users can do these
tests easily by using SPSS [19].

11.2 Two-way ANOVA

Two-way ANOVA is similar to one-way ANOVA except that it examines an additional
independent categorical variable. The two-way ANOVA involves three variables: one
quantitative (dependent variable) and two categorical variables. This test is not com-
104 Analysis of Variance (ANOVA)

Figure 11.1 Box and plot chart of income by religious groups

Figure 11.2 Error-bar chart of mean income for religious groups

monly used in health research. Use the data file <Data_3.dta> for practice.
Suppose that we want to compare the mean systolic BP (variable name is "sbp") in
different occupation and sex groups. Here, the dependent variable is systolic BP and
the independent variables are occupation and sex.
Since there are four levels (categories) of occupation (govt. job; private job; business;
and others) and two categories of sex (male and female) in the data, we will have a
factorial design with eight (4×2) data cells. The two-way ANOVA test answers the
following three questions:
 Analysis of Variance (ANOVA) 105

1. Does occupation influence the systolic BP (i.e., is the mean systolic BP equal
among the occupation groups)?
2. Does sex influence the systolic BP (i.e., is the mean systolic BP equal for males
and females)?
3. Does the influence of occupation on systolic BP depend on sex (i.e., is there
an interaction between occupation and sex)?
Questions one and two refer to the main effect, while the question three explains the
interaction of two independent variables (occupation and sex) on the dependent
variable (systolic BP). The primary objective of two-way ANOVA is to assess if there
is an interaction between the independent categorical variables on the dependent
variable.

Assumptions
1. The dependent variable (systolic BP) is normally distributed at each level of
the independent variables (occupation and sex);
2. The variances of the dependent variable (systolic BP) at each level of the inde-
pendent variables are equal; and
3. Subjects represent random samples from the population.
First, we need to check if the systolic BP is normally distributed in different categories
of occupation and sex separately. We can check this by constructing the histograms and
Q-Q plots, or by doing the Shapiro Wilk test (Chapter 8). We also need to check the
homogeneity of variances in each group of the independent variables (occupation and
sex) by using the Levene's test (Section 10.2.1).

11.2.1 Commands for two-way ANOVA

To get the two-way ANOVA test results for systolic BP (variable name is "sbp"), occu-
pation (variable name is "occupation"), and sex (variable name is "sex_1"), use any of
the following commands. ANOVA does not allow string variables in the analysis. You
need to change the format of a string variable to a numeric variable before including it
in the analysis.
anova sbp occupation##sex_1
Or,
anova sbp occupation sex_1 occupation#sex_1
Either of the above commands will generate Table 11.4.
106 Analysis of Variance (ANOVA)

Table 11.4 Two-way ANOVA table

. anova sbp occupation##sex_1

Number of obs = 210 R-squared = 0.0386

Root MSE = 20.0049 Adj R-squared = 0.0053

Source | Partial SS df MS F Prob > F

-----------------+----------------------------------------------------
Model | 3245.48737 7 463.641053 1.16 0.3283
|
occupation | 470.647297 3 156.882432 0.39 0.7589
sex_1 | 1308.03373 1 1308.03373 3.27 0.0721
occupation#sex_1 | 1735.11998 3 578.373327 1.45 0.2308
|
Residual | 80839.5793 202 400.195937
-----------------+----------------------------------------------------
Total | 84085.0667 209 402.320893

You can get the error graph of the adjusted mean of systolic BP for sex (males and
females) by occupation. To get the plot of mean systolic BP with error bars of occupa-
tion by sex, use the following commands successively (Fig 11.3):
margins sex_1, at(occupation=(1(1)4))
marginsplot

11.2.2 Interpretation
Table 11.4 shows the outputs of the two-way ANOVA test. The table shows the main
effects of the independent variables as well as their interaction. Look at the p-values
(Prob > F) for occupation and sex. They are 0.758 and 0.072, respectively. The findings
indicate that the mean systolic BP is not different in different occupation groups as well
as sex (males and females). Now, look at the p-value for "occupation#sex_1", which
indicates the significance of the interaction between these two variables (occupation
and sex) on systolic BP. A p-value of <0.05 indicates the presence of interaction. The
presence of interaction indicates that the systolic BP in different occupation groups is
influenced by (depends on) sex. In our example, the p-value for interaction is 0.230
(>0.05), which means that there is no interaction between occupation and sex to influ-
ence the systolic BP.

11.2.3 Post hoc test for two-way ANOVA

The post-hoc test (also discussed under one-way ANOVA) is performed if the main
effect is significant (i.e., the p-values for occupation and/or sex are <0.05), otherwise
 Analysis of Variance (ANOVA) 107

Adjusted Predictions of sex_1 with 95% CIs

140
Linear Prediction (sbp)

130
120
110

GOVT JOB PRIVATE JOB BUSINESS OTHERS

Occupation

Female Male

Figure 11.3 Error-bar chart of mean systolic BP of sex by occupation

Table 11.5 Pairwise comparisons table for mean systolic BP by occupation

. pwmean sbp, over(occupation) mcompare(bonferroni) effects

Pairwise comparisons of means with equal variances

over : occupation

---------------------------
| Number of
| Comparisons
-------------+-------------
occupation | 6
---------------------------

------------------------------------------------------------------------------------------
| Bonferroni Bonferroni
sbp | Contrast Std. Err. t P>|t| [95% Conf. Interval]
-------------------------+----------------------------------------------------------------
occupation |
PRIVATE JOB vs GOVT JOB | -3.036395 3.883548 -0.78 1.000 -13.38208 7.309289
BUSINESS vs GOVT JOB | -1.52619 3.883548 -0.39 1.000 -11.87187 8.819493
OTHERS vs GOVT JOB | -2.364103 3.821385 -0.62 1.000 -12.54419 7.81598
BUSINESS vs PRIVATE JOB | 1.510204 4.074798 0.37 1.000 -9.344964 12.36537
OTHERS vs PRIVATE JOB | .672292 4.015596 0.17 1.000 -10.02517 11.36975
OTHERS vs BUSINESS | -.8379121 4.015596 -0.21 1.000 -11.53537 9.859545
------------------------------------------------------------------------------------------

it is not necessary. Our data shows that there is no significant effect of occupation and
sex on systolic BP, since the p-values are 0.758 and 0.072, respectively. However, if
108 Analysis of Variance (ANOVA)

you want to perform the multiple comparisons test (say, Bonferroni) for “occupation”
after the two-way ANOVA, use the first command, while use the second command to
get the same for “sex_1”.
pwmean sbp, over(occupation) mcompare(bonferroni) effects
pwmean sbp, over(sex_1) mcompare(bonferroni) effects
The first command will provide Table 11.5. Interpretation of the results is the same as
discussed in Section 11.1.4.
 12
Repeated Measures ANOVA

Repeated measures design is commonly used in experimental studies. In repeated

measures design, measurements of the same variable are made on each subject on two
or more different occasions (either at different points in time or under different condi-
tions, such as different treatments). It is similar to a paired t-test, except that there are
more than two measurements in the repeated measures ANOVA. The repeated
measures ANOVA test compares the means across one or more variables that are based
on repeated observations. In this chapter, one-way (within-subjects) repeated measures
ANOVA is discussed.

12.1 One-way repeated measures ANOVA

The one-way repeated measures ANOVA test is analogous to the paired samples t-test
that we have discussed earlier (Chapter 10). The main difference is that, in a paired
samples t-test, we have two measurements on the same subjects at different times (e.g.,
before and after giving a drug, or pre-test and post-test results), while in a one-way
repeated measures ANOVA, there are three or more measurements on the same
subjects at different points in time (i.e., the subjects are exposed to multiple measure-
ments over a period of time or conditions). One-way repeated measures ANOVA is
also called one-way within-subjects ANOVA. Use the data file <Repeat anova_3.dta>
for practice.
Suppose that we are interested in assessing the mean blood sugar levels at four differ-
ent time intervals (e.g., at hour-0, hour-7, hour-14, and hour-24) after administration of
a drug on 15 study subjects. The objective is to assess whether the drug reduces blood
110 Repeated Measures ANOVA

sugar levels over time (i.e., whether the mean blood sugar levels over time are the same
or different).
To conduct this study, we have randomly selected 15 individuals from a population and
measured their blood sugar levels at the baseline, i.e., before administration of a drug
(hour-0). All the individuals are then administered a drug (say, drug A), and their blood
sugar levels are measured again after 7 hours, 14 hours, and 24 hours. We are interested
in knowing if the blood sugar levels over time, after giving the drug, are the same or
not (in other words, whether the drug is effective in reducing the blood sugar levels
over time). The variable "time" in the dataset indicates the times of measurement of
blood sugar levels in the subjects. In this example, we have only one treatment group
(received drug A) but have the outcome measurements (blood sugar) at four different
points in time on the same subjects (i.e., we have one treatment group with four levels
of measurement on the same subjects).

Hypothesis
H0: The mean blood sugar levels are equal (same) at each level of measurement
(i.e., the mean blood sugar levels at 0, 7, 14, and 24 hours in the population are the
same).
HA: The mean blood sugar levels are not equal at different levels of measurement
(i.e., the mean blood sugar levels at 0, 7, 14, and 24 hours in the population are
different).
Assumptions
1. The dependent variable (blood sugar level) is normally distributed in the popu-
lation at each level of within-subjects factor;
2. The population variances of the differences between all combinations of relat-
ed groups/levels are equal (called the Sphericity assumption); and
3. The subjects represent a random sample from the population.

12.1.1 Commands
The data file “Repeat anova_3.dta” has the following variables:

Subject: It indicates the study participants, like participants 1, 2, 3, and so on. You will
notice that there are 15 subjects enrolled in this study (you can check it by using the
command "tab", like "tab subject").
 Repeated Measures ANOVA 111

Time: The variable "time" indicates the times of measurement of blood sugar levels.
Blood sugar levels were measured on the same subjects at four different times, such as:
a) at the baseline, i.e., before treatment (coded as 0); b) 7 hours after treatment (coded
as 1); c) 14 hours after treatment (coded as 2); and d) 24 hours after treatment (coded
as 3). You can check the times of measurement of blood sugar levels on the subjects by
using the following command.
tab time
This will provide Table 12.1. The table shows that there are four categories of the
variable "time", and they are: a) before treatment; b) 7 hours after treatment; c) 14
hours after treatment; and d) 24 hours after treatment.

Table 12.1 Frequency distribution of the variable “time”

. tab time

Time of measurement | Freq. Percent Cum.

-----------------------+-----------------------------------
before treatment | 15 25.00 25.00
7 hrs after treatment | 15 25.00 50.00
14 hrs after treatment | 15 25.00 75.00
24 hrs after treatment | 15 25.00 100.00
-----------------------+-----------------------------------
Total | 60 100.00

Sugar: The variable “sugar” indicates the blood sugar levels of the study subjects at
different times of measurement.
There is another variable named “treatment” in the dataset that indicates in which treat-
ment group the subjects were enrolled in. This variable is not needed for the analysis
of one-way repeated measures ANOVA because we need only one treatment group for
the analysis.
To perform the one-way repeated measures ANOVA test, use the following command:
anova sugar subject time, repeated(time)
Once you use this command, the Stata results may show “matsize too small”. If you see
this message in the output window, you need to increase the “matsize” by using the
following command. Matsize indicates “maximum matrix size”, which influences the
number of variables that can be included in any of Stata's estimation commands.
set matsize 10000
112 Repeated Measures ANOVA

The above command will increase the “matsize” to 11,000 in Stata (you can only
increase the “matsize” up to 11,000). Now, use the following command to perform the
repeated measures ANOVA test:
anova sugar subject time, repeated(time)
While using the command, the computer may take some time to analyze the data,
depending on the speed and memory of your computer and the size of the data. It may
take 3-7 minutes to calculate the ANOVA test results. The results of the analysis are
shown in Table 12.2.

Table 12.2 Repeated measures ANOVA

. anova sugar subject time, repeated(time)

Panel 1.
Number of obs = 60 R-squared = 0.7643
Root MSE = 3.94868 Adj R-squared = 0.6689

Source | Partial SS df MS F Prob > F

-----------+----------------------------------------------------
Model | 2123.86667 17 124.933333 8.01 0.0000
|
subject | 1194.73333 14 85.3380952 5.47 0.0000
time | 929.133333 3 309.711111 19.86 0.0000
|
Residual | 654.866667 42 15.5920635
-----------+----------------------------------------------------
Total | 2778.73333 59 47.0971751

Between-subjects error term: subject

Levels: 15 (14 df)
Lowest b.s.e. variable: subject

Panel 2.

Repeated variable: time

Huynh-Feldt epsilon = 0.4463
Greenhouse-Geisser epsilon = 0.4232
Box's conservative epsilon = 0.3333

------------ Prob > F ------------

Source | df F Regular H-F G-G Box
-----------+----------------------------------------------------
time | 3 19.86 0.0000 0.0001 0.0001 0.0005
Residual | 42
----------------------------------------------------------------
 Repeated Measures ANOVA 113

12.1.2 Interpretation
Look at the “time” row in Table 12.2 (Panel 1). The p-value (Prob > F) is 0.000. This
indicates that the mean blood sugar levels significantly differ at different time points
(we have four time points). However, before explaining this table, we need to check
whether the sphericity assumption (assumption 2) has been violated or not. To check
whether the sphericity assumption is violated or not, we need to do the "Mauchly’s
test" for the sphericity assumption. If the assumption is violated (i.e., Mauchly’s test
p-value is <0.05), we commonly use the p-value of the Greenhouse-Geisser (G-G) test
as shown in the second panel of Table 12.2 to interpret the results of the repeated
measures ANOVA test.
If the within-subjects factor has more than two levels, three types of tests are available
in repeated measures ANOVA. In our dataset, the within-subjects factor is the times of
measurement of blood sugar levels (variable name is "time"), which has four levels —
before treatment, 7 hours after treatment, 14 hours after treatment, and 24 hours after
treatment (Table 12.1). The tests are:
1. Standard univariate test (when the sphericity assumption is not violated);
2. Alternative univariate tests (Greenhouse-Geisser, Huynh-Feldt, and Lower-
bound); and
3. Multivariate tests (Pillai's Trace, Wilks' Lambda, Hotelling's Trace, and Roy's
Largest Root).
All these tests evaluate the same hypothesis that the population means are equal at all
levels of measurement. The standard univariate test is based on the sphericity assump-
tion, i.e., the standard univariate test result is considered if the sphericity assumption is
not violated. In reality, and in most cases, the sphericity assumption is violated, and we
cannot use the standard univariate test results as provided in panel 1 of Table 12.2
(time row). It is, therefore, recommended to use the alternative univariate test, such as
the "Greenhouse-Geisser (G-G)" test (or the others), as provided under panel 2 of Table
12.2. The results show that the p-value of the G-G test is 0.0001 (in the time row of
panel 2), which is <0.05. This indicates that the mean blood sugar levels differ signifi-
cantly at different time points. To check the mean of which time points are statistically
different, see below (Table 12.5).
However, if you want to check the sphericity assumption, you need to do the Mauch-
ly’s test. To get the Mauchly’s test, you need to install the modules "Mauchly and
moremata" (commonly, they are not the built-in commands in Stata). To install the
modules, use the following commands:
114 Repeated Measures ANOVA

ssc install mauchly

ssc install moremata
After installation of the modules, use the following commands to get the Mauchly’s
test results:
xtset subject
mauchly sugar, m(time)
This will provide the results of Mauchly’s test of sphericity (Table 12.3). Our interest
is in the p-value of the test. The p-value of the test, as shown in the table, is zero. If the
p-value is >0.05, the sphericity assumption is not violated, and you can use the results
provided in the first panel of Table 12.2 (in the time row). If the p-value is <0.05 (i.e.,
the sphericity assumption is violated), use the results of the Greenhouse-Geisser (G-G)
or the other test [e.g., Huynh-Feldt (H-F)] as provided under panel 2 of Table 12.2.

Table 12.3 Mauchly’s test of sphericity assumption

. xtset subject
panel variable: subject (balanced)

. mauchly sugar, m(time)

Mauchly's Test of Sphericity

________________________________________________________________________________
Mauchly's W. Chi2. d.f. P-value. Epsilon_gg. Epsilon_ff. Lower-bound
______________________________________________________________________________

0.0689 34.0332 5 0 0.4232 0.4463 0.3333

________________________________________________________________________________

To get the means and standard deviations of blood sugar levels at four time points, use
the following command:
tabstat sugar, stat(n mean sd) by(time)
This will provide Table 12.4 with the means and standard deviations of blood sugar
levels at different time points, including the number of subjects (n). You can get the
pairwise comparison of mean blood sugar levels by using the following command
(with Bonferroni’s test) (Table 12.5):
pwmean sugar, over(time) mcompare(bonferroni) pveffects
Finally, to get the line graph of mean blood sugar levels over time with 95% CI, use the
 Repeated Measures ANOVA 115

following commands successively after the primary analysis (Fig 12.1):

anova sugar subject time, repeated(time)
margins time
marginsplot

Table 12.4 Mean and SD of blood sugar levels at different time points
. tabstat sugar, stat(n mean sd) by(time)

Summary for variables: sugar

by categories of: time (Time of measurement)

time | N mean sd
-----------------+------------------------------
before treatment | 15 110.5333 4.73387
7 hrs after trea | 15 105.2 4.427189
14 hrs after tre | 15 101.5333 6.300416
24 hrs after tre | 15 100.4667 7.099966
-----------------+------------------------------
Total | 60 104.4333 6.862738
----------------------------------------------------------------------------------

Predictive Margins of time with 95% CIs

115
Linear Prediction (blood sugar)

110
105
100
95

before treatment 7 hrs after treatment 14 hrs after treatment 24 hrs after treatment
Time of measurement

Figure 12.1 Mean blood sugar levels with 95% CIs at different time points
116 Repeated Measures ANOVA

Table 12.5 Pairwise comparison of mean blood sugar levels

. pwmean sugar, over(time) mcompare(bonferroni) pveffects

Pairwise comparisons of means with equal variances

over : time

---------------------------
| Number of
| Comparisons
-------------+-------------
time | 6
---------------------------

------------------------------------------------------------------------------------------
| Bonferroni
sugar | Contrast Std. Err. t P>|t|
--------------------------------------------------+---------------------------------------
time |
7 hrs after treatment vs before treatment | -5.333333 2.098526 -2.54 0.083
14 hrs after treatment vs before treatment | -9 2.098526 -4.29 0.000
24 hrs after treatment vs before treatment | -10.06667 2.098526 -4.80 0.000
14 hrs after treatment vs 7 hrs after treatment | -3.666667 2.098526 -1.75 0.516
24 hrs after treatment vs 7 hrs after treatment | -4.733333 2.098526 -2.26 0.168
24 hrs after treatment vs 14 hrs after treatment | -1.066667 2.098526 -0.51 1.000
------------------------------------------------------------------------------------------
 13

Association Between Two Categorical

Variables: Chi-square Test of Independence

The chi-square test is a commonly used statistical test for testing a hypothesis in health
and social sciences research. The chi-square value ranges from 0 to ∞ (infinity), and it
does not take any negative value. This test is suitable to determine the association
between two categorical variables, whether the data are from cross-sectional,
case-control, or cohort studies. On the other hand, in epidemiology, cross-tabulations
are commonly done to calculate the odds ratio (OR) [for a case-control study] and
relative risk (RR) [for a cohort study] with their 95% confidence intervals (CI). The
OR and RR are the measures of strength of association between two variables. In this
chapter, we have discussed the chi-square and the Fisher’s exact tests for hypothesis
testing and how to get the RR and OR using Stata. We have also demonstrated how to
perform a stratified analysis in this chapter. Use the data file <Data_3.dta> for practice.

13.1 Chi-square test of independence

The Chi-square test of independence is used to determine the association between two
categorical variables. Suppose that you have collected data on gender (sex) and diabe-
tes from a group of individuals selected randomly from a population. You are interest-
ed in knowing if there is an association between gender and diabetes. In such a situa-
tion, the chi-square test is the appropriate test for testing the hypothesis.
118 Association Between Two Categorical Variables: Chi-square Test of Independence

Hypothesis
H0: There is no association between gender and diabetes (it can also be stated as
gender and diabetes are independent).
HA: There is an association between gender and diabetes (or, gender and diabetes
are not independent).

Assumption
1. The data is a random sample drawn from a population.

13.1.1 Commands
The basic command to get the chi-square test results is to generate a cross-table using
the command "tab" with the option of chi-square statistics. Use any of the following
commands to find an association (chi-square test) between sex and diabetes.
tab sex diabetes, chi2
tab sex diabetes, row col chi2
The first command will generate a cross-table of sex (the first variable is placed on the
row) and diabetes with only the observed frequencies and the chi-square test results.
The second command will provide the row and column percentages in the cross-table,
including the observed frequencies and chi-square test results (Table 13.1).
You can also use the option “all” to get all the relevant statistics (chi-square, Cramer’s
V, and others) for the association between two categorical variables (Table 13.2), such
as:
tab sex diabetes, all
tab sex diabetes, col row all
tab sex diabetes, expected all
The last command will provide all the relevant statistics with the expected cell values.
The chi-square test is valid if no more than 20% of cells have expected values of less
than 5. If the expected value is less than 5 in more than 20% of the cells, we need to use
the Fisher’s exact test instead of the chi-square test. The command for the Fisher’s
exact test is (Table 13.3):
tab sex diabetes, exact
tab sex diabetes, expected exact
Association Between Two Categorical Variables: Chi-square Test of Independence 119

Table 13.1 Chi-square test results with row and column percentages
. tab sex diabetes, row col chi2

| Have diabetes
Sex: | mellitus
string | Yes No | Total
-----------+----------------------+----------
f | 20 113 | 133
| 15.04 84.96 | 100.00
| 44.44 68.48 | 63.33
-----------+----------------------+----------
m | 25 52 | 77
| 32.47 67.53 | 100.00
| 55.56 31.52 | 36.67
-----------+----------------------+----------
Total | 45 165 | 210
| 21.43 78.57 | 100.00
| 100.00 100.00 | 100.00

Pearson chi2(1) = 8.7995 Pr = 0.003

Table 13.2 Cross tabulation with chi-square and other test results
. tab sex diabetes, col row all

| Have diabetes
Sex: | mellitus
string | Yes No | Total
-----------+----------------------+----------
f | 20 113 | 133
| 15.04 84.96 | 100.00
| 44.44 68.48 | 63.33
-----------+----------------------+----------
m | 25 52 | 77
| 32.47 67.53 | 100.00
| 55.56 31.52 | 36.67
-----------+----------------------+----------
Total | 45 165 | 210
| 21.43 78.57 | 100.00
| 100.00 100.00 | 100.00

Pearson chi2(1) = 8.7995 Pr = 0.003

likelihood-ratio chi2(1) = 8.5367 Pr = 0.003
Cramér's V = -0.2047
gamma = -0.4618 ASE = 0.135
Kendall's tau-b = -0.2047 ASE = 0.071
120 Association Between Two Categorical Variables: Chi-square Test of Independence

13.1.2 Interpretation
Table 13.1 is a 2 by 2 table of sex and diabetes with row and column percentages and
the chi-square test results. The question is, which percentage should you report? It
depends on the situation/study design and what you want to report. For the data of a
cross-sectional study, it may provide better information to the readers if row percent-
ages are reported. In that case, the row percentages indicate the prevalence of the
condition (in this example, diabetes).

Table 13.3 Fisher’s exact test results with expected cell values
. tab sex diabetes, expected exact

| Have diabetes
Sex: | mellitus
string | Yes No | Total
-----------+----------------------+----------
f | 20 113 | 133
| 28.5 104.5 | 133.0
-----------+----------------------+----------
m | 25 52 | 77
| 16.5 60.5 | 77.0
-----------+----------------------+----------
Total | 45 165 | 210
| 45.0 165.0 | 210.0

Fisher's exact = 0.005

1-sided Fisher's exact = 0.003

For example, one can understand from Table 13.1 that the prevalence of diabetes
among males is 32.47% and that of females is 15.04% when row percentages are used.
However, the column percentages can also be reported for a cross-sectional data (most
of the publications use column percentages). If column percentages are used, the
meaning will be different. In our example (Table 13.1), the results indicate that 55.56%
of the diabetic patients are males, compared to 31.52% of the non-diabetic individuals.
If the data is from a case-control study, you must report column percentages (you
cannot use row percentages for case-control studies). On the other hand, for the data of
a cohort study, one should report the row percentages. In such a situation, it would
indicate the incidence (instead of the prevalence) of the disease among males and
females.
We can see in Table 13.1 (in the row of "Total") that the overall (irrespective of gender)
prevalence of diabetes is 21.43% (considering the data is from a cross-sectional study).
 Association Between Two Categorical Variables: Chi-square Test of Independence 121

Table 13.1 also shows that 32.47% of males have diabetes compared to only 15.04% of
females (i.e., the prevalence among males and females). The chi-square test actually
tests the hypothesis of whether the prevalence of diabetes among males and females in
the population is the same or not.
Now, look at the Pearson’s chi-square test results provided at the bottom of the table
[(Pearson chi2(1) = 8.7995 Pr = 0.003)]. The "Pearson chi2(1)" indicates the Pearson’s
chi-square test result with the degree of freedom (df) of 1, while "Pr" indicates the
p-value.
Before we conclude the chi-square test results, it is important to check if there is any
cell in the 2 by 2 table that has an expected value of less than 5. This can be checked
using the option "expected". Table 13.3 displays the expected cell values and the Fish-
er’s exact test p-value. The table shows that there is no cell in the 2 by 2 table with an
expected value of less than 5 (the minimum expected value found is 16.5). For the use
of the chi-square test results, it is desirable to have no cell in a 2 by 2 table with an
expected value of less than 5. If this is not fulfilled, we need to use the Fisher’s exact
test p-value to interpret the results.
Since we do not have the problem of an expected value of less than 5 in the 2 by 2 table,
we will consider the chi-square test results given at the bottom of Tables 13.1 and 13.2
for conclusion. The results show that the calculated chi-square value is 8.79 (df= 1) and
the corresponding p-value is 0.003. Since the p-value is <0.05, we will reject the null
hypothesis. This indicates that there is a significant association between gender and
diabetes. It can also be concluded that the prevalence of diabetes among males is
significantly higher than that of females, which is statistically significant at 95% confi-
dence level (p=0.003).

13.2 Relative risk and odds ratio

We calculate relative risk (RR) for the data of cohort studies and odds ratio (OR) for
case-control studies (OR is also sometimes calculated for cohort studies and cross-sec-
tional studies). To calculate the RR and OR with their confidence intervals (CI), we
need to use the "epidemiology and related" option of data analysis. We also need to
recode the outcome and exposure variables as 0/1 (0 for no disease/unexposed; 1 for
have disease/exposed) if they are not coded like this.
We want to calculate the RR from the cross-tabulation between diabetes (outcome
variable) and sex_1 (exposure variable). Since the variable "diabetes" (in our dataset)
122 Association Between Two Categorical Variables: Chi-square Test of Independence

is coded as 1/2 (1= have diabetes and 2= don’t have diabetes), we need to change the
coding structure of the variable "diabetes" as 0= don’t have diabetes and 1= have
diabetes using the command "recode" as discussed in Section 5.2. There is no problem
with the variable "sex_1" in the dataset since it is already coded as 0= female and 1=
male. Once it is done (there is a variable in the dataset "diabetes1", which is coded as
0/1), use the following command to get the RR and OR with their 95% CIs:
cs diabetes1 sex_1
cs diabetes1 sex_1, or
cs diabetes1 sex_1, or level(99)
The command “cs” indicates “cohort study”. This command will automatically
provide the RR and its 95% CI (by default) along with other statistics (Table 13.4). If
you want to get the OR, in addition to RR, use the second command. The third com-
mand is for getting the 99% CIs for both the RR and OR.

Table 13.4 Relative risk and 95% confidence interval

. cs diabetes1 sex_1

| Sex: numeric |
| Exposed Unexposed | Total
-----------------+------------------------+------------
Cases | 25 20 | 45
Noncases | 52 113 | 165
-----------------+------------------------+------------
Total | 77 133 | 210
| |
Risk | .3246753 .1503759 | .2142857
| |
| Point estimate | [95% Conf. Interval]
|------------------------+------------------------
Risk difference | .1742994 | .0533492 .2952495
Risk ratio | 2.159091 | 1.287892 3.619616
Attr. frac. ex. | .5368421 | .2235371 .7237276
Attr. frac. pop | .2982456 |
+-------------------------------------------------
chi2(1) = 8.80 Pr>chi2 = 0.0030

Instead of using the “cs” command, you can use the following commands to get the OR
without a 2 by 2 table:
tabodds diabetes1 sex_1, or
tabodds diabetes1 sex_1, base(2) or
The first command will provide the OR without the 2 by 2 table, and with the first
 Association Between Two Categorical Variables: Chi-square Test of Independence 123

category of sex (females; indicated by Odds Ratio = 1.0) as the comparison group
(Table 13.5). The second command will provide the same but the second category of
sex (males) as the comparison group (Table 13.5).

Table 13.5 Odds ratio with 95% CI and p-value

tabodds diabetes1 sex_1, or

---------------------------------------------------------------------------
sex_1 | Odds Ratio chi2 P>chi2 [95% Conf. Interval]
-------------+-------------------------------------------------------------
Female | 1.000000 . . . .
Male | 2.716346 8.76 0.0031 1.362845 5.414068
---------------------------------------------------------------------------
Test of homogeneity (equal odds): chi2(1) = 8.76
Pr>chi2 = 0.0031

Score test for trend of odds: chi2(1) = 8.76

Pr>chi2 = 0.0031

tabodds diabetes1 sex_1, base(2) or

---------------------------------------------------------------------------
sex_1 | Odds Ratio chi2 P>chi2 [95% Conf. Interval]
-------------+-------------------------------------------------------------
Female | 0.368142 8.76 0.0031 0.184704 0.733759
Male | 1.000000 . . . .
---------------------------------------------------------------------------
Test of homogeneity (equal odds): chi2(1) = 8.76
Pr>chi2 = 0.0031

Score test for trend of odds: chi2(1) = 8.76

Pr>chi2 = 0.0031

If you want to calculate the OR with 95% CI (default) for a case-control study, use the
first command (Table 13.6) as shown below. Use the second command if you want to
get the 99% CI.
cc diabetes1 sex_1
cc diabetes1 sex_1, level(99)
Another way of getting the OR is to use the command “logistic”, which is used for
logistic regression analysis. The command is:
logistic diabetes1 sex_1
The results of the above command (called univariate logistic regression analysis) are
shown in Table 13.7.
124 Association Between Two Categorical Variables: Chi-square Test of Independence

Table 13.6 Odds ratio and 95% confidence interval

. cc diabetes1 sex_1
Proportion
| Exposed Unexposed | Total Exposed
-----------------+------------------------+------------------------
Cases | 25 20 | 45 0.5556
Controls | 52 113 | 165 0.3152
-----------------+------------------------+------------------------
Total | 77 133 | 210 0.3667
| |
| Point estimate | [95% Conf. Interval]
|------------------------+------------------------
Odds ratio | 2.716346 | 1.311301 5.641061 (exact)
Attr. frac. ex. | .6318584 | .2373985 .8227284 (exact)
Attr. frac. pop | .3510324 |
+-------------------------------------------------
chi2(1) = 8.80 Pr>chi2 = 0.0030

Table 13.7 Odds ratio by using the “logistic” command

. logistic diabetes1 sex_1

Logistic regression Number of obs = 210

LR chi2(1) = 8.54
Prob > chi2 = 0.0035
Log likelihood = -104.84341 Pseudo R2 = 0.0391

------------------------------------------------------------------------------
diabetes1 | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
sex_1 | 2.716346 .9334131 2.91 0.004 1.385123 5.326991
_cons | .1769912 .0429362 -7.14 0.000 .1100168 .284737
------------------------------------------------------------------------------

In the above examples, both the outcome (dependent variable) and exposure (indepen-
dent variable; sex) variables are dichotomous variables with the coding schemes of 0/1.
If the exposure variable has more than two levels/categories (e.g., the variable "reli-
gion" has 3 categories in our dataset; 1= Muslim, 2= Hindu, and 3= Christian), the
commands "cc" or "cs" will not calculate the ORs. In that case (when the exposure
variable has more than two categories), use the command "logistic" to get the ORs,
such as (outputs are provided in Table 13.8):
logistic diabetes1 i.religion
The prefix "i." used for the variable "religion" (i.e., i.religion) is necessary when it is a
categorical variable with more than two levels. When the "i." prefix is used before the
exposure variable, Stata considers it a categorical variable. If the prefix "i." is not used,
 Association Between Two Categorical Variables: Chi-square Test of Independence 125

Stata will consider the variable as a continuous variable and the results will be mislead-
ing. With the above command, Stata will consider the first category (Muslims) as the
comparison group by default. If you want the other category of religion as the compari-
son group (say, last category or Christians), use any of the following commands:
logistic diabetes1 ib3.religion
tabodds diabetes1 religion, base(3) or

Table 13.8 Odds ratios for an exposure variable with more than 2 levels
. logistic diabetes1 i.religion

Logistic regression Number of obs = 210

LR chi2(2) = 3.01
Prob > chi2 = 0.2215
Log likelihood = -107.60445 Pseudo R2 = 0.0138

------------------------------------------------------------------------------
diabetes1 | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
religion |
HINDU | 1.465201 .5378927 1.04 0.298 .7135283 3.008732
Christian | .5016722 .3271554 -1.06 0.290 .1397418 1.801001
|
_cons | .26 .0572364 -6.12 0.000 .1688846 .4002734
------------------------------------------------------------------------------

13.2.1 Interpretation
Table 13.4 shows the cross-tabulation of diabetes and sex. The table shows that of the
45 cases of diabetes (cases), 25 are males (males are exposed since they are coded as
1). On the other hand, of the 165 non-cases (who do not have diabetes), 52 are males.
The RR (also called risk ratio) calculated from the data is 2.15, and the 95% CI is 1.28
- 3.61. Stata has also provided the chi-square test results at the bottom of the table,
including the p-value of the test (0.003). From the data, we can conclude that males are
at 2.15 times higher risk of having diabetes compared to females, which is statistically
significant at 95% confidence level (RR: 2.15; 95% CI of RR: 1.28 - 3.61; p=0.003).
Table 13.6 shows the OR (of being male) with the 95% CI. The OR, as calculated, is
2.71 and its 95% CI is 1.31 - 5.64. The analysis also provided the p-value of the
chi-square test (Pr>chi2), which is 0.003. We can, therefore, conclude that males are
more likely to have diabetes compared to females, which is statistically significant at
95% confidence level (OR: 2.71; 95% CI of OR: 1.31 - 5.64; p=0.003).
126 Association Between Two Categorical Variables: Chi-square Test of Independence

Table 13.7 shows the analysis of the data using the "logistic" command (univariate
logistic regression analysis). The table shows the OR (2.71), the 95% CI of OR (1.38 -
5.32) and the p-value (0.004). This is the OR for males compared to females (lower
value is considered as the comparison group).
Table 13.8 shows the association between diabetes and religion, which has three levels.
Note that Stata considers the lowest value/code (Muslims) of the exposure variable as
the comparison group during such an analysis (by default). The table shows the ORs,
95% CIs, and p-values for Hindus and Christians. The OR of diabetes for being a
Hindu is 1.46 (95% CI: 0.71 - 3.00; p=0.298) compared to Muslims, which is not statis-
tically significant since the p-value is greater than 0.05. Similarly, the OR of diabetes
for being a Christian is 0.50 (95% CI: 0.13 - 1.80; p=0.290) compared to Muslims,
which is also not statistically significant.

13.3 Stratified analysis

The stratified analysis is a statistical method that allows us to test for confounding and
interaction effects. It also provides the adjusted RR (or OR) for the association
between an exposure and an outcome after controlling for one or more variables.
For example, you may be interested in estimating the adjusted value of RR (or OR) for
the association between sex and diabetes after controlling (adjusting) for the variable
“family history of diabetes” (the variable name is “f_history”). To get the adjusted RR
(or OR), we will do the stratified analysis using the following commands (Tables 13.9
and 13.10):
cs diabetes1 sex_1, by(f_history)
cc diabetes1 sex_1, by(f_history)
The first command will provide the adjusted RR (Table 13.9), while the second com-
mand will provide the adjusted OR (Table 13.10). From these tables, we can also get
information on whether the family history of diabetes is a confounding factor in the
relationship between sex and diabetes, and whether or not there is an interaction
between sex and family history of diabetes on the outcome variable (diabetes).
 Association Between Two Categorical Variables: Chi-square Test of Independence 127

Table 13.9 Results of stratified analysis (RR)

. cs diabetes1 sex_1, by(f_history)

Family history o | RR [95% Conf. Interval] M-H Weight

-----------------+-------------------------------------------------
Yes | 2.907692 1.330149 6.356186 3.554688
No | 1.794872 .7021495 4.588146 1.902439
-----------------+-------------------------------------------------
Crude | 2.159091 1.287892 3.619616
M-H combined | 2.519746 1.36505 4.6512
-------------------------------------------------------------------
Test of homogeneity (M-H) chi2(1) = 0.631 Pr>chi2 = 0.4271

Table 13.10 Results of stratified analysis (OR)

. cc diabetes1 sex_1, by(f_history)

Family history o | OR [95% Conf. Interval] M-H Weight

-----------------+-------------------------------------------------
Yes | 3.818182 1.390158 11.51741 2.40625 (exact)
No | 2.192308 .4142173 9.721011 1.268293 (exact)
-----------------+-------------------------------------------------
Crude | 2.716346 1.311301 5.641061 (exact)
M-H combined | 3.257001 1.520939 6.974675
-------------------------------------------------------------------
Test of homogeneity (M-H) chi2(1) = 0.44 Pr>chi2 = 0.5056

Test that combined OR = 1:

Mantel-Haenszel chi2(1) = 9.53
Pr>chi2 = 0.0020

13.3.1 Interpretation
The stratified analysis in epidemiology is done to estimate the strength of association
(RR or OR) between an exposure (e.g., sex) and an outcome (diabetes) after
controlling (adjusting) for a third categorical variable (e.g., family history of diabetes).
It also enables us to examine whether: a) the third variable is a confounding factor, or
b) there is an interaction (also called effect modification) between exposure and the
third factor.
The stratified analysis can be done for the data of cross-sectional, cohort, or case-con-
trol studies. It is suitable to adjust for a single stratified variable, though more than one
variable can be used in the analysis.
In our example, we have examined the relationship between sex (exposure) and diabetes
(outcome) at two levels of the categorical stratified variable "family history of diabetes".
128 Association Between Two Categorical Variables: Chi-square Test of Independence

Table 13.9 shows the results of stratified analysis considering that the data is from a
cohort study. The results show that among those who have a family history of diabetes,
the RR for males (compared to females) is 2.90 (95% CI: 1.33 - 6.35), while the RR is
1.79 (95% CI: 0.70 - 4.58) when there is no family history of diabetes. The table also
shows the crude (Crude) and adjusted RR (M-H combined). The crude (or unadjusted)
RR is calculated without considering the third (stratified) variable (i.e., family history
of diabetes). The crude RR, as calculated by Stata, is 2.15 (95% CI: 1.28 - 3.61), while
the adjusted RR (M-H combined) is 2.51 (95% CI: 1.36 - 4.65). The adjusted RR
indicates the RR after controlling for the family history of diabetes.
In our data (Table 13.9), we can see that there is a difference between the crude (2.15)
and adjusted (2.51) RR. This (i.e., when there is a difference between crude and adjust-
ed RR) may indicate that the family history of diabetes has some confounding effect
(influence) on the relationship between sex and diabetes. Though there is no set rule for
deciding what amount of difference is considered significant, in general, more than
20% change is considered important (in our example, it is less than 20%). However, if
the crude (unadjusted) and adjusted RRs (or ORs) are close together, the third variable
(stratified variable) is not a confounding factor in the relationship between an exposure
and an outcome.
Now, the question is whether there is an interaction between sex and family history of
diabetes on the outcome. If there is an interaction, the RRs (or ORs) in two strata of the
third variable will be different. In our example, for males, if there is a family history of
diabetes, the RR is 2.90, while the RR is 1.79 if there is no family history of diabetes,
and they are different. This indicates that there may have an interaction between sex
and family history of diabetes on the outcome. However, before we conclude like this,
we need to check the statistical significance of the difference. Stata has provided the
statistical test (Test of homogeneity (M-H); chi2(1) = 0.631; Pr>chi2 = 0.4271) to
understand the significance of the difference at the bottom of the table (Table 13.9). It
shows that the p-value (Pr>chi2 = 0.4271) is 0.427. Since the p-value is >0.05, the
difference in RRs in the two strata is not statistically significant. We may, therefore,
conclude that there is no interaction between sex and family history of diabetes on
outcome even though the RRs are different in two strata of family history of diabetes.
If the homogeneity test is significant (less than 0.05), it is not appropriate to report the
adjusted RR (or OR). The results should be presented for each stratum separately.
Table 13.10 has provided similar information, where we have calculated the OR
(instead of RR) considering that the data is from a case-control study. We can see that
 Association Between Two Categorical Variables: Chi-square Test of Independence 129

the crude and adjusted ORs are 2.71 and 3.25, respectively. Since the ORs are substan-
tially different (exactly 20%), the family history of diabetes is a confounding factor in
the relationship between sex and diabetes. The p-value for the adjusted OR is provided
at the bottom of the table (Mantel-Haenszel chi2(1) = 9.53; Pr>chi2 = 0.0020). It
shows that the p-value is 0.002. We, therefore, conclude that males are 3.25 (M-H com-
bined OR) times more likely to have diabetes compared to females after controlling for
family history of diabetes, which is statistically significant (95% CI: 1.42 – 6.97;
p=0.002).
On the other hand, the ORs of diabetes for males with and without a family history of
diabetes are 3.81 and 2.19, respectively. Though they are different, the difference is not
statistically significant (p=0.505) as shown in the table [Test of homogeneity (M-H);
chi2(1) = 0.44; Pr>chi2 = 0.5056]. Therefore, there is no interaction between sex and
family history of diabetes on the outcome variable.
130
 14

Hypothesis Test of Proportions

Data is frequently collected in health and social sciences research to estimate the
proportions. We may have a situation where there is a single group of individuals and
a certain proportion of them have a particular characteristic (e.g., a disease). For exam-
ple, a researcher has collected data from a population by taking a random sample and
found that a certain percentage (proportion) of individuals have diabetes. The research-
er may be interested in testing the null hypothesis that the population proportion is
equal to a pre-specified value/proportion (one-sample test of proportion).
On the other hand, for the comparison of proportions of two independent samples or
the proportions of two groups of individuals, a two-sample test of proportions is used.
In this chapter, we have discussed how to test the null hypothesis for one-sample and
two-sample proportions. Use the data file “Data_3.dat”.

14.1 One-sample test of proportion

There is a variable “diabetes1” in the dataset. In the variable, there are individuals who
have diabetes (coded as 1) and those who do not have diabetes (coded as 0) [the
variable’s coding scheme must be 0/1, otherwise the command for the statistical test of
proportion will not work]. Let us assume that the data is a random sample from a
district. We can calculate the prevalence of diabetes from the data by using the follow-
ing command:
tab diabetes1
This command will provide Table 14.1. The table shows that there are 45 individuals
132 Hypothesis Test of Proportions

who have diabetes out of 210. Therefore, the prevalence of diabetes is 21.43%.
We are interested in testing the hypothesis of whether or not the prevalence of diabetes
in the district is different from the national prevalence of 19.0%. Here, the null hypoth-
esis is "the prevalence of diabetes in the district is not different from 19.0%", while the
alternative hypothesis is "the prevalence of diabetes in the district is different from
19.0%". This is a situation where we can apply the one-sample test of proportion.

Table 14.1 Prevalence of diabetes

. tab diabetes1

have |
diabetes 01 | Freq. Percent Cum.
------------+-----------------------------------
no | 165 78.57 78.57
yes | 45 21.43 100.00
------------+-----------------------------------
Total | 210 100.00

Table 14.2 One-sample test of proportion

. prtest diabetes1==.19

One-sample test of proportion diabetes1: Number of obs = 210

------------------------------------------------------------------------------
Variable | Mean Std. Err. [95% Conf. Interval]
-------------+----------------------------------------------------------------
diabetes1 | .2142857 .0283152 .158789 .2697824
------------------------------------------------------------------------------
p = proportion(diabetes1) z = 0.8971
Ho: p = 0.19

Ha: p < 0.19 Ha: p != 0.19 Ha: p > 0.19

Pr(Z < z) = 0.8152 Pr(|Z| > |z|) = 0.3697 Pr(Z > z) = 0.1848

The one-sample test of proportion is analogous to the one-sample t-test. The one-sam-
ple proportion test is used to compare the observed proportion with a hypothetical
value. To do the one-sample test of proportion, use the following command:
prtest diabetes1==.19
The results are shown in Table 14.2. The hypothesis that we have tested is a two-tailed
hypothesis. The p-value of the two-tailed test is provided under "Ha: p != 0.19", which
is 0.369 [Pr(|Z| > |z|) = 0.3697]. Since the p-value is >0.05, we cannot reject the null
 Hypothesis Test of Proportions 133

hypothesis. It can, therefore, be concluded that the prevalence of diabetes in the district
may not be different from the national prevalence of 19.0%.

14.2 Two-sample test of proportions

In section 14.1, we analyzed the variable "diabetes1" and found that the overall preva-
lence of diabetes is 21.43%. There is another variable in the dataset named "sex_1",
which is coded as "0" for females and "1" for males. For the analysis of two-sample
proportions, the categorical variable must be a numeric variable. If it is a string
variable, the command will not be executed. From the data, we can calculate the preva-
lence of diabetes among males and females by generating a cross-table. To generate a
cross-table of sex (variable name: sex_1) and diabetes (variable name: diabetes1) with
row percentages, use the following command:
tab sex_1 diabetes1, row
The above command will generate Table 14.3. The table shows that the prevalence of
diabetes among females and males is 15.04% and 32.47%, respectively.

Table 14.3 Cross-tabulation of sex and diabetes

. tab sex_1 diabetes1, row

+----------------+
| Key |
|----------------|
| frequency |
| row percentage |
+----------------+

Sex: | have diabetes 01

numeric | no yes | Total
-----------+----------------------+----------
Female | 113 20 | 133
| 84.96 15.04 | 100.00
-----------+----------------------+----------
Male | 52 25 | 77
| 67.53 32.47 | 100.00
-----------+----------------------+----------
Total | 165 45 | 210
| 78.57 21.43 | 100.00

Suppose that we are interested in examining whether the same proportion of males and
females has diabetes in the population, i.e., whether the prevalence of diabetes is the
134 Hypothesis Test of Proportions

same among females and males. This is a situation where we can apply the two-sample
test of proportions. For this example, the null hypothesis is "the proportion of males
who have diabetes is equal to the proportion of females who have diabetes in the popu-
lation". The alternative hypothesis is that the two proportions are not the same (differ-
ent) in the population.
To test the null hypothesis that the two proportions are the same in the population, use
the following command:
prtest diabetes1, by (sex_1)

Table 14.4 Two-sample test of proportions

. prtest diabetes1, by (sex_1)

Two-sample test of proportions Female: Number of obs = 133

Male: Number of obs = 77
------------------------------------------------------------------------------
Variable | Mean Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
Female | .1503759 .0309939 .0896289 .2111229
Male | .3246753 .0533624 .2200869 .4292638
-------------+----------------------------------------------------------------
diff | -.1742994 .0617104 -.2952495 -.0533492
| under Ho: .0587581 -2.97 0.003
------------------------------------------------------------------------------
diff = prop(Female) - prop(Male) z = -2.9664
Ho: diff = 0

Ha: diff < 0 Ha: diff != 0 Ha: diff > 0

Pr(Z < z) = 0.0015 Pr(|Z| < |z|) = 0.0030 Pr(Z > z) = 0.9985

The results are shown in Table 14.4. The table shows that the difference between the
two proportions (female to male) is -.174 (-17.4%). This means that the prevalence of
diabetes among females is 17.4% less than that of males (15.03% vs. 32.46%). The
95% CI of the difference is also given in the table, which is -.295 (29.5%) to -.053
(5.3%) ["diff" row in the table]. Our interest is in the two-sided p-value of the test,
which is 0.003 (Pr(|Z| < |z|) = 0.0030). Since the p-value is <0.05, we will reject the null
hypothesis and conclude that the proportion of males who have diabetes is different
from that of females (i.e., the prevalence of diabetes among males is significantly high-
er than that of females). The chi-square test of independence also tests the same
hypothesis.
 135

15

Association Between Two Continuous

Variables: Correlation

The nature and strength of the relationships between two or more continuous variables
can be examined by regression and correlation analysis. Correlation is concerned with
measuring the strength of a relationship between continuous variables. The correlation
model provides information about the relationship between two variables without
distinguishing which one is the dependent and which one is the independent variable.
But the basic procedure for regression and correlation models is the same.
Under the correlation model, we calculate the "r" value. The "r" is called the sample
correlation coefficient. It indicates the degree of linear relationship between the depen-
dent (Y) and independent (X) variables. The value of "r" ranges between “–1” and
“+1”. This chapter will cover the correlation model. Use the data file <Data_3.dta> for
practice.

15.1 Pearson’s correlation

Pearson’s correlation is used when the normality assumptions are met, i.e., when both
the variables involved in the correlation analysis are normally distributed. Suppose that
we want to explore if there is a correlation between systolic blood pressure (BP) (vari-
able name is "sbp") and diastolic BP (variable name is "dbp").

Hypothesis
H0: There is no correlation between systolic and diastolic BP.
136 Association Between Two Continuous Variables: Correlation

HA: There is a correlation between systolic and diastolic BP.

Assumptions
1. The variables (systolic and diastolic BP) are normally distributed in the popu-
lation;
2. The subjects represent a random sample from the population.

15.1.1 Scatter plot

The first step, before carrying out the correlation analysis, is to generate a scatter plot.
The scatter plot provides information about:
• Whether there is a correlation (relationship) between two variables;
• Whether the relationship (if there is any) is linear or non-linear; and
• Direction of the relationship, i.e., whether the relationship is positive (if the
value of one variable increases with the increase of the other variable) or nega-
tive (if the value of one variable decreases with the increase of the other
variable).
To generate a scatter plot of systolic and diastolic BP, use the following commands
(also see Section 7.2):
scatter dbp sbp
scatter dbp sbp || lfit dbp sbp
The first command will display a simple scatter plot of systolic and diastolic BP (Fig
15.1), while the second command will display the regression line (fit line) on the
scatter plot (Fig 15.2).

15.1.2 Commands for Pearson’s correlation

To determine the Pearson’s correlation coefficient between two continuous quantita-
tive variables, we use the command "correlate" or simply "corr" or "pwcorr". The com-
mands "correlate" or "corr" provide correlations of the listed variables based on the
non-missing observations. On the other hand, the command "pwcorr (pairwise correla-
tion)" reports on all the observations available for each variable pair.
corr sbp dbp
corr sbp dbp age income
pwcorr sbp dbp, sig
pwcorr sbp dbp age income, sig
 Association Between Two Continuous Variables: Correlation 137

120
120

100
100

Diastolic BP
Diastolic BP

80
80

60
100 120 140 160 180 200
60

Systolic BP
100 120 140 160 180 200 Diastolic BP Fitted values
Systolic BP

Figure 15.1 Scatter plot of systolic and Figure 15.2 Scatter plot of systolic and
diastolic BP diastolic BP with the regression line
120
100
Diastolic BP

80
60

10 20 30 40 50
Age

Diastolic BP Fitted values

Figure 15.3 Scatter diagram of diastolic BP and age

The first command will provide the Pearson’s correlation coefficient (r value) of
systolic and diastolic BP without the p-value (Table 15.1). The second command will
generate the correlation matrix of the variables (sbp, dbp, age, and income) included in
the command (Table 15.2). If you want to obtain the correlation coefficient of systolic
and diastolic BP along with the p-value, use the third command (Table 15.3). The
fourth command will provide a correlation matrix table for the variables included in the
command with the corresponding p-values (Table 15.4). The commands for the
correlation matrix are provided as an example.
138 Association Between Two Continuous Variables: Correlation

Table 15.1 Pearson’s correlation between sbp and dbp

. corr sbp dbp
(obs=210)
| sbp dbp
-------------+------------------
sbp | 1.0000
dbp | 0.8468 1.0000

15.1.3 Interpretation
In the first step, we have constructed the scatter plots of systolic and diastolic BP (Figs
15.1 and 15.2). Figure 15.1 shows that the data points are scattered around an invisible
straight line (indicating linear relationship) and there is an increase in the diastolic BP
(Y) as the systolic BP increases (X). This indicates that there may have a positive
correlation between these two variables.

Table 15.2 Correlation matrix of systolic BP, diastolic BP, age and income
. corr sbp dbp age income
(obs=210)
| sbp dbp age income
-------------+------------------------------------
sbp | 1.0000
dbp | 0.8468 1.0000
age | -0.0378 -0.0224 1.0000
income | 0.0163 0.0697 -0.1324 1.0000

Table 15.3 Pearson’s correlation between systolic and diastolic BP with the p-value
. pwcorr sbp dbp, sig

| sbp dbp
-------------+------------------
sbp | 1.0000
|
dbp | 0.8468 1.0000
| 0.0000

Look at Fig 15.2, which shows the regression line on the scatter plot. The regression
line has passed from near the lower left corner to the upper right corner, indicating a
positive correlation between systolic and diastolic BP. If the relationship was negative
(inverse), the regression line would have passed from the upper left corner to the lower
right corner.
 Association Between Two Continuous Variables: Correlation 139

Figure 15.3 shows the scatter plot of diastolic BP and age. It does not indicate any
correlation between diastolic BP and age since the dots are scattered around the regres-
sion line, which is more or less parallel to the X-axis.

Table 15.4 Correlation matrix with the corresponding p-values

. pwcorr sbp dbp age income, sig

| sbp dbp age income

-------------+------------------------------------
sbp | 1.0000
|
dbp | 0.8468 1.0000
| 0.0000
|
age | -0.0378 -0.0224 1.0000
| 0.5856 0.7465
|
income | 0.0163 0.0697 -0.1324 1.0000
| 0.8144 0.3146 0.0554

Table 15.1 shows the Pearson’s correlation coefficient (r value) of systolic and diastol-
ic BP, which is 0.846. Table 15.3 shows the same result, but with the p-value, which is
0.000. The correlation coefficient [r] indicates the strength or degree of the linear
relationship between two variables (systolic and diastolic BP). As the value of "r" is
positive and the p-value is <0.05, we can conclude that there is a significant positive
correlation between systolic and diastolic BP.
The value of “r” lies between “–1” and “+1”. Values near to “zero” indicate no correla-
tion, while values near to “+1” or “–1” indicate a strong correlation. The negative value
of “r” (– r) indicates an inverse relationship. A value of r ≥ 0.8 indicates a very strong
correlation; an “r” value between 0.6 and <0.8 indicates a moderately strong correla-
tion; an “r” value between 0.3 and <0.6 indicates a fair correlation; and an “r” value of
<0.3 indicates a poor correlation [8].

15.2 Spearman and Kendall’s tau-b correlations

The Spearman and Kendall’s tau-b are the nonparametric methods of obtaining the
correlation coefficients. The Spearman correlation is performed when the normality
assumption is violated (i.e., if the distribution of either the dependent or independent,
or both the variables, is not normally distributed). Spearman correlation is also applica-
ble for two categorical ordinal variables if they have ≥5 levels, such as intensity of pain
140 Association Between Two Continuous Variables: Correlation

(no, mild, moderate, severe, and very severe pain) and grade of cancer (stage 1, stage
2, stage 3, stage 4, and stage 5).
To obtain the Spearman correlation coefficient of systolic BP (variable name “sbp”)
and income, where income is not normally distributed, use the following command:
spearman sbp income
This command will report the Spearman correlation coefficient of systolic BP and
income along with the p-value as shown in Table 15.5.
The Kendall’s tau-b statistic is used to determine the correlation between two ordinal
variables, or an ordinal and a continuous variable (provided the ordinal variable has
less than 5 levels). To determine the correlation between age group (variable name is
age2) and systolic BP, use the following command (Table 15.6):
ktau age2 sbp

Table 15.5 Spearman correlation between systolic BP and income

. spearman sbp income

Number of obs = 210

Spearman's rho = 0.0070

Test of Ho: sbp and income are independent

Prob > |t| = 0.9192

15.2.1 Interpretation
Table 15.5 shows the number of pairwise observations (n=210) used to calculate the
Spearman correlation coefficient. Spearman’s rho indicates the Spearman correlation
coefficient. In our example, Spearman’s rho is 0.007, which is very small. The p-value
of this test is indicated by “Prob > |t|”. The p-value of this test is 0.9192, which is
>0.05. We cannot, therefore, reject the null hypothesis. This indicates that there is no
statistically significant correlation between systolic BP and income.
Table 15.6 shows the results of the Kendall’s tau correlation between age group and
systolic BP. The correlation coefficient (Kendall’s tau-b) of the variables is 0.0114 and
the corresponding p-value is 0.83. This indicates that there is no significant correlation
between age group and systolic BP.
 Association Between Two Continuous Variables: Correlation 141

Table 15.6 Kendall’s tau-b correlation between age group and systolic BP
. ktau age2 sbp

Number of obs = 210

Kendall's tau-a = 0.0092
Kendall's tau-b = 0.0114
Kendall's score = 201
SE of score = 946.641 (corrected for ties)

Test of Ho: age2 and sbp are independent

Prob > |z| = 0.8327 (continuity corrected)

15.3 Partial correlation

The purpose of performing partial correlation is to determine the correlation between
two variables after controlling for one or more other variables (continuous or categori-
cal). This means that, through partial correlation, we get the adjusted "r" value after
controlling for other variables included in the analysis.
For example, if we suspect that the relationship between diastolic and systolic BP may
be influenced (confounded) by other variables (such as age and diabetes), we should
use the partial correlation to exclude the influences of other variables (age and diabe-
tes). The partial correlation gives us the adjusted correlation coefficient (r value). If we
want to get the correlation coefficient of diastolic and systolic BP after adjusting for
age and diabetes, use the following command:
pcorr dbp sbp age diabetes
The command above will provide the correlation coefficient of diastolic and systolic
BP after controlling for age and diabetes, including the p-value as shown in Table 15.7.

15.3.1 Interpretation
The results of the partial correlation of diastolic and systolic BP after adjusting for age
and diabetes mellitus are displayed in Table 15.7. We can observe that the partial
correlation coefficient of diastolic and systolic BP is 0.847 and the p-value is 0.000.
This means that these two variables (diastolic and systolic BP) are significantly
correlated (p=0.000) even after controlling for age and diabetes mellitus. The table also
provides the results of semipartial correlation. In semipartial correlation, the correla-
tion coefficient is calculated holding the other variables (age and diabetes, in our
example) constant either for X or Y, but not for both. In partial correlation, the other
142 Association Between Two Continuous Variables: Correlation

Table 15.7 Correlation between systolic and diastolic BP after controlling for age
and diabetes
. pcorr dbp sbp age diabetes
(obs=210)

Partial and semipartial correlations of dbp with

Partial Semipartial Partial Semipartial Significance

Variable | Corr. Corr. Corr.^2 Corr.^2 Value
------------+-----------------------------------------------------------------
sbp | 0.8470 0.8468 0.7175 0.7171 0.0000
age | 0.0220 0.0117 0.0005 0.0001 0.7524
diabetes | 0.0409 0.0218 0.0017 0.0005 0.5574

variables are held constant for both X and Y. We should consider the partial correlation
for reporting.
The partial correlation provided for the age indicates the partial correlation coefficient
of age and diastolic BP (0.022; p=0.752) after controlling for systolic BP and diabetes.
 16
Linear Regression

Regression analysis is a useful statistical method of data analysis in health and social
sciences disciplines. The nature and strength of relationships between two or more
continuous variables can be ascertained by regression and correlation analyses.
We have discussed correlation in Chapter 15. While correlation is concerned with
measuring the strength of the linear relationship between variables, regression analysis
is useful in predicting or estimating the value of one variable corresponding to a given
value of another variable. For instance, we can use regression analysis to examine
whether systolic BP is a good predictor of diastolic BP and also to get the estimated
(predicted) value of diastolic BP corresponding to a value of systolic BP. In regression
analysis, our main interest is in regression coefficient (also called slope or β). The
regression coefficient indicates the strength of association between dependent (Y) and
independent (X) variables. Linear regression analyses can be done either as simple
linear regression or multiple linear regression methods.
In this chapter, both simple and multiple linear regression methods are discussed.
Multiple linear regression is a type of multivariable analysis. The multivariable analy-
sis is a statistical tool where multiple independent variables are considered for a single
outcome variable. The terms "multivariate analysis" and "multivariable analysis" are
often used interchangeably in health and social sciences research. In fact, multivariate
analysis refers to a statistical method for the analysis of multiple outcome variables.
Multivariable analyses are widely used in observational studies, intervention studies
(randomized or nonrandomized trials), and studies of diagnosis and prognosis. The
main purposes of multivariable analyses are to:
144 Linear Regression

• Adjust for the confounding factors;

• Predict the probability of an outcome when several characteristics are present
in an individual;
• Determine the relative contribution of independent variables to the outcome
variable; and
• Assess the interaction of multiple variables for the outcome.
There are several types of multivariable analysis methods. The choice of multivariable
analysis for the type of outcome variable is summarized in Table 9.3 (Chapter 9). The
commonly used multivariable analysis methods in health and social sciences research
include multiple linear regression, logistic regression, and proportional hazards regres-
sion (Cox regression), which are discussed in this book. Use the data file
<Data_4.dta> for practice.

16.1 Simple linear regression

In simple linear regression, there is one dependent (outcome) and one independent
(explanatory or predictor) variable. The objective of simple linear regression is to find
the population regression equation, which describes the true relationship between a
dependent variable (Y) and an independent variable (X). In a simple linear regression
model, two variables are involved – one is an independent variable (X) placed on the
X-axis, and the other is a dependent variable (Y) placed on the Y-axis. Then, we call it
"regression of Y on X".
Suppose that we want to perform a simple linear regression analysis of diastolic BP
(dependent variable) on systolic BP (independent variable). The objective is to find the
population regression equation to predict diastolic BP by systolic BP.

Assumptions
1. Normality: For any fixed value of X (systolic BP), the sub-population of Y
values (diastolic BP) is normally distributed;
2. Homoscedasticity: The variances of the sub-populations of “Y” are all equal;
3. Linearity: The means of the sub-populations of “Y” lie on the same straight
line; and
4. Independence: Observations are independent of each other.
The first step in analyzing the data for regression is to construct a scatter diagram to
 Linear Regression 145

visualize the relationship between the two variables, which is discussed in Chapter 15.
The scatterplot will provide an indication of the linear relationship between the
variables, diastolic and systolic BP. For example, to get the scatter plot of diastolic and
systolic BP with the fit-line, use the following command:
graph twoway lfit dbp sbp || scatter dbp sbp

16.1.1 Commands for simple linear regression

The command for linear regression analysis is “regress”. To do the regression of
diastolic BP (variable name is “dbp”) on systolic BP (variable name is “sbp”), use the
first of the following commands (Table 16.1):
regress dbp sbp
regress dbp sbp, vce(robust)
The first variable immediately after the command “regress” is the dependent variable.
The second command will provide robust estimates of the standard error of the regres-
sion coefficient (Table 16.1). Robust regression is an alternative to ordinary regression
(least squares regression; first command). It provides better estimates of regression
coefficients, especially when outliers are present in the data.
You can conduct the regression analysis on a subset of your data. For example, if you
want to perform the regression analysis only for the females (variable name is “sex”
and females are coded as 0), use the following command (Table 16.2):
regress dbp sbp if sex==0
You can get the predicted values of diastolic BP (outcome variable) of the individuals
after performing the regression analysis. Use the following command to get the
predicted values:
predict predbp
The above command will generate a new variable "predbp" containing the predicted
values of diastolic BP for all the individuals in the dataset. You can see the new
variable at the bottom of the variables window. You can also calculate the estimated
(predicted) diastolic BP of an individual whose systolic BP is 110 mmHg by using the
following command (outputs not shown):
margins, at(sbp=(110))
146 Linear Regression

16.1.2 Interpretation
The results of simple linear regression analysis are provided in Table 16.1. The table
shows the coefficient of determination (R-squared) of diastolic BP on systolic BP. The
coefficient of determination is the square of the correlation coefficient value (r value).
The table shows that the coefficient of determination (R-squared) of diastolic BP on
systolic BP is 0.717 and the p-value (Prob > F) is 0.000. The table also shows the value
for the adjusted coefficient of determination (Adj R-squared).

Table 16.1 Regression of diastolic BP on systolic BP

. regress dbp sbp

Source | SS df MS Number of obs = 210

-------------+------------------------------ F( 1, 208) = 527.20
Model | 20688.9902 1 20688.9902 Prob > F = 0.0000
Residual | 8162.57647 208 39.2431561 R-squared = 0.7171
-------------+------------------------------ Adj R-squared = 0.7157
Total | 28851.5667 209 138.045774 Root MSE = 6.2644

------------------------------------------------------------------------------
dbp | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
sbp | .4960326 .0216034 22.96 0.000 .4534429 .5386223
_cons | 19.40677 2.793132 6.95 0.000 13.90029 24.91325
------------------------------------------------------------------------------

. regress dbp sbp, vce(robust)

Linear regression Number of obs = 210

F( 1, 208) = 440.31
Prob > F = 0.0000
R-squared = 0.7171
Root MSE = 6.2644
------------------------------------------------------------------------------
| Robust
dbp | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
sbp | .4960326 .0236391 20.98 0.000 .4494297 .5426355
_cons | 19.40677 3.01498 6.44 0.000 13.46293 25.35061
------------------------------------------------------------------------------

The coefficient of determination, or the R-squared value, indicates the amount of varia-
tion in "Y" due to "X" that can be explained by the regression line. Here, the R-squared
value is 0.717 (~0.72), which indicates that 72% of the variation in diastolic BP can be
explained by systolic BP. The rest of the variation (28%) is due to other factors (unex-
plained variation). The adjusted R-squared value (0.715), as shown in the table, is the
adjusted value for better population estimation. The significance of the R-squared
 Linear Regression 147

value is assessed by the F-test [F (1, 208) = 527.20] as shown in the table. Since the
p-value (Prob > F) of the coefficient of determination is less than 0.05, it is statistically
significant. This finding indicates that the linear regression model is useful in predict-
ing the dependent variable (diastolic BP) by the independent variable (systolic BP) in
the model.

Table 16.2 Regression of diastolic BP on systolic BP among females

. regress dbp sbp if sex==0

Source | SS df MS Number of obs = 133

-------------+------------------------------ F( 1, 131) = 404.31
Model | 16168.3835 1 16168.3835 Prob > F = 0.0000
Residual | 5238.71425 131 39.9901851 R-squared = 0.7553
-------------+------------------------------ Adj R-squared = 0.7534
Total | 21407.0977 132 162.174983 Root MSE = 6.3238

------------------------------------------------------------------------------
dbp | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
sbp | .5177267 .025748 20.11 0.000 .466791 .5686625
_cons | 17.38358 3.380971 5.14 0.000 10.69521 24.07194
------------------------------------------------------------------------------

We can conclude from the data that there is a significant positive correlation between
diastolic and systolic BP since the regression coefficient (0.496; “Coef.” in Table 16.1)
is positive and statistically significant (p= 0.000)], and we can, therefore, use the
regression equation for prediction. Table 16.1 shows the regression (also called
explained) sum of squares (20688.99; in the row "Model" and column "SS") and the
residual (also called error) sum of squares (8162.57; in the row "Residual" and column
"SS"). The residual is the difference between the observed value and the predicted
value (i.e., the observed value and the value on the regression line). The residual sum
of squares provides an idea of how well the regression line actually fits into the data.
The smaller the value, the better the fit.
The strength of the relationship between variables is measured by the regression coef-
ficient (β or b) or slope. Table 16.1 shows the regression coefficient (row "sbp" and
column "Coef.”) of diastolic BP on systolic BP, which is 0.496 with a p-value of 0.000
(P> | t |). The table also shows the value for "a" (_cons) or Y-intercept, which is 19.40.
These values (regression coefficient and constant) are needed to construct the linear
regression equation.
148 Linear Regression

The value of “b” (regression coefficient) indicates the amount of change in “Y” for
each unit change in “X”. In our example, the value of “b” is 0.496. It indicates that if
the systolic BP increases (or decreases) by 1 mmHg, the diastolic BP will increase (or
decrease) by 0.496 mmHg. The table shows the significance (p-value) of “b”, which is
0.000. A p-value of <0.05 indicates that “b” is not equal to zero in the population (the
null hypothesis is that “b” is equal to zero in the population). For simple linear regres-
sion, if R-square is significant, “b” will also be significant.
On the other hand, the value of "a" (constant or Y-intercept) in our example is 19.407.
The value, a= +19.407, indicates that the regression line crosses or cuts the Y-axis
above the origin (zero) and at the point of 19.407 (a negative value indicates the regres-
sion line cuts the Y-axis below the origin). The value of "a" does not have any practical
meaning since it indicates the average diastolic BP of individuals if the systolic BP is
zero.
We know that the equation for simple linear regression is Y = a + bX, where "Y" is the
predicted value of the dependent variable; "a" is the Y-intercept or constant; "b" is the
regression coefficient or slope; and "X" is a value of the independent variable. There-
fore, the regression or prediction equation for this regression model is:

With this equation, we can estimate the diastolic BP of an individual by his/her systolic
BP. For example, what will be the estimated diastolic BP of an individual whose systol-
ic BP is 130 mmHg? Using the above equation, the answer is that the estimated diastol-
ic BP will be equal to 83.89 mmHg (19.407 + 0.496 × 130). Stata can calculate this for
you if you use the following command after performing the regression analysis:
margins, at(sbp=(130))
If we want to use the regression equation for the purpose of prediction, “b” needs to be
statistically significant (p<0.05). In our example, the p-value for “b” is 0.000. We can,
therefore, use the equation for the prediction of diastolic BP by systolic BP.
The analysis (Table 16.1) has actually evaluated whether "b" is zero or not in the popu-
lation by the t-test (Null hypothesis: the regression coefficient (b) is equal to "zero" in
the population; Alternative hypothesis: the population regression coefficient is not
equal to "zero"). We will reject the null hypothesis since the p-value is 0.000 (<0.05).
We can, therefore, conclude that the systolic BP can be considered in estimating the
diastolic BP by using the following regression equation:

Y = 19.407 + 0.496 × X.
 Linear Regression 149

16.2 Multiple linear regression

In simple linear regression, two variables are involved—one dependent (Y) and one
independent (X) variable. The independent variable is also called the explanatory or
predictor variable. In multiple linear regression, there is more than one explanatory
(independent) variable in the model. The explanatory variables may be quantitative or
categorical variables. The main purposes of multiple regression analysis are to:
• Obtain the adjusted estimates of the regression coefficients of the explanatory
variables in the model;
• Predict or estimate the value of the dependent variable by the explanatory vari
ables in the model; and
• Understand the amount of variation in the dependent variable explained by the
explanatory variables together in the model.
Suppose that we want to assess the contribution of four variables (age, systolic BP, sex,
and religion) in estimating (or predicting) the diastolic BP in a sample of individuals
selected randomly from a population. Here, the dependent variable is diastolic BP, and
the explanatory variables (independent variables) are age, systolic BP, sex, and
religion. Of the explanatory variables, two are quantitative (age and systolic BP) and
two are categorical variables (sex and religion). Of the categorical variables, sex has
two levels (0= female and 1= male) and religion has three levels (1= Muslim, 2=
Hindu, and 3= Christian). Regression does not allow string variables in the analysis. If
you want to include any string variable in the model, you need to convert the variable
into a numeric variable. How to convert a string variable into a numerical variable is
discussed in Chapter 5 (Section 5.1).
When an independent variable is a categorical variable with more than two levels (like
religion), we cannot simply include it in regression analysis because the code numbers
are arbitrary (Stata will consider it as a quantitative variable) and the regression
estimates will be misleading. We need to create dummy variables for such categorical
variables before including them in the analysis. The dummy variables are the dichoto-
mous indicator variables (coded as 0/1) representing the categories of a categorical
variable. The number of dummy variables generated for a categorical variable is equal
to the number of levels minus one. For example, if we want to include the variable
"religion" in the analysis, we need to generate two dummy variables since it has three
levels (Muslim, Hindu, and Christian).
We can generate the dummy variables in Stata as described in Section 5.11. Stata can
150 Linear Regression

can also generate the dummy variables automatically during regression analysis if the
prefix "i." is used before a variable that has more than two levels (in general, if the
prefix "i." is used for any variable, Stata considers it a categorical variable during
analysis).
For example, if we include religion in the regression analysis as "i.religion", Stata will
automatically generate two dummy variables for religion during the analysis, with the
first category (Muslim) as the comparison group by default. You can change the com-
parison group by using the prefix “.ib”. For example, if you want the second category
of religion (Hindu) as the comparison group, use the prefix “.ib2” (since Hindus are
coded as 2).
We always need to decide on a comparison group (e.g., a comparison group for religion
or other variables) before generating the dummy variables or entering a categorical
variable in the model with the prefix “.i”. Stata, by default, considers the first category
(e.g., Muslims for the variable “religion” since Muslims are in the first category) of a
variable as the comparison group if the variable is entered with the prefix “i.” (e.g.,
i.religion). If we want to consider the other category (e.g., the third category or Chris-
tians) as the comparison group, we should use the prefix “.ib3” since Christians are
coded as 3.
In our regression analysis, we will use the variable “religion” with Christians (coded as
3) as our comparison group by using the prefix “.ib3”. When a variable is coded as 0/1
(e.g., the variables “sex” and “diabetes”), the regression estimates in multiple regres-
sion analysis will be for the higher value, and the lower value will be the comparison
group.

16.2.1 Sample size for multiple regression

Multiple regression analysis should be done if the sample size is fairly large. The mini-
mum sample size needed for the analysis depends on how many independent variables
we want to include in the model. Different authors provided different guidelines. One
author recommends a minimum of 15 subjects for each of the independent variables in
the model. Other authors provided a formula (n= 50 + 8m) to estimate the number of
subjects required for the analysis. For example, if we intend to include five indepen-
dent variables in the model, we need to have at least 90 subjects (50 + 8×5). For
stepwise regression, there should be 40 cases for each of the independent variables in
the model.
 Linear Regression 151

16.2.2 Commands for multiple linear regression analysis

The basic command for regression analysis is “regress”. Use the following commands
for the multiple regression analysis where the dependent variable is diastolic BP (vari-
able name is “dbp”) and the explanatory (independent) variables are age, systolic BP
(variable name is “sbp”), sex, and religion. We will use the prefix “.ib3” before the
variable “religion” so that Stata automatically generates the dummy variables and
considers the third category (Christians) as the comparison group during analysis. We
will also use the prefix “i.” for sex to indicate it as a categorical variable. Note that the
variable immediately after the command “regress” is the dependent variable.
regress dbp age sbp i.sex ib3.religion
regress dbp age sbp i.sex i.religion
The first command will provide the outputs with Christians as the comparison group
(Table 16.3), while the second command will provide the outputs where Muslims are
the comparison group. If you want the second category (Hindus) to be the comparison
group, use the prefix “ib2” for religion.
You can get the standardized coefficients (beta) by using the following command:
regress dbp age sbp i.sex ib3.religion, beta
The outputs of the above command are provided at the bottom of Table 16.3. The
standardized coefficients are used to understand the relative influence of the indepen-
dent variables on the dependent variable. The higher the value, the greater the influ-
ence.
You can introduce interaction terms into the model. Suppose that you want to check if
there is an interaction between sex and religion. For this, use the first command:
regress dbp age sbp i.sex i.religion i.sex#i.religion
regress dbp age sbp i.sex i.religion i.sex#c.sbp
The second command will demonstrate the interaction between sex and systolic BP.
The prefix “c.” used for “spb” is to indicate that it is a continuous variable.
You can use the “margins” command, after the regression analysis, to get the predict-
ed values of dependent variable for the independent variables in the model (also see
Section 16.1.1). For instance, to get the predicted values of diastolic BP (i.e., adjusted
mean of diastolic BP) for religion, sex, and religion#sex, use the following commands
(results not shown):
152 Linear Regression

Table 16.3 Results of the multiple regression analysis

. regress dbp age sbp i.sex ib3.religion

Source | SS df MS Number of obs = 210

-------------+------------------------------ F( 5, 204) = 107.89
Model | 20934.8733 5 4186.97466 Prob > F = 0.0000
Residual | 7916.69336 204 38.8073204 R-squared = 0.7256
-------------+------------------------------ Adj R-squared = 0.7189
Total | 28851.5667 209 138.045774 Root MSE = 6.2296

------------------------------------------------------------------------------
dbp | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
age | -.0283828 .0562385 -0.50 0.614 -.139266 .0825004
sbp | .4892801 .0216686 22.58 0.000 .446557 .5320032
|
sex |
Male | -2.164135 .9128191 -2.37 0.019 -3.963905 -.3643654
|
religion |
MUSLIM | .212462 1.363372 0.16 0.876 -2.475645 2.900569
HINDU | -.229616 1.488835 -0.15 0.878 -3.165094 2.705862
|
_cons | 21.82239 3.426989 6.37 0.000 15.06553 28.57925
------------------------------------------------------------------------------

. regress dbp age sbp i.sex ib3.religion, beta

Source | SS df MS Number of obs = 210

-------------+------------------------------ F( 5, 204) = 107.89
Model | 20934.8733 5 4186.97466 Prob > F = 0.0000
Residual | 7916.69336 204 38.8073204 R-squared = 0.7256
-------------+------------------------------ Adj R-squared = 0.7189
Total | 28851.5667 209 138.045774 Root MSE = 6.2296

------------------------------------------------------------------------------
dbp | Coef. Std. Err. t P>|t| Beta
-------------+----------------------------------------------------------------
age | -.0283828 .0562385 -0.50 0.614 -.0185771
sbp | .4892801 .0216686 22.58 0.000 .8352803
|
sex |
male | -2.164135 .9128191 -2.37 0.019 -.0889736
|
religion |
MUSLIM | .212462 1.363372 0.16 0.876 .00888
HINDU | -.229616 1.488835 -0.15 0.878 -.0087588
|
_cons | 21.82239 3.426989 6.37 0.000 .
------------------------------------------------------------------------------

margins religion sex, atmeans

margins religion#sex, atmeans
margins sex, at(religion=(1 2 3)) atmeans
 Linear Regression 153

The first command will provide the average diastolic BP for different categories of
religion and sex, separately. The second and third command will provide the average
diastolic BP for religion and sex together (e.g., Muslim males, Muslim females, Hindu
males, Hindu females, Christian males, and Christian females).

16.2.3 Interpretation
In the analysis, we used diastolic BP (dbp) as the dependent variable and age, systolic
BP (sbp), sex, and religion as the explanatory variables.
Table 16.3 shows the outputs of the multiple regression analysis. The table shows that
data from 210 subjects were analyzed. It shows the values for R-squared (0.725) and
adjusted R-squared (0.718), including the p-value (Prob > F; 0.000).
The R-squared value of 0.725 indicates that all the independent variables (age, systolic
BP, sex, and religion) together in the model explain 72.5% of the variation in diastolic
BP, which is statistically significant (p= 0.000). If the sample size is small, the
R-squared value may overestimate the population value. The adjusted R-squared
(0.718) gives the R-squared value for better population estimation.
Table 16.3 also shows the regression coefficients (Coef.), p-values (P > | t |) and 95%
confidence intervals (95% Conf. Interval) for all the explanatory variables in the mod-
el, along with the constant (_cons). The regression coefficients as shown in the table
are for age (-0.028; p= 0.614), systolic BP (0.489; p<0.001), sex (-2.164; p= 0.019 for
males compared to females), Muslims (0.212; p= 0.876 compared to Christians) and
Hindus (-0.229; p= 0.878 compared to Christians).
From the analysis, we can conclude that the systolic BP and sex are the factors signifi-
cantly influencing the diastolic BP (since the p-values are <0.05). The other variables
in the model (age and religion) do not have any significant influence in explaining (or
predicting) the diastolic BP. The regression coefficient (Coef.) [also called multiple
regression coefficient] for systolic BP, in this example, is 0.489 (95% CI: 0.44 to
0.53; p<0.001). This indicates that the average increase (or decrease) in diastolic BP
is 0.489 mmHg if the systolic BP increases (or decreases) by 1 mmHg after adjusting
for all other variables (age, sex, and religion) in the model. On the other hand, the
regression coefficient for sex is –2.164 (95% CI: -3.96 to -0.36; p= 0.019), which
means that the average diastolic BP of males is 2.16 mmHg less (since the coefficient
is negative) than that of females after adjusting for all other variables (age, systolic BP,
and religion) in the model. If the regression coefficient was positive (e.g., +2.164), the
154 Linear Regression

average diastolic BP of males would be 2.16 mmHg higher than that of females, given
the other variables constant in the model.
Table 16.3 (at the bottom) shows the standardized coefficients (beta). The standardized
coefficients are used to understand the magnitude of the influence of independent
variables on the dependent variable. The higher the value, the greater the influence.
The table shows that the beta for systolic BP and sex are 0.835 and -0.088, respectively.
Therefore, systolic BP has the highest influence (also greater than sex) in predicting
diastolic BP.
Regression equation
The regression equation to estimate the average value of the dependent variable with
the explanatory variables is given below:

Y= a + B1X1 + B2X2 + B3X3 + B4X4 …….. BnXn

Here, "Y" represents the estimated mean value (predicted value) of the dependent
variable; "a" represents the constant (or Y-intercept); "B" represents the regression
coefficient(s) of the variables in the model; and "X" represents the value of the
variable(s) in the model.
Suppose that we want to estimate the diastolic BP of an individual who is 40 years old,
male, Muslim, and has a systolic BP of 120 mmHg. In Table 16.3, we can find the
regression coefficients for age [= -0.028 (B1)], systolic BP [= 0.489 (B2)], sex [= -2.164
(B3) for being male] and Muslims [= 0.212 (B4) for being Muslim] and the constant (=
21.82). Therefore, the estimated diastolic BP of the individual will be:

Y= 21.82 + (-0.028 × 40) + (0.489 × 120) + (-2.164 × 1) + (0.212 × 1) = 79.67.

16.2.4 Regression diagnostics

The regression analysis is based on certain assumptions. It is important to check the
underlying assumptions before considering whether the regression analysis is useful or
valid. In regression analysis, the assumptions are commonly checked on the residuals.
Residual is the difference between an observed value and a predicted value (value
given by the fitted line or regression line). Regression diagnostics are used to evaluate
whether the assumptions are true or not.
For practical purposes, after checking for multicollinearity of independent variables
and the assumption of linearity (the relationship between X and the mean of Y is
 Linear Regression 155

linear), we need to check the following four assumptions on residuals for a linear
regression model to be valid. The assumptions to be checked on the residuals are:
1) The residuals are normally distributed with the mean equal to zero;
2) The residuals have constant variance (homoscedasticity);
3) There is no outlier; and
4) The data points are independent.

16.2.4.1 Checking for multicollinearity

Before deciding about the multiple regression model, we need to check for multicol-
linearity (inter-correlations among the independent variables) of the independent
variables. If there are moderate to high inter-correlations among the independent
variables, two situations may occur. Firstly, the importance of a given explanatory
variable may be difficult to determine because of a biased (distorted) p-value; and
secondly, a dubious relationship may be obtained. For example, if there is multicol-
linearity among the independent variables, we may observe that the regression coeffi-
cient for sex is not significant (though it is actually significant) and that the systolic BP
has a negative relationship (though the relationship is positive) with the diastolic BP.
Another important sign of multicollinearity is a severe reduction of the adjusted R
squared value.
To determine the correlations among the independent variables, we can generate the
Pearson’s correlation matrix. For example, if we want to see the correlations among the
systolic BP, age, sex, and religion, use the following command to get the correlation
matrix (Table 16.4).
corr sbp age sex religion
Table 16.4 shows the correlation coefficients (r values) among the variables included
in the analysis. The highest correlation coefficient that we see in the table is -0.13,
which is for religion and sex. In general, if the r value is greater than 0.5, it is consid-
ered that the correlation may interfere with the regression analysis. However, in our
analysis, there is no such problem as shown in the table.
Pearson’s correlation can only check for collinearity between any two variables. Some-
times a variable may be multicollinear with a combination of variables. It is, therefore,
preferable to use the tolerance (or variance inflation factor) measure, which indicates
the strength of the linear relationships among the independent variables.
156 Linear Regression

Table 16.4 Correlation matrix of systolic BP, age, sex, and religion
. corr sbp age sex religion
(obs=210)

| sbp age sex religion

-------------+------------------------------------
sbp | 1.0000
age | -0.0281 1.0000
sex | -0.1207 0.0586 1.0000
religion | -0.0249 -0.0582 -0.1308 1.0000

To get the measures of tolerance and variance inflation factor (VIF), use the command
“vif” after performing the multiple regression analysis. A tolerance value indicates the
degree of collinearity. The tolerance value is the inverse of the VIF measure (1/VIF).
The tolerance value ranges from 0 to 1. A value of “zero” indicates that the variable is
almost in a linear combination (i.e., has a very strong correlation) with other indepen-
dent variables. To get the values for VIF and tolerance, use the following command
after performing the regression analysis:
vif
This command will provide both the VIF and Tolerance (1/VIF) values of the indepen-
dent variables included in the regression analysis (Table 16.5).
The table (Table 16.5) shows that the tolerance (1/VIF) values for sex, systolic BP, and
age are greater than 0.95. The tolerance values for both religion 1 (Muslims) and 2
(Hindus) are 0.41. Usually, the dummy variables have lower tolerance values. The
recommended tolerance level is greater than 0.6 before we include the variables in the
multiple regression model. However, a tolerance value of 0.40 and above is also
acceptable, especially if it is a dummy variable.
If there are variables that are highly correlated (tolerance value is <0.4), one way to
solve the problem is to exclude one of the correlated variables from the model. The
other way is to combine the explanatory variables together (e.g., by taking their sum).
Finally, to develop a model for multiple regression, we should first check for multicol-
linearity and then the other assumptions (see below). If the requirements are fulfilled,
then only we can finalize the regression model.

16.2.4.2 Checking for linearity

When we do a linear regression analysis, it is assumed that the relationship between the
response (dependent) variable and the predictor variables (independent variables) is
 Linear Regression 157

Table 16.5 Multicollinearity test results

vif

Variable | VIF 1/VIF

-------------+----------------------
age | 1.01 0.992729
sbp | 1.02 0.982955
1.sex | 1.05 0.955039
religion |
1 | 2.41 0.414242
2 | 2.40 0.417031
-------------+----------------------
Mean VIF | 1.58

linear. This is called the assumption of linearity. The best way to check the linearity
assumption is to construct a scatterplot and visually inspect the plot for linearity.
Checking linearity for a simple linear regression is straightforward since there is only
one predictor (independent) variable against the response (dependent or outcome)
variable. For a simple linear regression, we can check the linearity by generating a
scatterplot of the dependent variable against the independent variable, which is
discussed in Section 7.2. However, to check the linear relationship between diastolic
and systolic BP, use the following commands to get a scatterplot of diastolic BP against
systolic BP:
twoway (scatter dbp sbp)
twoway (scatter dbp sbp) (lfit dbp sbp)
twoway (scatter dbp sbp) (lfit dbp sbp) (lowess dbp sbp)
The first command will display a scatterplot of diastolic BP and systolic BP without the
regression line (fit-line). The second command will provide the scatterplot with the
fit-line (Fig 16.1), while the third command will provide all the above with a smooth
prediction line. Figure 16.1 shows that the relationship between diastolic and systolic
BP is linear since the scatter dots are lying more or less symmetrically around a straight
line.
Checking the linearity assumption in multiple linear regression is not straightforward.
It can be checked in several different ways. The most straightforward way is to draw
scatter plots of standardized residuals (z-residuals) against each of the predictor (inde-
pendent) variables included in the regression analysis. Standard residuals are the resid-
uals divided by the standard deviation of the residuals.
158 Linear Regression

120
100
Diastolic BP

80
60

100 120 140 160 180 200

Systolic BP

Diastolic BP Fitted values

Figure 16.1 Scatterplot of diastolic and systolic BP with the fit line

We have done the regression analysis of diastolic BP (dependent variable) on age,

systolic BP, sex, and religion. First, do the regression analysis, then generate the
standardized residual variable “zresid” by using the following command:
predict zresid, rstandard
This command will generate a new variable “zresid” with the z-values of the residuals.
Now, use the following commands to get the scatterplots of z-residuals against the
systolic BP (Fig 16.2) and age (Fig 16.3):
twoway (scatter zresid sbp) (lfit zresid sbp)
twoway (scatter zresid age) (lfit zresid age)
Both the plots (Figs 16.2 and 16.3) show that the scatter dots are symmetrical above
and below the straight lines, indicating that the relationships are linear. You can also
construct the same for z-residuals and religion by using the command:
twoway (scatter zresid religion) (lfit zresid religion)

16.2.4.3 Checking for normality of residuals

Normality of residuals is required in regression analysis to validate hypothesis testing
for R-squared values and regression coefficients, i.e., the normality assumption
ensures that the p-values of the F-test and t-test are valid. Normality is not required to
 Linear Regression 159

4
Standardized residuals
Standardized residuals

2
2

0
0

-2
-2

100 120 140 160 180 200 10 20 30 40 50

Systolic BP age in years

Standardized residuals Fitted values Standardized residuals Fitted values

Figure 16.2 Scatterplot of z-residuals Figure 16.3 Scatterplot of z-residuals

against systolic BP against age

obtain unbiased estimates of the regression coefficients (b values). Similarly, the

normality assumption of the independent variables is not required for multiple regres-
sion analysis.
After running the regression analysis, use the following command to generate the
residual variable:
predict residual, resid
This command will generate a new variable “residual” with the residuals in the data
file. Now, use any of the following commands to check for the normality of residuals.
kdensity residual, normal
histogram residual
pnorm residual
qnorm residual
The command “kdensity” stands for kernel density plot. Using the option “normal” will
provide the overlaid normal density curve on the plot (Fig 16.4). The command “histo-
gram” will generate a histogram of the residuals (Fig 16.5), while the commands
“pnorm” and “qnorm” will provide the P-P (Fig 16.6) and Q-Q (Fig 16.7) plots, respec-
tively. The P-P plot is sensitive to non-normality in the middle range of data, while the
Q-Q plot is sensitive to non-normality near the tails.
160 Linear Regression

You can also use formal statistical tests (Shapiro-Wilk test or Skewness-Kurtosis test)
to evaluate the normality of residuals by using any of the following commands (Table
16.6):

Kernel density estimate

.08
.04
.03

.06
Density

.02

Density

.04
.01

.02
0

-40 -20 0 20 40
Residuals

Kernel density estimate

0
Normal density
-20 -10 0 10 20
kernel = epanechnikov, bandwidth = 2.7696 Residuals

Figure 16.4 Kernel density estimate with Figure 16.5 Histogram of residuals
overlaying normal density curve
1.00

40
0.75

20
Normal F[(r-m)/s]

Residuals
0.50

0
0.25

-20
0.00

-40

0.00 0.25 0.50 0.75 1.00 -40 -20 0 20 40

Empirical P[i] = i/(N+1) Inverse Normal

Figure 16.6 P-P plot of residuals Figure 16.7 Q-Q plot of residuals

swilk residual
sktest residual
Table 16.6 shows that the p-values of both these tests are greater than 0.05, indicating
that the distribution of residuals is normal.
To get the mean of the residuals, use the following command (Table 16.7):
sum residual
All the plots (Kernel density, histogram, P-P, and Q-Q plots) indicate that the distribu-
tion of the residuals is normal (also see Chapters 5 and 8). The mean of the residuals,
 Linear Regression 161

as shown in Table 16.7, is -7.01e-09 (-7.01 x e-09), which is equal to “zero”. For practical
purposes, simply construct the histogram of the residuals and decide whether the distri-
bution is approximately normal or not.

Table 16.6 Normality test of residuals

. swilk residual

Shapiro-Wilk W test for normal data

Variable | Obs W V z Prob>z

-------------+--------------------------------------------------
residual | 210 0.99256 1.158 0.339 0.36734

. sktest residual
Skewness/Kurtosis tests for Normality
------- joint ------
Variable | Obs Pr(Skewness) Pr(Kurtosis) adj chi2(2) Prob>chi2
-------------+---------------------------------------------------------------
residual | 210 0.7485 0.9940 0.10 0.9499

16.2.4.4 Checking for homoscedasticity

To check for heteroscedasticity (i.e., the variances of the residuals are not homoge-
neous), after the regression analysis, use either of the following commands. The
outputs are shown in Table 16.8.
estat hettest
estat imtest

Table 16.7 Mean of the residuals

sum residual

Variable | Obs Mean Std. Dev. Min Max

-------------+--------------------------------------------------------
residual | 210 -7.01e-09 6.154585 -15.05615 18.24044

The first test of heteroscedasticity given by the command “hettest” is the Breusch-Pa-
gan test, and the second test given by the command “imtest” is the White’s test. Both
these tests test the null hypothesis that the variance of the residuals is homogeneous.
Therefore, to fulfill the assumption, we expect a p-value greater than 0.05. Table 16.8
shows that the p-values of both these tests are greater than 0.05 (use the p-value of
162 Linear
LinearRegression
Regression

Table 16.8 Test for heteroscedasticity

. estat hettest

Breusch-Pagan / Cook-Weisberg test for heteroskedasticity

Ho: Constant variance
Variables: fitted values of dbp

chi2(1) = 0.27
Prob > chi2 = 0.6059

. estat imtest

Cameron & Trivedi's decomposition of IM-test

---------------------------------------------------
Source | chi2 df p
---------------------+-----------------------------
Heteroskedasticity | 18.98 16 0.2697
Skewness | 4.25 5 0.5144
Kurtosis | 0.02 1 0.8930
---------------------+-----------------------------
Total | 23.24 22 0.3880
---------------------------------------------------

either test for interpretation). This indicates that the variances of the residuals are
homogeneous.
However, these tests are very sensitive to model assumptions. It is, therefore, a com-
mon practice to combine the tests with a diagnostic plot (a plot of residuals against the
predicted values) to make a judgment on the severity of heteroscedasticity. To generate
the plot of residuals against the fitted (predicted) values of the dependent variable
(diastolic BP), use either of the following commands:
rvfplot
rvfplot, yline(0)
The above commands will generate Figures 16.8 and 16.9, respectively. The figures are
similar except that Figure 16.9 has a reference line that represents Y= 0. If the scatters
of the points show no clear pattern (as seen in Fig 16.8 and Fig 16.9), we can conclude
that the variances of the sub-population of “Y” are constant (homoscedastic).
If there is evidence of heteroscedasticity (Fig 16.10), one of the solutions is to run a
regression with robust standard errors by using the following command:
regress dbp age sbp i.sex ib3.religion, vce(robust)
 Linear Regression 163

20

20
10

10
Residuals

Residuals
0

0
-10
-10

-20
-20

60 80 100 120 60 80 100 120

Fitted values Fitted values

Figure 16.8 No heteroscedasticity Figure 16.9 No heteroscedasticity

Figure 16.10 Presence of heteroscedasticity

16.2.4.5 Checking for outliers

The presence of outliers is checked on the standardized values (z-values) of the residu-
als. Outliers are the z-values of the residuals that are either less than -3.0 or greater than
+3.0. To assess the outliers, we need to generate (we have done it before; see Section
16.2.4.2) a variable for the z-values of residuals (or transform the variable “residual”
generated earlier into z-values; see Section 5.10) by using the following command:
predict zresid, rstandard
This command will generate a new variable “zresid” (if not done earlier) with the
z-values of residuals. Now, use the following commands to check if there are any
values that are greater than +3.0 or less than -3.0.
list sl zresid if (zresid<-3)
164 Linear Regression

list sl zresid if (zresid>3)

tab zresid
The first command will display the z-residual values if they are less than minus 3.0
along with the serial numbers (ID numbers). Note that if there is no value that is less
than minus 3.0, Stata will not show any output. Similarly, the second command will
display the z-residual values that are greater than +3.0 along with the ID numbers. If
there are any values that are less than -3.0 and/or greater than +3.0, consider them the
outliers. If you use the third command, Stata will provide a long table of z-residual
values (better to avoid this).
Instead of using the above commands, you can check the highest 10 and lowest 10
values of the variable “zresid” by using the following commands after sorting the
variable:
sort zresid
list zresid in 1/10
list zresid in -10/-1
The above commands will display the outputs as shown in Table 16.9. The table shows
that there are no values that are greater than +3.0 or less than minus 3.0. This indicates
that there is no outlier in the residuals. However, one or more outliers may be present
in a dataset. The outliers should be carefully checked for mistakes in data entry. If there
is no evidence of mistakes and the value is plausible, then the value should not be
altered. It is also discouraged to exclude outliers from analysis. The presence of outli-
ers in the data may influence the results of regression or other statistical analyses. A
recommended strategy for handling the outliers is to carry out the analysis with and
without the outliers. If there is little difference in the results, the outliers have minimal
effect. If the difference is substantial, it may be better to find an alternative method of
data analysis, such as data transformation or using the rank method [3]. You may also
run the regression with robust standard errors if there are outliers in the data by using
the following command:
regress dbp age sbp i.sex ib3.religion, vce(robust)

16.2.4.6 Test for independence

The test for checking the independence of observations (i.e., values of residuals are
independent or there is no autocorrelation) is needed for time-series data. The test for
independence is done to evaluate if there is any autocorrelation in the residuals. Auto-
correlation refers to the degree of correlation of the same variable between two succes-
 Linear Regression 165

Table 16.9 Upper and lower 10 z-values of the residuals

. list zresid in 1/10 . list zresid in -10/-1

+-----------+ +----------+
| zresid | | zresid |
|-----------| |----------|
1. | -2.462475 | 201. | 1.600753 |
2. | -2.258894 | 202. | 1.665619 |
3. | -2.158022 | 203. | 1.800676 |
4. | -2.071153 | 204. | 1.823561 |
5. | -2.065784 | 205. | 1.846748 |
|-----------| |----------|
6. | -2.003306 | 206. | 1.887106 |
7. | -1.93732 | 207. | 1.951255 |
8. | -1.908815 | 208. | 2.181215 |
9. | -1.884941 | 209. | 2.678532 |
10. | -1.83873 | 210. | 2.960519 |
+-----------+ +----------+

sive time intervals (i.e., the degree of similarity between a given time series). For
cross-sectional data, autocorrelation is not an issue.
The independence of residuals (autocorrelation) is assessed by the Durbin-Watson
(DW) statistic and is applicable for time-series data. The DW test is done after execut-
ing the multiple regression analysis. The DW statistic ranges from 0 to 4. A value of 2
indicates that there is no autocorrelation. Values less than 0 to 2 indicate the presence
of a positive autocorrelation, while values greater than 2 to 4 indicate the presence of a
negative autocorrelation. To perform the DW test, it requires a time variable. In our
dataset (Data_4.dta), there is no time variable. However, we can generate a time
variable (for the purpose of demonstration) by using the following command before
doing the DW test:
gen time = _n
This command will generate a time variable “time”. Now, use the following command
to let Stata know which variable is the time variable for this analysis (if the time variable
in your data file is “year”, use the command: tsset year):
tsset time
Finally, use the following command, after multiple regression analysis, to get the DW
test statistic (Table 16.10):
166 Linear Regression

Table 16.10 Durbin-Watson test

. gen time= _n

. tsset time
time variable: time, 1 to 210
delta: 1 unit

. estat dwatson

Durbin-Watson d-statistic( 6, 210) = 1.700676

estat dwatson
If the DW statistic value hovers around 2 (between 1.5 and 2.5), it indicates that the
data points are independent (there is no autocorrelation). Table 16.10 shows that the
value of the DW statistic is 1.70. Since the value is close to 2, it can be considered that
there is no autocorrelation in the residuals.

16.2.5 Variable selection for a model

In general, the independent variables to be selected for a multivariable analysis should
include the risk factors of interest and potential confounding factors (based on theory,
prior research, and empirical findings), while variables with lots of missing values
should be excluded.
We have used the "Enter" method (see Table 16.11) for the regression analysis of data
so far in this chapter. The "Enter" method uses all the independent variables in the
model as decided by the researcher. It does not remove any variables from the model
automatically during analysis. Automatic procedures can be used to determine which
independent variable(s) will be included (retained) in the model. The major reason for
using the automatic selection procedure (i.e., the stepwise method) is to identify the
useful independent variables necessary to estimate or predict the outcome variable.
The major limitations of automatic selection procedures (stepwise methods) are that
the analysis may provide invalid estimates and confidence intervals. Therefore, the
stepwise methods should be used cautiously [28, 40].
Stata and other data analysis software have the option to automatically select the inde-
pendent variables for a model. They use statistical criteria to select the variables and
their order in the model. The commonly used variable selection techniques are the
stepwise methods.
 Linear Regression 167

In simple terms, stepwise regression is a process that helps us to determine which

variables are important and which are not in explaining the outcome variable. Certain
variables may have high p-values (e.g., p-values greater than 0.05) and do not mean-
ingfully contribute to predicting the outcome. In stepwise regression, only the import-
ant variables that are statistically contributing to the outcome are kept to ensure the
best linear model for prediction. Different methods of stepwise approaches are
described in Table 16.11.

16.2.5.1 Backward selection method

For a backward selection method of variables in multiple regression, use the following
command (outputs are shown in Table 16.12):
xi: stepwise, pr(0.2): regress dbp sbp age i.sex i.religion
In this example, “pr(0.2)” indicates the removal (exclusion) criteria of independent
variables from the model, i.e., if the p-value of an independent variable is ≥0.02, the
variable will be automatically removed from the model. You can change the criteria of
removal based on your requirements. For example, you may decide the removal crite-
ria as 0.1 or 0.3 or others. The use of “xi:” before the command “stepwise” will consid-
er both the categorical and quantitative variables for the analysis.
Sometimes it is necessary to keep one or more variables in the model that are consid-
ered important for theoretical or practical reasons, though they are not statistically
significant (forced-entry of variables). In our example, the variable “age” is not signifi-
cantly associated with the outcome variable (p=0.614; Table 16.3). However, if you
want to force-entry the variable “age” into the model in a stepwise method, use the first
of the following commands (Table 16.13):
xi: stepwise, pr(0.2) lockterm1: regress dbp age sbp i.religion
xi: stepwise, pr(0.2) lockterm1: regress dbp (age i.religion) sbp
The first command will force-entry the variable “age” (the first variable immediately
after the dependent variable) into the model. The second command will force-entry the
variables “age” and “religion” into the model.
For the backward stepwise method with a removing criteria of p≥0.2 and adding
(entry) criteria of p<0.1, use the following command:
xi: stepwise, pr(.2) pe(.1): regress dbp sbp age i.sex i.religion
168 Linear Regression

Table 16.11 Variable selection methods for modeling

Technique Method Advantages and limitations
Backward In this method, all the variables are Better for assessing
selection initially included, and in each step, the (adjusting) for confounding
most statistically insignificant variable effect than the forward
(p>0.05; useless variable) is dropped. This selection method.
process is repeated until all the variables
left over are statistically significant.
Forward In this method, first a single independent Best suited for dealing with
selection variable that has the strongest association the studies where the sample
(smallest p-value) with the outcome size is small. Does not deal
variable is entered into the model. Then well with suppressor
(second step), the method identifies the (confounding) effects.
variables among those not in the model
that, when added to the model so far
obtained, explain the largest amount of
the remaining variability. The second step
is repeated until the addition of the extra
variable is not statistically significant.
Stepwise/ This is a combination of forward and Has the ability to manage
Remove backward selection methods. In the large number of potential
selection stepwise method, variables that are predictor variables, fine-
entered are checked at each step for tuning the model to choose
removal. Likewise, in the removal the best predictor variables
method, variables that are excluded will from the available options.
be checked for re-entry.
Enter (all Enters all the variables at the same time Including all variables may
variables) and does not remove any variable be problematic, if there are
automatically from the model. many independent variables
and the sample size is small.

16.2.5.2 Forward selection method

For a forward selection method with an entry-term p-value of ≤0.2, use the following
command [pe(#) indicates the level of significance for inclusion into the model]:
xi: stepwise, pe(0.2): regress dbp age sbp i.sex i.religion
 Linear Regression 169

Table 16.12 Modeling with backward selection method

. xi:stepwise, pr(0.2): regress dbp sbp age i.sex i.religion

i.sex _Isex_0-1 (naturally coded; _Isex_0 omitted)

i.religion _Ireligion_1-3 (naturally coded; _Ireligion_1 omitted)
begin with full model
p = 0.8763 >= 0.2000 removing _Ireligion_3
p = 0.6723 >= 0.2000 removing _Ireligion_2
p = 0.6181 >= 0.2000 removing age

Source | SS df MS Number of obs = 210

-------------+------------------------------ F( 2, 207) = 272.86
Model | 20917.406 2 10458.703 Prob > F = 0.0000
Residual | 7934.16062 207 38.3292784 R-squared = 0.7250
-------------+------------------------------ Adj R-squared = 0.7223
Total | 28851.5667 209 138.045774 Root MSE = 6.1911

------------------------------------------------------------------------------
dbp | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
sbp | .489694 .0215077 22.77 0.000 .4472918 .5320963
_Isex | -2.180167 .8930832 -2.44 0.015 -3.940872 -.4194618
_cons | 21.01581 2.838019 7.41 0.000 15.42068 26.61094
------------------------------------------------------------------------------

For a forward stepwise method with a removing criteria of p≥0.2 and adding (entry)
criteria of p<0.15, use the following command:
xi: stepwise, pr(.2) pe(.15) forward: regress dbp sbp age i.sex i.religion

16.2.5.3 Interpretation
The interpretation of the outputs of stepwise methods is the same as discussed in
Section 16.2.3. In the backward selection method of analysis, our dependent variable
was “dbp” (Table 16.12). The independent variables included in the analysis were
“sbp, age, sex, and religion”. The outputs of the analysis (Table 16.12) show that two
variables (systolic BP and sex) are significantly associated with diastolic BP and are
retained in the model.
The R-squared value, calculated in the analysis, is 0.725, indicating that the indepen-
dent variables (systolic BP and sex) together explain 72.5% of the variation in diastolic
BP, which is statistically significant [p=0.000 (Prob > F)]. The regression coefficients
(Coef.) and p-values (P > | t |) for systolic BP and sex are 0.489 (p=0.000) and -2.180
(for males compared to females; p=0.015), respectively. We can, therefore, conclude
from the analysis that the systolic BP and sex are the important factors significantly
influencing the diastolic BP (since the p-values are <0.05).
170 Linear Regression

The regression coefficient for systolic BP is 0.489 (95% CI: 0.44 to 0.53; p<0.001).
This indicates that the average increase (or decrease) in diastolic BP is 0.49 mmHg if
the systolic BP increases (or decreases) by 1 mmHg after adjusting for sex. The regres-
sion coefficient for sex, on the other hand, is –2.180 (95% CI: -3.940 to -0.419; p=
0.015), which means that the average diastolic BP of males is 2.18 mmHg less (since
the coefficient is negative) than the females after adjusting for systolic BP.

Table 16.13 Backward selection method with forced-entry of age in the model
. xi: stepwise, pr(0.2) lockterm1: regress dbp age sbp i.religion

i.religion _Ireligion_1-3 (naturally coded; _Ireligion_1 omitted)

begin with full model
p = 0.8181 >= 0.2000 removing _Ireligion_2
p = 0.6124 >= 0.2000 removing _Ireligion_3

Source | SS df MS Number of obs = 210

-------------+------------------------------ F( 2, 207) = 263.03
Model | 20704.4665 2 10352.2333 Prob > F = 0.0000
Residual | 8147.10016 207 39.3579718 R-squared = 0.7176
-------------+------------------------------ Adj R-squared = 0.7149
Total | 28851.5667 209 138.045774 Root MSE = 6.2736

------------------------------------------------------------------------------
dbp | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
age | -.0353995 .056452 -0.63 0.531 -.1466941 .0758952
sbp | .4956516 .0216435 22.90 0.000 .4529816 .5383216
_cons | 20.48269 3.281515 6.24 0.000 14.01322 26.95217
------------------------------------------------------------------------------
 17
Logistic Regression

Logistic regression is a commonly used multivariable method of data analysis in health

and social sciences research. This tool can be applied to analyze the data of cross-sec-
tional, case-control, or cohort studies. Logistic regression analysis is performed when
the outcome (dependent) variable is a categorical variable, either dichotomous (also
called binary variable, e.g., disease – present/absent), unordered polychotomous (e.g.,
type of food preferred – rice/bread/meat) or ordinal variable (severity of pain – mild/-
moderate/severe). The predictive (independent) variables can be either categorical or
continuous. Like other multivariable analyses, the purposes of multivariable logistic
regression analysis are to:
• Adjust the estimate of risk (odds ratio) for a number of independent variables
included in the model;
• Determine the relative contribution of independent variables to a single outcome;
• Predict the probability of an outcome for a number of independent variables in
the model; and
• Assess the interaction of multiple independent variables on the outcome variable.
Logistic regression analysis can be applied in several methods, depending on the type
of outcome variable and study design. They are:

1. Binary logistic regression: This method is used when the dependent variable is a
dichotomous (binary) categorical variable, such as a disease (present/absent), vaccinat-
ed (yes/no), or the outcome of a patient (died/survived). The binary logistic regression
can be applied as:
a) Unconditional binary logistic regression: This method is used when the depen-
172 Logistic Regression

dent variable is a dichotomous categorical variable in an unmatched study

design (e.g., unmatched case-control studies). The term “unconditional binary
logistic regression” is commonly expressed as “unconditional logistic regres-
sion or logistic regression”; and
b) Conditional binary logistic regression: This method is applied where the dep-
endent variable is a dichotomous variable and the cases are matched with
controls for one or more variables (e.g., matched case-control studies). The
word “binary” is commonly omitted from the terminology and is simply
expressed as “conditional logistic regression”.

2. Multinominal logistic regression: This method of data analysis is used when the
outcome (dependent) variable is a nominal categorical variable with more than two
levels, such as health-seeking behavior (did not seek treatment/ received treatment
from village doctors/ received treatment from pharmacists), type of cancer (stomach
cancer/ lung cancer/ skin cancer) and others.

3. Ordinal logistic regression (proportional odds regression): This method is used

when the outcome variable is an ordinal categorical variable, like severity of pain
(mild/ moderate/ severe) and stage of cancer (stage 1/ stage 2/ stage 3/ stage 4).

17.1 Mathematical concept of logistic regression model

In logistic regression analysis, odds are transformed into natural log (ln) of odds, i.e.,
“ln odds”. The “ln” is the log to the base of e. To reverse (antilog) the ln, we take the
exponential (ex) of the log value. When the odds are transformed into ln odds, it is
called the logit transformation. In logistic regression, ln odds of the outcome variable
are put on the Y-axis. The multivariable logistic regression model is given by the equa-
tion:

Where, P denotes the probability of the outcome, β0 is the intercept (constant) and βi is
the regression coefficient of the ith variable (i= 1, 2, …, n), and Xi represents the values
of the predictor (independent) variables in the model, Xi = (X1, X2, … Xn).
The regression coefficients (β) that we get in logistic regression analysis are the ln
odds and the exponential of the regression coefficients are the odds ratios (ORs) for the
 Logistic Regression 173

categorical independent variables after adjusting for other variables in the model. If the
independent variable is a continuous variable, the interpretation is different and is
discussed in Section 17.2.1.1. We can calculate the probability (p) of an outcome by
using the following formula:

Or

A detailed explanation of the model can be found in any standard biostatistics book. In
this chapter, we will discuss the unconditional and conditional logistic regression
methods, while the multinominal logistic regression method is discussed in Chapter
18.

Assumptions for logistic regression

Logistic regression does not make any assumption concerning the distribution of the
predictor (independent) variables. However, it is sensitive to high correlations among
the independent variables (multicollinearity). The outliers may also affect the results of
logistic regression analysis.

17.2 Binary logistic regression

Binary logistic regression analysis is appropriate when the outcome variable is a
dichotomous categorical variable (e.g., disease present/absent). It is commonly used
for the adjustment of one or more confounding factors or to model (identify) the
predictors for a dichotomous categorical outcome. For logistic regression analysis in
Stata, the dichotomous outcome variable must be coded as “0= disease absent” and
“1= disease present”. Stata will consider the higher value to be the predicted outcome
and the lower value as the comparison group.

17.2.1 Unconditional binary logistic regression

Unconditional binary logistic regression is simply called logistic regression. We will
use the term “logistic regression” to mean unconditional logistic regression throughout
this chapter.
174 Logistic Regression

Suppose that you have conducted an unmatched study (e.g., a cross-sectional or an

unmatched case-control study) to identify the factors (or predictors) associated with
diabetes mellitus (dependent or outcome variable). The independent (explanatory or
predictor) variables/factors that you have considered for logistic regression analysis
are sex (variable name: sex), age (variable name: age), peptic ulcer (variable name:
pulcer), family history of diabetes (variable name: fhistory), and religion (variable
name: religion).
To perform logistic regression analysis, recode all the dichotomous independent
variables as "0 for no" and "1 for yes" if they are not coded like this. If they are coded
like this, interpretation of the odds ratio (OR) is straightforward, otherwise, interpreta-
tion may be complicated (especially if you do not use the prefix "i." to the variable).
Use the data file <Data_4.dta> for practice.
The coding scheme of the variables to be used in logistic regression analysis is shown
in Table 17.1. You can also check the coding scheme (value labels of categorical
variables only) of the variables by using the following command:
label list diabetes sex pulcer fhistory religion
Or,
codebook diabetes sex pulcer fhistory religion
To perform the logistic regression analysis, use either of the following commands:
logistic diabetes i.sex age i.pulcer i.fhistory i.religion
logit diabetes i.sex age i.pulcer i.fhistory i.religion, or
logit diabetes i.sex age i.pulcer i.fhistory i.religion
The first two commands will report the results in terms of odds ratios (ORs) with their
95% confidence intervals (CIs) (Table 17.2), while the third command will present the
results in terms of logistic regression coefficients with their 95% CIs (Table 17.3). In
the second command, the option “or” indicates the odds ratio. If you use this option
with the “logit” command, Stata will report the ORs. If the prefix “i.” is used for the
independent variables, Stata will consider them as categorical variables during analy-
sis. Note that the first variable after the command is the dependent (outcome) variable.
We recommend using the first command for the analysis since our interest is to get the
ORs for easy interpretation of the categorical independent variables.
In logistic regression, Stata considers (by default) the first category (lowest value) of
the categorical independent variables as the comparison group. For example, we have
entered the variable “sex” in the analysis. The coding scheme of sex is 0 for females
 Logistic Regression 175

Table 17.1 Codebook

Variable name Variable label Variable codes
diabetes Have diabetes (dependent variable) 0= no; 1= yes
sex Sex 0= female; 1= male
age Age in years Actual value
pulcer Have peptic ulcer 0= no; 1= yes
fhistory Have family history of diabetes 0= no; 1= yes
religion Religion of the subjects 1= Muslim; 2= Hindu;
3= Christian

Table 17.2 Logistic regression analysis using the command “logistic”

. logistic diabetes i.sex age i.pulcer i.fhistory i.religion

Logistic regression Number of obs = 210

LR chi2(6) = 102.61
Prob > chi2 = 0.0000
Log likelihood = -57.807479 Pseudo R2 = 0.4702

------------------------------------------------------------------------------
diabetes | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
sex |
Male | 4.80143 2.555999 2.95 0.003 1.691379 13.63014
age | 1.262949 .0503621 5.85 0.000 1.168 1.365616
|
pulcer |
Yes | 5.937446 2.947017 3.59 0.000 2.244453 15.70684
|
fhistory |
Yes | 2.85935 1.531351 1.96 0.050 1.000918 8.168382
|
religion |
HINDU | 1.655078 .9067061 0.92 0.358 .5655909 4.843225
Christian | .8324862 .7949383 -0.19 0.848 .1281054 5.409867
|
_cons | .0000187 .0000296 -6.90 0.000 8.48e-07 .0004131
------------------------------------------------------------------------------

and 1 for males. Therefore, the analysis will give the OR for males compared to
females. On the other hand, the independent variable “religion” has three levels (Table
17.1). Since Muslims are coded as 1 (first category), Muslims will be the comparison
group for religion. The ORs that the Stata will provide for other religions (Hindus and
176 Logistic Regression

Christians) will be compared to Muslims. However, you can change the comparison
group by using the following commands:
logistic diabetes i.sex age i.pulcer i.fhistory ib3.religion
logistic diabetes ib1.sex age i.pulcer i.fhistory ib3.religion
The first command will consider the last category of religion (category 3 or Christians)
as the comparison group, while the second command will consider males (because
males are coded as 1) as the comparison group for sex, and Christians as the compari-
son group for religion during analysis.
Occasionally, in cross-sectional studies, data is collected through cluster sampling
methods. In such a situation, it is necessary to control (adjust) for the cluster effects
during analysis. To adjust for cluster effects (say, the name of the cluster variable is
“clus”), use the following command:
logistic diabetes i.sex age i.pulcer i.fhistory i.religion, vce(cluster clus)

17.2.1.1 Interpretation
In logistic regression analysis, we have entered five independent variables, among
which one is a continuous variable (age) and the others are categorical variables (sex,
peptic ulcer, family history of diabetes, and religion). Let us interpret the outputs
provided in Table 17.2.
The table shows that the data from 210 subjects were analyzed. The likelihood ratio
(LR) chi-square [LR Chi2(4)] value is 102.61 and its p-value (Prob > chi2) is 0.000.
We want the LR chi-square test to be significant (p<0.05). A significant LR chi-square
test indicates that the proposed model is better than the null model (i.e., a model with-
out any independent variable) in predicting the outcome variable. Furthermore, if the
LR chi-square test p-value is greater than 0.05, it means that the independent variables
are unable to predict the outcome variable (a situation where none of the independent
variables in the model are significant).
The pseudo R-square value indicates the amount of variation in the outcome variable
that can be explained by the independent variables in the model. In this example, the
pseudo R-square (Pseudo R2) value is 0.4702. This indicates that 47.02% of the varia-
tion in the outcome variable (diabetes) can be explained by all the independent
variables (sex, age, peptic ulcer, family history of diabetes, and religion) in the model.
However, the pseudo R-square value should be interpreted cautiously because it is not
equivalent to the R-squared value that we get in a linear regression model (Sections
 Logistic Regression 177

Table 17.3 Logistic regression analysis using the command “logit”

. logit diabetes i.sex age i.pulcer i.fhistory i.religion

Iteration 0: log likelihood = -109.11177

Iteration 1: log likelihood = -65.686705
Iteration 2: log likelihood = -58.065695
Iteration 3: log likelihood = -57.807622
Iteration 4: log likelihood = -57.807479
Iteration 5: log likelihood = -57.807479

Logistic regression Number of obs = 210

LR chi2(6) = 102.61
Prob > chi2 = 0.0000
Log likelihood = -57.807479 Pseudo R2 = 0.4702

------------------------------------------------------------------------------
diabetes | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
sex |
Male | 1.568914 .5323411 2.95 0.003 .5255445 2.612283
age | .2334492 .0398766 5.85 0.000 .1552926 .3116058
|
pulcer |
Yes | 1.781279 .4963443 3.59 0.000 .808462 2.754096
|
fhistory |
Yes | 1.050594 .5355591 1.96 0.050 .0009176 2.100271
|
religion |
HINDU | .5038483 .5478328 0.92 0.358 -.5698842 1.577581
Christian | -.1833386 .9548967 -0.19 0.848 -2.054902 1.688225
|
_cons | -10.88592 1.578628 -6.90 0.000 -13.97998 -7.79187
------------------------------------------------------------------------------

16.1.2 and 16.2.3). This information is not needed if the objective of the analysis is to
adjust for Odds Ratio.
The table (Table 17.2) also shows the ORs (Odds Ratios) for the independent variables
with their standard errors (Std. Err.), 95% CIs (95% Conf. Interval), and p-values
(P>|z|). In our analysis, there are five independent variables in the model. All of them
are categorical variables (sex, family history of diabetes, peptic ulcer, and religion),
except for age, which is entered as a continuous variable. The interpretation of the OR
for a categorical and a continuous variable is not the same.
Let us first interpret the ORs for categorical variables. Table 17.2 shows that the OR
for sex (being male) is 4.80 (95% CI: 1.69 - 13.63), which is statistically significant
(P=0.003). Here, the comparison group is female. The OR of 4.80 suggests that males
are 4.8 times more likely to have diabetes compared to females after adjusting (con-
178 Logistic Regression

trolling) for age, family history of diabetes, peptic ulcer, and religion. Note that the
ORs provided by logistic regression analysis are the adjusted ORs. Similarly, individu-
als who have a family history of diabetes are 2.85 times more likely (OR: 2.85; 95%
CI: 1.00 to 8.16; p=0.05) to have diabetes compared to those who do not have a family
history, after adjusting for age, sex, peptic ulcer, and religion. But the OR is not statisti-
cally significant (though the p-value is 0.05) as the null value (one) is included in the
95% CI. Lastly, the ORs for Hindus and Christians provided in the table are compared
to Muslims after controlling for age, sex, peptic ulcer, and family history of diabetes.
The interpretation of OR for age is different since the variable is entered as a continu-
ous variable. In our example, the OR for age is 1.262 (95% CI: 1.168 – 1.365;
p=0.000). This means that the odds of having diabetes will increase (since the value of
OR is greater than one) by 26.2% (calculated as OR – 1; i.e., 1.262 – 1 = 0.262 or
26.2%) (95% CI: 16.8% to 36.5%) with each year increase in age after adjusting for all
other variables in the model, which is statistically significant (p<0.001).
All the outputs in Table 17.3 are the same as the outputs in Table 17.2 except that Table
17.3 provides the regression coefficients (with their standard errors and 95% CIs)
rather than the ORs. Note that the z-values (Z) and p-values (P>|z|) are the same in both
tables. The coefficients are used to calculate the predicted probabilities of the outcome
variable with the independent variables in the model. In logistic regression, the odds
are transformed into ln odds (logit transformation) during analysis. Therefore, the
coefficients reported in the table are the ln odds and their exponentials (ex) are the ORs.
For example, the coefficient for males is 1.568914 and its exponential is 4.80, which is
the OR for males (Table 17.2).
The table (Table 17.3) also shows the iterations. The iterations indicate the log
likelihood values. The first iteration (iteration 0) indicates the log likelihood of the
null model (i.e., a model without any independent variable). The table shows that the
log likelihood has increased [from -109.11 (Iteration 0) to -57.80 (Iteration 5)] with
the inclusion of independent variables in the model. The improvement in log likeli-
hood value is statistically significant as indicated by the p-value (0.000) of the log
likelihood ratio chi-square test (Prob > Chi2). A significant p-value indicates that the
model is significantly better than the null model.

17.2.2 Logistic regression diagnostics

In order to check if the analysis is valid, the model has to satisfy certain assumptions
of logistic regression. When the assumptions are not met, the analysis may provide
 Logistic Regression 179

biased estimates of the coefficients, which may lead to misinterpretation of the results.
It is, therefore, important to check the underlying assumptions before considering the
logistic regression analysis valid. Regression diagnostics are used to evaluate whether
the assumptions are true or not. In this section, we will discuss how to assess some of
the important assumptions to validate the model, like multicollinearity, model fit, and
others.

17.2.2.1 Checking for multicollinearity

Multicollinearity occurs when one or more independent variables are in a linear combi-
nation (highly correlated) with other independent variables in the model. The degree of
multicollinearity can vary and can have different effects on the model. When multicol-
linearity is present, the model becomes dubious (i.e., the model may not provide a
correct estimate of the regression coefficients). It is, therefore, important to check for
multicollinearity of the independent variables in the model.
Multicollinearity can be checked by generating a correlation matrix of the independent
variables included in the model. To generate the correlation matrix, use the following
command (Table 17.4):
correlate sex age pulcer fhistory religion
Table 17.4 shows the correlation coefficients (r values) of the independent variables
included in the analysis. If an r value is greater than 0.5, it is generally considered that
the variable has a correlation with another variable that may affect the regression
analysis. We don’t have any such problem in our analysis because none of the r values
is greater than 0.5 (the highest value is 0.22).
Another simple and subjective way to examine multicollinearity is to check the
standard errors (SE) of the coefficients as provided in Table 17.3. If multicollinearity
is present and affects the model, the magnitude of the standard errors (SEs) of some of
the coefficients will be very high (greater than 5.0) or very low (less than 0.001). The
existence of multicollinearity means that the model is not statistically stable. To solve
the problem (in general), look at the SEs and omit the variable(s) with a very high (or
very low) SE until the magnitude of the SEs hovers between 0.001 and 5.0.
Sometimes an independent variable may have a strong correlation with the constant
(residuals or errors) in the model. If there is a strong correlation between the constant and
any of the independent variables, you can omit the constant from the model. To check the
correlation between constant and independent variables in the model, first generate a
residual variable (representing the constant) by using the following command:
180 Logistic Regression

Table 17.4 Correlation matrix of the independent variables

. correlate sex age pulcer fhistory religion
(obs=210)

| sex age pulcer fhistory religion

-------------+---------------------------------------------
sex | 1.0000
age | 0.0586 1.0000
pulcer | 0.0520 0.2153 1.0000
fhistory | -0.2222 0.1585 0.1282 1.0000
religion | -0.1308 -0.0582 -0.1038 0.1453 1.0000

predict residual, resid

This command will generate a new variable “residual” containing the residual values
(errors). Now, to get the correlation matrix of independent and residual variables, use
the following command (Table 17.5):
correlate sex age pulcer fhistory religion residual
Table 17.5 shows a moderate correlation (r = 0.68) between age and residual (con-
stant). However, it has not affected the results since none of the SEs of coefficients are
greater than 5.0 or less than 0.001 (Table 17.3). If it had affected the results, some of
the SEs of the coefficients would have been either greater than 5.0 or less than 0.001.
If there is a problem like this, it is suggested to run the logistic regression analysis
without (omitting) the constant. The following is the command to do the analysis with-
out the constant in the model:
logistic diabetes i.sex age i.pulcer i.fhistory i.religion, nocons

17.2.2.2 Checking for model fit

Logistic regression is commonly done to adjust the ORs for confounding factors and to
identify predictors for the outcome variable. When the intention of analysis is predic-
tion (i.e., to identify the predictors), then the question is “How good is the model for
prediction?” This is judged based on the Hosmer-Lemeshow goodness-of-fit test and
the positive and negative predictive values given in the classification table.

Hosmer-Lemeshow goodness-of-fit test

The Hosmer-Lemeshow test indicates how well the observed and predicted values fit
with each other (i.e., the observed and predicted probabilities match with each other).
The null hypothesis is “the model fits” and the p-value is expected to be >0.05
 Logistic Regression 181

Table 17.5 Correlation matrix of independent variables and residuals

. correlate sex age pulcer fhistory religion residual
(obs=210)

| sex age pulcer fhistory religion residual

-------------+------------------------------------------------------
sex | 1.0000
age | 0.0586 1.0000
pulcer | 0.0520 0.2153 1.0000
fhistory | -0.2222 0.1585 0.1282 1.0000
religion | -0.1308 -0.0582 -0.1038 0.1453 1.0000
residual | 0.1227 0.6824 0.3096 0.1749 0.1768 1.0000

(non-significant). If the p-value is not significant (>0.05), it suggests that the model is
good for prediction of the outcome variable (i.e., the observed and predicted values are
close together). This test is done after running the logistic regression analysis, and the
command is:
lfit, group(10)
The above command will generate Table 17.6. The table shows that the Hosmer-Leme-
show chi-square test p-value is 0.085. Since the p-value is greater than 0.05, we can
conclude that the model is useful for prediction of the outcome variable by the inde-
pendent variables included in the model. If the test is significant (p<0.05), we will
conclude that the model is not good enough to predict the outcome variable by the
independent variables in the model. This information is not needed if the objective of
doing the logistic regression analysis is to adjust for the confounding factors.

Classification table
The classification table provides us with the sensitivity, specificity, and positive and
negative predictive values, and overall accuracy of the model. The predictive values
indicate how well the model is able to predict the correct category of the dependent
variable (i.e., have or do not have the disease). To get the sensitivity, specificity, and
positive and negative predictive values, we need to generate the classification table by
using the following command (after running the logistic regression analysis). Usually,
the classification table is generated at a cut-off value of 0.5 (50%). You can, however,
change the cut-off value of your choice.
lstat, cutoff(0.5)
This command will generate Table 17.7. The table shows that the overall accuracy of
182 Logistic Regression

Table 17.6 Hosmer-Lemeshow goodness-of-fit test

. lfit, group(10)

Logistic model for diabetes, goodness-of-fit test

(Table collapsed on quantiles of estimated probabilities)

number of observations = 210

number of groups = 10
Hosmer-Lemeshow chi2(8) = 13.87
Prob > chi2 = 0.0851

this model to predict diabetes (with a predicted probability of 0.5 or greater) is 90.5%
(shown at the bottom of the table as correctly classified). The sensitivity and specificity
of the model are 71.11% [32 ÷ 45] and 96.36% [159 ÷ 165], respectively, while the
positive and negative predictive values are 84.2% [32 ÷ 36] and 92.4% (159 ÷ 172],
respectively. Interpretation of these findings is a little complicated and needs further
explanation, especially the explanation of sensitivity, specificity, and positive and
negative predictive values. For a detailed explanation, readers may refer to any
standard epidemiology book [10, 16]. You can see the sensitivity and specificity with
varying cut-off points (in a graph) by using the following command (not shown):
lsens
To conclude, the information provided in the classification table is needed if the inten-
tion of logistic regression analysis is to predict the outcome variable with independent
variables in the model. We can ignore the information if the objective of the analysis is
to adjust for the confounding factors.

17.2.2.3 ROC curve

We can construct the receiver-operating characteristic (ROC) curve to assess the model
discrimination, which is an indication of the accuracy of logistic regression model.
Discrimination is defined as the ability of the model to distinguish between those who
have the outcome (e.g., a disease) and those who do not have the outcome. Discrimina-
tion is evaluated by using the ROC curve analysis. In ROC curve analysis, the area
under the curve (called C-statistic) is measured.
The area under the ROC curve ranges from 0 to 1. A value of 0.5 indicates the model
is useless. Values between 0.7 and 0.8 are considered acceptable discrimination, values
between 0.8 and 0.9 indicate excellent discrimination, and values ≥0.9 indicate
 Logistic Regression 183

Table 17.7 Classification table

. lstat, cutoff(0.5)

Logistic model for diabetes

-------- True --------

Classified | D ~D | Total
-----------+--------------------------+-----------
+ | 32 6 | 38
- | 13 159 | 172
-----------+--------------------------+-----------
Total | 45 165 | 210

Classified + if predicted Pr(D) >= .5

True D defined as diabetes != 0
--------------------------------------------------
Sensitivity Pr( +| D) 71.11%
Specificity Pr( -|~D) 96.36%
Positive predictive value Pr( D| +) 84.21%
Negative predictive value Pr(~D| -) 92.44%
--------------------------------------------------
False + rate for true ~D Pr( +|~D) 3.64%
False - rate for true D Pr( -| D) 28.89%
False + rate for classified + Pr(~D| +) 15.79%
False - rate for classified - Pr( D| -) 7.56%
--------------------------------------------------
Correctly classified 90.95%
--------------------------------------------------

outstanding discrimination.
To get the ROC curve, after performing the logistic regression analysis, use the follow-
ing command (Fig 17.1):
lroc
Figure 17.1, generated by the command above, shows that the area under the curve is
0.915 (at the bottom of Figure 17.1). Since the value is >0.9, it is an excellent model
for prediction. However, we can separately calculate the area under the ROC curve
with its 95% CI. To calculate the area under the curve, we need to generate a classifier
variable (say, class). Then we will calculate the area under the curve with its 95% CI.
Use the following commands:
predict class
roctab diabetes class
The first command will generate the classifier variable “class”, while the second com-
mand will calculate the area under the curve with its 95% CI (Table 17.8).
184 Logistic Regression

1.00
0.75
Sensitivity
0.50
0.25
0.00

0.00 0.25 0.50 0.75 1.00

1 - Specificity
Area under ROC curve = 0.9154

Figure 17.1 ROC curve

17.2.2.4 Other postestimation commands

The postestimation commands are used after performing the logistic regression analy-
sis. Based on the analysis, you can calculate the estimated probability of the dependent
variable for each subject in the dataset using the following command:
predict pre1, pr
This command will generate a variable “pre1” with the probability of having the
outcome (diabetes) for each subject.
Stata can also provide the probability of having the outcome variable for the indepen-
dent variables in the model. Use the following command to get the predicted probabili-
ty (adjusted probability) of having diabetes for each level of sex and religion, consider-
ing the average values of the covariates in the model:
margins religion sex, atmeans
marginsplot
The first command will provide the predicted probability of diabetes for each level of
religion and sex (Table 17.9). The table shows that the predicted probability (Margin) of
diabetes (i.e., the probability of having diabetes) for being a Muslim is 0.0720, while it
is 0.113 for Hindus and 0.060 for Christians. The second command will display a graph
of predicted probability for all the categories of religion with their 95% CIs (Fig 17.2).
 Logistic Regression 185

Table 17.8 Area under the ROC curve with 95% CI

. predict class
(option pr assumed; Pr(diabetes))

. roctab diabetes class

ROC -Asymptotic Normal--

Obs Area Std. Err. [95% Conf. Interval]
--------------------------------------------------------
210 0.9154 0.0278 0.86078 0.96993

17.2.3 Variable selection for a model

We have discussed different methods of variable selection for a model in the previous
chapter in detail (Chapter 16). Like other multivariable analyses, independent
variables to be selected for a logistic regression should include the risk factors of inter-
est and potential confounders, while avoiding variables with lots of missing values.
So far in this chapter, we have used the “Enter” method for logistic regression analysis.
The “Enter” method uses all the independent variables in the model specified by the
researchers. We can use the automatic selection method (stepwise method) for analysis
as well. For logistic regression, the commonly used method for the automatic selection
of variables for a model is the “Backward LR” method. If there is evidence of multicol-
linearity, you may select the “Forward LR” method for analysis. However, stepwise
methods are nowadays discouraged from being used because of the biased estimates
provided by the analysis [28, 40]. The commands for stepwise logistic regression are
provided below:
xi: sw, pr(.1) pe(.05): logistic diabetes i.sex age i.religion i.fhistory
xi: sw, pr(.1) pe(.05) forward: logistic diabetes i.sex age i.religion i.fhistory
xi: sw, pr(.1) pe(.05) forward lockterm1: logistic diabetes (age i.religion) i.sex
i.fhistory
The first command is for the backward LR method, while the second one is for the
forward LR method. The third command is for the forced entry of variables “age” and
“religion” into the model (also see section 16.2.5.1). The “pr(0.1)” indicates the
removal criteria of independent variables from the model (i.e., if the p-value of a
variable is ≥0.10, the variable will be removed from the model), while the “pe(0.05)”
indicates the inclusion (entry) criteria of variables into the model (i.e., if the p-value of
a variable is <0.05, the variable will be added to the model). You can change the crite-
ria for removal and inclusion based on your needs. The interpretation of the outputs is
the same as discussed in section 17.2.1.1.
186 Logistic Regression

Table 17.9 Predicted probabilities for religion and sex

. margins religion sex, atmeans

Adjusted predictions Number of obs = 210

Model VCE : OIM

Expression : Pr(diabetes), predict()

at : 0.sex = .6333333 (mean)
1.sex = .3666667 (mean)
age = 29.01905 (mean)
0.pulcer = .7190476 (mean)
1.pulcer = .2809524 (mean)
0.fhistory = .5428571 (mean)
1.fhistory = .4571429 (mean)
1.religion = .6 (mean)
2.religion = .2761905 (mean)
3.religion = .1238095 (mean)

------------------------------------------------------------------------------
| Delta-method
| Margin Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
religion |
MUSLIM | .0720608 .0275834 2.61 0.009 .0179982 .1261233
HINDU | .11389 .0490783 2.32 0.020 .0176983 .2100817
Christian | .0607226 .0527785 1.15 0.250 -.0427214 .1641666
|
sex |
female | .0467855 .0199011 2.35 0.019 .00778 .085791
male | .1907179 .0637455 2.99 0.003 .065779 .3156569
------------------------------------------------------------------------------

17.2.4 Incorporating interaction terms in the model

Interaction and confounding are not the same. Interaction is also called “effect modifi-
cation”. Interaction is said to be present when the strength of association (OR or RR)
of an independent variable with an outcome variable is different at different levels
(categories) of a third variable. If the strength of association is the same at different
levels of the third variable, there is no interaction.
For example, a researcher is interested in determining an association between smoking
and heart disease. He found that there is a causal association between smoking and
heart disease. To determine if there was an interaction between smoking and hyperten-
sion (the third factor), data were stratified by hypertension (have hypertension and
don't have hypertension), and the strength of association (OR) between smoking and
heart disease was calculated in each stratum. If the strength of association (OR or RR)
is the same in these two strata (hypertensive and non-hypertensive), interaction is
 Logistic Regression 187

Adjusted Predictions with 95% CIs

.3
.2
Pr(Diabetes)

.1
0
-.1

MUSLIM HINDU Christian

Religion

female male
(mean)

Figure 17.2 Marginal values of religion with the 95% CI

absent. If the strength of association (OR or RR) is different in those two strata, there
is an interaction between smoking and hypertension for the causation of heart disease
(see Section 13.3).
We can include the interaction terms in logistic regression analysis with other variables
for adjustment in the model. Suppose that we are interested in assessing if there is an
association of sex and a family history of diabetes with diabetes mellitus. We are also
keen to know if there is an interaction between sex and family history of diabetes on
the outcome. In such a situation, we need to include both the independent variables
(sex and family history of diabetes) and their interaction terms in the model. Note that
the variables for which we are looking for interaction must be included in the model
independently. To add the interaction term (sex#fhistory) into the model, use the
following command:
logistic diabetes i.sex age i.pulcer i.religion i.fhistory i.sex#i.fhistory
The above command will generate Table 17.10. The interaction term included in the
model is indicated as “male#Yes”. The table shows that the p-value for the interaction
is 0.514 (>0.05), indicating that there is no interaction between sex and family history
of diabetes for the outcome (i.e., the effect of sex on diabetes is not dependent on the
family history of diabetes). If there is an interaction (p-value <0.05), data needs to be
analyzed separately at each level of sex or family history of diabetes, depending on the
objective.
188 Logistic Regression

17.2.5 Sample size for logistic regression

Sample size is always a concern for the analysis of data. The sample size needed for a
logistic regression analysis depends on the effect size (OR) you are trying to demon-
strate and the variability of the data. It is always better to calculate the sample size
during the design phase of the study by using an appropriate formula and selecting the
relevant parameters. However, a rule-of-thumb for planning a logistic regression
analysis is that for every independent variable in the model, you need to have at least
10 outcomes (some authors recommend a minimum of 15-25 cases for each indepen-
dent variable) [6, 23].

17.2.6 Conditional logistic regression

In section 17.2.1 we have discussed the unconditional logistic regression analysis,
which is used for an unmatched design. The conditional logistic regression analysis is
done when the cases and controls are matched for one or more variables (e.g., a
matched case-control design).
Suppose that we have conducted a matched case-control study to identify the risk
factors for death due to COVID infection. In this study, the cases are matched with the
controls by gender. The outcome of interest in this study is death due to COVID (case).
The controls are those who survived the COVID infection. Each case (1:1) is matched
with a person who survived by gender. The risk factors that will be evaluated in this
study are age, religion, diabetes, and hypertension. The data for this matched case-con-
trol study is given in the data file <Data_Ca-Co_matched.dta>. We will use this data
file for conditional logistic regression analysis.
In the dataset, the variable “mID” indicates the matching ID number of cases and
controls (we must have this variable), while the variable “death” indicates whether the
subject survived (control; coded as 0) or died (case; coded as 1) due to COVID. The
codebook of variables to be used in the analysis is provided in Table 17.11. You can
also check the coding scheme (value labels) of the variables by using the following
command:
codebook mID death gender religion diabetes htn
In logistic regression analysis, we will use the variables death (outcome variable), age,
religion, diabetes, and hypertension. The variable mID is needed to specify the groups.
Since the cases and controls are matched for gender, it is meaningless to include this
variable (gender) in the analysis. To perform the conditional logistic regression, use the
 Logistic Regression 189

Table 17.10 Logistic regression with interaction terms

logistic diabetes i.sex age i.fhistory i.pulcer i.religion i.sex#fhistory

Logistic regression Number of obs = 210

LR chi2(7) = 103.03
Prob > chi2 = 0.0000
Log likelihood = -57.595829 Pseudo R2 = 0.4721

------------------------------------------------------------------------------
diabetes | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
sex |
male | 3.421403 2.510496 1.68 0.094 .8121282 14.41398
age | 1.259393 .0504362 5.76 0.000 1.16432 1.362229
|
fhistory |
Yes | 2.104785 1.4767 1.06 0.289 .5321228 8.325375
|
pulcer |
Yes | 5.802878 2.886573 3.53 0.000 2.188887 15.3838
|
religion |
HINDU | 1.692301 .9342122 0.95 0.341 .5735658 4.993121
Christian | .8405074 .8056811 -0.18 0.856 .1284131 5.501405
|
sex#fhistory |
male#Yes | 1.978321 2.069448 0.65 0.514 .2546167 15.37117
|
_cons | .0000256 .0000418 -6.49 0.000 1.05e-06 .0006254
------------------------------------------------------------------------------

following command:
clogit death age ib3.religion i.htn i.diabetes, group(mID)
The variable immediately after the command (clogit) must be the outcome (dependent)
variable. The above command will generate Table 17.12. The table shows the logistic
regression coefficients, which are difficult to interpret. We prefer to get the ORs, which
are easier to interpret. To get the ORs, use the following command after the primary
analysis:
clogit, or
This command will give you Table 17.13 with the ORs. We will interpret the outputs
provided in this table to draw conclusions.
190 Logistic Regression

Table 17.11 Codebook for the data file “Data_Ca-Co_matched.dta”

Variable name Variable label Variable codes
mID Matching ID Actual value
death Death due to COVID 0= control/alive; 1=
case/died
gender Sex of the subject 0= female; 1= male
age Age in years Actual value
religion Religion of the subjects 1= Muslim; 2= Hindu;
3= Christian
diabetes Have diabetes 0= no; 1= yes
htn Have hypertension 0= no; 1= yes

17.2.6.1 Interpretation
The interpretation of the outputs is similar to unconditional logistic regression analysis
as discussed earlier. Table 17.12 shows that the model explains 34% of the variation in
the outcome variable by the independent variables (age, religion, hypertension, and
diabetes) included in the model (Pseudo R2 = 0.3421).
Our main interest is in ORs, 95% CIs, and p-values. Table 17.13 shows the adjusted
ORs for all the variables in the model. The table shows that age (p= 0.004), religion
(Muslims compared to Christians) (p= 0.020) and having hypertension (compared to
no hypertension) (p= 0.037) are the factors significantly associated with deaths due to
COVID. Data indicates that Muslims are more likely to die compared to Christians
(adjusted OR: 13.0; 95% CI: 1.49 – 113.2; p= 0.020) after adjusting for age, hyperten-
sion, and diabetes. On the other hand, those who have hypertension are 2.56 times
more likely to die compared to those who do not have hypertension (adjusted OR:
2.56; 95% CI: 1.06 – 6.22; p= 0.037) after controlling for age, religion, and diabetes.
The interpretation of OR for age is different since the variable is entered as a continu-
ous variable (see section 17.2.1.1). In this example, the OR for age is 1.076 (95% CI:
1.02 – 1.13; p=0.004). This means that the odds of dying increase by 7.6% (1.076 – 1.0
= 0.076 or 7.6%) with each year increase in age after adjusting for religion, hyperten-
sion, and diabetes, which is statistically significant at 95% confidence level.
 Logistic Regression 191

Table 17.12 Conditional logistic regression with coefficients

. clogit death age ib3.religion i.htn i.diabetes, group(mID)

Iteration 0: log likelihood = -47.461808

Iteration 1: log likelihood = -45.703307
Iteration 2: log likelihood = -45.60376
Iteration 3: log likelihood = -45.603507
Iteration 4: log likelihood = -45.603507

Conditional (fixed-effects) logistic regression Number of obs = 200

LR chi2(5) = 47.42
Prob > chi2 = 0.0000
Log likelihood = -45.603507 Pseudo R2 = 0.3421

------------------------------------------------------------------------------
death | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
age | .0741377 .0258347 2.87 0.004 .0235026 .1247727
|
religion |
MUSLIM | 2.565603 1.104299 2.32 0.020 .401217 4.729989
HINDU | 1.922886 1.091629 1.76 0.078 -.2166663 4.062439
|
htn |
Yes | .9436792 .4514828 2.09 0.037 .0587891 1.828569
|
diabetes |
Yes | .780711 .4194145 1.86 0.063 -.0413263 1.602748
------------------------------------------------------------------------------

17.3 Analysis of cross-sectional data: Estimation of prevalence

ratio
It is common practice to analyze data from a cross-sectional study using logistic
regression when the outcome variable is dichotomous. Logistic regression analysis,
when used for cross-sectional data, provides the adjusted ORs. The ORs provided by
the analysis are actually the prevalence odds ratios (PORs). We can use the prevalence
ratio (PR) to quantify the association between exposure and outcome in cross-sectional
studies.
In cross-sectional studies, when the prevalence of an outcome (e.g., a disease) is more
than 10%, POR overestimates the PR if PR is greater than one (i.e., if PR is greater than
1, the POR will be greater than PR). Estimates for the confounding factors are also not
equivalent for these two measures. As such, PR should be used in preference to POR
while analyzing the cross-sectional data. Therefore, when the prevalence of an
outcome of our interest is greater than 10%, our objective should be to calculate the PR
rather than the POR for the association of exposure to the outcome.
192 Logistic Regression

Table 17.13 Conditional logistic regression with OR

. clogit, or

Conditional (fixed-effects) logistic regression Number of obs = 200

LR chi2(5) = 47.42
Prob > chi2 = 0.0000
Log likelihood = -45.603507 Pseudo R2 = 0.3421

------------------------------------------------------------------------------
death | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
age | 1.076955 .0278228 2.87 0.004 1.023781 1.132891
|
religion |
MUSLIM | 13.0085 14.36527 2.32 0.020 1.493641 113.2943
HINDU | 6.840675 7.467477 1.76 0.078 .8051986 58.1159
|
htn |
Yes | 2.569417 1.160048 2.09 0.037 1.060552 6.224974
|
diabetes |
Yes | 2.183024 .9155918 1.86 0.063 .959516 4.966664
------------------------------------------------------------------------------

In this section, we will discuss how to analyze cross-sectional data to get the adjusted
PR. Several methods are available to obtain the adjusted PR with Stata. They are:
a) Poisson regression;
b) Cox regression (proportional hazards analysis) with constant time; and
c) Generalized linear model (GLM).
Of the above methods, Poisson regression is the easiest. We will, however, demonstrate
the use of other methods in this section.
We will use the data file <Data_4.dta> for the analysis (consider that the data is from
a cross-sectional study). Our interest is to estimate the PR of diabetes for males com-
pared to females after controlling for age, family history of diabetes (fhistory), and
peptic ulcer (pulcer).
 Logistic Regression 193

17.3.1 Poisson regression

We can obtain the adjusted PR using the Poisson regression technique. To get the PR
for sex after controlling for age, family history of diabetes (fhistory), and peptic ulcer
(pulcer), use the following command:
poisson diabetes i.sex age i.fhistory i.pulcer, irr
This command will generate Table 17.14. The IRR (incidence risk ratio) in the table
indicates the adjusted PR since we have analyzed the cross-sectional data.
The analysis reported the IRRs of the independent variables, which are the adjusted
PRs (Table 17.14). The table shows that the PR of diabetes for males (compared to
females) is 1.93 (95% CI: 1.07 – 3.49; p= 0.028) after controlling for age, family histo-
ry of diabetes, and peptic ulcer. The findings indicate that the prevalence of diabetes
among males is 1.9 times higher than that of females after adjusting for all other
variables in the model, which is statistically significant. The interpretation of PR for
other categorical variables is similar to this. The interpretation of PR for age is differ-
ent. The data shows that the IRR (PR) for age is 1.139 (~1.40). This indicates that with
each year increase in age, the prevalence ratio of diabetes will increase by 14% (1.14
minus 1) after adjusting for other variables in the model.

17.3.2 Cox regression with constant time

To analyze the data for Cox regression with constant time, first we need to generate a
new variable (say, ctime) that will have the same value (say, 1) for all the subjects (con-
stant time). Then we will use the command “stset” to let Stata recognize the time
(ctime) and failure (diabetes) variables for the Cox regression analysis. Use the follow-
ing commands to generate the constant time variable “ctime” and set the data for Cox
regression analysis:
gen ctime=1
stset ctime, failure(diabetes)
The first command will generate a new variable “ctime” with all the values equal to 1.
The second command will set the data for survival and Cox regression analyses (Chap-
ter 19). Now, use the following command to get the PR:
stcox i.sex age i.fhistory i.pulcer
The command above will generate Table 17.15. You may use the option “, nolog” to
suppress the iterations from the outputs. The hazard ratios (Haz. Ratio) reported in the
table indicate the adjusted PRs.
194 Logistic Regression

Table 17.14 Poisson regression

. poisson diabetes i.sex age i.fhistory i.pulcer, irr

Iteration 0: log likelihood = -81.30859

Iteration 1: log likelihood = -81.307121
Iteration 2: log likelihood = -81.307121

Poisson regression Number of obs = 210

LR chi2(4) = 66.03
Prob > chi2 = 0.0000
Log likelihood = -81.307121 Pseudo R2 = 0.2888

------------------------------------------------------------------------------
diabetes | IRR Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
sex |
male | 1.938541 .5830272 2.20 0.028 1.075156 3.495254
age | 1.139111 .0281784 5.27 0.000 1.085199 1.1957
|
fhistory |
Yes | 1.446838 .4649351 1.15 0.250 .770708 2.716125
|
pulcer |
Yes | 2.359057 .7257438 2.79 0.005 1.290843 4.311251
_cons | .0011379 .0010385 -7.43 0.000 .0001902 .0068073
------------------------------------------------------------------------------

17.3.3 Generalized linear model

You can also use the generalized linear model (GLM). But the GLM with the binomial
and log-link functions may suffer from convergence problems, especially when a
continuous variable(s) is included in the model for adjustment (i.e., as an independent
variable). However, to get the PR for sex after controlling for peptic ulcer and a family
history of diabetes, use the following command:
binreg diabetes i.sex i.fhistory i.pulcer, rr nolog
We have used the option "nolog" to suppress iterations. The risk ratios in the table
(Table 17.16) indicate the adjusted PRs since we have analyzed the cross-sectional
data. Whichever method is used, the interpretation is the same as described in Section
17.3.1.
 Logistic Regression 195

Table 17.15 Cox regression with constant time

. stcox i.sex age i.fhistory i.pulcer

failure _d: diabetes

analysis time _t: ctime

Iteration 0: log likelihood = -240.61984

Iteration 1: log likelihood = -207.96213
Iteration 2: log likelihood = -207.60698
Iteration 3: log likelihood = -207.60693
Refining estimates:
Iteration 0: log likelihood = -207.60693

Cox regression -- Breslow method for ties

No. of subjects = 210 Number of obs = 210

No. of failures = 45
Time at risk = 210
LR chi2(4) = 66.03
Log likelihood = -207.60693 Prob > chi2 = 0.0000

------------------------------------------------------------------------------
_t | Haz. Ratio Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
sex |
male | 1.938541 .5830272 2.20 0.028 1.075156 3.495254
age | 1.139111 .0281784 5.27 0.000 1.085199 1.1957
|
fhistory |
Yes | 1.446838 .4649351 1.15 0.250 .770708 2.716125
|
pulcer |
Yes | 2.359057 .7257438 2.79 0.005 1.290843 4.311251
------------------------------------------------------------------------------
196 Logistic Regression

Table 17.16 Generalized linear model (GLM)

. binreg diabetes i.sex i.fhistory i.pulcer , rr nolog

Generalized linear models No. of obs = 210

Optimization : MQL Fisher scoring Residual df = 206
(IRLS EIM) Scale parameter = 1
Deviance = 172.3863397 (1/df) Deviance = .8368269
Pearson = 199.3341071 (1/df) Pearson = .9676413

Variance function: V(u) = u*(1-u) [Bernoulli]

Link function : g(u) = ln(u) [Log]
BIC = -929.1178
------------------------------------------------------------------------------
| EIM
diabetes | Risk Ratio Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
sex |
male | 2.331071 .5153874 3.83 0.000 1.51133 3.595436
|
fhistory |
Yes | 2.233878 .5522354 3.25 0.001 1.376051 3.626472
|
pulcer |
Yes | 3.044956 .7708366 4.40 0.000 1.853948 5.001086
_cons | .0574094 .0163787 -10.02 0.000 .03282 .1004216
------------------------------------------------------------------------------
 18
Multinominal Logistic Regression

Multinominal logistic regression is an extension of binary logistic regression. It is used

when the dependent (outcome) categorical variable has more than two levels (catego-
ries) that cannot be arranged into an order, e.g., health-seeking behavior (did not seek
treatment, received treatment from the village doctors, received treatment from the
pharmacists) or marital status (unmarried, married, divorced, separated).
Suppose that a researcher has conducted a study to identify the factors associated with
health-seeking behavior of mothers for diarrhea among their children. Here the depen-
dent variable is “health-seeking behavior”, which has three levels (Table 18.1). The
independent (explanatory) variables included in the study are maternal age, religion,
sex of the child, and severity of diarrhea. The coding scheme of all these variables is
provided in Table 18.1.
Data for this study is provided in the data file <Data_5 multinominal>. We will use
this data to demonstrate multinominal logistic regression analysis.

Table 18.1 Codebook of data file “Data_5 multinominal”

Variable name Variable label Variable codes
age Maternal age in years Actual value
religion Religion of the subjects 1= Muslim; 2= Others
behavior Health seeking behavior 1= Did not receive treatment;
2= Treated by a village doctor;
3= Treated by a pharmacist
sex Sex of the child 0= female; 1= male
sdiar Severity of diarrhea 0= Not severe; 1= Severe
198 Multinominal Logistic Regression

For multinominal logistic regression analysis, it is necessary to select a reference group

(category) of the outcome variable for comparison. We will select the first category
(did not receive treatment) of the outcome variable as the reference category (base) for
our analysis. The analysis will, therefore, provide estimates for the categories “treated
by village doctors” compared to “did not receive treatment”, and “treated by pharma-
cists” compared to “did not receive treatment”. In this example, we will include two
categorical variables (religion and severity of diarrhoea) and a quantitative variable
(age) as explanatory variables in the model. To do the multinominal logistic regression
analysis, use the following command:
mlogit behavior ib2.religion i.sdiar age
mlogit behavior ib2.religion i.sdiar age, rrr
mlogit behavior ib2.religion i.sdiar age, base(2) rrr
The first command will provide the regression coefficients (outputs are not shown).
The second command will provide the relative risk ratios (RRRs) rather than the coef-
ficients. For both these commands (first and second), the first category of the outcome
variable (did not receive treatment) will be the comparison group (Stata considers the
first category as the comparison group by default). The third command will provide the
RRRs with the second category of the outcome variable as the comparison group (indi-
cated by the “base(2)” option). The prefix ".ib2", used for religion, is to indicate the
second category (other religion; coded as 2) of religion to be the comparison group.
Since it is easier to interpret the RRRs, we have shown the outputs of the second com-
mand in Table 18.2. The RRRs are the exponentials of the coefficients that we get by
using the first command.

18.1 Interpretation
Table 18.2 shows the results of the analysis of the second command. The first iteration
(iteration 0) indicates the log likelihood of the null model (i.e., a model without any
independent variable). The table shows that the log likelihood has increased [from
-223.23 (Iteration 0) to -174.89 (Iteration 5)] with the inclusion of independent
variables in the model. The log likelihood chi-square p-value (Prob > Chi2) as provid-
ed in the table is significant (p= 0.000). A significant p-value indicates that the model
is significantly better than the null model; i.e., the addition of independent variables
(religion, severity of diarrhea, and age) in the model has improved the ability to predict
the outcome variable compared to the null model.
 Multinominal Logistic Regression 199

Table 18.2 Multinominal logistic regression

. mlogit behavior ib2.religion i.sdiar age, rrr

Iteration 0: log likelihood = -223.23547

Iteration 1: log likelihood = -181.2718
Iteration 2: log likelihood = -175.03672
Iteration 3: log likelihood = -174.89221
Iteration 4: log likelihood = -174.89173
Iteration 5: log likelihood = -174.89173

Multinomial logistic regression Number of obs = 210

LR chi2(6) = 96.69
Prob > chi2 = 0.0000
Log likelihood = -174.89173 Pseudo R2 = 0.2166

-------------------------------------------------------------------------------------------
behavior | RRR Std. Err. z P>|z| [95% Conf. Interval]
--------------------------+----------------------------------------------------------------
Did_not_receive_treatment | (base outcome)
--------------------------+----------------------------------------------------------------
Treated_by_vill_doc |
religion |
Muslim | 2.627298 1.324947 1.92 0.055 .9777956 7.059443
|
sdiar |
severe | 4.8923 2.469855 3.14 0.002 1.818812 13.15947
age | 1.271572 .0517161 5.91 0.000 1.174145 1.377083
_cons | .0000683 .0001029 -6.37 0.000 3.56e-06 .0013092
--------------------------+----------------------------------------------------------------
Treated_by_pharmacist |
religion |
Muslim | 2.080179 .6647571 2.29 0.022 1.111948 3.891498
|
sdiar |
severe | .8920665 .3586054 -0.28 0.776 .4057134 1.961441
age | 1.002805 .0242082 0.12 0.908 .9564624 1.051392
_cons | .5782204 .395847 -0.80 0.424 .1511349 2.212188
-------------------------------------------------------------------------------------------

The Pseudo R-squared (Pseudo R2) value indicates how much variation in the depen-
dent variable can be explained by the independent variables in the model. The results
show that 21.6% of the variation in the dependent variable can be explained by the
independent variables together in the model (religion, severity of diarrhea, and age).
However, the readers should interpret the pseudo R-squared value cautiously as it is
not equivalent to the R-squared value that we get in linear regression analysis (Sections
16.1.2 and 16.2.3).
The main output has two parts, labelled with the categories of outcome variable (treat-
ed by village doctors and treated by pharmacists). The analysis provided the RRRs
with their 95% CIs and p-values.
The first half of the table has the results for “treated by village doctors” compared
to “did not receive treatment”. The results indicate that Muslims (compared to other
200 Multinominal Logistic Regression

religions) are more likely to receive treatment from village doctors after controlling for
mothers’ age and severity of diarrhea, but the association is not statistically significant
(adjusted RRR: 2.62; 95% CI: 0.97 – 7.05; p=0.055). However, severity of diarrhea
(i.e., if the baby has severe diarrhea) is significantly associated with seeking treatment
from village doctors after controlling for mothers’ age and religion (adjusted RRR:
4.89; 95% CI: 1.81 – 13.15; p=0.002).
For the quantitative variables, an RRR greater than one indicates an increased likeli-
hood of the response category (treated by village doctors) compared to the reference
category (did not receive treatment). The results show that with the increase in moth-
ers’ age, it is significantly more likely to receive treatment from the village doctors
after adjusting for religion and severity of diarrhea (adjusted RRR: 1.27; 95% CI: 1.17
– 1.37; p=0.000) [in other words, the likelihood of receiving treatment from village
doctors increases by 27% with each year increase in mother’s age].
The second half of the table shows the results for "treated by pharmacists" compared
to "did not receive treatment". The interpretations are similar to those mentioned
above. The results show that only religion (being Muslim) is significantly associated
with seeking treatment from pharmacists after adjusting for maternal age and severity
of diarrhea (adjusted RRR: 2.08; 95% CI: 1.11 – 3.89; p= 0.022).

18.2 Post-estimation commands

After performing the regression analysis, you can get the predicted probabilities of the
outcome variable by using the “margins” command, such as:
margins religion, atmeans predict (outcome(1))
margins religion, atmeans predict (outcome(2))
margins religion, atmeans predict (outcome(3))
The above commands will provide the predicted values for Muslims and other
religions for each level of the outcome variable (Table 18.3). The marginal (Margin)
values indicate the probabilities of the outcome. For example, the marginal value for
Muslims is 0.390 for the first category of the outcome variable (i.e., did not receive
treatment). This indicates that the probability of not seeking treatment for being a Mus-
lim is 0.39 (which is 0.58 for other religions), considering the average values of other
independent variables in the model. The interpretation of other outputs is similar to
this.
 Multinominal Logistic Regression 201

Table 18.3 Marginal values

. margins religion, atmeans predict (outcome(1))

Adjusted predictions Number of obs = 210

Model VCE : OIM

Expression : Pr(behavior==Did_not_receive_treatment), predict(outcome(1))

at : 1.religion = .5571429 (mean)
2.religion = .4428571 (mean)
0.sdiar = .7190476 (mean)
1.sdiar = .2809524 (mean)
age = 29.01905 (mean)

------------------------------------------------------------------------------
| Delta-method
| Margin Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
religion |
Muslim | .3902659 .0514549 7.58 0.000 .2894162 .4911156
Others | .5810089 .0563714 10.31 0.000 .470523 .6914949
------------------------------------------------------------------------------

. margins religion, atmeans predict (outcome(2))

Adjusted predictions Number of obs = 210

Model VCE : OIM

Expression : Pr(behavior==Treated_by_vill_doc), predict(outcome(2))

at : 1.religion = .5571429 (mean)
2.religion = .4428571 (mean)
0.sdiar = .7190476 (mean)
1.sdiar = .2809524 (mean)
age = 29.01905 (mean)

------------------------------------------------------------------------------
| Delta-method
| Margin Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
religion |
Muslim | .1166562 .0368918 3.16 0.002 .0443495 .1889628
Others | .0661029 .0283829 2.33 0.020 .0104735 .1217323
------------------------------------------------------------------------------

. margins religion, atmeans predict (outcome(3))

Adjusted predictions Number of obs = 210

Model VCE : OIM

Expression : Pr(behavior==Treated_by_pharmacist), predict(outcome(3))

at : 1.religion = .5571429 (mean)
2.religion = .4428571 (mean)
0.sdiar = .7190476 (mean)
1.sdiar = .2809524 (mean)
age = 29.01905 (mean)

------------------------------------------------------------------------------
| Delta-method
| Margin Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
religion |
Muslim | .4930779 .0527806 9.34 0.000 .3896299 .5965259
Others | .3528881 .0548775 6.43 0.000 .2453302 .460446
------------------------------------------------------------------------------
202 Multinominal Logistic Regression

We can use the command "margins" to plot the predicted probabilities of each category
of the outcome variable (behavior) by religion. Stata will create the plots based on the
last margins command used. We can also combine 3 margins plots into a single figure
by using the command "graph combine" (Fig 18.1). Use all the following commands
successively:
margins religion, atmeans predict (outcome(1))
marginsplot, name (not_treated)
margins religion, atmeans predict (outcome(2))
marginsplot, name (treated_vdoc)
margins religion, atmeans predict (outcome(3))
marginsplot, name (treated_pharm)
graph combine not_treated treated_vdoc treated_pharm
Pr(Behavior==Did_Not_Receive_Treatment)

Adjusted Predictions of religion with 95% CIs Adjusted Predictions of religion with 95% CIs
Pr(Behavior==Treated_By_Vill_Doc)
.7

.2
.15
.6
.5

.1
.05
.4
.3

Muslim Others Muslim Others

Religion Religion
Pr(Behavior==Treated_By_Pharmacist)

Adjusted Predictions of religion with 95% CIs

.6
.5
.4
.3
.2

Muslim Others
Religion

Figure 18.1 Marginal values of different categories of religion for seeking treatment
 19
Survival Analysis

There are situations when researchers are interested in knowing the progress of a
patient from a specific point in time (e.g., from the point of diagnosis or initiation of
treatment) until the occurrence of a certain outcome, such as death or recurrence of any
event (e.g., recurrence of cancer). The prognosis of a condition is commonly assessed
by estimating the: a) Median survival time and b) Cumulative probability of survival
after a certain time interval (e.g., 5-year or 3-year or others).
For instance, a researcher may be interested in determining the median survival time of
colonic cancer if the patient is treated (or not treated), and the estimated probability
that a patient with colonic cancer may survive for more than 5 years (5-year cumulative
survival probability). The methods employed to answer the above questions in a
follow-up study are known as survival analysis (or lifetable analysis) methods.
Survival analysis is done in follow-up studies, where subjects are usually followed
over a specified period of time and the focus is on the time at which the event of inter-
est occurs. To do the survival analysis, we need to have data (information) from each
of the subjects, at least on the following variables:
• Time: It is the length of time the patient was observed in the study (called
“survival time”). Time can be measured in days, weeks, months, years or other
units of time.
• Outcome: Whether the patient developed the outcome of interest (event) during
the study period, or whether the patient was either lost to follow-up or remained
alive at the end of the study (censored); and
• Treatment group: Which treatment (e.g., treatment A or B) did the patient receive
in the study (optional)?
204 Survival Analysis

Survival time is of two types – a) Censored time; and b) Event time. The censored time
is the amount of time contributed by the:
a) Patients who did not develop the outcome and remained in the study up to the
end of the study period, or
b) Patients who were lost to follow-up due to any reason, such as migration or
withdraw, or
c) Patients who developed the outcome for reasons other than the disease of interest
(e.g., died in a car accident)
On the other hand, the event time is the amount of time contributed by the patients who
developed the outcome of interest during the study period. In survival analysis, the
outcome measure of interest is survival time, which is a mixture of event time and
censored time.
If we have the above information, it is possible to estimate the median survival times
and cumulative survival probabilities for two or more treatment groups for compari-
son. Such a comparison allows us to answer the question, “which treatment delays the
time of occurrence of an event?” The method commonly used to analyze survival-time
data is the Kaplan-Meier method, and Stata can be used for the analysis of such data.
Use the data file <Data_survival_4.dta> for practice. The codebook for the data file
is provided in Table 19.1. Note that in this data, the event of our interest is death.

Table 19.1 Codebook for the data file “Data_survival_4.dta”

Variable name Variable label Variable codes
time Survival time in days Actual value in days
outcome Survival status 0= Censored; 1= Died/Event
treatment Treatment group 0= Placebo; 1= New treatment
age Age in years Actual value
sex Sex 0= Male; 1= Female

19.1 Survival analysis: Kaplan-Meier method

Suppose that a researcher has conducted a follow-up study (clinical trial) on patients
with acute heart attack (myocardial infraction) to determine the effectiveness of a new
drug (n=22) compared to a placebo (n=22) in reducing the time to death. The patients
 Survival Analysis 205

were followed until time to death or up to six months (180 days), whichever came first.
The outcome of interest in this study was the time to death (event) due to acute heart
attack. The objective was to assess whether the "new treatment" delays (increases) the
time to event (death) compared to placebo among patients with heart attack. Data from
this study is provided in the data file <Data_survival_4.dta>, which includes the
following necessary variables:
• Time: The variable “time” indicates the amount of time each patient has spent
in the study in days;
• Treatment: It specifies which treatment the patient received in this clinical trial
(coded as 0= received placebo; 1= received new treatment);
• Outcome (event): Whether the patient developed the event, i.e., died or not
(coded as 0= censored/did not die; 1= died).

Assumptions
• The probability of outcome is similar among the censored (lost to follow-up)
and under-observation individuals;
• There is no secular trend over the calendar period;
• Risk is uniform during the interval; and
• Loss to follow-up is uniform over the interval.

19.1.1 Preparing data for analysis

Before conducting the survival analysis, we need to prepare the dataset so that Stata
can automatically recognize the time variable and the event variable (censored/event)
during analysis. In our dataset, the time variable is “time” and the event variable is
“outcome”. The event variable must be coded as 1= event and 0= censored. Use the
following command to prepare the data for survival analysis:
stset time, failure(outcome)
When the command ”stset” is used, Stata generates some new variables, like “_st, _d,
_t, and _t0”. You can see them at the bottom of the variables window. If you save the
data file after using the “stset” command, you don’t need to prepare the data file again
for survival analysis in the subsequent sessions. But if you don’t save the data file, you
need to prepare it before survival analysis every time during the subsequent sessions.
In fact, we need to set the data using the “stset” command for any analysis that uses the
commands beginning with “st…”, for example, stdescribe, stcox, sts test, or others.
206 Survival Analysis

19.1.2 Commands for Kaplan-Meyer method

Before analyzing data for survival analysis, let us check our data for the number of
subjects in each treatment group, including the number of events that occurred. To get
the summary of the data, use the following command:
tab2 treatment outcome, row
This command will generate a two-by-two table (Table 19.2) of the treatment group
and outcome. The table shows that there are 44 subjects enrolled in this study (22 in
both the placebo and new treatment groups). In total, 27 patients died (events) during
the study period, of which 16 (59.26%) were in the placebo group and 11 (40.74%)
were in the new treatment group.

Table 19.2 Area under the ROC curve with 95% CI

. tab2 treatment outcome, col

-> tabulation of treatment by outcome

+-------------------+
| Key |
|-------------------|
| frequency |
| column percentage |
+-------------------+

Treatment | Survival status

group | Censored Died | Total
--------------+----------------------+----------
Placebo | 6 16 | 22
| 35.29 59.26 | 50.00
--------------+----------------------+----------
New treatment | 11 11 | 22
| 64.71 40.74 | 50.00
--------------+----------------------+----------
Total | 17 27 | 44
| 100.00 100.00 | 100.00

The primary objectives of survival analysis are to get the median survival time and
cumulative survival probability (survival functions) for each treatment group and the
significance of the difference (log-rank test) in survival functions between the treat-
ment groups. Once the data set is prepared (by the "stset" command), use the following
commands to get the median survival time and its 95% CI by treatment group. The
option "by(treatment)" will present the results separately for the placebo and new treat-
ment groups.
 Survival Analysis 207

stsum, by(treatment)
stci, by(treatment)
The first command will provide the median survival times of the placebo and new
treatment groups, while the second command will provide the same but with their 95%
CIs (Table 19.3).

Table 19.3 Median survival time by treatment group

. stsum, by(treatment)

failure _d: outcome

analysis time _t: time

| incidence no. of |------ Survival time -----|

treatm~t | time at risk rate subjects 25% 50% 75%
---------+---------------------------------------------------------------------
Placebo | 1422 .0112518 22 22 40 .
New trea | 2409 .0045662 22 89 146 .
---------+---------------------------------------------------------------------
total | 3831 .0070478 44 29 89 .

. stci, by(treatment)

failure _d: outcome

analysis time _t: time

| no. of
treatment | subjects 50% Std. Err. [95% Conf. Interval]
-------------+-------------------------------------------------------------
Placebo | 22 40 12.89864 22 71
New trea | 22 146 10.79461 89 .
-------------+-------------------------------------------------------------
total | 44 89 21.23218 41 168

To get the survival functions disaggregated by treatment group, use the following com-
mands:
sts list, by(treatment)
sts list, by(treatment) compare
The first command will provide a long table showing the survival functions (cumula-
tive survival probabilities) at different time points (Table 19.4), while the second com-
mand will provide a table showing a comparison of survival functions between placebo
and new treatment groups (Table 19.5).
208 Survival Analysis

Table 19.4 Survival functions at different time points by treatment group

. sts list, by (treatment)

failure _d: outcome

analysis time _t: time

Beg. Net Survivor Std.

Time Total Fail Lost Function Error [95% Conf. Int.]
-------------------------------------------------------------------------------
Placebo
2 22 1 0 0.9545 0.0444 0.7187 0.9935
3 21 1 0 0.9091 0.0613 0.6830 0.9765
4 20 1 0 0.8636 0.0732 0.6344 0.9539
7 19 1 0 0.8182 0.0822 0.5853 0.9276
10 18 1 0 0.7727 0.0893 0.5374 0.8985
22 17 1 0 0.7273 0.0950 0.4910 0.8671
28 16 1 0 0.6818 0.0993 0.4462 0.8338
29 15 1 0 0.6364 0.1026 0.4029 0.7988
32 14 1 0 0.5909 0.1048 0.3610 0.7621
37 13 1 0 0.5455 0.1062 0.3207 0.7239
40 12 1 0 0.5000 0.1066 0.2818 0.6843
41 11 1 0 0.4545 0.1062 0.2444 0.6433
54 10 1 0 0.4091 0.1048 0.2085 0.6007
61 9 1 0 0.3636 0.1026 0.1743 0.5567
63 8 1 0 0.3182 0.0993 0.1418 0.5111
71 7 1 0 0.2727 0.0950 0.1112 0.4637
127 6 0 1 0.2727 0.0950 0.1112 0.4637
140 5 0 1 0.2727 0.0950 0.1112 0.4637
146 4 0 1 0.2727 0.0950 0.1112 0.4637
158 3 0 1 0.2727 0.0950 0.1112 0.4637
167 2 0 1 0.2727 0.0950 0.1112 0.4637
180 1 0 1 0.2727 0.0950 0.1112 0.4637
New treatment
2 22 1 0 0.9545 0.0444 0.7187 0.9935
6 21 1 0 0.9091 0.0613 0.6830 0.9765
12 20 1 0 0.8636 0.0732 0.6344 0.9539
54 19 1 0 0.8182 0.0822 0.5853 0.9276
56 18 0 1 0.8182 0.0822 0.5853 0.9276
68 17 1 0 0.7701 0.0904 0.5325 0.8973
89 16 1 0 0.7219 0.0967 0.4822 0.8645
96 15 2 0 0.6257 0.1051 0.3883 0.7926
125 13 0 1 0.6257 0.1051 0.3883 0.7926
128 12 0 1 0.6257 0.1051 0.3883 0.7926
131 11 0 1 0.6257 0.1051 0.3883 0.7926
140 10 0 1 0.6257 0.1051 0.3883 0.7926
141 9 0 1 0.6257 0.1051 0.3883 0.7926
143 8 1 0 0.5475 0.1175 0.2979 0.7410
145 7 0 1 0.5475 0.1175 0.2979 0.7410
146 6 1 0 0.4562 0.1285 0.2047 0.6782
148 5 0 1 0.4562 0.1285 0.2047 0.6782
162 4 0 1 0.4562 0.1285 0.2047 0.6782
168 3 1 0 0.3041 0.1509 0.0676 0.5910
173 2 0 1 0.3041 0.1509 0.0676 0.5910
180 1 0 1 0.3041 0.1509 0.0676 0.5910
-------------------------------------------------------------------------------

We need to use a statistical test to objectively assess the significance of the difference
in survival functions between the treatment groups. The most commonly used statisti-
cal test is the log-rank test. However, there are alternatives to this test. They are the
Tarone-Ware and Peto tests. The commands to get these test statistics are:
 Survival Analysis 209

Table 19.5 Comparison of survival functions at different time points by treatment

groups
. sts list, by (treatment) compare

failure _d: outcome

analysis time _t: time

Survivor Function
treatment Placebo New treat
----------------------------------
time 2 0.9545 0.9545
24 0.7273 0.8636
46 0.4545 0.8636
68 0.3182 0.7701
90 0.2727 0.7219
112 0.2727 0.6257
134 0.2727 0.6257
156 0.2727 0.4562
178 0.2727 0.3041
200 . .
----------------------------------

sts test treatment

sts test treatment, tware
sts test treatment, peto
The above commands will provide the results of the log-rank, Tarone-Ware, and Peto
tests, respectively (Table 19.6).
The cumulative survival probabilities are usually portrayed visually by a graph called
the survival curve. You can generate the survival curve by using the following com-
mand:
sts graph, by(treatment)
sts graph, by(treatment) ci
The first command will display the survival curve by treatment groups, as shown in
Figure 19.1. The second command will also display the curve, but with 95% CIs. You
can also generate the cumulative incidence curve (1 – cumulative survival) by using
the following command (Figure 19.2):
sts graph, by(treatment) failure

19.1.3 Interpretation
In total, 44 subjects were enrolled in this study, of which 22 received a placebo and
210 Survival Analysis

Table 19.6 Significance tests for survival functions

. sts test treatment

failure _d: outcome

analysis time _t: time

Log-rank test for equality of survivor functions

| Events Events
treatment | observed expected
--------------+-------------------------
Placebo | 16 10.62
New treatment | 11 16.38
--------------+-------------------------
Total | 27 27.00

chi2(1) = 4.66
Pr>chi2 = 0.0309

. sts test treatment, tware

failure _d: outcome

analysis time _t: time

Tarone-Ware test for equality of survivor functions

| Events Events Sum of

treatment | observed expected ranks
--------------+--------------------------------------
Placebo | 16 10.62 33.847597
New treatment | 11 16.38 -33.847597
--------------+--------------------------------------
Total | 27 27.00 0

chi2(1) = 6.07
Pr>chi2 = 0.0138

. sts test treatment, peto

failure _d: outcome

analysis time _t: time

Peto-Peto test for equality of survivor functions

| Events Events Sum of

treatment | observed expected ranks
--------------+--------------------------------------
Placebo | 16 10.62 4.3863004
New treatment | 11 16.38 -4.3863004
--------------+--------------------------------------
Total | 27 27.00 0

chi2(1) = 6.03
Pr>chi2 = 0.0141

another 22 received the new treatment. There were a total of 27 deaths during the study
period, among which 16 (59.26%) were in the placebo group and 11 (40.74%) were in
the new treatment group (Table 19.2).
 Survival Analysis 211

1.00 Kaplan-Meier survival estimates Kaplan-Meier failure estimates

1.00
0.75

0.75
0.50

0.50
0.25

0.25
0.00

0.00
0 50 100 150 200 0 50 100 150 200
analysis time analysis time

treatment = Placebo treatment = New treatment treatment = Placebo treatment = New treatment

Figure 19.1 Kaplan-Meier survival curve Figure 19.2 Cumulative incidence curve

Table 19.3 shows the median survival times for both the placebo and new treatment
groups, including their 95% CIs. The median survival time is the time when the cumu-
lative survival probability is 0.50 (i.e., the time when 50% of the patients develop the
event). The table indicates that the median survival time, if a patient is in the placebo
group, is 40 days (95% CI: 22 to 71), while it is 146 days (95% CI: 89 to .), if a patient
is in the new treatment group. This means that the new treatment increases the survival
time, i.e., the new treatment is associated with a longer time-to-event (and the placebo
is associated with a shorter time-to-event). We, therefore, conclude that an individual
will live longer if s/he receives the new treatment compared to the placebo.
Table 19.4 shows the cumulative survival probability (survivor function) at different
points in time in the placebo and new treatment groups. In this table, we can see that
the cumulative survival probability at the end of 71 days (in the time column), in the
placebo group, is 0.272 (27.2%). Since there is no death after that, the cumulative
survival probability at the end of 180 days will be the same (0.272).
On the other hand, the cumulative survival probability is 0.304 (30.4%) at the end of
168 days for the patients who were in the new treatment group. As there is no death
after that, the cumulative survival probability at the end of 180 days will be the same
(0.304). In the new treatment group, the cumulative survival probability at the end of
71 days (68 days in the table) is about 0.770 (77.0%), which is much higher than the
placebo group (0.273). This indicates that the probability of surviving at the end of 71
days is higher among the patients who received the new treatment compared to place-
bo. This also indicates the benefit of the new treatment (i.e., the new treatment is better
than the placebo). A comparison of the survival functions at different points in time is
212 Survival Analysis

provided in Table 19.5. For instance, at the end of 90 days, the cumulative probability
of survival is 0.2727 in the placebo group compared to 0.7219 in the new treatment
group.
However, if we consider the cumulative survival probability of patients in both groups
at the end of 180 days, the probabilities are not that different (0.272 in the placebo
group and 0.304 in the new treatment group). This information suggests that though
survival experiences are significantly different between the treatment groups (as
indicated by the log rank test), the difference in survival probability at the end of 180
days is small.
Now, the question is whether the survival experiences of both these groups in the popu-
lation are different or not. For an objective comparison of survival experiences in two
groups, it is desirable to use a statistical method that will tell us whether the difference
in survival experiences in the population is statistically significant or not. Here, the
null hypothesis is that “the survival experiences in the placebo and new treatment
groups are the same in the population”. Such a null hypothesis can be tested by the
log-rank test. We have done the log-rank test and the results are provided in Table 19.6.
The p-value (Pr>chi2) of the test is 0.030, which is <0.05. This indicates that survival
experiences of both these groups in the population are not the same. In other words, it
tells us that the survival probability is better (since the median survival time is higher
in the new treatment group) if the patient is in the new treatment group compared to the
placebo group (i.e., the new treatment is more effective/better than placebo in improv-
ing the patients’ survival).
There are alternative procedures for testing the null hypothesis that the two survival
curves are identical. They are the Breslow test, the Tarone-Ware test, and the Peto test.
Here, we have performed the Tarone-Ware and Peto tests, and the results are shown in
Table 19.6. The log-rank test ranks all the deaths equally, while the other tests give
more weight to early deaths.

Survival curve: The cumulative survival probability is usually portrayed visually by a

graph called the survival curve (Fig 19.1). The steps in the graph represent the times
when events (deaths or any other event of interest) occurred. The graph allows us to
represent visually the median survival time and the cumulative survival probability for
any specific time period (e.g., 30-day, 60-day, or 90-day cumulative survival probabili-
ty). In general, the line above indicates the better survival probability. We can see that
the line for the new treatment is above the line for the placebo, indicating that the new
treatment delays (increases) the time to event compared to the placebo.
 Survival Analysis 213

19.2 Cox regression

Cox regression is also called proportional hazards analysis. In the previous section
(Section 19.1), we discussed the survival analysis using the Kaplan-Meier method.
Like other regression analyses (e.g., multiple linear regression and logistic regression),
Cox regression is a multivariable analysis technique where the dependent measure is a
mixture of time-to-event and censored-time observations. Cox regression is commonly
done in follow-up studies (e.g., randomized trials) to assess the prognosis. Cox regres-
sion with constant time can be used for the analysis of cross-sectional data to estimate
the prevalence ratio, which is discussed in Section 17.3.2. For the Cox regression
analysis, we will use the same data file (Data_survival_4.dta) that was used for the
survival analysis.
Returning to our previous example (survival analysis; Section 19.1), where we
analyzed the data to assess the effectiveness of a new treatment against a placebo. Our
objective was to determine whether the new treatment delays the time-to-death com-
pared to the placebo among patients with heart disease. We found that the new treat-
ment significantly delayed the time-to-death compared to the placebo, as indicated by
the higher median survival time and cumulative survival probability, and a significant
log-rank test. However, the effectiveness of the new treatment might be influenced
(confounded) by other factors, such as age, hypertension, diabetes, or other character-
istics. All these factors, therefore, need to be controlled during analysis to assess the
effectiveness of the new treatment. Cox regression is a statistical method that is used
to control the confounding factors (categorical, continuous, or discrete covariates) that
may influence the effectiveness of a treatment.
Cox regression gives us the hazard ratio (HR) after adjusting the variables included in
the model, which is analogous to relative risk (RR). A hazard ratio (also called a
relative hazard) is the ratio of the hazard rate if the individuals are exposed (e.g., to a
new treatment) compared to the individuals not exposed (e.g., to placebo). In Cox
regression, the dependent variable is the log of hazard.

19.2.1 Commands
In Cox regression analysis, we will consider the variables age and sex for adjustment.
The command for Cox regression is “stcox” and it reports the hazard ratios (HRs). We
only use the independent variables with the “stcox” command. In survival analysis,
once the data are set with the “stset” command (see Section 19.1.1), Stata automatical-
ly recognizes the time and event variables. To perform the Cox regression analysis, use
214 Survival Analysis

the following command:

stcox i.treatment ib1.sex age
stcox ib1.treatment ib1.sex age
stcox ib1.treatment ib1.sex age, nolog nohr
The first command will provide HR for the new treatment compared to the placebo
(Table 19.7). The second command will report the HR for the placebo group compared
to the new treatment group (Table 19.8). In the third command, we have used the
options “nolog” and “nohr” to suppress the iteration history and to get the coefficients
instead of HRs, respectively. We have used the prefix “.ib1” for the variables “treat-
ment” (in the second command) and “sex” to indicate the new treatment and females
as the comparison groups (since new treatment and females are both coded as 1).
You may prefer using the second command for the analysis. Because if the new treat-
ment is the comparison group, we expect an HR greater than one for the placebo group
(since we assume that the new treatment is better than placebo), which is easier to
interpret.
Stepwise methods are also available for modeling with Cox regression. For the back-
ward stepwise method with a removing criteria of p≥0.2 and an adding (entry) criteria
of p<0.1, use the following command:
xi: stepwise, pr(.2) pe(.1): stcox ib1.treatment ib1.sex age
For the forward stepwise method with a removing criteria of p≥0.2 and an adding
(entry) criteria of p<0.10, use the following command:
xi: stepwise, pr(.2) pe(.10) forward: stcox ib1.treatment ib1.sex age
The outputs of these commands are not provided. For further information, see Section
16.2.5. You can generate the survival curve for the treatment groups after adjusting for
age and sex by using the following command:
sts graph, by(treatment) adjust(age sex)
This command will display the survival curve for the treatment groups after controlling
for age and sex (Figure 19.3). You can also get the cumulative incidence curve adjusted
for age and sex by using the following command (Fig 19.4):
sts graph, by(treatment) failure adjust(age sex)
 Survival Analysis 215

Table 19.7 Cox regression with placebo as comparison group

. stcox i.treatment ib1.sex age

failure _d: outcome

analysis time _t: time

Iteration 0: log likelihood = -88.962207

Iteration 1: log likelihood = -84.097071
Iteration 2: log likelihood = -83.983381
Iteration 3: log likelihood = -83.983329
Refining estimates:
Iteration 0: log likelihood = -83.983329

Cox regression -- Breslow method for ties

No. of subjects = 44 Number of obs = 44

No. of failures = 27
Time at risk = 3831
LR chi2(3) = 9.96
Log likelihood = -83.983329 Prob > chi2 = 0.0189

--------------------------------------------------------------------------------
_t | Haz. Ratio Std. Err. z P>|z| [95% Conf. Interval]
---------------+----------------------------------------------------------------
treatment |
New treatment | .3665283 .1678311 -2.19 0.028 .1493989 .8992233
|
sex |
male | 2.433823 1.060213 2.04 0.041 1.036315 5.715921
|
age | 1.008859 .0229496 0.39 0.698 .9648665 1.054857
--------------------------------------------------------------------------------

19.2.2 Interpretation
We have analyzed the data of 44 subjects (22 in the new treatment group and 22 in the
placebo group). The variables included in the analysis are “treatment”, “sex”, and
“age” to get the estimated effect of the new treatment (compared to placebo) after
adjusting for sex and age. Table 19.7 shows the results of the Cox regression analysis.
The table at the beginning shows the iteration history. The first iteration (iteration 0)
indicates the log likelihood of the null model (i.e., a model without any independent
variable). The table shows that the log likelihood has increased (from -88.96 to -83.98)
with the inclusion of independent variables in the model. The log likelihood ratio
chi-square p-value (Prob > chi2), as shown in the table, is significant (p= 0.018). A
significant p-value indicates that the addition of independent variables in the model has
improved the ability to predict the outcome compared to the null model.
216 Survival Analysis

Table 19.8 Cox regression with new treatment as comparison group

. stcox ib1.treatment ib1.sex age

failure _d: outcome

analysis time _t: time

Iteration 0: log likelihood = -88.962207

Iteration 1: log likelihood = -84.097071
Iteration 2: log likelihood = -83.983381
Iteration 3: log likelihood = -83.983329
Refining estimates:
Iteration 0: log likelihood = -83.983329

Cox regression -- Breslow method for ties

No. of subjects = 44 Number of obs = 44

No. of failures = 27
Time at risk = 3831
LR chi2(3) = 9.96
Log likelihood = -83.983329 Prob > chi2 = 0.0189

------------------------------------------------------------------------------
_t | Haz. Ratio Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
treatment |
Placebo | 2.728302 1.249273 2.19 0.028 1.112071 6.693488
|
sex |
male | 2.433823 1.060213 2.04 0.041 1.036315 5.715921
|
age | 1.008859 .0229496 0.39 0.698 .9648665 1.054857
------------------------------------------------------------------------------

Table 19.7 shows that out of a total of 44 subjects included in the analysis, 27 died
(number of failures). In our analysis, placebo is the comparison group for the variable
“treatment” and females are the comparison group for “sex”. We will, therefore, get the
HRs for the new treatment group compared to placebo and for the males compared to
females.
The table (Table 19.7) shows the HRs and their corresponding p-values (P>|z|) with the
95% CIs. The HR for the new treatment is 0.366 (95% CI: 0.149 to 0.899) with a p-val-
ue of 0.028. This finding indicates that compared to the placebo, patients in the new
treatment group are less likely (63.4%; one minus 0.366) to have a shorter time to
event (i.e., have greater survival time or survive longer) after controlling for age and
sex, which is statistically significant (p=0.028) at 95% confidence level. Furthermore,
males are more likely (2.43 times) to have a shorter time to event (have a shorter
survival time) compared to females (p=0.041) after controlling for the variables “treat-
 Survival Analysis 217

ment” and “age”. Age, independently, does not have any significant effect on the
survival time since the p-value is greater than 0.05.

Survivor functions Failure functions

adjusted for age sex adjusted for age sex
1.00

1.00
0.75

0.75
0.50

0.50
0.25

0.25
0.00

0.00
0 50 100 150 200 0 50 100 150 200
analysis time analysis time
treatment = Placebo treatment = New treatment treatment = Placebo treatment = New treatment

Figure 19.3 Survival curve after Figure 19.4 Cumulative incidence curve
controlling for age and sex after controlling for age and sex

If you consider the new treatment as the comparison group (second command; Table
19.8), you will get the HR for the placebo group, which is 2.728. This value (2.728) is
actually the inverse of the HR (0.366) that we had for the new treatment group com-
pared to the placebo. Interpretation is simple. An HR of 2.7 indicates that the patients
in the placebo group are 2.7 times more likely to have a shorter time to event (i.e., the
patients in the placebo group are more likely to die early) compared to patients in the
new treatment group.
The interpretation of a continuous variable (e.g., age) when included in the model is
different. If the HR for age is greater than 1 (say, 1.3), it indicates that the HR will
increase (or decrease) by 30% (1.3 minus 1) with each year of increase (or decrease) in
age. On the other hand, if the HR is less than 1 (say, 0.7), it means that with each year
increase in age, the HR will be reduced by 30% (1 minus 0.7).

19.2.3 Checking for assumptions

Before we conclude the results of Cox regression, we have to check for the important
assumptions, such as:
a) There is no multicollinearity among the independent variables; and
b) Relative hazards over time are proportional (also called the proportionality
assumption of proportional hazards analysis).
218 Survival Analysis

To check for the presence of multicollinearity, look at the standard errors (std. err.) of
the coefficients of the variables included in the model (Table 19.9). Since there is no
value which is very small (<0.001) or very large (>5.0) (also see Section 17.2.2.1),
there is no problem of multicollinearity in the model.
The second assumption (the proportionality assumption) is the major one. If this
assumption is violated, the simple Cox regression model is invalid, and more sophisti-
cated analyses are required. Formal statistical tests and graphical methods (log-mi-
nus-log plot) can be used for detecting violation of this assumption.
The statistical test for checking the proportional hazards assumption is performed by
the following command (this command needs to be used after performing the Cox
regression analysis since this test is based on the most recent use of the Cox regression
results):
estat phtest, detail
This command will provide Table 19.10. The table shows the p-values for the variables
“treatment” (new treatment), “sex” (male) and “age”. The p-value for age is less than
0.05, while the p-values for other variables are greater than 0.05. A p-value of less than
0.05 indicates that the assumption is violated. This suggests that there are some poten-
tial problems with age, while there is no problem with sex and treatment groups (since
the p-values are greater than 0.05). The overall test (global test) p-value is also <0.05.
In such a situation, you can either omit age from the model (since it is not significant)
or use the time-dependent Cox regression method. For further details, readers are
referred to any standard text book.
As an alternative, we can also check the assumption by using the graphical method (by
a log-minus-log plot). To generate the log-minus-log plot for the treatment groups,
after adjusting for age and sex, use the following command:
stphplot, by(treatment) adjust(age sex)
The above command will display the log-minus-log plot for the treatment groups after
adjusting for age and sex (Fig 19.5). If there is a constant vertical difference between
the two curves (i.e., the curves are parallel to each other), it means that the relative
hazards over time are proportional. If the curves cross each other or are much closer
together at some points in time and much further apart at other points in time, then the
assumption is violated. In our example (Fig 19.5), the two lines are not really parallel,
indicating that the assumption may be violated. When the proportional hazards
assumption is violated, it is recommended to use the Cox regression with a time-depen-
dent covariate.
 Survival Analysis 219

Table 19.9 Cox regression with coefficients of HR

. stcox ib1.treatment ib1.sex age, nolog nohr

failure _d: outcome

analysis time _t: time

Cox regression -- Breslow method for ties

No. of subjects = 44 Number of obs = 44

No. of failures = 27
Time at risk = 3831
LR chi2(3) = 9.96
Log likelihood = -83.983329 Prob > chi2 = 0.0189

------------------------------------------------------------------------------
_t | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
treatment |
Placebo | 1.003679 .4578939 2.19 0.028 .1062239 1.901135
|
sex |
male | .8894631 .4356164 2.04 0.041 .0356707 1.743255
|
age | .0088198 .022748 0.39 0.698 -.0357655 .0534051
------------------------------------------------------------------------------

Table 19.10 Test for proportional hazards assumption

. estat phtest, detail

Test of proportional-hazards assumption

Time: Time
----------------------------------------------------------------
| rho chi2 df Prob>chi2
------------+---------------------------------------------------
0b.treatment| . . 1 .
1.treatment | 0.05039 0.07 1 0.7892
0.sex | 0.20439 1.12 1 0.2901
1b.sex | . . 1 .
age | -0.53299 9.22 1 0.0024
------------+---------------------------------------------------
global test | 12.40 3 0.0061
----------------------------------------------------------------
220 Survival Analysis

3
-ln[-ln(Survival Probability)]
2
1
0

1 2 3 4 5
ln(analysis time)

treatment = Placebo treatment = New treatment

Figure 19.5 Log-minus-log plot for checking the proportionality assumption

 20
Nonparametric Methods

Nonparametric methods, in general, are used when the continuous dependent variable
is not normally distributed. Nonparametric tests are also used when the data is
measured on nominal and ordinal scales. Table 20.1 shows the types of nonparametric
methods recommended against parametric tests when the dependent variable is not
normally distributed in the population. Nonparametric tests are less sensitive compared
to parametric tests and may, therefore, fail to detect differences between groups that
may actually exist. Use the data file <Data_4.dta> for practice.

Table 20.1 Nonparametric methods against the alternative parametric methods

Nonparametric test Alternative parametric test
Mann-Whitney U test Independent-samples t-test
Wilcoxon Signed Ranks test Paired t-test
Kruskal-Wallis test One-way ANOVA
Friedman test One-way repeated measures ANOVA
Chi-square test of independence None
Spearmen’s correlation Pearson’s correlation

20.1 Mann-Whitney U test

The Mann-Whitney U test is also called the Wilcoxon rank-sum test. This test is the
nonparametric equivalent of the independent samples t-test. This test compares the
differences between two samples (groups) on a continuous measure (variable) when
222 Nonparametric Methods

the sample distributions are not normal or the sample size is small (<30). This test is
based on ranks of observations and is more efficient than the median test. This test tests
the null hypothesis that the two populations have the same median. For example, we
may want to know whether the median systolic BP (where the distribution of systolic
BP is non-normal) of males and females is the same. To test this hypothesis, use the
following command:
ranksum sbp, by(sex)

Table 20.2 Rank-sum (Mann-Whitney) test

. ranksum sbp, by(sex)

Two-sample Wilcoxon rank-sum (Mann-Whitney) test

sex | obs rank sum expected

-------------+---------------------------------
female | 133 14616.5 14031.5
male | 77 7538.5 8123.5
-------------+---------------------------------
combined | 210 22155 22155

unadjusted variance 180070.92

adjustment for ties -130.90
----------
adjusted variance 179940.02

Ho: sbp(sex==female) = sbp(sex==male)

z = 1.379
Prob > |z| = 0.1679

. tabstat sbp, by(sex) stat(median)

Summary for variables: sbp

by categories of: sex (Sex: numeric)

sex | p50
-------+----------
female | 124
male | 122
-------+----------
Total | 123
------------------

This command will provide Table 20.2. The table shows the p-value (Prob > |z|) of the
test, which is 0.167 (greater than 0.05). This indicates that the distribution of systolic BP
among males and females is not different (or, the median systolic BP of males and
females is not different). With this test result, the median systolic BP of females and males
should be reported. To get the median systolic PB by sex, use the following command:
 Nonparametric Methods 223

tabstat sbp, by(sex) stat(median)

This command will report the median (p50) systolic BP for males and females (Table
20.2).

20.2 Median test

The median test is an alternative to the Mann-Whitney U test. Like the Mann-Whitney
U test, this test also compares the difference in medians between two categories/levels
on a continuous variable. This test is based on the number of observations below and
above the common median. Suppose that we want to determine whether or not the
median age of diabetics and nondiabetics is the same in the population. Here, the null
hypothesis is “the median age of diabetics and nondiabetics is the same in the popula-
tion”. Use the following commands to get the median test results, and the median age
for diabetic and nondiabetic individuals:
median age, by(diabetes)
tabstat age, by(diabetes) stat(median)
The results are shown in Table 20.3. The table (at the bottom) shows the median (p50)
age of diabetic (39 years) and nondiabetic (26 years) individuals. The table also shows
the frequency distribution of diabetic and nondiabetic individuals above and below the
common median, which is 28 years (provided as Total under the “tabstat” command).
For instance, 41 individuals with diabetes are aged over 28 years (common median),
while only 4 are below 28 years. However, our interest is in the p-value given by the
median test. The median test p-value is 0.000 (<0.05), which indicates that the median
age of diabetics and nondiabetics is different in the population.

20.3 Wilcoxon signed ranks test

This test is the nonparametric alternative to the paired samples t-test. This test com-
pares the distribution of two related samples (e.g., pre-test and post-test results). The
Wilcoxon test converts the scores into ranks and then compares them. For example, in
order to evaluate the impact of a training session, you have taken the pre- and post-tests
before and after the training session. You want to assess if there is an improvement in
the post-test scores compared to pre-test scores due to the training session. Use the
following command for the Wilcoxon signed rank test:
signrank post_test = pre_test
224 Nonparametric Methods

Table 20.3 Median test

. median age, by(diabetes)

Median test

Greater |
than the | Diabetes mellitus
median | No Yes | Total
-----------+----------------------+----------
no | 106 4 | 110
yes | 59 41 | 100
-----------+----------------------+----------
Total | 165 45 | 210

Pearson chi2(1) = 43.4324 Pr = 0.000

Continuity corrected:
Pearson chi2(1) = 41.2416 Pr = 0.000

. tabstat age, by(diabetes) stat(median)

Summary for variables: age

by categories of: diabetes (Diabetes mellitus)

diabetes | p50
---------+----------
No | 26
Yes | 39
---------+----------
Total | 28
--------------------

This command will provide Table 20.4. Just look at the p-value of the test, which is
0.000. This indicates that the pre- and post-test scores (medians) are significantly
different. We may, therefore, conclude that the training has significantly improved the
post-test scores compared to the pre-test scores. You can get the medians of the pre-
and post-test scores by using the “tabstat” command as shown earlier (Section 20.2).

20.4 Kruskal-Wallis test

It is the nonparametric equivalent of the one-way ANOVA test. In this test, scores are
converted into ranks, and the mean rank of each group is compared. Suppose that we
want to test the hypothesis of whether or not the systolic BP (variable name "sbp") is
different among religious groups (Muslim, Hindu, and Christian). Here the null
hypothesis is "the systolic BP is not different across the religious groups". Use any of
the following commands to test the hypothesis:
 Nonparametric Methods 225

Table 20.4 Wilcoxon signed ranks test

. signrank post_test = pre_test

Wilcoxon signed-rank test

sign | obs sum ranks expected

-------------+---------------------------------
positive | 32 528 264
negative | 0 0 264
zero | 0 0 0
-------------+---------------------------------
all | 32 528 528

unadjusted variance 2860.00

adjustment for ties -1.50
adjustment for zeros 0.00
----------
adjusted variance 2858.50

Ho: post_tes = pre_test

z = 4.938
Prob > |z| = 0.0000

kwallis sbp, by(religion)

Or,
median sbp, by(religion)
The first command will provide Table 20.5. Since the p-value of the Chi-square test is
0.973 (greater than 0.05), we are unable to reject the null hypothesis. We may, there-
fore, conclude that the median systolic BP among the religious groups is not signifi-
cantly different. The median systolic BP in different religious groups can be obtained
by using the "tabstat" command.

20.5 Friedman test

The Friedman test is the nonparametric alternative to the one-way repeated measures
ANOVA test. Suppose that we are interested in assessing the mean blood sugar levels
at four different time intervals (e.g., at hour-0, hour-7, hour-14, and hour-24) after
administration of a drug on 15 study subjects. The objective of this study is to deter-
mine whether or not the drug reduces blood sugar levels over time (i.e., whether the
average blood sugar levels over time are the same or different).
To conduct this study, we randomly selected 15 individuals from a population and
226 Nonparametric Methods

Table 20.5 Kruskal-Wallis test

. kwallis sbp, by(religion)

Kruskal-Wallis equality-of-populations rank test

+----------------------------+
| religion | Obs | Rank Sum |
|-----------+-----+----------|
| MUSLIM | 126 | 13298.50 |
| HINDU | 58 | 6175.00 |
| Christian | 26 | 2681.50 |
+----------------------------+

chi-squared = 0.054 with 2 d.f.

probability = 0.9733

chi-squared with ties = 0.054 with 2 d.f.

probability = 0.9733

measured their blood sugar levels at the baseline, i.e., before administration of a drug
(hour-0). All the individuals were then administered a drug (say, drug A), and their
blood sugar levels were measured again after 7 hours, 14 hours, and 24 hours. We are
interested in knowing if the blood sugar levels over time after giving the drug are the
same or not (in other words, whether the drug is effective in reducing the blood sugar
levels over time). The variable "time" in the dataset indicates the times of measurement
of blood sugar levels. In this example, we have only one treatment group (received
drug A) but the outcome is measured (blood sugar) at four different points in time on
the same subjects (i.e., we have one treatment group with four levels of measurements
on the same subjects). Use the data file <Data_Repeat_anova_3.dta> for the exercise.
To perform the Friedman test, we need to install a module (emh package), which does
not come preinstalled in Stata. To install the module, use the following command:
ssc install emh
For the Friedman test, Stata needs a dataset which is in long format (our dataset is in
long format). If your dataset is in wide format, you need to convert it into long format
as discussed in Section 5.12. After installing the module, let us first check the number
of study subjects and the mean and median blood sugar levels at different time points
by using the following command (Table 20.6):
tabstat sugar, by(time) stat(n mean p50)
 Nonparametric Methods 227

Table 20.6 Friedman test

. tabstat sugar, by(time) stat(n mean p50)

Summary for variables: sugar

by categories of: time (Time of measurement)

time | N mean p50

-----------------+------------------------------
before treatment | 15 110.5333 110
7 hrs after trea | 15 105.2 105
14 hrs after tre | 15 101.5333 100
24 hrs after tre | 15 100.4667 98
-----------------+------------------------------
Total | 60 104.4333 105
------------------------------------------------

. emh sugar time, strata(subject) anova transformation(rank)

Extended Mantel-Haenszel (Cochran-Mantel-Haenszel) Stratified Test of Association

ANOVA (Row Mean Scores) Statistic:

Q (3) = 27.5625, P = 0.0000
Transformation: Ranks

Now, use the following command to perform the Friedman test:

emh sugar time, strata(subject) anova transform(rank)
The basic syntax of the command is:
emh outcome_variable explanatory_variable, strata(repeated variable) anova
transform(rank).
The output of the Friedman test is provided at the bottom of Table 20.6.
Table 20.6 shows the summary of the blood sugar levels at different time points of
measurement. It shows that the mean and median (p50) blood sugar levels have gradu-
ally decreased over time. For instance, the median blood sugar at the baseline was 110,
while it was 98 at 24 hours after treatment.
The output of the Friedman test is provided at the bottom of Table 20.6. In the table,
Q(3) (=27.56) is the test statistic of the Friedman test, and the p-value of the test is
0.000. This indicates that there is a significant difference in blood sugar levels across 4
time points (p<0.001). In other words, the findings suggest that the drug is effective in
reducing blood sugar levels since the median blood sugar levels have reduced over
time.
228
 21
Analysis of Covariance (ANCOVA)

ANCOVA, or analysis of covariance, is a useful technique to statistically control the

extraneous variable(s) [called covariate] for the comparison of means of two or more
groups. It is similar to ANOVA. In ANOVA, one can incorporate only the categorical
independent variables to have the main effect and interaction. But in ANCOVA, one
can incorporate both categorical and quantitative variables in the model, including the
interaction between categorical and quantitative independent variables. ANCOVA can
be performed as a one-way, two-way, or multivariate ANCOVA technique. Use the
data file <Data_3.dta> for practice.

21.1 One-way ANCOVA

The purpose of using the one-way ANCOVA test is to assess the difference in the mean
of the dependent variable (e.g., systolic BP) against a categorical variable (e.g., sex, or
effect of a drug) after controlling for one or more quantitative variables [called covari-
ates, such as age and diastolic BP] in the model. The one-way ANCOVA test involves
at least three variables, such as:
• One quantitative dependent variable (e.g., systolic BP, post-test score, or blood
sugar level);
• Only one categorical independent variable with two or more levels (e.g., sex,
type of intervention, or type of drug); and
• One (or more) covariate (continuous quantitative variable), e.g., diastolic BP,
age, pre-test score or baseline blood sugar level.
The covariates to be selected for a model should be one or more continuous variables
230 Analysis of Covariance (ANCOVA)

that are significantly correlated with the dependent variable. One can also include
categorical variables as covariates in the model.
Suppose that a researcher is interested in comparing the effectiveness of 3 drugs (drug
A, drug B, and drug C) in reducing systolic BP. To conduct the study, the researcher
randomly selected three groups of people and assigned a drug to each group. In this
scenario, one-way ANOVA could be used. However, it was observed that the mean age
and pre-treatment systolic BP of these three groups were not the same. Since age and
pre-treatment systolic BP may influence the effectiveness of the drugs in reducing
systolic BP, it requires adjustment for these variables (age and pre-treatment systolic
BP) to conclude the results. In such a situation, one-way ANCOVA can be used. In
ANCOVA, the independent variable must be a categorical variable (here it is "type of
drug"). ANCOVA can adjust more than one covariate, either continuous or categorical.

Another example: Assume that you have organized a staff training. You have taken the
pre-and post-tests of the participants to evaluate the effectiveness of the training. Now,
you want to conclude if males and females (independent variable "sex") have similar
performance in the post-test (dependent variable), after controlling for age and pre-test
scores (covariates). If the assumptions are met, one-way ANCOVA is the appropriate
test for this situation as well.

Hypothesis
Assume that you want to assess if the mean systolic BP (dependent variable; variable
name is “sbp”) is the same among males and females (independent variable; variable
name is “sex_1”) after controlling for diastolic BP (covariate; variable name is “dbp”).
H0: There is no difference in mean systolic BP between males and females in the
population, after controlling for diastolic BP.
HA: The mean systolic BP of males and females is different in the population, after
controlling for diastolic BP.

Assumptions
1. The dependent variable is normally distributed at each level of the independent
variable;
2. The variances of the dependent variable at each level of the independent
variable are the same (homogeneity of variances);
3. The covariates (if more than one) are not strongly correlated with each other
(r<0.8);
 Analysis of Covariance (ANCOVA) 231

4. There is a linear relationship between the dependent variable and the covariates
at each level of the independent variable;
5. There is no interaction between the covariate (diastolic BP) and the independent
variable (sex) [called homogeneity of regression slopes].

21.1.1 Commands
Before performing the ANCOVA, it is better to check the descriptive statistics of the
dependent variable (systolic BP) at each level of the independent variable (sex). Use
the following command to get the descriptive statistics of systolic BP by sex (Table
21.1):
tabstat sbp, by(sex_1) stat(n mean sd)

Table 21.1 Descriptive statistics (unadjusted) of systolic BP by sex

. tabstat sbp, by(sex_1) stat(n mean sd)

Summary for variables: sbp

by categories of: sex_1 (Sex: numeric)

sex_1 | N mean sd
-------+------------------------------
Female | 133 129.5714 21.37695
Male | 77 124.5584 17.22108
-------+------------------------------
Total | 210 127.7333 20.05794
--------------------------------------

Now, to perform the ANCOVA, use the first of the following commands. Here, the
dependent variable is systolic BP (sbp), the independent variable is sex (sex_1), and
the covariate is diastolic BP (dbp).
anova sbp i.sex_1 c.dbp i.sex_1#c.dbp
regress
margins sex_1, atmeans
The outputs of the above commands are displayed in Tables 21.2 (first and second
commands) and 21.3 (third command). The prefix “i.” when used before the name of
an independent variable, tells Stata that it is a categorical variable. It also tells Stata to
generate dummy variables (a set of dichotomous or indicator variables) if the categori-
cal variable has more than two levels (see Sections 5.11 and 16.2). On the other hand,
the prefix “c.” is used before a variable to indicate that the variable is a continuous
232 Analysis of Covariance (ANCOVA)

Table 21.2 ANCOVA table: between-subjects effects

. anova sbp i.sex_1 c.dbp i.sex_1#c.dbp

Number of obs = 210 R-squared = 0.7188

Root MSE = 10.7136 Adj R-squared = 0.7147

Source | Partial SS df MS F Prob > F

-----------+----------------------------------------------------
Model | 60440.1719 3 20146.724 175.52 0.0000
|
sex_1 | 1.82547561 1 1.82547561 0.02 0.8998
dbp | 41991.2066 1 41991.2066 365.84 0.0000
sex_1#dbp | .024723806 1 .024723806 0.00 0.9883
|
Residual | 23644.8947 206 114.781042
-----------+----------------------------------------------------
Total | 84085.0667 209 402.320893
. regress

Source | SS df MS Number of obs = 210

-------------+------------------------------ F( 3, 206) = 175.52
Model | 60440.1719 3 20146.724 Prob > F = 0.0000
Residual | 23644.8947 206 114.781042 R-squared = 0.7188
-------------+------------------------------ Adj R-squared = 0.7147
Total | 84085.0667 209 402.320893 Root MSE = 10.714

------------------------------------------------------------------------------
sbp | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
sex_1 |
Male | 1.569876 12.44837 0.13 0.900 -22.97267 26.11242
dbp | 1.458842 .0732245 19.92 0.000 1.314476 1.603207
|
sex_1#c.dbp |
Male | .0022405 .1526608 0.01 0.988 -.2987373 .3032184
|
_cons | 6.348646 6.254367 1.02 0.311 -5.982131 18.67942
------------------------------------------------------------------------------

variable. The follow-up “regress” command without any argument will display results
in the form of a regression table (Table 21.2). The last command will provide the
predicted (adjusted) mean of systolic BP at each level of sex considering the average
value of the covariate (diastolic BP) (Table 21.3).
You can get the pairwise comparison of predicted means of the dependent variable
when the main effect of the independent variable (with more than two levels) is statisti-
cally significant. If the main effect is not significant (in our example, sex is not signifi-
cantly associated with systolic BP; Table 21.2), it is not really necessary. However, for
the purpose of demonstration, we have performed the pairwise comparison with the
Bonferroni option, using the following command:
 Analysis of Covariance (ANCOVA) 233

Table 21.3 Mean systolic BP by sex after adjustment for diastolic BP

. margins sex_1, atmeans

Adjusted predictions Number of obs = 210

Expression : Linear prediction, predict()

at : 0.sex_1 = .6333333 (mean)
1.sex_1 = .3666667 (mean)
dbp = 82.76667 (mean)

------------------------------------------------------------------------------
| Delta-method
| Margin Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
sex_1 |
Female | 127.0921 .9372843 135.60 0.000 125.2442 128.94
Male | 128.8475 1.282687 100.45 0.000 126.3186 131.3763
------------------------------------------------------------------------------

pwcompare sex_1, mcompare(bon) effects

This command will provide a comparison of adjusted means of the dependent variable
(systolic BP) between the levels (categories) of sex (Table 21.4). The option “effects”
used with the command is to get the p-values.

21.1.2 Interpretation: One-way ANCOVA

Table 21.1 shows the mean and SD of systolic BP by sex. The table shows that the
unadjusted mean systolic BP of females is 129.5 mmHg and that of males is 124.5
mmHg.
Table 21.2 shows the results of one-way ANCOVA. This is the main table to interpret
the results. We have tested the null hypothesis that the population mean of systolic BP
in males and females is the same after controlling for diastolic BP. Look at the p-value
(Prob > F) for sex (sex_1) in the table, which is 0.899 (main effect). Since the p-value
is >0.05, we cannot reject the null hypothesis. We may, therefore, conclude that the
mean systolic BP of males and females in the population is not different after
controlling for diastolic BP.
We can also assess the influence (association) of the covariate (diastolic BP) on the
dependent variable (systolic BP). The p-value for diastolic BP is 0.000, which is highly
significant. This indicates that there is a significant association between systolic and
diastolic BP after controlling for sex.
234 Analysis of Covariance (ANCOVA)

However, before we conclude the results, it is important to check whether the assump-
tion of homogeneity of regression slopes is violated or not. To check this, we need to
look at the significance level (p-value) of the interaction term (sex_1#dbp) in the table
(Table 21.2). We can see that the p-value for interaction is 0.988 (>0.05), suggesting
that there is no interaction. This indicates that the homogeneity of regression slopes
assumption is not violated. A p-value of <0.05 (i.e., if there is an interaction) suggests
the regression slopes are not homogeneous and the ANCOVA test is inappropriate.

Table 21.4 Pairwise comparison of adjusted means

. pwcompare sex_1, mcompare(bon) effects

Pairwise comparisons of marginal linear predictions

Margins : asbalanced

note: option bonferroni ignored since there is only one comparison

---------------------------------------------------------------------------------
| Unadjusted Unadjusted
| Contrast Std. Err. t P>|t| [95% Conf. Interval]
----------------+----------------------------------------------------------------
sex_1 |
Male vs Female | 1.569876 12.44837 0.13 0.900 -22.97267 26.11242
---------------------------------------------------------------------------------

Table 21.3 shows the outputs of the “margins” command. Margins are the statistics
calculated from predictions of a previously analyzed model. In our example, we have
used the option “atmeans” with the “margins” command (there are other options like
“asbalanced” and “asobserved”) to get the predicted values (adjusted values) consider-
ing the average value of the covariate. Therefore, the “margins” command has provid-
ed us with the adjusted (predicted) mean of systolic BP for each level of sex, consider-
ing the average value of diastolic BP (82.76 mmHg). We can see that the mean systolic
BP of females is 127.09 mmHg, while it is 128.84 mmHg for males after adjusting for
the average value of diastolic BP (the adjusted means are different from the unadjusted
means as shown in Table 21.1).
Table 21.4 shows the pairwise comparison of the predicted means. This analysis is not
necessary in our example since the independent variable (sex) is not significantly asso-
ciated with the dependent variable. If the independent variable has more than two
levels and is statistically significant, then the table for pairwise comparison is import-
ant.
Table 21.4 shows that there is no significant difference in mean systolic BP between
 Analysis of Covariance (ANCOVA) 235

males and females as the p-value is 0.90 (>0.05). The analysis ignored the Bonferroni
option since the variable “sex” has only two levels. The Bonferroni test is a type of
multiple comparison test (there are other tests for pairwise comparisons, such as Schef-
fe’s, Sidak’s, and Tukey’s tests) used for pairwise comparison of means. The Bonfer-
roni correction is done to adjust the type I errors when multiple pairwise tests are
performed on a single variable.

21.2 Two-way ANCOVA

In two-way ANCOVA, there are two independent categorical variables with two or
more levels/categories, while in one-way ANCOVA, there is only one independent
categorical variable with two or more levels. At least four variables are involved in the
analysis of two-way ANCOVA. They are:
• One continuous dependent variable (e.g., diastolic BP, blood sugar, or post-test
score);
• Two categorical independent variables (with two or more levels) [e.g., occupation,
diabetes, or type of drug]; and
• One or more continuous covariates (e.g., age, systolic BP, or income).
Two-way ANCOVA provides information, after controlling for the covariate(s), on:
• Whether there is a significant main effect of the first independent variable
(e.g., occupation) on the dependent variable;
• Whether there is a significant main effect of the second independent variable
(e.g., diabetes) on the dependent variable; and
• Whether there is an interaction between the independent variables (e.g., occupation
and diabetes).
Suppose that we want to assess, after controlling for age (covariate):
1. Whether or not occupation influences the diastolic BP (i.e., is the mean diastolic
BP same in different occupational groups?);
2. Whether or not diabetes influences the diastolic BP (i.e., is the mean diastolic
BP same for diabetics and non-diabetics?); and
3. Does the influence of occupation on diastolic BP depend on the presence of
diabetes (i.e., is there an interaction between occupation and diabetes)?
Questions 1 and 2 refer to the main effect, while question 3 explains the interaction
of two independent variables (occupation and diabetes) on the dependent variable
236 Analysis of Covariance (ANCOVA)

(diastolic BP). For the analysis, we will use the data file <Data_3.dta>. The variable
names are for diastolic BP is “dbp”, occupation is “occupation (1= govt. job; 2= private
job; 3= business; 4= others)”, diabetes is “diabetes1 (0= no diabetes; 1= have diabe-
tes)”, and age is “age”.

Assumptions
All the assumptions for the one-way ANCOVA are applicable to two-way ANCOVA.

21.2.1 Commands
To perform the two-way ANCOVA, use the following command. The variables includ-
ed in the analysis are dbp (diastolic BP; as dependent variable), occupation, diabetes1,
and age.
anova dbp i.occupation i.diabetes1 c.age i.occupation#i.diabetes1
regress
The first command is the basic command for ANCOVA. The follow-up command
"regress" will display the outputs in the form of a regression table, as we have seen
earlier in one-way ANCOVA. The outputs of both the commands are displayed in
Table 21.5.
To get the predicted (adjusted) mean of diastolic BP at each level of occupation and
diabetes considering the average values of age (covariate), use the following com-
mand:
margins occupation diabetes1, atmeans
The outputs are displayed in Table 21.6. You can also get the marginal means for a
combination of occupation and diabetes (e.g., govt. job with diabetes, govt. job without
diabetes, etc.), if you use the following command (outputs not shown):
margins occupation#diabetes1
Or,
margins occupation#diabetes1, atmeans
You can generate a plot of predicted values (adjusted means) of the dependent variable
for the independent variables in the model. To generate a plot of adjusted mean diastol-
ic BP for occupation and diabetes, use the following commands successively (Fig
21.1):
 Analysis of Covariance (ANCOVA) 237

margins occupation, within(diabetes1) atmeans

marginsplot

Table 21.5 Results of two-way ANCOVA

. anova dbp i.occupation i.diabetes1 c.age i.occupation#i.diabetes1

Number of obs = 210 R-squared = 0.0133

Root MSE = 11.9008 Adj R-squared = -0.0260

Source | Partial SS df MS F Prob > F

---------------------+----------------------------------------------------
Model | 384.062807 8 48.0078509 0.34 0.9499
|
occupation | 161.42117 3 53.8070567 0.38 0.7676
diabetes1 | 5.41875169 1 5.41875169 0.04 0.8451
age | 15.420277 1 15.420277 0.11 0.7418
occupation#diabetes1 | 126.776448 3 42.2588161 0.30 0.8265
|
Residual | 28467.5039 201 141.629372
---------------------+----------------------------------------------------
Total | 28851.5667 209 138.045774

. regress

Source | SS df MS Number of obs = 210

-------------+------------------------------ F( 8, 201) = 0.34
Model | 384.062807 8 48.0078509 Prob > F = 0.9499
Residual | 28467.5039 201 141.629372 R-squared = 0.0133
-------------+------------------------------ Adj R-squared = -0.0260
Total | 28851.5667 209 138.045774 Root MSE = 11.901

--------------------------------------------------------------------------------------
dbp | Coef. Std. Err. t P>|t| [95% Conf. Interval]
---------------------+----------------------------------------------------------------
occupation |
PRIVATE JOB | -1.370739 2.507417 -0.55 0.585 -6.314954 3.573477
BUSINESS | -.9578347 2.602651 -0.37 0.713 -6.089838 4.174168
OTHERS | -3.433968 2.561338 -1.34 0.182 -8.484508 1.616571
|
diabetes1 |
yes | -1.513519 4.127483 -0.37 0.714 -9.652241 6.625204
age | -.0365282 .1107028 -0.33 0.742 -.2548161 .1817597
|
occupation#diabetes1 |
PRIVATE JOB#yes | -1.536793 6.179707 -0.25 0.804 -13.72217 10.64858
BUSINESS#yes | 2.673676 5.64161 0.47 0.636 -8.450656 13.79801
OTHERS#yes | 3.312635 5.554671 0.60 0.552 -7.640267 14.26554
|
_cons | 85.12535 3.322763 25.62 0.000 78.57341 91.67729
--------------------------------------------------------------------------------------

You can perform the pairwise comparison of predicted means of the dependent
variable when the main effects of the independent variables are statistically significant.
If the main effect is not significant for an independent variable, it is not necessary to do
the pairwise comparison test. In our example, the main effects of both occupation (p=
0.76) and diabetes (p= 0.84) are not statistically significant. However, for the purpose
238 Analysis of Covariance (ANCOVA)

Table 21.6 Adjusted means (predicted values) of diastolic BP by occupation and

diabetes
. margins occupation diabetes1, atmeans

Adjusted predictions Number of obs = 210

Expression : Linear prediction, predict()

at : 1.occupation = .2857143 (mean)
2.occupation = .2333333 (mean)
3.occupation = .2333333 (mean)
4.occupation = .247619 (mean)
0.diabetes1 = .7857143 (mean)
1.diabetes1 = .2142857 (mean)
age = 26.51429 (mean)

------------------------------------------------------------------------------
| Delta-method
| Margin Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
occupation |
GOVT JOB | 83.83251 1.549116 54.12 0.000 80.7779 86.88711
PRIVATE JOB | 82.13245 1.717086 47.83 0.000 78.74664 85.51827
BUSINESS | 83.4476 1.712003 48.74 0.000 80.07181 86.82339
OTHERS | 81.10839 1.663185 48.77 0.000 77.82886 84.38792
|
diabetes1 |
no | 82.76318 .92934 89.06 0.000 80.93067 84.59569
yes | 82.33521 1.830683 44.98 0.000 78.7254 85.94501
------------------------------------------------------------------------------

of demonstration, we have performed the pairwise comparisons with the Bonferroni

option by using the following command:
pwcompare occupation diabetes1, mcompare(bon) effects
This command will provide a comparison of adjusted mean diastolic BP within the
levels (categories) of occupation and diabetes (Table 21.7). The option “effects”, used
with this command, is for obtaining the p-values of the test.

21.2.2 Interpretation: Two-way ANCOVA

Table 21.5 shows the results of the two-way ANCOVA test. We have tested the null
hypothesis that:
• The mean diastolic BP (in the population) among the occupational groups is
the same after controlling for age and diabetes;
• The mean diastolic BP (in the population) among diabetics and non-diabetics
is the same after controlling for age and occupation; and
 Analysis of Covariance (ANCOVA) 239

Adjusted Predictions of occupation with 95% CIs

90
85
Linear Prediction
80
75
70

GOVT JOB PRIVATE JOB BUSINESS OTHERS

Occupation

no yes

Figure 21.1 Adjusted mean diastolic BP at different levels of occupation by diabetes

• There is no interaction between occupation and diabetes after controlling for

age.
Look at the p-values for occupation, diabetes, and the interaction term “occupation#di-
abetes1” in Table 21.5. They are 0.767, 0.845, and 0.826, respectively, indicating that
none of them is statistically significant (we are unable to reject any of the null hypothe-
ses). This means that occupation (after controlling for age and diabetes) and diabetes
(after controlling for age and occupation) do not have any significant influence on
diastolic BP. There is also no interaction between occupation and diabetes after
controlling for age. However, we should always check the p-value of the interaction
first. If the interaction is significant (p-value <0.05), then the main effects (of occupa-
tion and diabetes) are not important because the effect of one independent variable is
dependent on the levels of the other independent variable.
We also have information about the influence of covariates on the dependent variable.
We can see (Table 21.5) that the p-value for age is 0.741, which is not statistically
significant. This indicates that there is no significant association between age and
diastolic BP after controlling for occupation and diabetes.
Table 21.6 shows the adjusted (predicted) mean (Margin column) diastolic BP (depen-
dent variable) at different levels of the independent variables (occupation and diabe-
tes). As an example, the adjusted mean diastolic BP of the government job holders is
83.8 mmHg and that of diabetics (diabetes1 yes) is 82.3 mmHg.
240 Analysis of Covariance (ANCOVA)

Table 21.7 Pairwise comparison test for occupation and diabetes

. pwcompare occupation diabetes1, mcompare(bon) effects

Pairwise comparisons of marginal linear predictions

Margins : asbalanced

---------------------------
| Number of
| Comparisons
-------------+-------------
occupation | 6
diabetes1 | 1
---------------------------

------------------------------------------------------------------------------------------
| Bonferroni Bonferroni
| Contrast Std. Err. t P>|t| [95% Conf. Interval]
-------------------------+----------------------------------------------------------------
occupation |
PRIVATE JOB vs GOVT JOB | -2.139135 3.089479 -0.69 1.000 -10.37143 6.093161
BUSINESS vs GOVT JOB | .3790031 2.821173 0.13 1.000 -7.138358 7.896364
OTHERS vs GOVT JOB | -1.77765 2.778096 -0.64 1.000 -9.18023 5.624929
BUSINESS vs PRIVATE JOB | 2.518138 3.00248 0.84 1.000 -5.482339 10.51862
OTHERS vs PRIVATE JOB | .3614848 2.963006 0.12 1.000 -7.53381 8.256779
OTHERS vs BUSINESS | -2.156654 2.677556 -0.81 1.000 -9.291331 4.978024
|
diabetes1 |
yes vs no | -.4011393 2.050795 -0.20 0.845 -4.444971 3.642693
------------------------------------------------------------------------------------------

Table 21.7 depicts the pairwise comparison of adjusted mean diastolic BP within
different occupational groups and diabetes. When the independent variable(s) with
more than two levels is significantly associated with the dependent variable, a pairwise
comparison is required. Examine the p-values (P> |t|) in Table 21.7. Since all the p-val-
ues are >0.05, there is no significant difference in mean diastolic BP among the occu-
pational groups and diabetes.
Figure 21.1 plotted the adjusted mean diastolic BP with 95% CI of different occupa-
tional groups disaggregated by diabetes. Finally, from the data, we can conclude that
the diastolic BP is not influenced (there is no association) by occupation and diabetes
after controlling for age and the independent variables (diabetes for occupation and
occupation for diabetes) included in the model.
 22
Miscellaneous

In this chapter, we will discuss two important and useful statistical methods that are
frequently needed during data management and analysis, such as how to test the
reliability of a scale and develop the wealth quintiles.

22.1 Reliability of scales: Cronbach’s alpha

When researchers select a scale (e.g., a scale to measure depression) for their study, it
is important to check that the scale is reliable. One of the ways to check the internal
consistency (reliability) of a scale is to calculate the Cronbach’s alpha coefficient.
Cronbach’s alpha indicates the degree to which the items that make up the scale
correlate with each other in the group.
Ideally, Cronbach’s alpha coefficient should have a value above 0.7 to indicate that the
scale is reliable. However, this value is sensitive to the number of items on the scale. If
the number of items on the scale is less than 10, Cronbach’s alpha coefficient tends to
be low. In such a situation, it is appropriate to use the “average interitem correlation”.
The optimum range of the average (mean) interitem correlation value is between 0.2
and 0.4. Use the data file <Data Cronb. dta> for practice.
Before using the procedure, be sure that all the negatively worded values are
“reversed” by recoding (see Section 5.2). If this is not done, it will produce a very low
(or negative) value of Cronbach’s alpha coefficient. Hopefully, Stata can automatically
reverse the negatively worded values if an appropriate option is used with the main
command (see below). Assume that a researcher has used a scale to measure depres-
sion. The scale has 4 items (questions), q1, q2, q3, and q4. To get the Cronbach’s alpha
coefficient, use the following commands:
242 Miscellaneous

alpha q1-q4
alpha q1-q4, std
alpha q1-q4, std item detail
alpha q1-q4, item reverse(q3 q4)
The first command will provide Cronbach’s alpha coefficient based on unstandardized
items (Table 22.1) (q1-q4 indicates the variables from q1 to q4). When the option “std”
is used (second command), Stata will provide the coefficient based on standardized
values of the items (Table 22.1). This option is used when items are not measured on
the same scale. It also provides an unbiased estimate. Our suggestion is to use the
option “std” to get the alpha coefficient even though items are on the same scale of
measurement.
The third command (with the use of options "std", "item", and "detail"), Stata will
provide the item-wise values of coefficients in a table and an interitem correlation
matrix (Table 22.2). The last command [with the use of option "reverse(q3 q4)"] will
reverse the coding of variables q3 and q4. This option is used when one or more nega-
tively worded variables need to be reversed (we don’t have such a problem in our data).

Table 22.1 Cronbach’s alpha coefficients

. alpha q1-q4

Test scale = mean(unstandardized items)

Average interitem covariance: .6564501

Number of items in the scale: 4
Scale reliability coefficient: 0.8390

. alpha q1-q4, std

Test scale = mean(standardized items)

Average interitem correlation: 0.5675

Number of items in the scale: 4
Scale reliability coefficient: 0.8400

22.1.1 Interpretation
Table 22.1 shows the unstandardized (0.839) and standardized (0.840) values of Cron-
bach’s alpha coefficients. The standardized value is especially important when all the
items are not on the same scale of measurement. It also provides an unbiased estimate.
 Miscellaneous 243

In our example, the standardized Cronbach’s alpha coefficient is 0.840, which

indicates a very good correlation among items on the scale (i.e., the scale is reliable).

Table 22.2 Item-wise values of Cronbach’s alpha

. alpha q1-q4, std item detail

Test scale = mean(standardized items)

average
item-test item-rest interitem
Item | Obs Sign correlation correlation correlation alpha
-------------+-----------------------------------------------------------------
q1 | 60 + 0.7761 0.5991 0.6178 0.8290
q2 | 60 + 0.8444 0.7101 0.5429 0.7808
q3 | 60 + 0.8257 0.6789 0.5633 0.7947
q4 | 60 + 0.8416 0.7054 0.5460 0.7830
-------------+-----------------------------------------------------------------
Test scale | 0.5675 0.8400
-------------------------------------------------------------------------------

Interitem correlations (obs=60 in all pairs)

q1 q2 q3 q4
q1 1.0000
q2 0.5122 1.0000
q3 0.4911 0.6346 1.0000
q4 0.5483 0.6295 0.5892 1.0000

However, before considering the value of Cronbach’s alpha coefficient, look at the
"Interitem correlations matrix" displayed at the bottom of Table 22.2. All the values in
the matrix must be positive (all the values are positive in our example). The presence
of one or more negative values indicates that some of the items have not been "reverse
scored" correctly. This information is also provided in the first part of the table under
the "Sign" column (all the items have a positive sign).
The “item-rest correlation” in Table 22.2 indicates the degree to which each item
correlates with the total score. In our example, the values for q1 to q4 are 0.60, 0.71,
0.67, and 0.70, respectively. A small value (<0.30) for any item could be a problem. If
the Cronbach’s alpha coefficient (alpha) and item-rest correlation values for any item
are small (<0.7 and <0.3, respectively), one may consider omitting the item from the
scale that has a small value. In our example, there is no such problem.
If the number of items is small on the scale (fewer than 10), it may be difficult to get
a reasonable Cronbach’s alpha coefficient value. In such a situation, consider the aver-
age interitem correlation value provided in Table 22.2. In this example, the average
244 Miscellaneous

interitem correlation value ranges from 0.542 to 0.617, and the scale mean (average) is
0.567).

22.2 Constructing wealth quintiles

The wealth index is an indicator of the economic status of households or individuals
that is commonly used in demographic health and other surveys. The measure of the
wealth index may be linked with inequalities in individual characteristics, use of health
and other services, and health outcomes. The wealth index is commonly calculated
from information on dwelling and household characteristics, access to a variety of
consumer goods and services, and assets, which together are used as a measure of the
economic status of an individual. The wealth index is constructed using the household
asset data via principal component analysis (PCA).
After calculating the composite wealth index scores for each individual, they are
categorized into wealth quintiles. A quintile of a dataset represents 20% (one-fifth) of
a given sample. Therefore, the calculated wealth index, after arranging them into
ascending order, is classified into five categories, or quintiles. The lowest index group
(the first quintile) is the poorest section of the sample, while the highest index group
(the fifth quintile) is the richest section of the sample. In this section, we will discuss
how to construct the wealth quintiles from household information.
We will use the data file <Wealth.dta> for this exercise. There are 18 variables in this
data file with information on household assets and other characteristics. We will use all
this information to construct the wealth quintiles using the PCA technique. Follow the
succeeding steps to construct the wealth quintiles.

Step 1: All the variables to be used for constructing wealth quintiles must be dichoto-
mous variables with the coding scheme of 0/1. In our dataset, there are 18 variables, of
which two are categorical variables with more than two levels (water and toilet). We
need to dichotomize them with a 0/1 coding scheme. Let us first check the value labels
and coding schemes of the variables “water” and “toilet” by using the following com-
mand:
label list water toilet
Table 22.3 shows the code numbers and value labels of the variables. The table shows
that the variable "water" has 14 categories and the variable "toilet" has 11 categories.
We need to dichotomize (0/1) both these variables using some guidelines (e.g., demo-
 Miscellaneous 245

graphic health survey guidelines). Let us consider that the categories (code numbers)
11 to 13, 21, and 91 are the safe sources of drinking water and will code them as 1. The
other categories will be coded as 0. Similarly, for the variable "toilet (use of latrine)",
let us consider the code numbers 11 to 13, and 22 as hygienic practices, and will code
them as 1, while the others will be coded as 0. Use the following commands to recode
the variable "water":
gen water1=0
replace water1=1 if water==11
replace water1=1 if water==12
replace water1=1 if water==13
replace water1=1 if water==21
replace water1=1 if water==91

Table 22.3 Coding schemes of water and toilet

. label list water toilet
water:
11 Piped water into dwelling
12 Piped to yard/plot
13 Public tap/stand pipe
21 Tubewell or borehole
31 Protected well
32 Unprotected well
41 Protected spring
42 Unprotected spring
51 Rainwater
61 Tanker truck
71 Cart with small tank
81 Surface water (River/Lake/pond/stream/canal)
91 Bottled water
96 Other
toilet:
11 Flush or pour slush toilet flush to piped sewer system
12 Flush to septic tank
13 Flush to pit latrine
14 Flush to somewhere else
15 Flush don't know where
22 Pit latrine with slab
23 Pit latrine without slab/open pit
31 Bucket toilet
41 Hanging toilet/hanging latrine
51 No facility/bush/field
96 Other

All these commands will generate a new variable “water1” with the coding scheme of
0/1 as stated before. Now to label the new variable (water1) and put the value labels,
246 Miscellaneous

use the following commands:

lab var water1 "drinking water source"
la de water1 0"unsafe " 1"safe"
la values water1 water1
In the same manner, recode the second variable “toilet” as “toilet1”.

Step 2: Once the multinominal variables are dichotomized, check the prevalence (rela-
tive frequency) of all the variables to be included in the PCA by using the following
command:
sum electricity-toilet1
This command will provide Table 22.4 (“electricity-toilet1” as used with the command
indicates all the variables from electricity to toilet1). We can see that all the variables
are coded as 0/1 (columns Min and Max). The column "Mean" in the table indicates the
relative frequency (proportion) of code 1. For example, the mean of the variable "elec-
tricity" is 0.42. This indicates that 42% of the subjects use electricity. Now, identify the
variables with very small proportions (<0.01 or <1%). In our example, the variables
"car" and "boat" have very low proportions (0.005 or 0.5% and 0.004 or 0.4%, respec-
tively) and we will not include them in the PCA.

Step 3: Now we will do the principal component analysis (PCA) of all the variables
except for “car” and “boat” and calculate the wealth index by using the following com-
mands:
pca electricity radio tv mobile refrigerator almirah table chair watch ///
cycle motorcycle rikshow hhland firmland water1 toilet1, factor(1)
Or,
pca electricity - motorcycle rikshow - toilet1, factor(1)
predict comp1
ren comp1 w_index
The first or second command will do the PCA (output not shown). The third command
(which must be used after performing the PCA) will generate a new variable "comp1"
with a wealth index for all the study subjects. The last command will rename "comp1"
to "w_index" (we did it for our understanding).
 Miscellaneous 247

Table 22.4 Relative frequency of the variables

. sum electricity-toilet1

Variable | Obs Mean Std. Dev. Min Max

-------------+--------------------------------------------------------
electricity | 1185 .4236287 .4943416 0 1
radio | 1185 .1316456 .3382478 0 1
tv | 1185 .2962025 .4567742 0 1
mobile | 1185 .8185654 .3855406 0 1
refregerator | 1185 .0270042 .1621641 0 1
-------------+--------------------------------------------------------
almirah | 1185 .4194093 .4936707 0 1
table | 1185 .5738397 .4947264 0 1
chair | 1185 .5848101 .4929628 0 1
watch | 1185 .5797468 .4938079 0 1
cycle | 1185 .5409283 .4985325 0 1
-------------+--------------------------------------------------------
motorcycle | 1185 .0801688 .271669 0 1
car | 1185 .0059072 .0766631 0 1
boat | 1185 .0042194 .0648472 0 1
rikshow | 1185 .1063291 .3083885 0 1
hhland | 1185 .9611814 .1932439 0 1
-------------+--------------------------------------------------------
firmland | 1185 .3772152 .4848941 0 1
water1 | 1185 .8700422 .336399 0 1
toilet1 | 1185 .8624473 .3445754 0 1

Table 22.5 Wealth quintiles

. la de w_quintile 1"poorest" 2"poorer" 3"middle" 4"richer" 5"richest"

. la values w_quintile w_quintile

. tab w_quintile

5 quantiles |
of w_index | Freq. Percent Cum.
------------+-----------------------------------
poorest | 237 20.00 20.00
poorer | 237 20.00 40.00
middle | 238 20.08 60.08
richer | 236 19.92 80.00
richest | 237 20.00 100.00
------------+-----------------------------------
Total | 1,185 100.00

Step 4: We will now construct the wealth quintiles from the wealth index variable
(w_index) by using the following command:
xtile w_quintile=w_index, nq(5)
248 Miscellaneous

The above command will construct the wealth quintiles in a new variable “w_quintile”.
Finally, label the new variable and give the value levels by using the first two com-
mands below, and check the wealth quintiles by generating a frequency distribution
table using the last command. The outputs are provided in Table 22.5.
lab var w_quintile "wealth quintile"
la de w_quintile 1"poorest" 2"poorer" 3"middle" 4"richer" 5"richest"
la values w_quintile w_quintile
tab w_quintile
References
1. Acock AC. (2014). A Gentle Introduction to Stata. (4th Edition). Stata Press,
StataCorp LP, Texas
2. Advanced Research Computing: Statistical Methods and Data Analytics.
Regression with Stata chapter 2 – regression diagnostics: UCLA.
https://stats.oarc.ucla.edu/stata/webbooks/reg/chapter2/stata-webbooksregres
sionwith-statachapter-2-regression-diagnostics/
3. Altman DG. (1992). Practical Statistics for Medical Research (1st Edition).
Chapman & Hill.
4. Anderson M, Nelson A. Data analysis: Simple statistical tests. FOCUS on
Field Epidemiology: UNC School of Public Health. North Carolina Centre for
Public Health Preparedness; Vol 3(6).
5. Barros AJD, Hirakata VN. Alternatives for logistic regression in cross-sectional
studies: an empirical comparison of models that directly estimate the preva-
lence ratio. BMC Medical Research Methodology 2003; 3(21):1-13. http://w-
ww.biomedcentral.com/1471-2288/3/21
6. Bergmire-Sweat D, Nelson A, FOCUS Workgroup. Advanced Data Analysis:
Methods to Control for Confounding (Matching and Logistic Regression).
Focus on Field Epidemiology: UNC School of Public Health. North Carolina
Centre for Public Health Preparedness; Volume 4, Issue 1.
7. Chan YH. Biostatistics 103: Qualitative Data – Tests of Independence. Singapore
Med J 2003; Vol 44(10):498-503.
8. Chan YH. Biostatistics 104: Correlational Analysis. Singapore Med J 2003;
Vol 44(12):614-619.
9. Chan YH. Biostatistics 201: Linear Regression Analysis. Singapore Med J
2004; Vol 45(2):55-61.
10. Chan YH. Biostatistics 202: Logistic regression analysis. Singapore Med J
2004; Vol 45(4):149-153.
11. Chan YH. Biostatistics 203. Survival analysis. Singapore Med J 2004; Vol
45(6):249-256.
12. Chan YH. Biostatistics 3.5. Multinominal logistic regression. Singapore Med
J 2005; 46(6):259-268.
13. Daniel WW. (1999). Biostatistics: A Foundation for Analysis in the Health
Science (7th Edition). John Wiley & Sons, Inc.
250

14. Daniels L, Minot N. An introduction to statistics and data analysis using Stata.
(2020). SAGE Publications, Inc: USA.
15. Eric Vittinghoff E, Glidden DV, Shiboski SC, McCulloch CE. Regression
methods in biostatistics: Linear, Logistic, Survival, and Repeated Measures
Models. (2012). 2nd ed. Springer New York.
16. Gordis L. (2014). Epidemiology (5th Edition). ELSEVIER Sounders.
17. Hamilton LC. (2013). Statistics with Stata (8th Edition). Brooks/Cole Cengage
Leaning: USA.
18. Hess KR. Graphical methods for assessing violations of the proportional
hazards assumption in Cox regression. Stat Med 1995;14(15):1707-23. doi:
10.1002/sim.4780141510.
19. Islam MT, Kabir R, Nisha M. (2021). Learning SPSS without Pain (2 nd
Edition. ASA publications, Dhaka.
20. Juul S, Frydenberg M. (2014). An introduction to Stata for health researchers
(5th Edition). Stata press, StataCorp, Texas.
21. Katz MH. (2009). Study Design and Statistical Analysis: A Practical Guide for
Clinicians. Cambridge University Press.
22. Katz MH. (2010).Evaluating Clinical and Public Health Interventions – A
Practical Guide to Study Design and Statistics. Cambridge University Press.
23. Katz MH. (2011). Multivariable Analysis: A Practical Guide for Clinicians and
Public Health Researchers (3rd Edition). London, Cambridge University Press.
24. Katz MH. Multivariable Analysis: A Primer for Readers of Medical Research.
Ann Intern Med 2003; 138:644–650.
25. Khamis H. Measures of Association: How to Choose? JDMS 2008; 24:155–162.
26. Lee J, Chia KS. Estimation of prevalence rate ratios for cross sectional data: an
example in occupational epidemiology. British Journal of Industrial Medicine
1993; 50:861-864.
27. Long JS, Freese J. Regression models for categorical dependent variables
using Stata. Stata publication. StataCorp LP, Texas.
28. Malek MH, Berger DE, Coburn JW. On the inappropriateness of stepwise
regression analysis for model building and testing. Eur J Appl Physiol 2007;
101: 263–264. https://doi.org/10.1007/s00421-007-0485-9.
 251

29. Pfaff T. A brief introduction to Stata with 50+ basic commands. (2009). Institute
for Economic Education, University of Münster.
30. Rabe-Hesketh S, Everitt B. A Handbook of Statistical Analyses using Stata.
(2004). 3rd ed. Chapman & Hall/CRC.
31. Reboldi G, Angeli F, Verdecchia P. Multivariable Analysis in Cerebrovascular
Research: Practical Notes for the Clinician. Cerebrovasc Dis 2013;
35:187–193. DOI: 10.1159/000345491.
32. Schlesselman JJ, Stolley PD. (1982). Case-Control Studies: Design, Conduct,
Analysis. Oxford University Press, Oxford, New York.
33. Stata press. (2009). Stata multivariate statistics reference manual release 11.
Stata Publication. StataCorp LP, Texas.
34. Stata press. (2012). Data Analysis Using Stata (3rd Edition). StataCorp LP, Texas.
35. Szklo M, Nieto FJ. (2007). Epidemiology: Beyond the Basics (2nd Edition).
Jones and Bartlett Publishers.
36. Tabachnik BG, Fidell LS. (2007). Using multivariate statistics (5th Edition).
Boston: Pearson Education.
37. Thompson ML, Myers JE, Kriebel D. Prevalence odds ratio or prevalence
ratio in the analysis of cross sectional data: what is to be done? Occup Environ
Med 1998; 55:272-277.
38. UCLA. Advanced Research Computing: Statistical Methods and Data Analytics;
Repeated Measures Analysis with Stata. https://stats.oarc.ucla.edu/stata/semi-
nars/repeated-measures-analysis-with-stata/
39. UCLA. Institute for Digital Research & Institution: Statistical Consulting.
https://stats.idre.ucla.edu/stata/output/descriptive-statistics-using-the-summar
ize-command/
40. Whittingham M, Stephens P, Bradbury R, Freckleton R. Why do we still use
stepwise modelling in ecology and behavior? Journal of Animal Ecology
2006; 75(5):1182–1189.
41. Williams R. Using the margins command to estimate and interpret adjusted
predictions and marginal effects. The Stata Journal 2012;12(2):308–331.
42. Zwiener I, Blettner M, Hommel G: Survival analysis— part 15 of a series on
evaluation of scientific publications. Dtsch Arztebl Int 2011; 108(10):163–9.
DOI: 10.3238/arztebl.201 1.0163
252
Subject index with chapters and sections

Adjusted odds ratio, 13.3, 13.3.1, Correlation, 15.1

17.2.1.1, 17.2.6.1 Correlation coefficient, 15.1, 15.1.2,
Adjusted R-squared, 6.2.4.1, 16.1.2, 16.2.4.1
16.2.3 Correlation matrix, 15.1.2, 16.2.4.1,
Analysis of covariance (ANCOVA), 9, 17.2.2.1, 22.1
21 Cox regression, 17.3, 17.3.2, 19.2,
Analysis of variance (ANOVA), 9, 11.1, 19.2.1, 19.2.2
11.2 Cronbach's alpha, 22.1
Average interitem correlation, 22.1, Cross-tabulation, 6.1, 13.1
22.1.1 Cumulative probability of survival, 19,
19.1.2, 19.1.3, 19.2, 19.2.1
Backward LR, 17.2.3
Backward, 16.2.5.1, 17.2.3, 19.2 Data cleaning 4
Bar graph, 7.4, 7.4.1, 7.4.2 Data file generation, 2.1
Bartlett's test, 11.1.2, 11.1.5 Data file import, 2.1.3
Beta, 16.2.2, 16.2.3 Data file, 3.1, 3.1.1
Binary logistic regression, 17, 17.2 Data screening, 4
Bland Altman test, 9 Data transformation, 3.1.2.5, 5.5
Bonferroni's test, 11.1.3, 11.1.5, 11.2.3, Description, 3.3.1
12.1.2, 21.1.1, 21.1.2, 21.2.1 Dispersion, 6.2
Box and Plot, 4.2, 7.3, 11.1.3 Do-file execution, 3.1.3.3
Breslow's test, 19.1.3 Do-file, 3.1, 3.1.3
Dummy variable, 5.11, 16.2, 16.2.2,
Censored time, 19 16.2.4.1, 21.1.1
Central tendency, 6.2 Durbin-Watson test, 16.2.4.6
Chi-square test of independence, 9, 13.1
Class interval, 5.3 Egen, 5.10
Classification table, 17.2.2.2 Event time, 19
Codebook, 3.3.2 Exponential, 17.1, 17.2.1.1
Combine data, 5.4 Extraction of duration, 5.7
Conditional logistic regression, 17,
17.2.6 Fisher's exact test, 9, 13.1.1, 13.1.2
Confidence interval, 6.2 Forward LR, 17.2.3
Confounding factor, 13.3, 16.2.5, 17.2, Frequency, 6.1
17.2.4 Friedman test, 9, 20.5
Converting variables, 5.1
254

Generalized linear model, 17.3.3 Log rank test, 19.1.2, 19.1.3

Generating data files, 2.1 Logistic regression diagnostics, 17.2.2
Graph, 7 Logistic regression model, 17.1
Greenhouse-Geisser test, 12.1.2 Logistic regression, 17
Logit transformation, 17.1
Hazard ratio, 19.2, 19.2.1 Log-minus-log plot, 19.2.4
Help, 3.6 Long format, 5.12, 20.5
Histogram, 7.1, 8.1, 8.1.1
Homogeneity of regression slopes, 21.1, Mann-Whitney U test, 9, 20.1
21.1.2 Marginal value, 18.2, 21.2.1
Homogeneity of variances, 11, 11.1.2, Margins plot, 18.2
11.2, 21.1 Mauchly's test, 12.1.2
Homoscedasticity, 16.1, 16.2.4, 16.2.4.4 McNemer test, 9
Hosmer-Lemeshow test, 17.2.2.2 Median survival time, 19, 19.1.2, 19.1.3,
Huynh-Feldt test, 12.1.2 19.2
Hypothesis testing, 9 Median test, 20.2
Multicollinearity, 16.2.4, 16.2.4.1,
Independent samples t-test, 9, 10.2, 16.2.4.4, 17.1, 17.2.2.1, 19.2.4
10.2.2, 10.2.3 Multinominal logistic regression, 9, 17,
Interaction, 11.2, 11.2.2, 13.3, 13.3.1, 18
16.2.2, 17.2.4, 21.1, 21.1.2, 21.2, Multiple comparisons, 11.1.3, 11.1.4,
21.2.2 11.2.3, 21.1.2
Iteration, 19.2.1, 19.2.2 Multiple linear regression, 9, 16.2,
Interitem correlation, 22.1, 16.2.2, 16.2.3
22.1.1
Negative predictive value, 17.2.2.2
Kaplan-Meier method, 19, 19.1, 19.1.2 Non-parametric methods, 20
Kappa estimates, 9 Normality of data, 4.3, 5.5, 6.2.1, 7.1,
Kendall's tau-b, 15.2 8.1, 8.1.1
Kolmogorov-Smirnov test Numeric variable, 5.1.1, 5.1.2
Kruskal-Wallis test, 9, 20.4
Kurtosis, 6.2, 6.2.1 Odds ratio (OR), 13.2
One-sample test of proportion, 14.1
Levene's test, 10.2.1 One-sample t-test, 9, 10.1
Likelihood ratio, 17.2.1.1 One-way ANCOVA 21.1
Line graph, 7.5 One-way ANOVA 11.1
Linear regression, 16 One-way repeated measures ANOVA,
Linearity, 16.2.4.2 12.1
Log file, 3.1.2
 255

Outliers, 4.2, 6.2.1, 7.2, 7.3, 8.1.1, Ordinal logistic regression, 17

16.2.4.5 16.2.3
Out-of-range error, 4.1 Relative risk (RR), 13.2
Output file, 3.1, 3.1.2 Reliability of scales, 22.1
Relocation of variables, 5.8
Paired t-test, 9, 10.3, 10.3.1, 10.3.2 ROC curve, 17.2.2.3
Partial correlation, 15.3 R-squared, 16.1.2, 16.2.3, 16.2.4.3,
Pearson's chi-square, 13.1.2 16.2.5.3
Pearson's correlation coefficient, 9, 15.1
Percentile, 6.2, 6.2.1 Sample size, 8.1.1, 11.1, 16.2.1, 16.2.3,
Peto test, 19.1.2, 19.1.3 17.2.5, 20.1
Pie chart, 7.6 Scatterplot, 7.2
Poisson regression, 9, 17.3, 17.3.1 Scheffe's test, 11.1.3, 21.1.2
Positive predictive value, 17.2.2.2 Score calculation, 5.6
Post hoc test, 11.1.3, 11.1.4, 11.2.3, Sdtest, 10.2.1
11.1.5 Search, 3.6
Post-estimation commands, 17.2.2.4, Sensitivity, 17.2.2.2
18.2, Shapiro Wilk test, 8.1, 8.1.1
Prevalence odds ratio, 17.3 Simple linear regression, 16.1
Prevalence ratio, 17.3 Skewness, 6.2, 6.2.1
Proportion test, 14 Skewness-kurtosis (S-K) test, 8.1, 8.1.1
Proportional hazards analysis, 9, 17.3, Sorting, 3.4, 6.3, 16.2.4.5
19.2 Spearman correlation, 9, 15.2
Proportional odds regression, 9, 17 Specificity, 17.2.2.2
Proportionality assumption, 19.2.4 Sphericity assumption, 12.1
Pseudo R-squared, 17.2.1.1, 17.2.6.1, Stata files, 3
18.1 Stata windows, 1.2
Stepwise, 16.2.5, 16.2.5.1, 16.2.5.2,
Q-Q plot, 8.1, 8.1.1 16.2.5.3, 17.2.3, 19.2.1
Quartile, 6.2.1 Stratified analysis, 13.3
Quintile, 22.2 String variable, 5.1.1
Sub-group, 5.9
Repeated measures ANOVA, 12 Survival analysis, 19, 19.1
Recoding, 5.2 Survival curve, 19.1.2, 19.1.3, 19.2.1
Regression coefficient, 16.2, 16.2.3, Symbols, 3.5
16.2.4.1, 16.2.5.3, 17.2.1.1, Syntax, 3.2
17.2.2.1, 17.2.6
Regression diagnostics, 16.2.4, 17.2.2 Tarone-Ware test, 19.1.2, 19.1.3
Regression equation, 16.1, 16.1.2, Test for independence, 16.2.4.6
256

Time extraction, 5.7 Testing of hypothesis, 9

Tolerance, 16.2.4.1 Unequal variances, 11.1.5
Total score calculation, 5.6
Transformation of data, 5.5 Variable insertion, 2.1.4.2
t-test, 9, 10, 10.1, 10.2, 10.3 Variance inflation factor (VIF), 16.2.4.1
Tukey's test, 11.1.3, 11.1.5, 21.1.2 Version, 1.1
Two-sample test of proportion, 14.2
Two-way ANCOVA 21.2 Wealth quintile, 22.2
Two-way ANOVA 11.2 Wide format, 5.12, 20.5
Two-way repeated measures ANOVA Wilcoxon rank-sum test, 20.1
Wilcoxon Signed Ranks test, 9, 20.3
Unconditional logistic regression, 17, Wilk's lambda, 12.1.2
17.2.1