SPSS Without Pain

!"#$!!$%&'$()*&+,*$!
Learning SPSS without Pain

A Comprehensive Manual for Data Analysis and Interpretation of Outputs
First Edition
Mohammad Tajul Islam

MBBS (Dac), DTM&H (Thai), MSc (CTM) (Thai), MPH (Thai)
Professor (Adjunct), Department of Public Health

North South University and State University of Bangladesh
ASA Publications
Dhaka, Bangladesh
ISBN 978-984-34-8254-9
Copyright © 2020 by Author

First published 2020
This e-manual is free for everyone. Anyone can download this manual from the
links, print it out for personal use and share with others, but strictly prohibited to
use it for any kind of commercial interest. The datasets provided in the links and
used in this manual are hypothetical. It is also free for everyone to download and
use the datasets for practice.
Publisher
ASA Publications
Dhaka, Bangladesh
Email: [email protected]
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: Tk. 250
Production credits
Publisher: ASA Publications
Production Director: Md Altaf Hossain
Marketing Manager: Md Altaf Hossain
Composition: Zariath Al-Mamun Badhon and Md Mamunur Rasel
Cover Design: Zariath Al-Mamun Badhon
Printing: ASA Publications
Binding: Rahim Bindings
Printed in Bangladesh
To all my family members, students and young researchers in
health and social sciences
Preface
This manual is intended for the students (MPH, FCPS, MD, MS, MPhil, and
others), teachers and young researchers in health and social sciences. It is written
in a very simple language and used health related data as examples. This manual
answers three basic questions related to data analysis. They are: a) what to do
(what statistics to be used for the data analysis to achieve the objectives); b) how
to do (how to analyze data by SPSS); and c) what do the outputs mean (how to
interpret the outputs)? All these questions are answered in a very simple and
understandable manner with examples. This manual covers the basic statistical
methods of data analysis used in health and social science research. It is the gate-
way to learn SPSS and will help the users to go further. This manual is organized
in 22 sections that covers the data management, descriptive statistics, hypothesis
testing using bivariate and multivariable analysis and others. It is easier to learn
through exploration rather than reading. The users can explore further once the
basics are known. From my understanding, using the statistics as covered in this
manual, the students and researchers will be able to analyze most of their data on
epidemiological studies and publish them in the international peer review journals.
I am optimistic that this manual will make the students’ and researchers’ life easier
for analyzing data and interpreting the outputs meaningfully. The users can down-
load the datasets used in this manual from the following links. If you have any
comments about the manual, feel free to write at the e-mail address below.
Links for the e-manual and data files:

Link 1: https://github.com/rubyrider/Learning-SPSS-without-Pain
Link 2: https://jmp.sh/F65fcni
Link 3: https://drive.google.com/drive/folders/1t9QjNMBV-
bI-oyAwEBQEm8auuYv3sP7KW?usp=sharing
M. Tajul Islam
[email protected]
[email protected]
i
Acknowledgements
I am particularly indebted and grateful to Harun Or Rashid, Monjura Khatun Nisha

and Shakil Ahmed, who reviewed all the sections and provided their critical views
and constructive suggestions. Nisha and Shakil helped me in editing the final man-
uscript. Shakil took the lead role for publishing this manual. Many of my students,
who continually pushed and encouraged me to write this manual deserve a share of
the credit, including SM Anwar Sadat, Md. Naymul Hasan, Md. Golam Kibria,
Sinthia Mazumder, Humayun Kabir and many others. Finally, I express my sincer-
est thanks for the assistance that I have received from the ASA Publications for
publishing the manual. Being the author, I accept full responsibility for any defi-
ciencies that the manual may have.
To the users
Of course, this e-manual is free for all. However, if you can afford, please donate
Tk.100 only (US$ 2 for the users outside Bangladesh) to any charity or a needy
person. This little amount is sufficient to offer a meal to an orphan in the develop-
ing countries.
ii
Contents
Section 1 Introduction 1
1.1 Steps of data analysis 1
Section 2 Generating Data File 3

2.1 Generating data file 3
2.1.1 Defining variables 6
2.1.2 Data entry in SPSS 9
2.2 Data used in this manual 9
Section 3 Data Cleaning and Data Screening 11

3.1 Checking for out-of-range errors 11
3.2 Checking for outliers 12
3.3 Assessing normality of a dataset 13
Section 4 Data Analysis: Descriptive Statistics 14

4.1 Frequency distribution 14
4.2 Central tendency and dispersion 16
4.2.1 Outputs 18
4.2.2 Interpretation 18
4.3 Alternative method of getting measures of central tendency
and dispersion 19
4.3.1 Outputs 19
4.4 Descriptive statistics and histogram disaggregated by Sex 22
4.5 Checking for outliers 23
Section 5 Checking Data for Normality 25

5.1 How to understand that the data have come from a normally
distributed population 25
5.1.1 Outputs 27
Section 6 Data Management 30

6.1 Recoding of data 30
6.1.1 Recoding into same variable 30
6.1.2 Recoding into different variable 32
6.2 Making class intervals 33
6.3 Combine data into a new variable 36
6.4 Data transformation 39
iii
6.5 Calculation of total score 41
6.6 Calculation of duration 44
6.7 Selecting a sub-group for analysis 45
Section 7 Testing of Hypothesis 46

7.1 Association between quantitative and qualitative or quanti-
tative variables 47
7.2 Association between two qualitative variables 48
7.3 Multivariable analysis 49
7.4 Agreement analysis 50
Section 8 Student’s t-test for Hypothesis Testing 51

8.1 One-sample t-test 51
8.1.1 Outputs 52
8.2 Independent samples t-test 52
8.2.1 Commands 53
8.2.2 Outputs 53
8.3 Paired t-test 55
8.3.1 Commands 56
8.3.2 Outputs 56
Section 9 Analysis of Variance (ANOVA): One-way ANOVA 58

9.1 One-way ANOVA 58
9.1.1 Commands 59
9.1.2 Outputs 59
9.1.3.1 Interpretation of multiple comparisons table 61
9.1.4 Graph on distribution of medians/means 61
9.1.5 What to do if the variances are not homogeneous 62
9.1.5.1 Outputs 62
9.1.5.2 Interpretation 63
Section 10 Two-way ANOVA 64

10.1 Two-way ANOVA 64
10.1.1 Commands 65
10.1.2 Outputs 65
iv
Section 11 Repeated Measures ANOVA: One-way 69
11.1 One-way repeated measures ANOVA 69
11.1.1 Commands 70
11.1.2 Outputs 70
Section 12 Repeated Measures ANOVA: Within and Between-

Subjects 75
12.1 Within and between-subjects ANOVA 75
12.1.1 Commands 76
12.1.2 Outputs 77
Section 13 Association between Two Categorical Variables:

Chi-Square Test of Independence 85
13.1 Chi-square test of Independence 85
13.1.1 Commands 85
13.1.2 Outputs 86
Section 14 Association between Two Continuous Variables:

Correlation 89
14.1 Pearson correlation 89
14.1.1 Commands for scatter plot 90
14.1.2 Commands for Pearson correlation 91
14.1.3 Outputs 92
14.2 Spearman’s correlation 93
14.2.1 Commands for Spearman’s correlation 93
14.2.2 Outputs 93
14.3 Partial correlation 93
14.3.1 Commands 94
14.3.2 Outputs 94
Section 15 Linear Regression 96

15.1 Simple linear regression 97
15.1.1 Commands 97
15.1.2 Outputs 98
v
15.2 Multiple linear regression 100
15.2.1 Creating dummy variables 101
15.2.2 Changing string variable into numeric variable 102
15.2.3 Sample size for multiple regression 103
15.2.4 Commands for multiple linear regression 103
15.2.5 Outputs 103
15.2.7 Regression equation 105
15.2.8 Problem of multicollinearity 106
15.2.9 Checking for assumptions 108
15.2.9.1 Checking for outliers and independent data points 108
15.2.9.2 Checking for normality assumption of the residuals and
constant variance 110
15.2.10 Variable selection for the model 111
15.2.10.1 Outputs 113
15.2.10.2 Interpretation 113
Section 16 Logistic Regression 115

16.1 Logistic regression analysis 115
16.1.1 Commands 116
16.1.2 Outputs 116
16.1.3 Interpretation: Basic tables 117
16.1.4 Interpretation: Outputs under Block 0 118
16.1.5 Interpretation: Outputs under Block 1 120
16.1.6 ROC curve 123
16.1.7 Sample size for logistic regression 125
16.1.8 Variable selection for a model 125
Section 17 Survival Analysis 128

17.1 Survival analysis: Kaplan-Meier method 129
17.1.1 Commands 129
17.1.2 Outputs 130
Section 18 Cox Regression 135

18.1 Cox Regression or Proportional Hazards Regression 135
18.1.1 Commands 135
18.1.2 Outputs 136
Section 19 Non-parametric Methods 140

19.1 Mann-Whitney U test 140
vi
19.1.1 Commands 141
19.1.2 Outputs 141
19.2 Wilcoxon Signed Ranks test 142
19.2.1 Commands 142
19.2.2 Outputs 142
19.3 Kruskal-Wallis test 143
19.3.1 Commands 143
19.3.2 Outputs 144
19.4 Friedman test 144
19.4.1 Commands 145
19.4.2 Outputs 145
19.5 Chi-square test for goodness-of-fit 146
19.5.1 Commands 146
19.5.2 Outputs 146
Section 20 Checking Reliability of Scale: Cronbach’s Alpha 148

20.1 Cronbach’s alpha 148
20.1.1 Outputs 149
Section 21 Analysis of Covariance (ANCOVA): One-way ANCOVA 151

21.1 One-way ANCOVA 151
21.1.1 Commands 153
21.1.2 Outputs: Homogeneity of regression slopes 153
21.1.3 Interpretation: Homogeneity of regression slopes 153
21.1.4 Outputs: One-way ANCOVA 154
21.1.5 Interpretation: One-way ANCOVA 156
Section 22 Two-way ANCOVA 158

22.1 Two-way ANCOVA 158
22.1.1 Commands 159
22.1.2 Outputs 159
Annex 165
References 166
vii
Section 1
Introduction
SPSS stands for Statistical Package for Social Sciences. It is a powerful

window-based statistical data analysis software. The menu and dialog-box system
of SPSS have made the program user-friendly. SPSS is particularly useful to the
researchers in public health, medicine, social science and other disciplines. It
supports a wide range of univariate, bivariate, multivariable and multivariate data
analysis procedures.
This manual is intended for the students, teachers and researchers involved in
health and social science research. It provides practical guidance for using SPSS
for basic statistical analysis of data. Once the data file is loaded in SPSS, the users
can select items from a dropdown menu to analyze data, make graphs, transform
variables, and others. SPSS, in general, has made the life of the researchers easier
for data analysis. This manual is based on SPSS version 16.0. Although higher
versions are available, they offer little extra advantage to the users for commonly
used statistical methods of data analysis. Rather, if the user does not have the right
computer to support higher version, the execution of the program may become
difficult.
This manual is primarily developed targeting the Master of Public Health
(MPH) and post-graduate students in medicine (FCPS, MD, MS, and MPhil),
keeping in mind their needs. The aim of this manual is to provide a concise but
clear understanding on how to conduct a range of statistical analyses using SPSS
and interpret the outputs. Special emphasis is given on understanding the SPSS
outputs, which is a problem for many of the users. For better understanding of the
users, examples and data related to health research are used. However, to use the
manual effectively and to understand the outputs, it requires basic knowledge on
biostatistics and epidemiology. The users will find it easier if they review the
relevant statistical concepts and procedures, and epidemiological methods before
using this manual.
1.1 Steps of data analysis

We collect data for our studies using various tools and methods. The commonly
1
used tools for data collection are questionnaire and record sheet, while the com-
monly used data collection methods are face-to-face interview, observation, physi-
cal examination, lab test and others. Sometimes we use the available data (second-
ary data) for our research studies, for example, hospital records, and data of other
studies (e.g., Bangladesh Demographic and Health Survey data). Once data is
collected, the steps of data analysis are:
• Data coding, if pre-coded questionnaire or record sheet is not used
• Development of data file and data entry
• Data cleaning (checking for errors in data entry)
• Data screening (checking assumptions for statistical tests)
• Data analysis
• Interpretation of results
In the following sections, I have discussed the development of data file, data
management, data analysis and interpretation of the outputs.
2
Section 2
Generating Data File
Like other data analysis programs, SPSS has to read a data file to analyze data. We,
therefore, need to develop a data file for the use of SPSS. The data file can be
generated by SPSS itself or by any other program. Data files generated in other
programs can be easily transferred to SPSS for analysis. Here, I shall discuss how
to generate a data file in SPSS.
2.1 Generating data file

The first step in creating a data file is to give a “name” and “define” variables
included in the questionnaire/ record sheet. The next step is entering data in SPSS.
Suppose, we have collected data using a pre-coded questionnaire (codes are shown
in the parenthesis) with the following variables.
Categorical variables:
• Sex (m= male; f = female)
• Religion (1= Islam; 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 (0= No; 1= Yes; 3= Don’t know)
Quantitative variables (numerical variables):
• Age of the respondent
• Monthly family income
• Systolic blood pressure (BP)
• Diastolic BP
Suppose, we have decided to use V1 as the SPSS variable name for age, V2 for
sex, V3 for religion, etc. (table 2.1). Instead of V1, V2, V3, you can use any other
variable name (e.g., age for age, sex for sex, etc.) for your variables. It is always
better to develop a codebook in MS Word or MS Excel before entering data, as
shown in table 2.1. This is helpful during data analysis.
3
Table 2.1. Codebook
SPSS variable name Actual variable name Variable code
V1 Age in years Actual value
V2 Sex m= Male
f= Female
V3 Religion 1= Islam
2= Hindu
3= Others
V4 Occupation 1= Business
2= Government job
3= Private job
4= Others
V5 Monthly family income Actual value
V6 Marital status 1= Married
2= Unmarried
3= Others
V7 Have diabetes mellitus 1= Yes
2= No
V8_a Systolic blood pressure Actual value
V8_b Diastolic blood pressure Actual value
Note: Instead of V1, V2, etc., you can use any other name as SPSS variable name.
For example, you can use the variable name “age” instead of V1, “sex” instead of
V2, etc.
Now, open the SPSS program by double clicking the SPSS icon. You will see
the following dialogue box (fig 2.1). Click on cancel box ( ) to close the “SPSS
for Windows”. Now we have the dialogue box as shown in fig 2.2 (SPSS Data
Editor).
4
Figure 2.1. Dialogue box for defining variables
5
Figure 2.2. SPSS data editor: dialogue box for defining variables
The “SPSS Data Editor (fig 2.2)” shows Name, Type, Width, Decimals, Label,
Values, Missing, Columns, Align, Measure, at the top row. If you do not see this,
click on “Variable View” at the left-bottom corner of the window.
2.1.1 Defining variables:

We shall use SPSS data editor to define the variables. Before entering data, all the
study variables need to be defined including their coding information. The lower
versions of SPSS (version 12 and below) allow only 8 characters to name a vari-
able. The higher versions (13 and above) allow up to 64 characters to name a vari-
able. While writing the variable names, we need to follow certain rules. They are:
• The variables must be unique (all variables should have different names)
• Variables must begin with a letter (small or capital) rather than a number
• Cannot include full stop (.), space, or symbols like, ?, *, μ, λ, etc.
• Cannot include words that are used as commands by SPSS, such as ALL,
AND, BY, EQ, GE, GT, LE, LT, NE, NOT, OR, TO, WITH, etc.
• Cannot exceed 64 characters
To define the variables, follow the instructions below:
Name: In this column, type the brief SPSS variable name as shown in the code-
book. For example, type V1 (you can also use “age” as the variable name) for the
first variable, age. Note that this short name will be used to identify the variable in
the data file.
6
Type: This column indicates the characteristic of the variable, whether it is numer-
ic or string. Numeric means expressed by numbers (e.g., 1, 2, 3, etc.), while string
means expressed by alphabets (e.g., m, f, y, n, etc.). In SPSS, the default value for
“Type” is numeric. If the nature of the variable is string (alphabet or text variable),
we need to change it. To change the variable type into string, follow the following
steps:
Click on the cell under the column “Type” (you will see a box with three dots)
> Click on the “three-dot box” (you will see the options in a separate dialogue
box) > Select “String” from the options > Click OK
Similarly, if it is a date variable (e.g., date of hospital admission), you have to
change the variable type into a date format in the same manner.
Width: The default value for width is 8. In most of the cases, it is sufficient and
serves the purpose. However, if the variable has very large value, then we need to
increase it using the arrow (up and down) button in the box. For practical purpose,
keep the width 8 unless the variable values are larger than 8 characters.
Decimals: This is applicable for the numeric variables and the default value is 2.
If your data does not have any decimal value, you can make it “0” using the down
arrow or keep it as it is.
Label: This is the space where we write the longer description of the variable
(actual variable name as shown in the codebook). For example, we have used “V1”
to indicate age in years. We should, therefore, write “Age in years” in the label
column for the variable name “V1”.
Values: This is applicable for the variables to define their levels using code num-
bers (such as 1, 2 or m, f, etc). This allows the SPSS to retain the meaning of values
(code numbers) you have used in the data set. For example, our variable 2 is sex
and is defined by “V2”. It has two levels, male (coded as “m”) and female (coded
as “f”). Follow the commands below to put the value labels.
Click on cell under the column “Value” (you will see a box with three dots) >
Click on “three-dot box” > Click in the box “Value” > Type “m” > Click in the
box “Value label” > Type “male” > Click on “Add” > Repeat the same process
7
for female (value “f”, value label “female”, add) > OK
In this way, complete value labels for all the variables, if applicable. Note that
value labels are needed only for the variables that have been coded.
Missing: If there is any missing value in the dataset, SPSS has the option to indi-
cate that. If we want to set a missing value for a variable, we have to select a value
that is not possible (out of range) for that variable. For example, we have conduct-
ed a study and the study population was women aged 15-49 years. There are sever-
al missing values for age in the data (i.e., age was not recorded on the question-
naire or respondent did not tell the age). First, we have to select a missing value for
age. We can select any value which is outside the range 15-49 as the missing value.
Say, we have decided to use 99 as the missing value for age. Now, to put the miss-
ing value for age in SPSS, use the following commands.
Click on the cell under the column “Missing” (you will see a box with three
dots) > Click on the “three-dot box” > Select “Discrete missing values” > Click
on the left box > type “99” > Ok
However, you may omit this. Just keep the cell blank while entering data in the
data file. SPSS will consider the blank cells in the data file as missing values (sys-
tem missing).
Columns: The default value for this is 8, which is sufficient for most of the cases.
If you have a long variable name, then only change it as needed. For practical
purpose, just keep it as it is.
Align: You do not need to do anything for this.
Measure: This cell indicates the measurement scale of the data. If the variable is
categorical use “Nominal” for nominal or “Ordinal” for ordinal scale of measure-
ment. Otherwise use “Scale” for interval or ratio scale of measurement. You can
also keep it as it is.
In this way, define all the variables of your questionnaire/record sheet in the
SPSS data editor. The next step is data entry.
8
2.1.2 Data entry in SPSS:
Once all the variables are defined, click on the “Data View” tab at the bottom-left
corner of the window. You will see the following dialogue box (fig 2.3) with the
variable names at the top row. This is the spreadsheet for data entry. Now, you can
enter data starting from row 1 for each of the variables. Complete your data entry
in this spreadsheet and save the data file at your desired location/folder (save the
file as you save your file in MS Word, such as click on File> click on Save as ….
etc.).
If you want to open the data file later, then use the following steps:
Click on File > Open > Data > Select the folder you have saved your SPSS data
file > Select the file > Click “Open”
2.2 Data used in this manual

Following data files have been used in this manual as examples. All these datasets
are available at the following links. The users can download them for practice. The
data files (with hypothetical data) used in this manual include:
• Data_3.sav
• Data_4.sav
• Data_HIV.sav
• Data_repeat_anova_2.sav
• Data_survival_4.sav
• Data_cronb.sav
Links for the data files:

You can download the e-manual and data files from any of the links below.
Link 1: https://github.com/rubyrider/Learning-SPSS-without-Pain
Link 2: https://jmp.sh/F65fcni
Link 3: https://drive.google.com/drive/folders/1t9QjNMBV-
bI-oyAwEBQEm8auuYv3sP7KW?usp=sharing
9
Figure 2.3 Spreadsheet for data entry (SPSS data editor)
10
Section 3
Data Cleaning and Data Screening
Once data is entered into the SPSS, we need to be sure that there are no errors in
the dataset (i.e., there were no errors during data entry). Data cleaning is common-
ly done by generating frequency distribution tables of all the variables to see the
out-of-range values, and by cross tabulations (or by other means) for checking the
conditional values. If errors are identified, they need to be corrected. Simultane-
ously, we also need to check the data if it fulfils the assumptions of the desired
statistical test (data screening), e.g., is data normally distributed to do a t-test? The
users may skip this section for the time being and go to section 4. Once the users
develop some skills in data analysis, they can come back to this section. Use the
data file <Data_3.sav> for practice. The codebook of this data file can be seen in
the annex (table A.1).
3.1 Checking for out-of-range errors

We can check out-of-range errors by making a frequency distribution of the vari-
able or using the statistics option for minimum and maximum values. For example,
you want to see if there are any out-of-range errors in the variable “religion” (note
that the variable “religion” has 3 levels/values: 1= Islam; 2= Hindu; 3= Others). To
do this, use the following commands:
Analyze > Descriptive statistics > Frequencies > Select the variable “religion”
& push it into the “Variable(s)” box > Statistics > Select “minimum” and
“maximum” > Continue > Ok
Look at the first table (table 3.1) of SPSS output. If there is any value which is
out of the range 1-3 in the dataset, you can see it in the table (3.1) as shown below.
Table 3.1. Statistics
religion
N Valid 210
Missing 0
Minimum 1
Maximum 3
The table shows that the values range from 1 to 3 (minimum 1 and maximum
11
3), which are within the range of our code numbers. Therefore, there is no
out-of-range error in this variable.
3.2 Checking for outliers

Outliers can be checked by constructing the box and plot chart. Outliers are indi-
cated by ID numbers on the chart. Outliers are 1.5 box length distance from the
edge (upper or lower) of the box. The extreme values are indicated by “*” (3 box
length distance from the edge of the box). To construct the box and plot chart for
the variable “systolic blood pressure” (SPSS variable name: sbp), use the follow-
ing commands.
Analyze > Descriptive statistics > Explore > Select “sbp” and push it into the
“Dependent List” box > Ok
You will find the box and plot chart (fig 3.1) along with other outputs (table
3.2). The figure 3.1 shows that there are 3 outliers in “systolic blood pressure” as
indicated by the ID numbers (20, 54 & 193).
We can also examine the influence of outliers in the data comparing the 5%
trimmed mean (mean of the data after excluding upper 5% and lower 5% of the
values) with the mean of the whole dataset. If these two means are close together,
there is no influence of outliers in the dataset. Look at the table 3.2. The mean of
the systolic blood pressure (BP) is 127.7, while the 5% trimmed mean is 126.5.
Since the values are not that different (close to each other), there is no influence of
outliers in the data of systolic BP.
Figure 3.1. Box and Plot chart of systolic blood pressure
12
Table 3.2. Descriptives
Systolic BP Statistics Std. Error

Mean 127.73 1.384
95% Confidence Interval for Lower Bound 125.00
Mean Upper Bound 130.46
5% Trimmed Mean 126.58
Median 123.00
Variance 402.321
Std. Deviation 20.058
Minimum 91
Maximum 195
Range 104
Interquartile Range 28
Skewness .736 .168
Kurtosis .336 .334
3.3 Assessing normality of a dataset

One of the major assumptions for parametric tests is that the dependent quantita-
tive variable is normally distributed. Whether the data has come from normally
distributed population or not, can be checked in different ways. Commonest meth-
ods of checking normality of a dataset are through:
• Histogram
• Q-Q plot
• Formal statistical test (Kolmogorov Smirnov (K-S) test or Shapiro Wilk test)
This issue is discussed in detail in section 5.
13
Section 4
Data Analysis: Descriptive Statistics
Descriptive statistics are always used at the beginning of data analysis. The objec-
tive of using the descriptive statistics is to organize and summarize data. Common-
ly 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 a measure of dispersion provides information
about the amount of variation present in the dataset. Other descriptive statistics
include quartile and percentile. Use the data file <Data_3.sav> for practice.
4.1 Frequency distribution

Suppose, you want to find the frequency distribution of the variables “sex” and
“religion”. To do this, use the following commands:
Analyze > Descriptive Statistics > Frequencies > Select the variables “sex”
and “religion” and push them into the "Variable(s)” box > OK (fig 4.1 & 4.2)
Figure 4.1. Commands for frequency distribution of variables
14
Figure 4.2. Selection of variables for frequency distribution
You will see the following outputs (I have shown only the table of sex) (table
4.1). The table indicates that there are in total 210 subjects, out of which 133 or
63.3% are female and 77 or 36.7% are male. If there is any missing value, the table
will show it. In that case, use the “valid percent” instead of “percent” for reporting.
For example, table 4.2 shows 4 missing values. You should, therefore, report 130
or 63.1% are female and 76 or 36.9% are male. Note that the “Percent” and “Valid
Percent” will be the same, if there is no missing value.
Table 4.1. Frequency distribution of sex with no missing value
Sex
Frequency Percent Valid Percent Cumulative Percent
Valid Female 133 63.3 63.3 63.3
Male 77 36.7 36.7 100.0
Total 210 100.0 100.0
Table 4.2. Frequency distribution of sex with 4 missing values
Sex
Valid Female 130 61.9 63.1 63.1
Male 76 36.2 36.9 100.0
Total 206 98.1 100.0
Missing 9 4 1.9
Total 210 100.0
15
4.2 Central tendency and dispersion
We calculate the central tendency and dispersion for the quantitative variables.
Suppose, you want to find the mean, median, mode, standard deviation (SD), vari-
ance, standard error (SE), skewness, kurtosis, quartile, percentile (e.g., 30th and
40th percentile), minimum and maximum values of the variable “age” of the study
subjects. All these statistics can be obtained in several ways. However, using the
following commands is the easiest way to get them together (fig 4.3-4.5).
Analyze > Descriptive statistics > Frequency > Select the variable “age” and
push it into the "Variable(s)” box > Statistics > Select all the descriptive mea-
sures you desire (mean, median, mode, SD, SE, quartile, skewness, kurtosis) >
Select "Percentiles" > Write “30” in the box > Add > Write “40” in the box >
Add > Continue > OK
Figure 4.3. Commands for obtaining central tendency and dispersion
16
Figure 4.4. Selection of variable(s)
Figure 4.5. Selection of statistics for the variable(s)
17
4.2.1 Outputs:
The SPSS will produce the following output (table 4.3).
Table 4.3. Descriptive statistics of age
AGE
N Valid 210
Missing 0
Mean 26.5143
Std. Error of Mean .51689
Median 27.0000
Mode 26.00
Variance 56.10745
Skewness -.092
Std. Error of Skewness .168
Kurtosis -.288
Std. Error of Kurtosis .334
Minimum 6.00
Maximum 45.00
Percentiles 25 21.0000
30 22.3000
40 25.0000
50 27.0000
75 32.0000
4.2.2 Interpretation:
We can see all the descriptive statistics (central tendency and dispersion) that we
have selected for the variable “age” including the statistics for Skewness and Kur-
tosis in table 4.3. Hope, you understand the mean (average), median (middle value
of the data set), mode (most frequently occurring value), SD (average difference of
individual observation from the mean), variance (square of SD) and SE of the
mean. As presented in table 4.3, the mean age is 26.5 and SD is 7.49 years. Let me
discuss the other statistics provided in table 4.3, especially the skewness, kurtosis,
quartile and percentile.
Skewness and Kurtosis: These two statistics are used to judge whether the data
have come from a normally distributed population or not. In table 4.3, we can see
the statistics for Skewness (- .092) and Kurtosis (- .288). Skewness indicates the
spreadness of the distribution. Skewness “>0” indicates data is skewed to the right;
skewness “<0” indicates data skewed to the left, while skewness “~0” indicates
data is symmetrical (normally distributed). The acceptable range for normality of
18
a data set is skewness lying between “-1” and “+1”. However, normality should
not be judged based on skewness alone. We need to consider the statistics for
kurtosis as well. Kurtosis indicates “peakness” or “flatness” of the distribution.
Like skewness, the acceptable range of kurtosis for a normal distribution is
between “-1” and “+1”. Data for “age” has skewness -.092 and kurtosis -.288,
which are within the normal limits of a normal distribution. We may, therefore,
consider that the variable “age” in the population may be normally distributed.
Quartile and Percentile: When a dataset is divided into four equal parts after
arranging into 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 data is divided
into 100 equal parts (after ordered array), each part is called a percentile. We can
see in table 4.3 that, Percentile 25 (means Q1), Percentile 50 (Q2) and Percentile
75 (Q3) for age are 21, 27 and 32 years, respectively. Q1 or the first quartile is 21
years, means that 25% of the study subjects’ age is less than or equal to 21 years.
On the other hand, 30th percentile (P30) is 22.3 years, which means that 30% of the
study subjects’ age is less than or equal to 22.3 years. Hope, you can now interpret
the P40.
4.3 Alternative method of getting measures of central tendency and dispersion

If you want to get all the descriptive statistics (central tendency and dispersion)
and charts (such as histogram, stem and leaf, and box and plot charts) of the vari-
able “age”, use the following commands:
Analyze > Descriptive statistics > Explore > Select the variable “age” and push
it into the "Dependent List” box > Plots > Select "Stem and leaf” and “Histo-
gram" > Continue > OK
4.3.1 Outputs:
The outputs are shown in table 4.4 and figs 4.6 to 4.9.
19
Table 4.4. Descriptive statistics of age
Statistics Std. Error

AGE Mean 26.5143 .51689
Mean
Upper Bound 27.5333
Median 27.0000
Variance 56.107
Minimum 6.00
Maximum 45.00
Range 39.00
Interquartile Range 11.0000
Skewness -.092 .168
Kurtosis -.288 .334
Figure 4.6. Histogram of age
Figure 4.7. Stem and leaf chart of age

age Stem-and-Leaf Plot
Frequency Stem & Leaf
2.00 0. 66
10.00 1. 0122344444
24.00 1. 555566666777788888889999
44.00 2. 00000000000000111112222222233333334444444444
63.00 2. 555555556666666666666666777777777788888888888888999999999999999
34.00 3. 0000111111112222222333333333444444
26.00 3. 55556666666677777788888999
6.00 4. 001133
1.00 4. 5
Stem width: 10.00

Each leaf: 1 case(s)
20
Figure 4.8. Box and plot chart of age
Q3
Q2
Q1
Figure 4.9. Box and plot chart of diastolic BP
Before we understand the graphs, let us see the descriptive statistics provided in
table 4.4. We can see that the SPSS has provided mean and 5% trimmed mean of
age. Five percent trimmed mean is the mean after discarding 5% of the upper and
5% of the lower values of age. The extent of the effect of outliers can be checked
by comparing the mean with the 5% trimmed mean. If they are close together (as
we see in table 4.4; mean= 26.51 and 5% trimmed mean= 26.56), there is no signif-
icant influence of the outliers (or there are no outliers) on age in the dataset. If they
are very different, it means that the outliers have significant influence on the mean
value, and suggests for checking the outliers and extreme values in the dataset. The
21
table 4.4 also shows the 95% Confidence Interval (CI) for the mean of “age”,
which is 25.49-27.53. The 95% CI for the mean indicates that we are 95% confi-
dent/sure that the mean age of the population lies between 25.49 and 27.53 years.
The SPSS has provided several graphs (figs 4.6 to 4.9), such as histogram,
stem and leaf, and box and plot charts. Histogram gives us information about – a)
distribution of the dataset (whether symmetrical or not); b) concentration of
values; and c) range of values. Looking at the histogram (fig 4.6), it seems that the
data is more or less symmetrical. This indicates that age may be normally (approx-
imately) distributed in the population.
Stem and leaf chart (fig 4.7) provides information similar to a histogram, but
retains the actual information on data values. Looking at the stem and leaf chart,
we can have an idea about the distribution of the dataset (whether symmetrical or
not). Data displayed in figure 4.7 shows that the data is more or less symmetrical.
Stem and leaf charts are suitable for small datasets.
The box and plot chart (fig 4.8) 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/extreme values. The lower bound-
ary of the box indicates 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
smallest and largest values are indicated by the horizontal lines of the whiskers.
In the box and plot chart, presence of outliers is indicated by the ID number
and circle, while the presence of extreme values is indicated by “*”. Outliers are
the values lying between 1.5 and <3 box length distance from the edge (upper or
lower) of the box. On the other hand, the extreme values are 3 or more box length
distance from the upper or lower edge of the box. Fig 4.8 shows that there is no
outlier in the data for age. I have provided another box and plot chart, which is for
the variable “diastolic BP” (fig 4.9). Figure 4.9 shows that there are 3 outliers (ID
no. 19, 27 and 121) in the data of diastolic BP, but does not have any extreme
value.
4.4 Descriptive statistics and histogram disaggregated by Sex

If you want to get the outputs (measures of central tendency and dispersion of age)
by sex (males and females separately), use the following commands. The SPSS
will produce the outputs separately for males and females (table 4.5). Note that
22
there are other ways of doing this.
Analyze > Descriptive statistics > Explore > Select “age” and push it into the
"Dependent List" box > Select “sex” and push it into the "Factor List" box >
Plots > Deselect "Stem and leaf”, and select “Histogram" > Continue > OK
Only the table with descriptive statistics is provided below (table 4.5).
Table 4.5. Descriptive statistics of age by sex
Descriptives
Age Sex Statistics Std. Error
Female Mean 26.8872 .58981
Median 27.0000
Variance 46.267
Minimum 10.00
Maximum 45.00
Range 35.00
Skewness .074 .210
Kurtosis -.212 .417
Male Mean 25.8701 .97549
Median 26.0000
Variance 73.272
Minimum 6.00
Maximum 41.00
Range 35.00
Skewness -.153 .274
Kurtosis -.606 .541
4.5 Checking for outliers

Outliers and extreme values can be checked looking at the box and plot chart, as
discussed earlier. We can also check the presence of outliers using the following
commands. For example, we want to understand if there are any outliers present in
the variable “age”.
Analyze > Descriptive Statistics > Explore > Select the variable “age” and
23
push it into the "Dependent List” box > Select “ID_no” (ID no.) and push it
into the "Label cases by” box > Select “Statistics” under “Display” > Statistics
> Select “Outliers” > Continue > OK
The SPSS will provide 5 upper and 5 lower values with the ID (serial no.) num-
bers, as shown in table 4.6.
Table 4.6. Extreme values of the variable age
Extreme Values
Case Number id no. Value
age Highest 1 210 210 45.00
2 209 209 43.00
3 208 208 43.00
4 207 207 41.00
5 206 206 41.00
Lowest 1 3 3 10.00
2 1 1 10.00
3 4 4 11.00
4 2 2 11.00
5 6 6 12.00
a. Only a partial list of cases with the value 12.00 are shown in the table of lower
extremes.
24
Section 5
Checking Data for Normality
It is important to know the nature of distribution of a continuous random variable

before using statistical methods for hypothesis testing. To use parametric methods
for testing hypotheses (e.g., t-test, ANOVA), one of the important assumptions is
that the data of the dependent variable are normally distributed. It is, therefore,
necessary to check whether the data have come from a normally distributed popu-
lation or not, before we use the parametric methods. Use the data file
<Data_3.sav> for practice.
5.1 How to understand that the data have come from a normally distributed
population
This is an important assumption for doing a parametric test. Whether the data have
come from a normally distributed population or not, can be assessed in three
different ways. They are by:
a) Graphs, such as histogram and Q-Q chart;
b) Descriptive statistics, using skewness and kurtosis; and
c) Formal statistical tests, such as 1-sample Kolmogorov Smirnov (K-S) test
and Shapiro Wilk test.
Now, let us see how to get the histogram and Q-Q chart, and do the formal
statistical tests (K-S test and Shapiro Wilk test).
Suppose, we want to know whether the variable “systolic BP (SPSS variable
name: sbp)” is normally distributed in the population or not. We shall first
construct the histogram and Q-Q chart. To construct a histogram for systolic BP,
use the following commands:
Graphs > Legacy dialogs > Histogram > Select the variable “sbp” and push it
into the “Variable” box > Select “Display normal curve” clicking at the box >
Ok
The SPSS will produce a histogram of systolic BP, as shown in fig 5.1.
25
Figure 5.1. Histogram of systolic BP
To get the Q-Q plot for systolic BP, use the following commands:
Analyze > Descriptive statistics > Q-Q Plots > Select the variable “sbp” and
push it into the “Variables” box > For Test Distribution select “Normal” (usual-
ly remains as default) > Ok
The computer will produce a Q-Q plot of systolic BP, as shown in figure 5.2.
Figure 5.2. Q-Q plot of Systolic BP
26
To do the formal statistical tests (K-S test and Shapiro Wilk test) to understand
the normality of the data, use the following commands.
Analyze > Descriptive Statistics > Explore > Select the variable “sbp” and
push it into the "Dependent List" box > Plots > Deselect "Stem-and-leaf" and
select “Histogram” > Select “Normality plots with test” > Continue > OK
Note that these commands will also produce the histogram and Q-Q plot. You
may not need to develop histogram and Q-Q plot separately as mentioned earlier.
5.1.1 Outputs:
You will get the following tables (table 5.1 and 5.2) along with the histogram, Q-Q
plot and box and plot chart. The histogram, Q-Q plot and box and plot chart, gener-
ated by the commands, have been omitted to avoid repetition.
Table 5.1 Descriptive statistics of Systolic BP

Descriptives
Statistics Std. Error
SYSTOLIC BP Mean 127.8333 1.38161
Median 123.0000
Variance 400.857
Minimum 91.00
Maximum 195.00
Range 104.00
Skewness .728 .168
Kurtosis .343 .334
Table 5.2. Statistical tests for normality (of Systolic BP)
Tests of Normality
Kolmogorov-Smirnova Shapiro-Wilk
Statistic df Sig. Statistic df Sig.
SYSTOLIC BP .119 210 .000 .956 210 .000
a. Lilliefors Significance Correction
The SPSS has generated the histogram and Q-Q plot for “systolic BP” (fig 5.1 and
5.2) and tables 5.1 and 5.2. While getting the specific statistical tests (KS test and
27
Shapiro Wilk test) to check the normality of a dataset, the SPSS automatically
provides the descriptive statistics of the variable (systolic BP) (table 5.1). I have
already discussed the measures of skewness and kurtosis to assess the normality of
a dataset earlier (section 4).
Histogram (fig 5.1) provides an impression about the distribution of the dataset
(whether 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., distribution is not symmetrical).
The Q-Q plot (fig 5.2) also provides information on whether data have come
from a normally distributed population or not. The Q-Q plot compares the distribu-
tion of data with the standardized theoretical distribution from a specified family
of distribution (in this case normal distribution). If data are normally distributed,
all the points (dots) lie on the straight line. Note that our interest is in the central
portion of the line. Deviation from the central portion of the line means non-nor-
mality. Deviations at the ends of the plot indicate the existence of outliers. We can
see (in fig 5.2) that there is a slight deviation at the central portion as well as at the
ends. This may indicate that the data may not have come from a normally distribut-
ed population.
The specific tests (objective tests) to assess if the data have come from a
normally distributed population are the K-S (Kolmogorov-Smirnov) test and
Shapiro Wilk test. The results of these two tests are provided in table 5.2.
Look at the Sig (significance) column of table 5.2. Here, Sig (indicates the
p-value) is 0.000 for both the tests. A p-value of <0.05 indicates that the data have
not come from normally distributed population. In our example, the p-value is
0.000 for both the tests, which is <0.05. This means that the data of systolic BP
have not come from a normally distributed population. The null hypothesis here is
“data have come from a normally distributed population”. The alternative hypoth-
esis is “data have not come from a normally distributed population”. We will reject
the null hypothesis, since the p-value is <0.05.
Note that the K-S test is very sensitive to sample size. The K-S test may be
significant for slight deviations of a large sample data (n>100). Similarly, the
likelihood of getting a p-value <0.05 for a small sample (n<20, for example) is
low. Therefore, the rules of thumb for normality checking are:
1) Sample size <30: Assume non-normal;
2) Moderate sample size (30-100): If the formal test is significant (p<0.05),
consider non-normal distribution, otherwise check by other methods, e.g.,
28
histogram, Q-Q plot, etc.; and
3) Large sample size (n>100): If the formal 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
distribution is approximately symmetrical, consider that the data have come from
a normally distributed population.
29
Section 6
Data Management
While analyzing data, you may require to make class intervals, classify a group of
people with specific characteristic using a cutoff value (e.g., you may want to clas-
sify people who have hypertension using a cutoff value of either systolic or diastol-
ic BP), and recode data for other specific purposes. In this section, I shall discuss
data manipulations that are commonly needed during data analysis. For example,
• Recoding of data
• Making class intervals
• Combine data to form an additional variable
• Data transformation
• Calculation of total score
• Extraction of time
• Selection of a subgroup for data analysis
Use the data file <Data_3.sav> for practice.
6.1 Recoding of data

For example, you have the variable “sex” coded as “m” and “f”. You want to
replace the existing code “m” by 1 and “f” by 2. There are two options for recoding
data:
a) Recoding into same variable; and
b) Recoding into different variable.
My suggestion would be to use “recoding into different variable” option all the
times. This will keep the original data of the variable as it is.
6.1.1 Recoding into same variable:

Note that if you recode data into same variable, the original data/coding would be
lost. To do this follow the following commands:
Transform > Recode into Same Variables > Select “sex” and push it into the
“Variables” box > Click on “Old and new values” > Select “Value” (usually
30
default) under “Old Value” > Type m in the box below > type 1 in the “Value”
area under “New Value” > Click “Add” > Type f in the value area under “Old
Value” > type 2 in the “Value” area under “New Value” > Click Add > Contin-
ue > Ok (Fig 6.1 and 6.2)
Figure 6.1
Figure 6.2
31
Check the data file in the “data view” option. You will notice that all the “m”
has been replaced by 1 and “f” by 2. Now, go to the “variable view” of the data file
to replace the codes. Click in the “Values” box against the variable “sex”. Replace
the codes as 1 is “Male” and 2 is “Female”.
6.1.2 Recoding into different variable:

This option of recoding requires formation of a new variable. The original variable
and data will remain intact. To do this, follow the following commands:
Transform > Recode into Different Variables > Select “sex” and push it into
“Input Variable –Output Variables” box > Type “sex1” in the “Name” box and
type “Gender” in the “Label” box under “Output Variable” > Click “Change”
> Click “Old and New Values” > Type m in the “Value” box under the “Old
Value” > Type 1 in the “Value” box under the “New Value” > Click “Add” >
Type f in the “Value” box under the “Old Value” > Type 2 in the “Value” box
under the “New Value” > Click “Add” > Continue > Ok (Fig 6.3 and 6.4)
Here, the new variable generated is “sex1” (do not give any space between sex
and 1, while typing the variable name in the Name box). Follow the rules of writ-
ing variable names as mentioned in section 2 (2.1.1).
Figure 6.3
32
Figure 6.4
Check on the “variable view” option of the data file. You will notice that SPSS
has generated a new variable “sex1” (the last variable both in the variable view and
data view options). Like before, in the variable view, define the value labels of the
new variable sex1 as 1 is “male” and 2 is “female”.
6.2 Making class intervals (categories)

Suppose, you want to categorize the variable “age” into the following categories:
• ≤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)
We shall use the option “Recode into Different Variable” for this exercise. As
before, we need to generate a new variable. Suppose, the new variable we want to
generate is “age1”. I would suggest the users to use the option “Recode into Differ-
ent Variable” all the times. If you use the option “Recode into Same Variable”, you
will loss the original data that cannot be recovered once the data file is saved. To
do this, use the following commands:
33
Transform > Recode into Different Variable > Select “age” and push it into the
“Input Variable –Output Variables” box > Type “age1” in the “Name” box and
type “age group” in the “Label” box under “Output Variable” > Click
“Change” > Click on “Old and New Values” > Select “System-missing” under
“Old value” > Select “System-missing” under “New Value” > Click “Add” >
Select “Range, LOWEST through value” and type “20” in the box below >
Select “Value” under “New Value” and type “1” > Add > Now select “Range”
under “Old Value” > Type “21” in the upper box and “30” in the lower box >
Select “Value” under “New Value” and type “2” > Add > Again, type “31” in
the upper box and “40” in the lower box > Select “Value” under “New Value”
and type “3” > Add > Select “All other values” under “Old Value” > Select
“Value” under “New Value” and type “4” > Add > Continue > Ok (fig 6.5 and
6.6)
Figure 6.5
34
Figure 6.6
Go to the data file in the “data view” option and then to the “variable view”
option. You will notice that SPSS has generated a new variable “age1” (the last
variable both in the data view and variable view options). Like before, in the “vari-
able view” option, define the value labels of the variable “age1” as 1 is “≤ 20
years”, 2 is “21-30 years”, 3 is “31-40 years” and 4 is “>40 years”.
Using transformation, you can also classify people who have hypertension and
who do not have hypertension (for example). To do this, you shall have to use a
cutoff point to define hypertension. For example, we have collected data on
diastolic BP (SPSS variable name is “dbp”). We want to classify those as “hyper-
tensive” if the diastolic BP is >90 mmHg. Now, recode the variable diastolic BP
into a new variable (say, d_hyper) using “Recode into Different Variables” option
as ≤ 90 (normal BP) and > 90 (as hypertensive). Hope, you can do it now. If you
cannot, use the following commands:
Transform > Recode into Different Variable > Select “dbp” and push it into the
“Input Variable –Output Variables” box > Type “d_hyper” in the “Name” box
and type “diastolic hypertension” in the “Label” box under “Output Variable”
> Click “Change” > Click on “Old and New Values” > Select “System-miss-
ing” under “Old value” > Select “System-missing” under “New Value” > Add
> Select “Range, LOWEST through value” under “Old value” and type “90” in
35
the box below > Select “Value” under “New Value” and type “1” > Add >
Select “All other values” under “Old value” > Select “Value” under “New
Value” and type “2” > Add > Continue >Ok
This will create a new variable “d_hyper” with code numbers 1 and 2 (the last
variable both in the variable and data view options). Code 1 indicates the persons
without hypertension (diastolic BP ≤ 90) and code 2 indicates the persons with
hypertension (diastolic BP >90). As done before, in the “variable view” option,
define the value labels of the new variable “d_hyper” as 1 is “Do not have hyper-
tension” and 2 is “Have hypertension”.
6.3 Combine data into a new variable

Sometimes, the cutoff point of a measurement (e.g., hemoglobin, blood pressure,
etc.) for defining a condition (e.g., anemia, hypertension) may vary according to
gender. In such a situation, a single cutoff pint for defining a condition is not
appropriate.
For example, we have collected data on diastolic BP (SPSS variable name is
dbp) both for males and females. We have defined hypertension as diastolic BP
>85 mmHg if it is a female, and diastolic BP >90 mmHg if it is a male. Now, how
to classify those who have hypertension considering the gender?
To do this, first, we shall create a new variable, say “HTN” for which all the
values would be 0 (zero). Use the following commands to do this:
Transform > Compute Variable > Type “HTN” in “Target Variable” box >
Click in the box under “Numeric Expression” > Click “0” (zero) from the
number pad or keyboard > Ok (fig 6.7)
36
Figure 6.7
This will generate the new variable “HTN” with all the values 0 (you can check
it in the “data view” option; the last variable). Now use the following commands:
Transform > Compute Variable > Click in the box under “Numeric Expres-
sion” > Delete 0 > Click on 1 > Click If (optional case selection condition) >
Select “Include if case satisfies condition” > Select “dbp” and push it into the
box > Click “greater than sign (>)” then write “90” (always use the number
pad) > Click “&” on the “number pad” > Select “sex_1” and push it into the
box > Click on “=” and then “1” (note: 1 is the code no. for male) > Continue
> Ok > (SPSS will provide the message “Change existing variable” > Click on
“Yes” (fig 6.8 and 6.9)
Again,
Transform > Compute Variable > Click “If (optional case selection condition)”
> Delete “90” and write “85” (for dbp) and delete “1” and click “0” (for sex_1,
since 0 is the code for female) > Continue > Ok > (SPSS will give you the mes-
sage “Change existing variable” > Click “Yes”
37
Go to the “data view” option of the data file. You will notice that the new vari-
able “HTN” (the last variable both in the “data view” and “variable view” options)
has values either “0” or “1”. “0” indicates “no hypertension”, while “1” indicates
“have hypertension”. Like before, go to the “variable view” option and define the
value labels of the variable “HTN” as “0” is “No hypertension” and “1” is “Have
hypertension”.
Figure 6.8
38
Figure 6.9
6.4 Data transformation

In many situations, the data that have been collected for a study are not normally
distributed. Since parametric methods (in general) for testing hypotheses are better
than the non-parametric methods, data transformations are occasionally needed to
make the distribution normal and to meet the assumptions for a parametric test.
Depending on the shape of the data distribution, there are several transformation
options. Following table (table 6.1) shows some of the options for data transforma-
tion.
Table 6.1. 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 Zero values and negative
and big values smaller values
39
Commonly used method of data transformation is Log transformation. Let us
see how to get the Log transformation of data. Suppose, you want to transform
diastolic BP (variable name is dbp) into Log of diastolic BP. Use the following
commands:
Transform > Compute Variable > Type “log_dbp” under “Target Variable” >
Click on “Arithmetic” in the “Function Group” box > In “Functions and
special variables” box select “Lg10” > Click on the “up arrow” (left side of the
box. You will see LG10(?) appears in the Numeric Expression box) > Select
“dbp” in the “Type and Label” box > push it into the “Numeric Expression”
box > Ok (fig 6.10)
This will create a variable “log_dbp”, with the values “log of diastolic BP” (the
last variable). Similarly, you can transform your data into square root using the
option “sqrt” in the “Functions and special variables” box.
Figure 6.10
40
6.5 Calculation of total score
Suppose, you have conducted a study to assess the knowledge of the secondary
school children on how HIV is transmitted. To assess their knowledge, you have
set the following questions (data file: HIV.sav).
HIV is transmitted through:

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 syringe (variable name: k3) 1. Yes 2. No
4) Accidental needle stick injury (variable name: k4) 1. Yes 2. No
Note: All the correct answers are coded as 1.
To calculate the total knowledge score, use the following commands:
Transform > Count values within cases > Write “t_know” in the box under
“Target variable” > Write “total knowledge on HIV” in the box under “Target
level” > Select “k1, k2, k3 and k4” and push them into the “Variables” box >
Click “Define values” > Select “Value” and write “1” (since 1 is the correct
answer) in the box below > Click “Add” > Continue > OK (fig 6.11 and 6.12)
Figure 6.11
41
Figure 6.12
SPSS will generate a new variable “t_know (total knowledge on HIV)” (look
at the “variable view”). This variable has the total score of knowledge of the
students. Now, you can get the descriptive statistics and frequency of the variable
“t_know” by using the following commands:
Analyze > Descriptive statistics > Frequencies > Select “t_know” and push it
into the “Variables” box > Statistics > Select “Mean, Median and Std. devia-
tion” > Continue > Ok
You will get the tables (table 6.2 and 6.3), showing the descriptive statistics
(mean, median, etc.) and frequency distribution of total knowledge of the students.
Table 6.2 shows that the mean of the total knowledge is 2.18 (SD 0.63) and the
median is 2. Table 6.3 shows that there are 2 (1%) students who do not have any
knowledge on HIV transmission (since the score is 0, i.e., could not answer any
question correctly). One hundred and twenty five (63.8%) students know 2 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 using a cutoff value based on the total score.
42
Table 6.2. Descriptive Statistics
total knowledge
N Valid 196
Missing 0
Mean 2.18
Median 2.00
Std. Deviation .638
Table 6.3. Total knowledge

Valid 0 2 1.0 1.0 1.0
1 16 8.2 8.2 9.2
2 125 63.8 63.8 73.0
3 50 25.5 25.5 98.5
4 3 1.5 1.5 100.0
Total 196 100.0 100.0
There is an alternative way of getting the total score. In that case, the correct
answers have to be coded as 1, while the incorrect answers must be coded as 0
(zero). The commands are as follows:
Transform > Compute variable > Write “t_know” in the “Target variable” box
> Select “k1” under ‘Type and label” and push it into the “Numeric expres-
sion” box > From the key pad click “+” > Select “k2” and push it into the “Nu-
meric expression” box > From the key pad click “+” > Select “k3” and push it
into the “Numeric expression” box > From the key pad click “+” > Select “k4”
and push it into the “Numeric expression” box > OK (fig 6.13)
You will get the same results.
43
Figure 6.13
6.6 Calculation of duration

SPSS can extract time duration from dates. Suppose, you have the data on date of
admission (variable name is date_ad) and date of discharge (date_dis) of patients
admitted in a hospital. Now, you want to calculate the duration of hospital stay
(date of discharge minus date of admission). SPSS can calculate this for you. Use
the following commands:
Transform > Compute Variable > Type “dura” under “Target Variable” > Click
on “Time Duration Extraction” in the “Function Group” box > From “Func-
tions and special variables” box select “Ctime.Days” > Click on the up arrow
(at the left side of the box. You will see CTIME.DAYS(?) appears in the “Nu-
meric Expression” box > Select “date_dis” from “Type and Label” box > Push
it into the “Numeric Expression” box > Click on – (minus sign from the pad) >
44
Select “date_ad” from “Type and Label” box and push it into the “Numeric
Expression” box > Ok
You will notice that SPSS has generated a new variable “dura” (the last vari-
able) that contains the duration of hospital stay of each subject in the dataset.
6.7 Selecting a sub-group for analysis

You can select a specific sub-group for the analysis of your data. Suppose, you
want to analyze your data only for those who have diabetes mellitus. In the dataset,
the variable “diabetes” is coded as “1= yes (have diabetes)” and “2= no (do not
have diabetes)”. To select the group who have diabetes (i.e., diabetes=1), use the
following commands:
Data > Select cases > Select “If condition is satisfied” > Click on “If” > Select
the variable “diabetes” and push it into the empty box > Click “=” and then “1”
from the number pad > Continue > OK
This would exclude the subjects who do not have diabetes from the analysis.
The analysis will be only for those who have diabetes (n=45). If you make a
frequency distribution for “sex”, you will see that n=45 (table 6.4). Now, to get all
the subjects for the analysis (i.e., to deselect the subgroup), use the commands:
Data > Select cases > Select “All cases” > Ok
Table 6.4. Distribution of sex among diabetic patients

Sex: string
Cumulative
Frequency Percent Valid Percent Percent
Valid Female 20 44.4 44.4 44.4
Male 25 55.6 55.6 100.0
Total 45 100.0 100.0
45
Section 7
Testing of Hypothesis
The current and following sections provide basic information on how to select
statistical tests for testing hypotheses, perform the statistical tests in SPSS and
interpret the results of common problems related to health and social science
research. Before I proceed, let me discuss a little bit about the hypothesis.
A hypothesis is a statement about one or more populations. The hypothesis is
usually concerned with the parameters of the populations about which the state-
ment is made. There are two types of statistical hypothesis, Null (H0) and Alterna-
tive (HA) hypothesis. The null hypothesis is the hypothesis of equality or no differ-
ence. The null hypothesis always says that the two or more quantities (parameters)
are equal. Note that, we always test the null hypothesis, not the alternative hypoth-
esis. We either reject or do not reject the null hypothesis. If we can reject the null
hypothesis, then only we can accept the alternative hypothesis. It is, therefore,
necessary to have very clear understanding about the null hypothesis.
Suppose, we are interested to determine the association between coffee drink-
ing and stomach cancer. In this situation, the null hypothesis is "there is no associa-
tion between coffee drinking and stomach cancer (or, coffee drinking and stomach
cancer are independent)", while the alternative hypothesis is "there is an associa-
tion between coffee drinking and stomach cancer (or, coffee drinking and stomach
cancer are not independent) ". If we can reject the null hypothesis by a statistical
test (i.e., if the test is significant; p-value <0.05), then only we can say that there is
an association between coffee drinking and stomach cancer.
Various statistical tests are available to test hypothesis. Selecting an appropri-
ate statistical test is the key to analyze data. What statistical test to be used to test
the hypothesis depends on study design, data type, distribution of data, and objec-
tive of the study. It is, therefore, important to understand the nature of the variable
(categorical or quantitative), measurement type, as well as the study design.
Following table (table 7) provides basic guideline about the use of statistical tests
depending on the type of data and situation.
46
Table 7. Selecting statistical test for hypothesis testing
7.1 Association between quantitative and qualitative or quantitative variables
Situation for hypothesis testing Data normally Data
distributed non-normal
1 Comparison with single popula- 1-sample t-test Sign test/
tion mean (with a fixed value) Wilcoxon
Signed Rank
Example: You have taken a random
test
sample from a population of diabetic
patients to assess the mean age. Now,
you want to test the hypothesis
whether the mean age of diabetic
patients in the population is 55 years
or not.
2 Comparison of means of two Paired t-test Sign test/
related samples Wilcoxon
Signed Rank
Example: You want to test the
test
hypothesis whether the drug “Inderal”
reduces blood pressure (BP) or not.
To do this study, you have selected a
group of subjects and measured their
BP before administration of the drug
(measurements before treatment; or
pre-test). Then you have given the
drug “Inderal” to all the subjects 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) administration of the
drug is same or not.
3 Comparison between two inde- Independent samples Mann-Whitney
pendent sample means [associa- t-test U test (also
tion between quantitative and called Wilcox-
qualitative variable with 2 levels] on Rank-Sum
test)
sample of students of a university.
Now, you want to test the hypothesis
if the mean systolic blood pressure of
male and female students is same or
not.
47
Situation for hypothesis testing Data normally Data
distributed non-normal
4 Comparison of more than two One way ANOVA Kruskal Wallis
independent sample means test
[association between quantita-
tive and a categorical variable
with more than 2 levels]
sample from a population. You want
to test the hypothesis if the mean
income of different religious groups
(e.g., Muslim, Hindu and Christian) is
same or not. Another example, you
have three drugs, A, B & C. You want
to investigate whether all these three
drugs equally reduce the BP or not.
5 Association between two quanti- Pearson’s correlation Spearman’s
tative variables correlation
(Also valid for
Example: You want to test the
ordinal qualita-
hypothesis if there is a correlation
tive data)
between systolic BP and age.
7.2 Association between two qualitative variables
Situation for hypothesis testing Test statistics

1 Association between two qualitative variables (inde- Chi-square test/
pendent samples) Fisher’s Exact
test
Example: You have taken a random sample from a population
and want to test the hypothesis if 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 qualitative variables (relat- McNemar test
ed samples, such as data of a matched case-control
study design)
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 (a matched case-control
study design).
48
7.3 Multivariable analysis
Type of outcome/dependent variable Type of multivariable

analysis
1 Outcome variable (also called dependent variable) Multiple linear regres-
is in interval or ratio scale – e.g., blood pressure, sion;
birth weight, blood sugar, etc. Analysis of variance
(ANOVA)
2 Dependent variable is a dichotomous categorical Multiple logistic regres-
variable (i.e., nominal categorical variable with two sion
levels) – e.g., disease (present or absent); ANC
(taken or not taken); outcome (cured or not cured),
etc.
3 Dependent variable is a nominal categorical Multi-nominal logistic
variable with more than two levels – e.g., treatment regression
seeking behaviour (e.g., treatment not received;
treatment received from un-qualified doctor;
treatment received from qualified doctor); cause of
death (cancer, heart disease, pneumonia), etc.
4 Dependent variable is an ordinal categorical Proportional odds regres-

variable – e.g., severity of anaemia (no anaemia, sion (Ordinal regression)
mild to moderate anaemia, severe anaemia); stage
of cancer (stage1, stage 2, stage 3); severity of pain
(mild, moderate, severe), etc.
5 Dependent variable is time to outcome (time to Proportional hazards

death, time to recurrence, time to cure), etc. analysis (Cox regression)
6 Dependent variable is the counts – e.g., number of Poisson regression
post-operative infections; number of MI patients
admitted in a hospital, etc.
7 Incidence rates – incidence rate of tuberculosis; Poisson regression
incidence rate of pneumonia; incidence rate of car
accidents, etc.
49
7.4 Agreement analysis
Situation for hypothesis testing Test statistics

1 Agreement between two quantitative variables Bland Altman
Example: You want to test the hypothesis if two methods of test/plots
blood sugar measurements agree with each other or not.
2 Agreement between two categorical variables Kappa
Example: You want to test the hypothesis if diagnosis of estimates
cataract agree between two physicians.
50
Section 8
Student’s t-test for Hypothesis Testing
Student’s t-test is commonly known as t-test. It is a commonly used statistical

method to test hypothesis. There are several types of t-tests used in different situa-
tions (table 7 of section 7), such as: a) one-sample t-test; b) Independent samples
t-test; and c) Paired t-test. In this section, I shall discuss all these t-tests and inter-
pretation of the results. Use the data file <Data_3.sav> for practice.
8.1 One-sample t-test

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: dbp) of students
of the State University of Bangladesh taking a random sample. We are interested
to know if the mean diastolic BP of the students is 80 mmHg or not. Here,
Null hypothesis (H0): The mean diastolic BP of students is equal to 80 mmHg
in the population.
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.
The first job, before hypothesis testing, is to check whether the distribution of
diastolic BP is normal or not in the population (assumption 1). To do this, check
the histogram and/or Q-Q plot of diastolic BP and do the formal statistical test of
normality (K-S test or Shapiro Wilk test) as discussed in section 5. If the assump-
tion is met (diastolic BP is at least approximately normal), do the 1-sample t-test,
otherwise we have to use the non-parametric test (discussed later). Suppose,
diastolic BP is normally distributed in the population. Use the following com-
mands to do the 1-sample t-test:
51
Analyze > Compare means > One sample t-test > Select the variable “dbp”
and push it into the “Test variable(s)” box > Click in the “Test value” box and
write “80” > OK
8.1.1 Outputs:
The computer will provide the following tables (table 8.1 and 8.2).
Table 8.1. Descriptive statistics of diastolic BP
One-Sample Statistics
N Mean Std. Deviation Std. Error Mean
DIASTOLIC BP 210 83.0429 12.45444 .956
Table 8.2. One-sample t-test results
One-Sample Test
Test Value = 80
t df Sig. Mean 95% Confidence Interval of the
(2-tailed) Difference Difference
Lower Upper
DIASTOLIC BP 3.541 209 .000 3.04286 1.3486 4.7371
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 of the students is 83.04 mmHg with an SD of 12.4
mmHg (table 8.1). One-sample t-test results are shown in table 8.2. The calculated
value of “t” is 3.541 and the p-value (Sig. 2-tailed) is 0.000. Since the p-value is
<0.05, we can reject the null hypothesis at 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). The SPSS has also provided the
difference between the observed value (83.04) and hypothetical value (80.0) as
mean difference (which is 3.042) and its 95% confidence interval (1.34 – 4.73)
(table 8.2).
8.2 Independent samples t-test

Independent samples t-test involves one categorical variable with two levels (2
categories) and one quantitative variable. This test is done to compare the means
of two categories of the categorical variable. For example, we are interested to
52
know if the mean age of diabetic and non-diabetic patients is same or not. Here, the
test variable (dependent variable) is age (quantitative) and the categorical variable
is diabetes, which has two levels/categories (have diabetes and do not have diabe-
tes).
Hypothesis:
H0: The mean age of the diabetic and non-diabetic patients is same in the
population.
HA: The mean age of the diabetic and non-diabetic patients is different (not
same) in the population.
Assumptions:
1. The dependent variable (age) is normally distributed at each level of the
independent (diabetes) variable;
2. The variances of the dependent variable (age) at each level of the indepen-
dent variable (diabetes) are same/equal;
3. Subjects represent random samples from the populations.
8.2.1 Commands:
Use the following commands to do the independent samples t-test. Before doing
the test, we have to remember/check (from code book or variable view) the catego-
ry code numbers of diabetes. In our example, we have used code “1” for defining
“have diabetes” and “2” for “do not have diabetes”.
Analyze > Compare means > Independent samples t-test > Select “age” and
push it into the “test variable(s)” box and select “diabetes” for “grouping vari-
able” box > Click on “define groups” > Type 1 in “Group 1” box and type 2 in
“Group 2” box > Continue > OK
Note: You shall have to use exactly the same code numbers as it is in the data-
set for the grouping variable. Otherwise, SPSS cannot analyze the data.
8.2.2 Outputs:
The SPSS will produce the outputs as shown in table 8.3 and 8.4.
53
Table 8.3. Descriptive statistics of age by grouping variable (having diabetes)
Group Statistics
DIABETES N Mean Std. Deviation Std. Error
MELLITUS Mean
AGE Yes 45 27.9111 8.46335 1.26164
No 165 26.1333 7.18360 .55924
Table 8.4. Independent sample t-test results

Independent Samples Test
Levene's Test t-test for Equality of Means
for Equality of
Variances
F Sig. t df Sig. Mean Std. Error 95% Confidence

(2-tailed) Difference Difference Interval of the
Difference
Lower Upper
AGE Equal 3.218 .074 1.415 208 .159 1.7778 1.25671 -.699 4.255
variances
assumed
Equal 1.288 62.34 .202 1.7778 1.38003 -.980 4.536

variances
not
assumed
Table 8.3 shows the descriptive measures of age by grouping variable (diabetes).
We can see that there are 45 persons with diabetes and 165 persons without diabe-
tes. The mean age of the diabetic persons is 27.9 (SD 8.46) and that of the non-dia-
betic persons is 26.1 (SD 7.18) years.
Table 8.4 shows the t-test results. The first portion of the table indicates the
Levene’s test results. This test is done to understand if the variances of age in the
two categories of diabetes are homogeneous (equal) or not (assumption 2). Look at
the p-value (Sig.) of the Levene’s test, which is 0.074. Since the p-value is >0.05,
it indicates that the variances of age of the diabetic and non-diabetic persons are
equal (assumption 2 is fulfilled).
Now, look at the other portion of the table, the t-test for equality of means.
Here, we have to decide which p-value we shall consider. If Levene’s test p-value
is >0.05, take the t-test results at the upper row, i.e., t-test for “Equal variances
assumed”. If the Levene’s test p-value is ≤0.05, take the t-test results at the lower
54
row, i.e., t-test for “Equal variances not assumed”.
In this example, as the Levene’s test p-value is >0.05, we shall consider the
t-test results of “Equal variances assumed”, i.e., the upper row. Table 8.4 shows
that the t-value (calculated) is 1.415, and the p-value (2-tailed) is 0.159 (which is
>0.05) with 208 degrees of freedom. We cannot, therefore, reject the null hypothe-
sis. This means that the mean age of diabetic and non-diabetic persons in the popu-
lation from where samples are drawn is not different (p=0.159).
8.3 Paired t-test

The paired t-test is done to compare the difference between two means of related
samples. Related samples indicate measurements taken from the same subjects in
two or more different times/situations. For example, you have organized training
for 32 staff of your organization. To evaluate the effectiveness of the training, you
have taken a pre-test before the training to assess the current knowledge of the
participants. At the end of the training, you have again taken an examination
(post-test). Now you want to compare if the training has increased their knowledge
or not. Another example is “You want to understand the effectiveness of a drug
(e.g., Inderal) in reducing the systolic blood pressure (BP). To do this you have
selected a random sample from a population. You have measured the systolic BP
of all the individuals before giving the drug (pre-test or baseline). You have again
measured their systolic BP one-hour after giving the drug (post-test or endline)”.
Paired t-test is the appropriate test to compare the means in both these situations.
Hypothesis:
H0: There no difference of the mean scores before and after the training (for
example 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.
55
8.3.1 Commands:
Analyze > Compare means > Paired-samples t-test > Select the variables
“post-test” and “pre-test” and push them into the “Paired variables” box > OK
8.3.2 Outputs:
The SPSS will produce the following outputs (tables 8.5-8.7).
Table 8.5. Descriptive statistics of pre- and post-test results
Paired Samples Statistics

Mean N Std. Deviation Std. Error
Mean
Pair 1 Post test score 90.9844 32 8.44096 1.49216
Pre test score 53.5781 32 15.42835 2.72737
Table 8.6. Correlation between pre- and post-test results
Paired Samples Correlations

N Correlation Sig.
Pair 1 Post test score & Pre
32 .433 .013
test score
Table 8.7. Paired samples t-test results
Paired Samples Test

Paired Differences t df Sig.
(2-tailed)
Mean Std. Std. 95% Confidence
Deviation Error Interval of the
Mean Difference
Lower Upper
Pair 1 Post test
score - Pre 37.406 14.02040 2.47848 32.3514 42.4611 15.092 31 .000
test score
Table 8.5 shows the means of both the pre- (53.5) and post-test (90.9) scores along
with the standard deviations and SEs. Looking at the mean scores, we can have an
impression on 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. To understand, if
the difference between post-test mean and pre-test mean is significant or not, we
have to check the paired samples t-test results (table 8.7). Table 8.7 shows that the
mean difference between the post- and pre-test scores is 37.4. The calculated t-test
56
value is 15.09 and the p-value (sig.) is 0.000. As the p-value is <0.05, reject the
null hypothesis. This indicates that the mean knowledge score has been increased
significantly after the training (p<0.001). Note that for conclusion, we do not need
table 8.6.
57
Section 9
Analysis of Variance (ANOVA): One-way ANOVA
Analysis of variance or ANOVA is a commonly used statistical method for testing

hypothesis. ANOVA is done to compare the means when the categorical indepen-
dent variable has more than 2 levels. There are several types of ANOVA tests, such
as one-way ANOVA, two-way ANOVA, repeated-measures ANOVA and others.
In this section, I shall discuss the one-way ANOVA. Use the data file
9.1 One-way ANOVA

The one way-ANOVA test is done to compare the means of more than two groups,
while t-test compares the means of two groups. The ANOVA test involves two
variables, one categorical variable with more than two levels/categories (for exam-
ple, in our data the variable “religion” [variable name “religion_2”] has 4 catego-
ries – Muslim, Hindu, Christian and Buddhism) and a quantitative variable (e.g.,
income, age, blood pressure, etc.). Suppose, you want to assess if the mean income
of all the religious groups is 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 same/equal.
HA: Not all the means (of income) of religious groups are same.
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) for each level of the inde-
pendent variable (religion) are same; and
If the variances of the dependent variable in all the categories are not equal
(violation of assumption 2), but sample size in all the groups is large and similar,
58
ANOVA can be used.
9.1.1 Commands:
Analyze > Compare means > One-way ANOVA > Select “income” and push it
into the “Dependent list” box > Select “religion_2” for the “Factor” box >
Options > Select “Descriptive” and “Homogeneity of variance test” > Contin-
ue > OK
9.1.2 Outputs:
The SPSS will generate the following outputs (table 9.1-9.3).
Table 9.1. Descriptive statistics of income by religious groups
Descriptives
INCOME
N Mean Std. Std. Error 95% Confidence Minimum Maximum
Deviation Interval for Mean
Lower Upper
Bound Bound
MUSLIM 126 88180.90 17207.61 1532.976 85146.95 91214.85 55927 117210
HINDU 36 79166.03 17804.63 2967.439 73141.81 85190.25 53435 110225
CHRISTIAN 26 79405.62 19857.02 3894.282 71385.19 87426.04 52933 114488
BUDDHISM 22 84796.59 14447.34 3080.185 78391.00 91202.19 56249 109137
Total 210 85194.49 17724.03 1223.074 82783.34 87605.63 52933 117210
Table 9.2. Levene’s test for homogeneity of variances of income in different religious groups
Test of Homogeneity of Variances

INCOME
Levene Statistic df1 df2 Sig.
2.056 3 206 0.107
Table 9.3. ANOVA test results
ANOVA
INCOME
Sum of Squares df Mean Square F Sig.
Between Groups 3306848581.156 3 1102282860.385 3.642 .014
Within Groups 62348694323.301 206 302663564.676
Total 65655542904.457 9
59
In this example, I have used “income” as the dependent variable and “religion” as
the independent (or factor) variable. The independent variable (religion) has 4
categories (levels) – Muslim, Hindu, Christian and Buddhism.
Table 9.1 provides all the descriptive measures (mean, SD, SE, 95% CI, etc.)
of income by religion. For example, the mean income of Muslims is 88,180.9 with
SD of 17,207.6.
The second table (table 9.2) shows the test results of homogeneity of variances
(Levene’s test). This test was done to understand if all the group variances of
income are equal or not (assumption 2). Look at the p-value (Sig.), which is 0.107.
The p-value is >0.05, which means that the variances of income in all the religious
groups are equal (i.e., assumption 2 is not violated).
Now, look at the ANOVA table (table 9.3). The value of F-statistic is 3.642 and
the p-value is 0.014. Since the p-value is <0.05, reject the null hypothesis. This
means that, not all group means (of income) are same.
However, the ANOVA test does not provide information about which group
means are different. To understand which group means are different, we need to
use the post hoc multiple comparison test, such as Tukey’s test or Bonferroni test.
Use the following commands to get the post-hoc test results. Note that if the ANO-
VA test (F-test) is not significant (i.e., p-value is >0.05), we do not need the
post-hoc test.
Options > Select “Descriptive”, and “Homogeneity of variance test” > Contin-
ue > Post Hoc > Select “Tukey” (or Bonferroni) > Continue > OK
The SPSS will produce the following table (table 9.4) in addition to others.
60
Table 9.4. Comparisons of mean income between the religious groups
Multiple Comparisons
Dependent Variable: INCOME
Tukey HSD
(I) (J) Mean Std. Error Sig. 95% Confidence Interval
RELIGION RELIGION Difference (I-J)
Lower Upper
Bound Bound
MUSLIM HINDU 9014.88(*) 3287.767 .033 499.15 17530.60
CHRISTIAN 8775.29 3747.399 .092 -930.94 18481.52
BUDDHISM 3384.31 4019.891 .834 -7027.71 13796.33
HINDU MUSLIM -9014.88(*) 3287.767 .033 -17530.60 -499.15
CHRISTIAN -239.59 4477.525 1.000 -11836.93 11357.76
BUDDHISM -5630.56 4707.946 .630 -17824.73 6563.60
CHRISTIAN MUSLIM -8775.29 3747.399 .092 -18481.52 930.94
HINDU 239.59 4477.525 1.000 -11357.76 11836.93
BUDDHISM -5390.98 5039.677 .708 -18444.37 7662.41
BUDDHISM MUSLIM -3384.31 4019.891 .834 -13796.33 7027.71
HINDU 5630.56 4707.946 .630 -6563.60 17824.73
CHRISTIAN 5390.98 5039.677 .708 -7662.41 18444.37
* The mean difference is significant at the .05 level.
9.1.3.1 Interpretation of multiple comparisons table:

Table 9.4 shows the mean difference of income in different religious groups. We
can see that the difference of mean income of Muslims and Hindus is 9,014.88 (the
minus sign in row 4 indicates that Muslims have greater income than Hindus). The
p-value (Sig.) of this difference is 0.033, which is <0.05. This indicates that the
mean income of Muslims and Hindus may be different in the population (Muslims
have higher mean income compared to Hindus). The difference of means of other
religious groups are not significant as p-values are >0.05. The table has also
provided the 95% CI of the mean differences.
9.1.4 Graph on distribution of medians/means:

You can generate the box and plot charts to see the distribution of medians/means
of the dependent variable across the groups (at each level of the independent vari-
able, i.e., in different religious groups). To have the box and plot chart, use the
following commands:
Graphs > Legacy dialogs > Boxplots… > Simple > Select “Summaries for
groups of cases (already selected by default)” > Define > Select “income” for
“Variable” box and select “religion_2” for “Category axis’ box > Ok
61
The SPSS will generate the following box and plot chart (fig 9.1). The horizontal
line within the box indicates the median.
Figure 9.1. Box and plot chart of income by religious groups

140000
120000
100000
80000
60000
40000
N = 126 36 26 22
MUSLIM HINDU CHRISTIAN BUDDAH
RELIGION
9.1.5 What to do if the variances are not homogeneous?

If the group variances are not equal (i.e., Levene’s test p-value is <0.05), for the
comparison of group means, we have to use the Welch test (or Browne-Forsythe
test). Similarly, for the comparison of individual group means, instead of Tukey’s
(or Bonferroni) test use the Games-Howell test. Use the following commands to
get these test results:
Options > Select “Descriptive”, “Homogeneity of variance test” and “Welch”
> Continue > Post Hoc > Select “Games-Howell” under the “Equal Variances
not Assumed” > Continue > OK
9.1.5.1 Outputs:
The SPSS will produce the following additional tables (table 9.5 and 9.6).
Table 9.5. Welch test for equality of means
Robust Tests of Equality of Means

INCOME
Statistica df1 df2 Sig.
Welch 3.292 3 56.236 .027
a. Asymptotically F distributed.
62
Table 9.6. Comparison of means of income between the religious groups
income
Games-Howell
(I) religion (J) religion Mean Std. Error Sig. 95% Confidence Interval
Difference (I-J) Lower Upper
Bound Bound
MUSLIM HINDU 9014.877* 3340.016 .044 166.37 17863.38
CHRISTIAN 8775.289 4185.146 .175 -2541.95 20092.53
BUDDHISM 3384.314 3440.575 .760 -5931.90 12700.53
HINDU MUSLIM -9014.877* 3340.016 .044 -17863.38 -166.37
CHRISTIAN -239.588 4896.032 1.000 -13248.23 12769.05
BUDDHISM -5630.563 4277.059 .557 -16986.14 5725.02
CHRISTIAN MUSLIM -8775.289 4185.146 .175 -20092.53 2541.95
HINDU 239.588 4896.032 1.000 -12769.05 13248.23
BUDDHISM -5390.976 4965.176 .700 -18635.83 7853.88
BUDDHISM MUSLIM -3384.314 3440.575 .760 -12700.53 5931.90
HINDU 5630.563 4277.059 .557 -5725.02 16986.14
CHRISTIAN 5390.976 4965.176 .700 -7853.88 18635.83
*. The mean difference is significant at the 0.05 level.
9.1.5.2 Interpretation:
Table 9.5 shows the Welch test results of comparison of means (Robust Tests of
Equality of Means). Just look at the p-value (Sig.). The p-value is 0.027, which is
<0.05. This means that the mean income of all the religious groups is not same in
the population.
Table 9.6 conveys the same information as of Tukey’s test that I have discussed
earlier. Here, the difference of mean income of Muslims and Hindus is significant-
ly different as indicated by the p-value (Sig.) (p=0.044). The difference of means
among the other religious groups is not significant. The table has also provided the
95% CI of the differences.
63
Section 10
Two-way ANOVA
Two-way ANOVA is like one-way ANOVA except that it examines an additional

independent categorical variable. Therefore, the two-way ANOVA involves three
variables – one quantitative (dependent variable) and two categorical variables.
This test is not commonly used in health research. Use the data file <Data_3.sav>
for practice.
10.1 Two-way ANOVA

Suppose, we want to compare the mean systolic BP (SPSS variable name is “sbp”)
in different occupation and sex (male and female) groups. Here, the dependent
variable is systolic BP and the independent variables are occupation and sex.
Since we have 4 levels/categories in occupation (govt. job; private job; busi-
ness and others) and two categories in sex (male and female), we have a factorial
design with 8 (4X2) data cells. The two-way ANOVA test answers the following 3
questions:
1. Does occupation influence the systolic BP (i.e., is mean systolic BP among
the occupation groups same)?
2. Does sex influence the systolic BP (i.e., is the mean systolic BP same for
males and females)?
3. Does the influence of occupation on systolic BP depends on sex (i.e., is
there interaction between occupation and sex)?
Questions 1 and 2 refer to the main effect, while question 3 explains the inter-
action of two independent variables (occupation and sex) on the dependent vari-
able (systolic BP).
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
independent variables are same; and
64
First of all, we have to check for normality of data (systolic BP) in different
categories of occupation and sex separately using histogram, Q-Q plot and Shapiro
Wilk test (or, K-S test). We also need to check the homogeneity of variances in
each group of the independent variables (occupation and sex) using the Levene's
test.
10.1.1 Commands:
Analyze > General linear model > Univariate > Select “sbp” for "Dependent
variable" box and select “occupation” and “sex” for “Fixed factors” box >
Options > Select “Descriptive statistics, Estimates of effect size and Homoge-
neity test" > Continue > OK
10.1.2 Outputs:
The SPSS will give you the following outputs (table 10.1-10.4).
Table 10.1. Frequency distribution of independent variables
Between-Subjects Factors
Value Label N
OCCUPATION 1 GOVT JOB 60
2 PRIVATE JOB 49
3 BUSINESS 49
4 OTHERS 52
SEX f FEMALE 133
m MALE 77
65
Table 10.2. Descriptive statistics of systolic BP by occupation and sex
Descriptive Statistics
Dependent Variable: SYSTOLIC BP
OCCUPATION SEX Mean Std. Deviation N
GOVT JOB FEMALE 130.84 21.264 38
MALE 126.86 19.548 22
Total 129.38 20.574 60
PRIVATE JOB FEMALE 131.26 21.534 31
MALE 117.89 13.394 18
Total 126.35 19.894 49
BUSINESS FEMALE 131.10 24.023 31
MALE 123.44 14.448 18
Total 128.29 21.178 49
OTHERS FEMALE 125.73 18.772 33
MALE 129.26 19.084 19
Total 127.02 18.778 52
Total FEMALE 129.73 21.309 133
MALE 124.56 17.221 77
Total 127.83 20.021 210
Table 10.3. Levene’s test result for equality of variances
Levene's Test of Equality of Error Variances(a)

F df1 df2 Sig.
1.794 7 202 .090
Tests the null hypothesis that the error variance of the dependent variable is equal across
groups.
a. Design: Intercept+OCCUPATION+SEX+OCCUPATION * SEX
Table 10.4. The two-way AVOVA table
Tests of Between-Subjects Effects

Type III Sum of df Mean f Sig. Partial
Source Squares Square Eta
Squared
Corrected Model 3370.426(a) 7 481.489 1.210 .299 .040
Intercept 3127267.594 1 3127267.594 7856.211 .000 .975
OCCUPATION 470.413 3 156.804 .394 .758 .006
SEX 1394.678 1 1394.678 3.504 .063 .017
OCCUPATION * SEX 1769.393 3 589.798 1.482 .221 .022
Error 80408.741 202 398.063
Total 3515465.000 210
Corrected Total 83779.167 209
a R Squared = .040 (Adjusted R Squared = .007)
66
Table 10.1 (between-subjects factors) shows the frequencies of occupation and
sex. Table 10.2 (descriptive statistics) shows the descriptive measures of systolic
BP at different levels of occupation and sex. For example, the mean systolic BP of
females doing the government job is 130.84 (SD 21.2) and that of males doing the
government job is 126.8 (SD 19.5).
Table 10.3 shows the Levene’s test results for homogeneity of variances. The
p-value (Sig.) of the test, as shown in the table, is 0.090. A p-value >0.05 indicates
that the variances of systolic BP at each level of the independent variables (occu-
pation and sex) are not different. Thus, the assumption 2 is not violated.
The table of “Tests of between-subjects effects" (table 10.4) shows the main
effects of the independent variables. Look at the p-values (Sig.) of occupation and
sex. They are 0.758 and 0.063, respectively. This indicates 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”, which indicates the
significance of the interaction between these two variables on systolic BP. A p-val-
ue ≤0.05 indicates the presence of interaction, that means that the systolic BP of
different occupation groups is influenced by (depends on) sex. In our example, the
p-value is 0.221 (>0.05), which means that there is no interaction between occupa-
tion and sex to influence the systolic BP. The Partial Eta Squared (last column of
the table) indicates the effect size. The Eta statistics for occupation and sex are
0.006 and 0.017, which are very small. These values are equivalent to R2 (Coeffi-
cient of Determination). Eta 0.006 indicates that only 0.6% variance of systolic BP
can be explained by occupation (and 1.7% by sex). However, most of the research-
ers do not report this in their publications.
The Post-hoc test (as 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 it is not necessary. To have a clearer picture of the presence of interac-
tion, it is better to get a graph of the mean systolic BP for occupation and sex. Use
the following commands to get the graph.
Graphs > Line > Select "Multiple" > Select "Summarizes for groups of cases"
> Define > Select "Other summary function" > Move the dependent variable
(sbp) into the "Variable" box > Move one independent variable (occupation)
with most categories (here occupation has more categories than sex) into the
67
"Category axis" box > Move the other independent variable (sex) into "Define
line by" box > OK
This will produce the following graph (fig 10.1). The graph shows that there is
a greater difference in mean systolic BP between males (117.89) and females
(131.26) among private job holders, compared to other occupations. However, this
difference is not statistically significant to show an interaction between occupation
and sex. This means that there is no significant variation of systolic BP in the occu-
pation groups by sex.
Figure 10.1. Interaction between occupation and sex on systolic BP

134
132
130
128
126
124
122
120
SEX
118 FEMALE
116 MALE
GOVT JOB PRIVATE JOB BUSINESS OTHERS
OCCUPATION
68
Section 11
Repeated Measures ANOVA: One-way
The one-way repeated measures ANOVA test is analogous to paired samples t-test
that I have discussed earlier. The main difference is that, in paired samples t-test
we have two measurements at different times (e.g., before and after giving a drug,
or pre-test and post-test results) on the same subjects, while in 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 measurements
over a period of time or conditions). The one-way repeated measures ANOVA is
also called one-way within-subjects ANOVA. Use the data file <Data_re-
peat_anova_2.sav> for practice.
11.1 One-way repeated measures ANOVA

Suppose, we are interested to assess the mean blood sugar levels at 4 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 assess whether the drug
reduces the blood sugar levels over time.
To conduct this study, we have selected 15 individuals randomly from a popu-
lation and measured their blood sugar levels at the baseline (hour-0). All the indi-
viduals were then provided with the drug (say, drug A) and their blood sugar levels
were again measured at hour-7, hour-14 and hour-24. We are interested to know if
the blood sugar levels over time, after giving the drug, are same or not (in other
words, whether the drug is effective in reducing the blood sugar levels over time).
The variables hour-0, hour-7, hour-14, and hour-24 are named in SPSS as sugar_0,
sugar_7, sugar_14 and sugar_24, respectively. Note that, in this example, we have
only one treatment group (received drug A) but have the outcome measurements
(blood sugar) at 4 different points in time on the same subjects (i.e., we have one
treatment group with 4 levels of measurements).
Hypothesis:
H0: The mean blood sugar level is same/equal at each level of measurement
(i.e., the population mean of blood sugar at 0, 7, 14 and 24 hours is same).
69
HA: The mean blood sugar is not same at different levels of measurement (that
is, population mean of blood sugar at 0, 7, 14 and 24 hours is different).
Assumptions:
1. The dependent variable (blood sugar level) is normally distributed in the
population at each level of within-subjects factor;
2. The population variances of the differences between all combinations of
related groups/levels are equal (called Sphericity assumption); and
3. The subjects represent a random sample from the population.
11.1.1 Commands:
Analyze > General linear model > Repeated measures > In “Within subject
factor name” box write “time” (give any other name) after deleting factor1 >
in “Number of levels” box write “4” (since there are 4 time factors) > Add >
Write “blood_sugar” in “Measures Name” box > Add > Define > Select vari-
ables “sugar_0, sugar_7, sugar_14 and sugar_24” and push them into "With-
in-Subjects Variables" box > Options > Select "Descriptive statistics, Esti-
mates of effect size and Homogeneity tests" > Select “time” and push it into the
“Display means for” box > Select “Compare main effects” > Select “Bonfer-
roni” in box “Confidence interval adjustment” > Continue > Pots > Select
“time” and push it into “Horizontal axis” box > Add > Continue > OK
11.1.2 Outputs:
The SPSS will produce several tables. However, we need only the following tables
(table 11.1-11.7) for interpreting the results. The tables are set chronologically for
easier interpretation (not in the order as provided by SPSS).
Table 11.1. Codes for different levels of measurements of blood sugar
Within-Subjects Factors
Measure: Blood_sugar
Time Dependent Variable
1 sugar_0
2 sugar_7
3 sugar_14
4 sugar_24
70
Table 11.2. Descriptive statistics of blood sugar at different levels (times) of measurement
Mean Std. Deviation N
Blood sugar at hour 0 109.200 5.12975 15
Table 11.3. Descriptive statistics of blood sugar at different levels of measurement with 95% CI
Estimates
Time Mean Std. Error 95% Confidence Interval
Lower Bound Upper Bound
1 109.200 1.324 106.359 112.041
2 103.733 .963 101.667 105.800
3 97.867 1.055 95.604 100.129
4 98.133 1.515 94.885 101.382
Table 11.4. Mauchly’s test for Sphericity assumption
Mauchly's Test of Sphericityb

Within Mauchly's W Approx. df Sig. Epsilona
Subjects Chi-Square Green- Huynh-Feldt Lower-bound
Effect house-Geisser
Time .095 29.998 5 .000 .436 .463 .333
Tests the null hypothesis that the error covariance matrix of the orthonormalized transformed dependent
variables is proportional to an identity matrix.
a. May be used to adjust the degrees of freedom for the averaged tests of significance. Corrected tests are
displayed in the Tests of Within-Subjects Effects table.
b. Design: Intercept
Within Subjects Design: Time
Table 11.5. Test results of within subject effects (alternative univariate tests)
Tests of Within-Subjects Effects

Source Type III df Mean F Sig. Partial
Sum of Square Eta
Squares Squared
Time Sphericity Assumed 1299.933 3 433.311 29.023 .000 .675
Greenhouse-Geisser 1299.933 1.309 993.198 29.023 .000 .675
Huynh-Feldt 1299.933 1.389 935.638 29.023 .000 .675
Lower-bound 1299.933 1.000 1299.933 29.023 .000 .675
Error(Time) Sphericity Assumed 627.067 42 14.930
Greenhouse-Geisser 1299.933 18.324 34.222
Huynh-Feldt 1299.933 19.451 32.238
Lower-bound 1299.933 14.000 44.790
71
Table 11.6. Multivariate test results
Multivariate Testsb
Effect Value F Hypothesis Error df Sig. Partial Eta
df Squared
Time Pillai's Trace .918 44.724a 3.000 12.000 .000 .918
Wilks' Lambda .082 44.724a 3.000 12.000 .000 .918
Hotelling's Trace 11.181 44.724a 3.000 12.000 .000 .918
Roy's Largest
11.181 44.724a 3.000 12.000 .000 .918
Root
a. Exact statistic
b. Design: Intercept
Within Subjects Design: Time
Table 11.7. Pair-wise comparison of mean blood sugar at different times of measurement
Pairwise Comparisons
Measure:Blood_sugar
(I) Time (J) Time Mean Std. Error Sig.a 95% Confidence Interval
Difference (I-J) Differencea
1 2 5.467* 1.064 .001 2.202 8.732
3 11.333* 1.260 .000 7.467 15.200
4 11.067* 2.256 .001 4.143 17.990
2 1 -5.467* 1.064 .001 -8.732 -2.202
3 5.867* .608 .000 4.000 7.734
4 5.600* 1.492 .013 1.021 10.179
3 1 -11.333* 1.260 .000 -15.200 -7.467
2 -5.867* .608 .000 -7.734 -4.000
4 -.267 1.240 1.000 -4.072 3.539
4 1 -11.067* 2.256 .001 -17.990 -4.143
2 -5.600* 1.492 .013 -10.179 -1.021
3 .267 1.240 1.000 -3.539 4.072
Based on estimated marginal means

*. The mean difference is significant at the .05 level.
a. Adjustment for multiple comparisons: Bonferroni.
Figure 11.1. Mean blood sugar at different times of measurement
72
The outputs of the analysis are shown in tables 11.1-11.7 and figure 11.1. Table
11.1 shows the value labels (times) of the blood sugar measurements. Table 11.2
and 11.3 (descriptive statistics and estimates) shows the descriptive statistics
(mean, standard deviation, no. of study subjects, SE of the means, 95% CI, etc.) of
the blood sugar levels at different times of measurement, such as at hour-0, hour-7,
hour-14 and hour-24.
One of the important issues for the repeated measures ANOVA test is the Sphe-
ricity assumption, as mentioned earlier under “assumptions”. Table 11.4 shows the
test results of “Mauchly’s test of Sphericity” to understand whether the Sphericity
assumption is correct or violated. The table shows that the Mauchly’s W is 0.095
and the p-value is 0.000. Since the p-value is less <0.05, the Sphericity assumption
is violated (not correct).
Three types of tests are conducted if within-subjects factors (here, it is the
times of measurement of blood sugar, which has 4 levels – hour-0, hour-7, hour-14
and hour-24) have more than 2 levels (here, we have 4 levels). The tests are:
1. Standard univariate test (Sphericity Assumed) [table 11.5];
2. Alternative univariate tests (Greenhouse-Geisser; Huynh-Feldt; Lower-
bound) [table 11.5]; and
3. Multivariate tests (Pillai's Trace; Wilks' Lambda; Hotelling's Trace; Roy's
Largest Root) [table 11.6]
All these tests evaluate the same hypothesis – i.e., the population means are
equal at all levels of the measurement. The standard univariate test is based on the
Sphericity assumption, i.e., the standard univariate test result is considered, if
Sphericity assumption is correct (not violated). In reality and in most of the cases
(also in our example), the Sphericity assumption is violated, and we cannot use the
standard univariate test (Sphericity assumed as given in table 11.5) result.
In our example, we see that the Sphericity assumption is violated, since the
Mauchly’s test p-value is 0.000 (table 11.4). Therefore, we shall have to pick up
the test results either from alternative univariate tests (table 11.5) or multivariate
tests (table 11.6). For practical purpose, it is recommended to use the multivariate
test results for reporting, since it does not depend on Sphericity assumption.
However, for the sake of better understanding, let me discuss table 11.5, which
indicates the standard and alternative univariate test results. Table 11.5 shows the
73
univariate test results of within-subjects effects. The standard univariate ANOVA
test result is indicated by the row “Sphericity Assumed”. Use this test result, when
Sphericity assumption is correct/not violated (i.e., Mauchly’s test p-value is
>0.05). Since, our data shows that the Sphericity assumption is violated, we cannot
use this test result.
When Sphericity assumption is violated (not correct), you can use the results
of one of the alternative univariate tests (i.e., Greenhouse-Geisser, Huynh-Feldt or
Lower-bound) for interpretation. It is commonly the Greenhouse-Geisser test,
which is reported by the researchers. Table 11.5 shows that the test (Green-
house-Geisser) provided the same F-value and p-value like other tests. Since the
test’s p-value is 0.000, reject the null hypothesis. This means that, the mean blood
sugar levels at different time factors (i.e., at different levels of measurement) are
not same.
To make it simple, I would suggest to use the multivariate test results, which
are not dependent on the Sphericity assumption. Table 11.6 shows the multivariate
test results. In the multivariate tests table, the SPSS has given several test results,
such as Pillai's Trace, Wilks’ Lambda, Hotelling’s Trace and Roy’s Largest Root.
All these multivariate tests have given the same results. It is recommended to use
the Wilks’ Lambda test results for reporting. In our example, the multivariate test
indicates significant time effect on blood sugar levels, as the p-value of Wilks’
Lambda is 0.000. This means that the population means of blood sugar levels at
different time factors (different times of measurement) are not same.
The last table (table 11.7) shows pairwise comparison of means at different
times of measurement. It shows the results as we have seen under one-way ANO-
VA (table 9.4; Tukey HSD). Look at the p-values. It is better to assess the differ-
ences of adjacent measurements, such as the difference of blood sugar levels
between “time 1 & 2”, “time 2 & 3” and “time 3 & 4”. The table shows that all the
differences have p-values <0.05, except for “time 3 and 4” (p=1.0). This means
that mean blood sugar levels are significantly different in all adjacent time periods
except for the time between 3 and 4. The mean blood sugar levels at different times
of measurement are depicted in figure 11.1.
Note that if the overall test is not significant (i.e., p-value of Wilks’ Lambda is
>0.05), the table for pairwise comparison is not necessary.
74
Section 12
Repeated Measures ANOVA: Within and Between-Subjects
The within and between-subjects ANOVA is also called two-way repeated mea-
sures ANOVA. In the previous section, I have discussed the one-way repeated
measures ANOVA, which is also called within-subjects ANOVA. In within-sub-
jects ANOVA, we have only one treatment (intervention) group. On the other hand,
the within and between-subjects ANOVA is used when there are more than one
treatment group. In this method, at least 3 variables are involved – one dependent
quantitative variable, and two independent categorical variables with two or more
levels. Use the data file <Data_repeat_anova_2.sav> for practice.
12.1 Within and between-subjects ANOVA

Suppose, the researcher wants to do an experiment to compare the efficacy of two
drugs (to answer which one is more effective) in reducing the blood sugar levels
over time. In such a situation, the researcher may have the following questions to
answer:
1. Is there a difference in mean blood sugar levels between drug A and drug B?
This is termed Between-Subjects Factor – a factor that divides the subjects
into two or more distinct subgroups.
2. Is there a reduction in mean blood sugar levels over a time period? This is
termed Within-Subjects Factor – distinct measurements are made on the sa-
me subjects over time. For example, blood sugar levels over time or blood
pressure over time, etc.
3. Is there a group-time interaction? If there is a time trend, and whether this
trend exists for all groups or only for certain groups?
To answer these questions, we have to use within and between-subjects repeat-
ed measures ANOVA.
Suppose, the researcher has decided to compare the efficacy of Daonil (Gliben-
clamide) and Metformin (these drugs are used for the treatment of diabetes melli-
tus) on the reduction of blood sugar levels. In this example, there are 2 treatment
groups (SPSS variable name is “treatment”) – Daonil and Metformin. To do the
75
experiment, the researcher has selected 10 subjects and randomly allocated the
treatments (5 in each group). Blood sugar levels of the subjects were measured at
the baseline (sugar_0), after 7 hours (sugar_7), after 14 hours (sugar_14) and after
24 hours (sugar_24). Data is provided in the data file <Data_Re-
peat_anova_2.sav>.
Hypothesis:
We test two hypotheses here. One is for within-subjects effects and the other is for
between-subjects effects.
H0: Daonil and Metformin are equally effective in reducing the blood sugar
levels over time (between-subjects effects).
HA: Both these drugs are not equally effective in reducing the blood sugar level
over time (you cans also use one sided hypothesis, such as “Daonil is more
effective in reducing the blood sugar levels over time compared to Met-
formin”.
We can also test the hypothesis whether these drugs are effective in reducing
the blood sugar levels over time (within-subjects effects; discussed in section 11).
The assumptions of two-way repeated measures ANOVA are same as one-way
repeated measures ANOVA.
12.1.1 Commands:
Analyze > General linear model > Repeated measures > Write “time” in
“Within subject factor name” box > Write “4” in “Number of levels” box
(since we have 4 time levels) > Add > Write “blood_sugar” in “Measures
name” box > Add > Define > Select variables “sugar_0, sugar_7, sugar_14 and
sugar_24” and push them into "Within-Subjects Variables" box > Select “treat-
ment” and push it into “Between-subjects factors” box > Options > Select
“treatment” and push it into the “Display means for” box> Select “Compare
main effects” > Select “Bonferroni” in “Confidence interval adjustment” box
> Select "Descriptive statistics, and homogeneity tests" > Continue > Contrasts
> Select “time” > Select “Repeated” in the “Contrast” box under “Change
contrast” > Change > Continue > Plots > Select “time” and push it into “Hori-
zontal axis” box > Select “treatment” and push it into the “Separate lines” box
76
12.1.2 Outputs:
The SPSS will provide many tables, but only the relevant ones are provided below.
The outputs are arranged according to – A) Basic tables; B) Tables related to With-
in-subjects effects; C) Tables related to Between-subjects effects; D) Tables to
check the assumptions; and E) Additional tables.
A. Basic tables (table 12.1-12.3):
Table 12.1. Codes for different levels of measurement
Within-Subjects Factors
Measure: Bloodsugar
Sugar Dependent Variable
1 sugar_0
2 sugar_7
3 sugar_14
4 sugar_24
12.2. Codes of different treatment groups
Between-Subjects Factors
Value Label N
treatment groups 1 Daonil 5
2 Metformin 5
Table 12.3. Descriptive statistics of blood sugar at different levels and treatment groups
treatment groups Mean Std. Deviation N
Blood sugar at hour 0 Daonil 112.8000 2.16795 5
Metformin 108.4000 7.09225 5
Total 110.6000 5.46097 10
Metformin 103.0000 4.69042 5
Total 103.5000 4.22295 10
Metformin 98.6000 3.91152 5
Total 98.0000 3.52767 10
Metformin 97.6000 2.50998 5
Total 96.0000 2.98142 10
77
B. Within-subjects effects (table 12.4-12.6):
Table 12.4. Within-subjects multivariate test results
Multivariate Testsb
Effect Value F Hypothesis Error df Sig. Partial Eta
df Squared
time Pillai's Trace .955 42.767a 3.000 6.000 .000 .955
Wilks' Lambda 42.767a
.045 44.724 3.000 6.000 .000 .955
Hotelling's Trace 21.384 42.767a 3.000 6.000 .000 .955
Roy's Largest
21.384 42.767a 3.000 6.000 .000 .955
Root
time * treatment Pillai's Trace .452 1.649 a
3.000 6.000 .275 .452
Wilks' Lambda .548 1.649a 3.000 6.000 .275 .452
Hotelling's Trace .825 1.649a 3.000 6.000 .275 .452
Roy's Largest
.825 1.649a 3.000 6.000 .275 .452
Root
a. Exact statistic
b. Design: Intercept + treatment
Within Subjects Design: time
Table 12.5. Descriptive measures of blood sugar at different levels of time
Estimates
Measure: Bloodsugar
Time Mean Std. Error 95% Confidence Interval
1 110.600 1.658 106.776 114.424
2 103.500 1.405 100.259 106.741
3 98.000 1.164 95.316 100.684
4 96.000 .825 94.098 97.902
Table 12.6. Pairwise comparisons of blood sugar levels at different time intervals
Tests of Within-Subjects Contrasts

Measure:bloodsugar
Type III Sum of df Mean F Sig. Partial
Source time Squares Square Eta
Squared
time Level 1 vs. Level 2 504.100 1 504.100 48.010 .000 .857
Level 2 vs. Level 3 302.500 1 302.500 99.180 .000 .925
Level 3 vs. Level 4 40.000 1 40.000 4.000 .081 .333
time * treatment Level 1 vs. Level 2 28.900 1 28.900 2.752 .136 .256
Level 2 vs. Level 3 12.100 1 12.100 3.967 .082 .332
Level 3 vs. Level 4 10.000 1 10.000 1.000 .347 .111
Error(time) Level 1 vs. Level 2 84.000 8 10.500
Level 2 vs. Level 3 24.400 8 3.050
Level 3 vs. Level 4 80.000 8 10.000
78
C. Between-subjects effects (table 12.7-12.9 & fig 12.1):
Table 12.7. Test results of between-subjects effects

Measure:Bloodsugar
Transformed Variable:Average
Source Type III Sum of df Mean F Sig. Partial Eta
Squares Square Squared
Intercept 416364.025 1 416364.025 9211.593 .000 .999
treatment .625 1 .625 .014 .909 .002
Error 361.600 8 45.200
Table 12.8. Descriptive statistics of treatment groups
Estimates
Measure:Bloodsugar
treatment Mean Std. Error 95% Confidence Interval
groups Lower Bound Upper Bound
Daonil 102.150 1.503 98.683 105.617
Metformin 101.900 1.503 98.433 105.367
Table 12.9. Pairwise comparisons by treatment groups
Measure:Bloodsugar
(I) treatment (J) treatment Mean Std. Error Sig.a 95% Confidence Interval
groups groups Difference (I-J) Differencea
Daonil Metformin .250 2.126 .909 -4.653 5.153
Metformin Daonil -.250 2.126 .909 -5.153 4.653

Figure 12.1. Blood sugar levels by treatment group (Daonil and Metformin)
79
D. Tables for checking assumptions (table 12.10-12.12):
Table 12.10. Box’s M test
Box's Test of Equality of Covariance Matricesa

Box's M 14.734
F .633
df1 10
df2 305.976
Sig. .785
Tests the null hypothesis that the observed covariance matrices of the
dependent variables are equal across groups.
a. Design: Intercept + treatment
Table 12.11. Levene’s test of equality of variances
Levene's Test of Equality of Error Variancesa

F df1 df2 Sig.
Blood sugar at hour 0 3.805 1 8 .087
Blood sugar at hour 7 .076 1 8 .790
Tests the null hypothesis that the error variance of the dependent variable is equal across
groups.
a. Design: Intercept + treatment
Table 12.12. Mauchly’s test for Sphericity assumption
Mauchly's Test of Sphericityb

Measure: Bloodsugar
Within Mauchly's W Approx. df Sig. Epsilonb
Subjects Chi-Square Green- Huynh-Feldt Lower-bound
Effect house-Geisser
Time .124 14.007 5 .017 .534 .731 .333
Tests the null hypothesis that the error covariance matrix of the orthonormalized transformed dependent
variables is proportional to an identity matrix.
a. May be used to adjust the degrees of freedom for the averaged tests of significance. Corrected tests are
displayed in the Tests of Within-Subjects Effects table.
E. Additional tables (12.13-12.17 and fig 12.2):

I have provided some additional tables to show the results, when two treatment
groups are significantly different. Here, I have compared the blood sugar levels of
Daonil with Placebo.
80
Table 12.13. Descriptive statistics
treatment groups Mean Std. Deviation N
Blood sugar at hour 0 placebo 110.4000 3.36155 5
Daonil 112.8000 2.16795 5
Total 111.6000 2.95146 10
Daonil 104.0000 4.18330 5
Total 106.3000 4.08384 10
Daonil 97.4000 3.43511 5
Total 103.0000 6.91215 10
Daonil 94.4000 2.70185 5
Total 101.9000 8.29257 10
Table 12.14. Multivariate test results of within-subjects effects
Multivariate Testsb
Effect Value F Hypothesis Error df Sig.
df
time Pillai's Trace .949 37.505a 3.000 6.000 .000
Wilks' Lambda 37.505a
.051 44.724 3.000 6.000 .000
Hotelling's Trace 18.752 37.505a 3.000 6.000 .000
Roy's Largest
18.752 37.505a 3.000 6.000 .000
Root
time * treatment Pillai's Trace .927 25.566a 3.000 6.000 .001
Wilks' Lambda .073 25.566a 3.000 6.000 .001
Hotelling's Trace 12.783 25.566a 3.000 6.000 .001
Roy's Largest
12.783 25.566a 3.000 6.000 .001
Root
a. Exact statistic
Table 12.15. Pairwise comparison between time adjacent blood sugar levels

Measure:bloodsugar
Source time Type III Sum of df Mean F Sig.
Squares Square
time Level 1 vs. Level 2 280.900 1 280.900 66.881 .000
Level 2 vs. Level 3 108.900 1 108.900 19.274 .002
Level 3 vs. Level 4 12.100 1 12.100 1.066 .332
time * treatment Level 1 vs. Level 2 122.500 1 122.500 29.167 .001
Level 2 vs. Level 3 108.900 1 108.900 19.274 .002
Level 3 vs. Level 4 36.100 1 36.100 3.181 .112
Error(time) Level 1 vs. Level 2 33.600 8 4.200
Level 2 vs. Level 3 45.200 8 5.650
Level 3 vs. Level 4 90.800 8 11.350
81
Table 12.16. Test results of between-subjects effects
Tests of Between-Subjects Effects

Measure:Bloodsugar
Transformed Variable:Average
Source Type III Sum of df Mean F Sig.
Squares Square
Intercept 111724.900 1 111724.900 15324.461 .000
treatment 126.025 1 126.025 17.286 .003
Error 58.325 8 7.291
Table 12.17. Descriptive statistics by treatment type
Estimates
Measure:Bloodsugar
treatment Mean Std. Error 95% Confidence Interval
groups Lower Bound Upper Bound
placebo 109.250 1.208 106.465 112.035
Daonil 102.150 1.208 99.365 104.935
Figure 12.2. Blood sugar levels by treatment group (Placebo & Daonil)
A. Basic tables (outputs under A)
The outputs of the analysis are shown in tables and graphs. Table 12.1 and 12.2
show the value labels of the blood sugar measurements and treatment groups,
respectively. Table 12.3 shows the descriptive statistics (mean, standard deviation
and no. of study subjects) of blood sugar levels at different times of measurement
by treatment groups.
82
B. Within-subjects effects (outputs under B)
Table 12.4 shows the multivariate test results of within-subjects effects. Since the
Sphericity assumption is frequently violated, we would consider the multivariate
test results (table 12.4) as discussed in section 11. First, look at the interaction term
(time*treatment) in the row of Wilks’ Lambda. The p-value (Sig.) is 0.275, which
is not significant. This means that there is no interaction between time and treat-
ment (i.e., blood sugar levels over time are not dependent on the treatment groups).
Now, look at the row of Wilks’ Lambda at time. The p-value is 0.000, which is
statistically significant. This means that the mean blood sugar levels measured at
different times are significantly different (i.e., there is a significant reduction of
blood sugar levels over time in both the treatment groups; fig 12.1). Table 12.5
shows the means and 95% confidence intervals of blood sugar levels at different
times of measurement. Table 12.6 shows the difference in blood sugar levels
between the adjacent measurements. The table shows that there is significant
difference in blood sugar levels between time 1 and 2, and time 2 and 3 (p=0.000),
but not between time 3 and 4 (p=0.081). Note that interaction for any comparison
is not significant.
C. Between-subjects effects (outputs under C)

Table 12.7 shows the results of between-subjects effects (between two treatment
groups – Daonil and Metformin) on blood sugar levels. The p-value for treatment
is 0.909, which is >0.05. This indicates that there is no significant difference of
mean blood sugar levels over time, between Daonil and Metformin groups (also
see fig 12.1); i.e., we are unable to reject the null hypothesis. We can see in table
12.8 that the overall mean (without considering the times of measurement) of
blood sugar levels for Daonil and Metformin are not that different (102.1 vs
101.9). This indicates that both the drugs are equally effective in reducing the
blood sugar levels (no one is better than the other).
The pairwise comparison (table 12.9) shows that the difference in mean blood
sugar levels between Daonil and Metformin (0.250) is very small to reject the null
hypothesis (p=0.909). Figure 12.1 shows the mean blood sugar levels at different
times by treatment groups. It shows that the blood sugar levels have been reduced
by both the drugs over time, but the difference in reduction between the groups is
not significant.
83
D. Test of assumptions (outputs under D)
Whether the assumptions are violated or not are checked by: a) Box’s M test (table
12.10); b) Levene’s test (table 12.11); and c) Mauchly’s test (table 12.12). If the
assumptions are met, the p-values of all these tests would be >0.05. We can see that
the p-values of all these tests are >0.05 except for Mauchly’s test (p=0.017). Note
that the Mauchly’s test tests the Sphericity assumption. As discussed earlier, to
interpret the results, it is recommended to use the multivariate test, which is not
dependent on Sphericity assumption.
E. Additional tables (outputs under E)

Additional tables (table 12.13-12.17) are provided to demonstrate the results,
when the treatment groups are different (one is better than the other). Here, I have
compared the effectiveness of Daonil compared to placebo. Table 12.13 shows the
descriptive statistics of blood sugar levels at different time intervals by treatment
group (placebo and Daonil). Though there is no significant difference in mean
blood sugar levels at hour 0 (baseline), but they are different over time.
The multivariate test results of within-subjects effects (table 12.14) show that
there is interaction (time*treatment) between time and treatment (p=0.001) [see
the row of Wilks’ Lambda under time*treatment]. The p-value of Wilks’ Lambda
under “time” is also significant (p=0.000). This means that there is a significant
reduction in mean blood sugar levels over time, and it depends on the treatment
group (since there is interaction between time and treatment). Look at fig 12.2. It
shows that the blood sugar levels have been reduced significantly over time,
among the subjects under the treatment of Daonil, but there is no significant
change in the placebo group. The test of between-subjects effects (table 12.16) also
conveys the information that the difference in blood sugar levels over time is not
same for Daonil and placebo groups (p=0.003). We, therefore, conclude that
Daonil is effective in reducing the blood sugar level and is superior to (better than)
placebo (p=0.003).
84
Section 13
Association between Two Categorical Variables:
Chi-Square Test of Independence
The Chi-square test is a commonly used statistical test for testing hypothesis in
health research. This test is suitable to determine the association between two cate-
gorical 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 case-control studies] or Relative Risk (RR)
[for cohort studies] with 95% Confidence Intervals (CI). Odds Ratio and RR are
the measures of strength of association between two variables. Use the data file
13.1 Chi-square test of Independence

Suppose, you have collected data on gender (sex) and diabetes from a group of
individuals selected randomly from a population. You are interested to know if
there is any association between gender and diabetes. In such a situation,
Chi-square test is the appropriate test for testing the hypothesis.
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 diabe-
tes are not independent).
Assumption:
1. Data have come from a random sample drawn from a selected population.
13.1.1 Commands:
Analyze > Descriptive statistics > Crosstabs > Select “sex” and push it into the
“Row(s)” box > Select “diabetes” for the “Column(s)” box > Statistics > Select
“Chi-square” and “Risk” > Continue > Cells > Select “Raw” and “Column”
under percentages > Continue > OK
85
Note: We have selected “risk” to get the OR and RR including their 95% CIs.
13.1.2 Outputs:
Table 13.1. Cross-tabulation between sex and diabetes mellitus
Sex * Diabetes mellitus Crosstabulation

Diabetes mellitus Total
Yes No
Sex Male Count 25 52 77
% within Sex 32.5% 67.5% 100.0%
% within Diabetes mellitus 55.6% 31.5% 36.7%
Female Count 20 113 133
% within Sex 15.0% 85.0% 100.0%
Total Count 45 165 210
% within Sex 21.4% 78.6% 100.0%
Table 13.2. Chi-square test result with p-value
Chi-Square Tests
Value df Asymp. Sig. Exact Sig. Exact Sig.
(2-sided) (2-sided) (1-sided)
Pearson Chi-Square 8.799a 1 .003
Continuity Correctionb 7.795 1 .005
Likelihood Ratio 8.537 1 .003
Fisher's Exact Test .005 .003
Linear-by-Linear 8.758 1 .003
N of Valid Casesb 210
a. 0 cells (.0%) have expected count less than 5. The minimum expected count is 16.50.
b. Computed only for a 2x2 table
Table 13.3. Odds Ratio (OR) and Relative Risk (RR) with 95% Confidence Interval (CI)
Risk Estimate
Value 95% Confidence Interval
Lower Upper
Odds Ratio for Sex (Male / Female) 2.716 1.385 5.327
For cohort Diabetes mellitus = Yes 2.159 1.288 3.620
For cohort Diabetes mellitus = No .795 .670 .943
N of Valid Cases 210
Table 13.1 is a 2 by 2 table of sex and diabetes with row (% within sex) and column
(% within diabetes mellitus) percentages. The question is which percentage should
86
you report? It depends on the situation and what do you want to report. For the data
of a cross-sectional study, it may provide better information to the readers if row
percentages are reported. For example, one can understand from the table 13.1 that
the prevalence of diabetes among males is 32.5% and that of the females is 15.0%,
when row percentages are used. However, column percentages can also be report-
ed in cross-sectional studies (most of the publications use column percentages). If
column percentage is used, the meaning would be different. In this example (table
13.1), it means that among those who have diabetes, 55.6% are male, compared to
31.5% who do not have diabetes. If data are from a case-control study, you must
report the column percentages (we 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 this case it would indicate the incidence of the disease among
males and females.
We can see in table 13.1 (in the row of total) that the overall prevalence (irre-
spective of sex) of diabetes is 21.4% (consider the data is from a cross-sectional
study). Table 13.1 also shows that 32.5% of the males have diabetes compared to
only 15.0% among females (i.e., the prevalence among males and females). The
Chi-square test actually tests the hypothesis whether the prevalence of diabetes
among males and females is same in the population or not.
Table 13.2 shows the Pearson Chi-square test results, including the degree of
freedom (df) and p-value (Asymp. Sig). The table also shows other test results,
such as Continuity Correction and Fisher’s Exact test. Before we look at the
Chi-square test results, it is important to check if there is any cell in the 2 by 2 table
with expected value <5. This information is given at the bottom of the table at “a”
as “0 cells (0%) have expected count less than 5”. For the use of the Chi-square
test, it is desirable to have no cell (in a 2 by 2 table) with expected count less than
5. If this is not fulfilled, we have to use the Fisher’s Exact test p-value to interpret
the result (see table 13.2). In fact, to use the Chi-square test, no more than 20%
cells should have expected frequency <5. You can have the expected frequencies
for all the cells if you select “Expected” under “Count” in “Cell” option during
analysis.
For the Chi-square test, consider the Pearson Chi-square value (see table 13.2).
In our example, Chi-square value is 8.799 and the p-value is 0.003 (table 13.2).
Since the p-value is <0.05, there is an association between sex and diabetes. It can,
therefore, be concluded that the prevalence of diabetes among males is significant-
87
ly higher than the females, which is statistically significant at 95% confidence
level (p=0.003).
Table 13.3 shows the OR (2.716) and it’s 95% CI (1.385-5.327). Use the OR if
the data are from a case-control study. Odds Ratio is also sometimes used for
cross-sectional data. The table also provided the RR (2.159) and it’s 95% CI
(1.288-3.620) [take the RR and 95% CI for diabetes= yes]. Use RR if the data are
from a cohort study. Note that both the OR and RR is statistically significant as
they do not include 1 in the 95% CI. Odds ratio 2.71 indicates that males are 2.7
times more likely to have diabetes compared to the females. On the other hand, RR
2.15 indicates that the risk of having diabetes is 2.1 times higher in males com-
pared to females. SPSS will not provide the OR and RR, if there are more than 2
categories in any of the variables (e.g., a 2 by 3 table). In such a situation, you have
to get the OR and RR in different ways.
Table 13.4. Decision for using Chi-square test

Situation Right test
Sample size >100 and expected cell Pearson’s Chi-square (uncor-
value >10 rected)
Sample size 31-100 and expected cell Pearson’s Chi-square with
count between 5-9 Yate’s correction (continuity
correction row in table 13.2)
Sample size less than 30 and/or Fisher’s Exact test
expected cell value <5
88
Section 14
Association between Two Continuous Variables:
Correlation
Nature and strength of relationship between two or more continuous variables can
be determined by regression and correlation analysis. Correlation is concerned
with measuring the strength of relationship between continuous variables. The
correlation model provides information on the relationship between two variables,
without distinguishing which is dependent and which is independent variable. But
the basic procedure for regression and correlation model is the same.
Under the correlation model, we calculate the “r” value. The “r” is called the
sample correlation coefficient. It (“r” value) indicates the degree of linear relation-
ship between dependent (Y) and independent (X) variable. Value of “r” ranges
from +1 to –1. In this section, I shall discuss the correlation model. Use the data
file <Data_3.sav> for practice.
14.1 Pearson correlation

Person correlation is done when the normality assumption is met (i.e., both the
dependent and independent variables are normally distributed; assumption 1).
Suppose, we want to explore if there is a correlation/association between systolic
BP (variable name is “sbp”) and diastolic BP (variable name is “dbp”).
Hypothesis:
H0: There is no correlation between systolic and diastolic BP.
HA: There is a correlation between systolic and diastolic BP.
Assumptions:
1. The variables (systolic and diastolic BP) are normally distributed in the
population;
2. The subjects represent a random sample from the population.
The first step, before doing correlation, is to generate a scatter diagram. The
scatter diagram provides information/ideas about:
• Whether there is any correlation between the variables;
89
• Whether the relationship (if there is any) is linear or non-linear; and
• 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).
14.1.1 Commands for scatter plot:

To get the scatter plot of systolic and diastolic BP, use the following commands:
Graphs > Legacy dialogues > Scatter… > Select “Simple” > Define > Select
“sbp” for “X-axis” and “dbp” for “Y-axis” > Select “ID_no” for “Level cases
by” (this would label the outliers, if there is any, by its id number) > Ok
This will give you the following scatter plot (fig 14.1).
Figure 14.1. Scatter plot of systolic and diastolic BP
If you want to get the regression line on the scatter plot, use the following
commands:
Graphs > Legacy dialogues > Interactive > Scatterplot… > Select “sbp” and
drag it into the “X-axis” box > select “dbp” and drag it into the “Y-axis” box >
Click on “Fit” tab > Select “Regression” after clicking the dropdown arrow >
Ok
The SPSS will produce the following scatter plot (fig 14.2) with the regression
line on it. In the same manner, you can produce the scatter diagram of age and
diastolic BP (fig 14.3).
90
Figure 14.2. Scatter diagram of systolic and diastolic BP with regression line
Figure 14.3. Scatter diagram of diastolic BP and age with regression line
14.1.2 Commands for Pearson correlation:

Analyze > Correlate > Bivariate > Select the variables “sbp” and “dbp” and
push them into the "Variables" box > Select “Pearson” under the “Correlation
coefficients” (usually set as default) > Ok
91
14.1.3 Outputs:
Table 14.1. Pearson correlation between systolic and diastolic BP
Levene's Test of Equality of Error Variancesa

SYSTOLIC BP DIASTOLIC BP
SYSTOLIC BP Pearson Correlation 1 .858(**)
Sig. (2-tailed) . .000
N 210 210
DIASTOLIC BP Pearson Correlation .858(**) 1
Sig. (2-tailed) .000 .
N 210 210
** Correlation is significant at the 0.01 level (2-tailed).
In the first step, I have constructed the scatter plot of systolic and diastolic BP (fig
14.1 and 14.2). Figure 14.1 shows that the data points are scattered around an
invisible straight line 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. Look at fig 14.2, which shows the regression line in the scatter
plot. The regression line has passed from near to the lower left corner to the upper
right corner, indicating a positive correlation between systolic and diastolic BP. If
the relationship were negative (inverse), the regression line would have passed
from the upper left corner to lower right corner. Figure 14.3 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 regression line, which is more or
less parallel to the X-axis.
For correlation, look at the value of correlation coefficient [r] (Pearson
Correlation). Table 14.1 shows that the correlation coefficient of systolic and
diastolic BP is 0.858 and the p-value is 0.000. Correlation coefficient “r” indicates
the strength/degree of linear relationship between the two variables (systolic and
diastolic BP). As the value of “r” is positive and the p-value is <0.05, 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
correlation, while values near to “+1” or “–1” indicate strong correlation. Negative
(– r) value indicates an inverse relationship. A value of r ≥ 0.8 indicates very strong
correlation; “r” value between 0.6 and 0.8 indicates moderately strong correlation;
“r” value between 0.3 and 0.5 indicates fair correlation and “r” value < 0.3 indi-
cates poor correlation.
92
14.2 Spearman’s correlation
Spearman’s correlation (instead of Pearson correlation) is done when the normali-
ty assumption is violated (i.e., if the distribution of either the dependent or inde-
pendent or both the variables are not normal). Spearman’s correlation is also appli-
cable for two categorical ordinal variables, such as intensity of pain (mild, moder-
ate, severe pain) and grade of cancer (stage 1, stage 2, stage 3, etc.).
Suppose, we want to explore if there is any correlation between systolic BP
(variable name is “sbp”) and income (income is not normally distributed).
14.2.1 Commands for Spearman’s correlation:

Analyze > Correlate > Bivariate > Select the variables “sbp” and “income” and
push them into the "Variables" box > Select “Spearman” under the “Correla-
tion coefficients” > Ok
14.2.2 Outputs:
Table 14.2. Correlation (Spearman’s) between systolic BP and income
Correlations
Systolic BP Monthly income
Spearman's rho Systolic BP Correlation
1.000 .007
Coefficient
Sig. (2-tailed) .919
N 210 210
Monthly income Correlation
.007 1.000
Coefficient
Sig. (2-tailed) .919
N 210 210
Table 14.2 shows the Spearman’s correlation coefficient between systolic BP and
income. The results indicate that there is no correlation between systolic BP and
income (r= 0.007; p=0.919), since the “r” value is very small and the p-value is
>0.05.
14.3 Partial correlation

The purpose of doing the partial correlation is to assess the correlation (indicated
93
by the “r” value) between two variables after adjusting for one or more other vari-
ables (continuous or categorical). This means that through partial correlation, we
get the adjusted “r” value after controlling for the confounding factors. For exam-
ple, if we assume that the relationship between systolic and diastolic BP may be
influenced (confounded) by other variables (such as, age, and diabetes), we should
do the partial correlation to exclude the influence of other variables (age, and
diabetes). The partial correlation will provide the correlation (r value) between
systolic and diastolic BP after controlling/ adjusting for age and diabetes.
14.3.1 Commands:
Analyze > Correlate > Partial > Select “sbp” and “dbp” for “Variables” box >
Select “age” and “diabetes” for “Controlling for” box > Ok
14.3.2 Outputs:
Table 14.3. Correlation between systolic and diastolic BP after controlling for age and diabetes mellitus
Correlations
Control Variables Systolic BP Monthly income
Age & Diabetes Systolic BP Correlation 1.000 .847
mellitus Significance
. .000
(2-tailed)
df 0 206
Diastolic BP Correlation .847 1.000
Significance
.000 .
(2-tailed)
df 206 0
Table 14.3 shows the results of partial correlation between systolic and diastolic
BP after adjusting for age and diabetes mellitus. We can see that r=0.847 and
p=0.000. This means that these two variables (systolic and diastolic BP) are
significantly correlated (p=0.000), even after controlling for age and diabetes
mellitus. If the relationship between systolic and diastolic BP was influenced by
age and diabetes mellitus, the crude (unadjusted) and adjusted “r” values would be
different. Look at table 14.1, which shows the crude “r” value (0.858). After
adjusting for age and diabetes, the “r” value becomes 0.847 (table 14.3). Since the
crude and adjusted “r” values are almost similar, there is no influence of age and
94
diabetes mellitus in the relationship between systolic and diastolic BP (i.e., age and
diabetes mellitus are not the confounding factors in the relationship between
systolic and diastolic BP).
95
Section 15
Linear Regression
Regression analysis is a commonly used statistical method for data analysis.

Nature and strength of relationship between two or more continuous variables can
be determined by regression and correlation analysis. I have already discussed
about correlation in the previous section. While correlation is concerned about
measuring the direction and strength of linear relationship between the variables,
regression analysis is helpful to predict or estimate the value of one variable corre-
sponding to a value of another variable(s) (e.g., to understand whether systolic BP
is a good predictor of diastolic BP). In regression analysis, our main interest is in
regression coefficient (also called slope or β), which indicates the strength of asso-
ciation between dependent (Y) and independent (X) variable. Regression can be
done as: a) Simple linear regression, and b) Multiple linear regression.
In this section, both simple and multiple linear regressions are discussed. Mul-
tiple linear regression is a type of multivariable analysis. Multivariable analysis is
a statistical tool where multiple independent variables are considered for a single
outcome. The terms “multivariate analysis” and “multivariable analysis” are often
used interchangeably in health research. Multivariate analysis actually refers to the
statistical methods for the analysis of multiple outcomes. Multivariable analyses
are widely used in observational studies, intervention studies (randomized and
nonrandomized trials), and studies of diagnosis and prognosis. The main purposes
of multivariable analysis are to:
a) Determine the relative contribution of independent variables to the outcome
variable;
b) Adjust for the confounding factors;
c) Predict the probability of an outcome, when several characteristics are pres-
ent in an individual; and
d) Assess interaction of multiple variables for the outcome.
There are several types of multivariable analysis methods. The choice of multi-
variable analysis, for the type of outcome variable, is summarized in table 7.3 (sec-
tion 7). The commonly used multivariable analysis methods in health research
include multiple linear regression, logistic regression and proportional hazards
96
regression (Cox regression) that are discussed in this manual. Use the data file
15.1 Simple linear regression

In simple linear regression, there is one dependent and one independent variable.
The objective of simple linear regression is to find the population regression equa-
tion, which describes the true relationship between the dependent variable (Y) and
independent variable (X). In simple linear regression model, two variables are
involved – one is independent variable (X), placed on X-axis, and the other is
dependent variable (Y), placed on Y-axis. Then, we call it “regression of Y on X”.
Suppose, we want to do 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 the 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;
4. Independence: Observations are independent of each other.
The first step in analyzing the data for regression is to construct a scatter
diagram, which has already been discussed in section 14. This would give an idea
about the linear relationship between the variables, systolic and diastolic BP.
15.1.1 Commands:
Analyze > Regression > Linear > Select “dbp” for “Dependent" box and “sbp”
for "Independent(s)” box > Method “Enter” (usually the default) > Statistics >
Select “Estimates, Descriptive, Confidence interval, and Model fit” > Contin-
ue > Ok
97
15.1.2 Outputs:
Table 15.1. Mean and standard deviation of the variables
DIASTOLIC BP 83.04 12.454 210
SYSTOLIC BP 127.83 20.021 210
Table 15.2. Correlation between systolic and diastolic BP
Correlations
DIASTOLIC BP SYSTOLIC BP
Pearson Correlation DIASTOLIC BP 1.000 .858
SYSTOLIC BP .858 1.000
Sig. (1-tailed) DIASTOLIC BP . .000
SYSTOLIC BP .000 .
N DIASTOLIC BP 210 210
SYSTOLIC BP 210 210
Table 15.3. Correlation coefficient (R) and coefficient of determination (R-square)
Model Summary
Model R R Square Adjusted R Std. Error of the
Square Estimate
1 .858(a) .737 .736 6.403
a Predictors: (Constant), SYSTOLIC BP
Table 15.4. ANOVA table for significance of “R”
ANOVA(b)
Model Sum of df Mean F Sig.
Squares Square
1 Regression 23890.586 1 23890.586 582.695 .000(a)
Residual 8528.028 208 41.000
Total 32418.614 209
a Predictors: (Constant), SYSTOLIC BP
b Dependent Variable: DIASTOLIC BP
Table 15.5. Constant (a) and regression coefficient (b)
Coefficients(a)
Model Unstandardized Standardized t Sig. 95% Confidence
Coefficients Coefficients Interval for B
B Std. Beta Lower Upper
Error Bound Bound
1 (Constant) 14.779 2.862 5.164 .000 9.136 20.422
SYSTOLIC BP .534 .022 .858 24.139 .000 .490 .578
a Dependent Variable: DIASTOLIC BP
98
Table 15.1 and 15.2 provides the descriptive statistics (mean and standard devia-
tion) and correlation coefficient (r-value) of the diastolic and systolic BP. The
model summary table (table 15.3) shows the Pearson’s correlation coefficient “R”
(r = 0.858) and coefficient of determination “R-square” (r2 = 0.737). The value of
“R” is same, as we have seen in section 14.
It is important to note the value of R-square (coefficient of determination)
given in the model summary table (table 15.3). R-square indicates the amount of
variation in “Y” due to “X”, that can be explained by the regression line. Here, the
R-square value is 0.737 (~0.74), which indicates that 74% variation in diastolic BP
can be explained by the systolic BP. The rest of the variation (36%) is due to other
factors (unexplained variation). The adjusted R-square value (0.736), as shown in
the table, is the value when the R-square is adjusted for better population estimate.
The ANOVA table (table 15.4) indicates whether the correlation coefficient (R)
is significant or not (i.e., whether the linear regression model is useful to explain
the dependent variable by the independent variable). As the p-value (Sig.) is 0.000,
R is statistically significant at 95% confidence level. We can, therefore, conclude
that there is a significant positive correlation (because R value is positive) between
the diastolic and systolic BP, and we can use the regression equation for prediction.
The table also shows the regression (also called explained) sum of squares
(23890.586) and residual (also called error) sum of squares (8528.028). The resid-
ual indicates the difference between the observed value and predicted value (i.e.,
value on the regression line). Residual sum of squares provides an idea about how
well the regression line actually fits into the data.
Table 15.5 (coefficients) provides quantification of the relationship between
the diastolic and systolic BP. The table shows the values for “a” or Y-intercept
(also called constant) and “b” (unstandardized coefficients) or slope (also called
regression coefficient, β). Note that for a single independent variable, standardized
coefficient (Beta) is equal to Pearson’s correlation value.
Here, the value of “a” is 14.779 and “b” is 0.534 (both are positive). The value,
a= +14.78, indicates that the regression line crosses/cuts the Y-axis above the
origin (zero) and at the point 14.78 (a negative value indicates that the regression
line crosses the Y-axis below the origin). This value (value for a) does not have any
practical meaning, since it indicates the average diastolic BP of individuals, if the
systolic BP is 0 (zero).
99
The value of “b” (the regression coefficient or slope) indicates the amount of
variation/change in “Y” (here it is diastolic BP) for each unit change in “X” (sys-
tolic BP). Here, the value of “b” is 0.534, which means that if the systolic BP
increases (or decreases) by 1 mmHg, the diastolic BP will increase (or decrease)
by 0.534 mmHg. The table also shows the significance (p-value) of “b”, which is
0.000. Note that for simple linear regression, if R is significant, “b” will also be
significant and will have the same sign (positive or negative).
We know that the simple linear regression equation is, Y = a + bX (“Y” is the
predicted value of the dependent variable; “a” is the Y-intercept or constant; “b” is
the regression coefficient and “X” is a value of the independent variable). There-
fore, the regression/prediction equation for this regression model is
Y = 14.737 + 0.534X.
With this equation, we can estimate the diastolic BP by the systolic BP. For
example, what would be the estimated diastolic BP of an individual whose systolic
BP is 130 mmHg? The answer is, the estimated diastolic BP would be equal to
(14.727 + 0.534*130) 84.1 mmHg.
Note that, if we want to use the regression equation for the purpose of predic-
tion/estimation, “b” has to be statistically significant (p<0.05). In our example, the
p-value for “b” is 0.000, and we can, therefore, use the equation for the prediction
of diastolic BP by systolic BP.
Table 15.5 has actually evaluated whether “b” in the population is zero or not
by the t-test (Null hypothesis: “b” is equal to “zero” in the population; Alternative
hypothesis: the population regression coefficient is not equal to “zero”). We can
reject the null hypothesis, since the p-value is <0.05. It can, therefore, be conclud-
ed that the systolic BP can be used to predict/estimate the diastolic BP using the
regression equation, Y = 14.737 + 0.534X.
15.2 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 explanatory
or predictor variable. In multiple regression, there are more than one explanatory
(independent) variables in the model. The explanatory variables may be quantita-
tive or categorical. The main purposes of multiple regression analysis are to:
100
• Get the adjusted estimates of regression coefficients (B) of the explanatory
variables in the model;
• Predict or estimate the value of the dependent variable by the explanatory
variables in the model; and
• Understand the amount of variation in the dependent variable explained by
the explanatory variables in the model.
Suppose, we want to assess the contribution of four variables (age, systolic BP,
sex and religion) in explaining the diastolic BP in a sample of individuals selected
randomly from a population. Here, the dependent variable is the diastolic BP and
the explanatory variables (independent variables) are age, systolic BP, sex and
religion. Of the explanatory (independent) variables, two are quantitative (age and
systolic BP) and two are categorical variables (sex and religion). Of the categorical
variables, sex has two levels (male and female) and religion has 3 levels (Islam,
Hindu and Christian). When the independent variable is categorical with more
than two levels (e.g., religion), we need to create dummy variables for that vari-
able. For example, if we want to include the variable “religion” in the regression
model, we shall have to create dummy variables for religion.
15.2.1 Creating dummy variables:

In our example, the variable “religion” has 3 levels and are coded as 1= Islam; 2=
Hindu and 3= Christian. We cannot simply put “religion” as one of the explanatory
variables in the regression model, because the coding is arbitrary and the regres-
sion estimates obtained for religion would be meaningless. We need to create
dummy variables for religion.
The number of dummy variables to be created for “religion” is two (no. of
levels minus 1). Before creating the dummy variables, we have to decide about the
comparison group. Let us consider “Christian” as the comparison group, and
assign “0” (zero) as its code number. We shall create two dummy variables – one
is “reli_1” and the other is “reli_2” for religion. To create the dummy variables, we
shall have to recode the variable “religion” using the following commands.
Step 1: Create the first dummy variable “reli_1” for religion
Transform > Recode into different variables > Select “religion” and push it into
the “Input variable – Output variable” box > Write “reli_1” in the “Output vari-
101
able name” box > write “Dummy variable 1 for religion” in the “label” box >
Change > Click on “Old and New Values..” > Select “Value” under “Old
value” and write 1 in the box > Select “Value” under “New value” and write 1
in the box > Add > Select “All other values” under “Old value” > Write 0
(zero) in the box “Value” under the “New value” > Add > Continue > Ok
Step 2: Create the second dummy variable “reli_2” for religion
Transform > Recode into different variables > Select “religion” and push it into
the “Input variable – Output variable” box > Write “reli_2” in the “Output vari-
able name” box > Write “Dummy variable 2 for religion” in the “Label” box >
Change > Click on “Old and New Values..” > Select “Value” under “Old
value” and write 2 in the box > Select “Value” under “New value” and write 1
in the box > Add > Select “All other values” under “Old value” > Write 0 in the
box “Value” under the “New value” > Add > Continue > Ok
The above commands will create two dummy variables for religion, the
“reli_1” (for which code 1= Islam and 0= other religions, i.e., Hindu and Chris-
tian)” and “reli_2” (for which code 1= Hindu and 0= other religions, i.e., Islam and
Christian)”. You can see the new variables in the variable view of the data file.
Don’t forget to provide the value labels for the dummy variables.
15.2.2 Changing string variable into numeric variable:

If we want to include the variable “sex” in the model, we need to check its coding.
If the variable is coded as string variable (e.g., m= male and f= female, as is done
in our data), we need to recode it as a numeric variable, say 0= female and 1= male.
In this case, when multiple regression will be performed, the regression estimate in
the model will be for males compared to females (lower value will be the compari-
son group). Use the following commands to create a numeric variable for sex.
Transform > Recode into different variables > Select “sex” and push it into the
“Input variable – Output variable” box > Write “sex_1” in the “Output variable
name” box > Write “Sex numeric” in the “Label” box > Click on “Change” >
Click on “Old and New Values..” > Select “Value” under “Old value” and write
f in the box > Select “Value” under “New value” and type 0 in the box > Add
> Select “Value” under “Old value” and write m in the box > Write 1 in the box
“Value” under the “New value” > Add > Continue > Ok
102
This would create a new variable “sex_1” (last variable in the variable view)
with codes 0= female and 1= male. Go to the “variable view” of the data file and
set these code numbers in the column “Value” of the variable sex_1.
15.2.3 Sample size for multiple regression:

Multiple regression should be done if the sample size is fairly large. What would
be the minimum sample size, 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 vari-
ables in the model. Other authors provided a formula (n= 50 + 8m) to estimate the
number of subjects required for the model. For example, if we intend to include 5
independent variables in the model, we need to have at least 90 subjects (50 +
8*5). For stepwise regression, there should have 40 cases for each of the indepen-
dent variables in the model.
15.2.4 Commands for multiple linear regression:

Use the following commands for multiple regression analysis, where the depen-
dent variable is dbp (diastolic BP) and the explanatory (independent) variables are
age, sbp (systolic BP), sex_1, reli_1 and reli_2.
Analyze > Regression > Linear > Select “dbp” for “Dependent" box and “age,
sbp, sex_1, reli_1 and reli_2” for "Independent(s)” box > Method “Enter”
(usually the default) > Statistics > Select “Estimates, Confidence interval, and
Model fit” > Continue > Ok
15.2.5 Outputs:
Table 15.6. Multiple R, R-square and adjusted R-square values
Model Summaryb
Model R R Square Adjusted R Std. Error of the
Square Estimate
1 .852a .725 .719 6.233
a. Predictors: (Constant), Reli_2, Systolic BP, age, Sex, Reli_1
b. Dependent Variable: Diastolic BP
103
Table 15.7. ANOVA table for significance of R
ANOVAb
Model Sum of df Mean F Sig.
Squares Square
1 Regression 20925.182 5 4185.036 107.710 .000a
Residual 7926.385 204 38.855
Total 28851.567 209
a. Predictors: (Constant), Reli_2, Systolic BP, age, Sex, Reli _1
Table 15.8. Adjusted regression coefficients of explanatory variables and their significance
Coefficientsa
Error Bound Bound
1 (Constant) 20.894 3.420 6.110 .000 14.151 27.637
age .004 .058 .003 .070 .944 -.110 .119
Systolic BP .490 .022 .836 22.559 .000 .447 .532
Sex -2.179 .916 -.090 -2.380 .018 -3.985 -.374
Islam .162 1.369 .007 .119 .906 -2.537 2.862
Hindu -.271 1.492 -.010 -.182 .856 -3.212 2.670
a. Dependent Variable: Diastolic BP
Table 15.6 (model summary) shows the values for R (0.852), R-square (0.725) and
adjusted R-square (0.719) [adjusted for better population estimation]. In multiple
regression, the R measures the correlation between the observed value of the
dependent variable and the predicted value based on the regression model. The
R-square may overestimate the population value, if the sample size is small. The
adjusted R-square gives the R-square value of better population estimation. The
R-square value 0.725 indicates that all the independent variables (age, systolic BP,
sex and religion) together in the model explains 72.5% variation in diastolic BP,
which is statistically significant (p=0.000), as shown in the ANOVA table (table
15.7).
The Coefficients table (table 15.8) shows regression coefficients (unstandard-
ized and standardized), p-values (Sig.) and 95% confidence intervals (CI) for
regression coefficients of all the explanatory variables in the model along with the
constant. This is the most important table for interpretation of results. The unstan-
dardized regression coefficients (B) are shown in the table for age (0.004;
p=0.944), systolic BP (0.490; p<0.001), sex (-2.179; p=0.018 for males compared
104
to females), Islam (0.162; p=0.906 compared to Christian) and Hindu (-0.271;
p=0.856 compared to Christian).
From this output (table 15.8), we conclude that the systolic BP and sex are the
factors significantly 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 the diastolic BP. The unstandardized coefficient (B) [also
called multiple regression coefficient] for systolic BP, in this example, is 0.490
(95% CI: 0.45 to 0.53). This means 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 all other variables (age, sex and religion) in the model. On the
other hand, the unstandardized coefficient (B) for sex is -2.179 (95% CI: -3.985 to
-0.374), which means that (on an average) the diastolic BP of males is 2.2 mmHg
less (since the coefficient is negative. If it is positive, it would be more) than that
of the females, given the other variables constant.
The standardized coefficients (Beta) (table 15.8) indicate which independent
variables have more influence on the dependent variable (diastolic BP). Bigger the
value more is the influence. We can see in table 15.8 that the standardized coeffi-
cients for systolic BP and sex are 0.83 and -0.09, respectively. This means that
systolic BP has greater influence in explaining the diastolic BP than the sex.
15.2.7 Regression equation:

The regression equation to estimate the average value of the dependent variable
with the explanatory variables is as follows:
Y= a + B1X1 + B2X2 + B3X3 + B4X4 …….. BnXn
Here, “Y” is the estimated mean value of the dependent variable; “a” is the
constant (or Y-intercept); “B” is the regression coefficient(s) and X is the value of
the variable(s) in the model.
Suppose, we want to estimate the average diastolic BP of an individual who is
40 years old, male, Muslim and has systolic BP 120 mmHg. In table 15.8 (coeffi-
cients), we see that the regression coefficients are for age (=0.004 [B1]), systolic
BP (=0.49 [B2]), sex (= -2.179 [B3] for being male) and Islam (=0.162 [B4] for
being Muslim) and the constant is 20.894. Therefore, the estimated diastolic BP of
an individual would be:
Y= 20.894 + 0.004*40 + 0.49*120 + (-2.179*1) + 0.162*1 = 77.84.
105
15.2.8 Problem of multicollinearity:
Before deciding about the multiple regression model, we need to check for multi-
collinearity (inter-correlations among the independent variables) of the indepen-
dent variables. If there are moderate to high inter-correlations among the indepen-
dent variables, two situations may occur. Firstly, the importance of a given explan-
atory variable may be difficult to determine because of biased (distorted) p-value;
and the other is dubious relationships may be obtained. For example, if there is
multicollinearity, we may observe that the regression coefficient for sex is not
significant and the systolic BP has a negative relationship with the diastolic BP.
Another important sign of multicollinearity is a severe reduction of the Adjusted R
Square value.
To determine the correlations among the independent variables, we can gener-
ate the Pearson’s correlation matrix. For example, we want to see the correlations
among the systolic BP, age, sex and religion. Use the following commands to get
the correlation matrix.
Analyze > Correlate > Bivariate... > Select the variables “sbp, age, sex_1,
reli_1 and reli_2” for the “Variables” box > Ok
The SPSS will produce the following correlation matrix table (table 15.9).
Table 15.9. Correlation matrix of independent variables
Correlations
Systolic BP Sex Reli_1 Reli_2 age
Systolic BP Pearson Correlation 1 -.125 .026 -.011 -.042
Sig. (2-tailed) .071 .710 .870 .542
N 210 210 210 210 210
Sex Pearson Correlation -.125 1 .077 .038 -.066
Sig. (2-tailed) .071 .269 .581 .344
N 210 210 210 210 210
Reli_1 Pearson Correlation .026 .077 1 -.757* .073
Sig. (2-tailed) .710 .269 .000 .292
N 210 210 210 210 210
Reli_2 Pearson Correlation -.011 .038 -.757* 1 -.020
Sig. (2-tailed) .870 .581 .000 .776
N 210 210 210 210 210
Age Pearson Correlation -.042 -.066 .073 -.020 1
Sig. (2-tailed) .542 .344 .292 .776
N 210 210 210 210 210
**. Correlation is significant at the 0.01 level (2-tailed).
106
We can see in the table (15.9) that there is a moderately strong correlation
between reli_1 and reli_2 (r = -0.757), while the correlation coefficients for other
variables are low. However, the correlation between reli_1 and reli_2 did not affect
our regression analysis.
Pearson’s correlation can only check collinearity between any two variables.
Sometimes a variable may be multicollinear with a combination of variables. In
such a situation, it is better to use the tolerance measure, which gives the strength
of the linear relationships among the independent variables (usually the dummy
variables have higher correlation). To get the tolerance measure (another measure
for multicollinearity), use the following commands:
sbp, sex_1, reli_1 and reli_2” for "Independent(s)” box > Method “Enter” >
Statistics > Select “Estimates, Confidence interval, Model fit, and Collineari-
ty diagnostics” > Continue > Ok
This would provide the collinearity statistics in the coefficients table as shown
in table 15.10.
Table 15.10. Collinearity statistics for multicollinearity diagnosis
Coefficientsa
Model Unstandardized Standardized t Sig. 95% Confidence Collinearity
Coefficients Coefficients Interval for B Statistics
B Std. Beta Lower Upper Tolerance VIF
Error Bound Bound
1 (Constant) 20.894 3.420 6.110 .000 14.151 27.637
age .004 .058 .003 .070 .944 -.110 .119 .983 1.018
Systolic BP .490 .022 .836 22.559 .000 .447 .532 .981 1.019
Sex -2.179 .916 -.090 -2.380 .018 -3.985 -.374 .950 1.053
Islam .162 1.369 .007 .119 .906 -2.537 2.862 .411 2.432
Hindu -.271 1.492 -.010 -.182 .856 -3.212 2.670 .416 2.404
The tolerance value ranges from 0 to 1. A value close to “zero” indicates that
the variable is almost in a linear combination (i.e., has strong correlation) with
other independent variables. In our example (table 15.10), the tolerance values for
age, systolic BP, and sex are more than 0.95. However, the tolerance values of
Islam (reli_1) and Hindu (reli_2) [the dummy variables] are a little more than 0.4.
The recommended tolerance level is more than 0.6 before we put the variable in
the multiple regression model. However, a tolerance of 0.4 and above is accept-
107
able, especially if it is a dummy variable. The other statistics provided in the last
column of the table are the VIF (Variance Inflation Factor). This is the inverse of
the tolerance value.
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., taking
their sum).
Finally, for developing a model for multiple regression, we should first check
for multicollinearity and then the residual assumptions (see below). If they fulfil
the requirements, then only we can finalize the regression model.
15.2.9 Checking for assumptions:

For practical purpose, there are three assumptions that need to be checked on the
residuals for the linear regression model to be valid. The assumptions are:
a. There is no outlier;
b. The data points are independent;
c. The residuals are normally distributed with mean = 0 and have constant vari-
ance.
15.2.9.1 Checking for outliers and independent data points (assumptions a

and b):
Use the following commands to check for outliers (casewise diagnostics) and data
points are independent (Durbin-Watson statistics):
Statistics > Select “Estimates, Confidence interval, Model fit, Casewise diag-
nostics and Durbin-Watson” > Continue > Ok
The SPSS will produce the Model Summary table (table 15.11), casewise diag-
nostics table (table 15.12, if there are outliers, otherwise not) and residuals statis-
tics table (table 15.13).
108
Table 15.11. Durbin-Watson statistics for checking data points are independent
Model Summaryb
Model R R Square Adjusted R Std. Error of the Durbin-Watson
Square Estimate
1 .852a .725 .719 6.233 1.701
a. Predictors: (Constant), Religion: dummy var_2, Systolic BP, age, Sex:
numeric code, Religion: dummy var_1
Table 15.12. Case number and the outlier
Casewise Diagnosticsa
Case Number Std. Residual Diastolic BP Predicted Value Residual
204 -3.625 62 85.11 -23.115
Table 15.13. Residuals statistics with outliers in the data set
Residuals Statisticsa
Minimum Maximum Mean Std. Deviation N
Predicted Value 65.46 115.85 82.77 9.917 210
Residual -23.115 18.678 .000 6.301 210
Std. Predicted Value -1.745 3.336 .000 1.000 210
Std. Residual -3.625 2.929 .000 .988 210
Table 15.14. Residuals statistics without any outliers in the data set
Residuals Statisticsa
Minimum Maximum Mean Std. Deviation N
Predicted Value 65.32 116.16 82.77 10.006 210
Residual -15.446 18.452 .000 6.158 210
Std. Predicted Value -1.743 3.338 .000 1.000 210
Std. Residual -2.478 2.960 .000 .988 210
Look at the residuals statistics table (table 15.13). Our interest is in the Std.
Residual value. The “minimum” and “maximum” values should not exceed “+3”
or “–3”. Table 15.13 shows that the minimum value is “–3.625” (exceeded –3).
This means that there are outliers. Now, look at the casewise diagnostics table
(table 15.12). The table shows that there is an outlier in the diastolic BP, the value
of which is 62 and the case number (ID number) is 204 (if there is no outlier in the
data, this table will not be provided). If no outlier is present in the data, we shall
get a Residuals Statistics table like table 15.14.
109
The Durbin-Watson test is done to check whether data points are independent.
The Model Summary table (table 15.11) shows the Durbin-Watson statistics
results in the last column. The Durbin-Watson estimate ranges from 0 to 4. Values
around 2 indicate that the data points are independent. Values near zero indicate a
strong positive correlation and values near 4 indicate a strong negative correlation.
The table shows that the value of the Durbin-Watson statistics, in our example, is
1.701, which is close to 2 (i.e., data points are independent)
15.2.9.2 Checking for normality assumption of the residuals and constant

variance:
Commands:
Statistics > Select “Estimates, Confidence interval, Model fit” > Continue >
Plots > Select “Histogram and normal probability plot” under “Standard-
ized residual plots” > Place “*ZRESID on Y; and *ZPRED on X” under
“Scatter 1 of 1” > Continue > Ok
The SPSS will produce the histogram (fig 15.1), normal probability plot (fig
15.2) and a scatter plot (fig 15.3) for the residuals. The distribution of the residuals
is normal as seen in the histogram (fig 15.1) and P-P plot (fig 15.2). The constant
variance (homoscedasticity) is checked in the scatter plot (fig 15.3). If the scatter
of the points shows no clear pattern (as seen in fig 15.3), we can conclude that the
variance is constant.
Figure 15.1. Histogram
110
Figure 15.2. Normal probability plot
Figure 15.3. Scatter plot of standardized residual vs standardized predicted value
15.2.10 Variable selection for the model:

In general, independent variables to be selected for multivariable analysis should
include the risk factors of interest and potential confounders (based on theory,
prior research findings and empirical findings), while variables with lots of miss-
ing values should be excluded.
We have used the “Enter” method for modelling earlier in this section. The
“Enter” method uses all the independent variables in the model included by the
researcher. It does not exclude any variable automatically from the model. Auto-
matic procedures can be used to determine which independent variables will be
111
included in the model. The major reason for using the automatic selection proce-
dure is to minimize the number of independent variables necessary to estimate or
predict the outcome. SPSS and other data analysis software have the option to
automatically select the independent variables in the model. They use statistical
criteria to select the variables and their order in the model. The commonly used
variable selection techniques are provided in table 15.15.
Table 15.15. Methods of variable selection
Technique Method Advantages and disad-
vantages
Forward This method enters variables in the Best suited for dealing
model sequentially. The order is deter- with the studies where
mined by the variable’s association the sample size is small.
(significance) with the outcome (vari- Does not deal well with
ables with strongest association are suppressor (confounding)
entered first) after adjustment for the effects.
other variables already in the model.
Backward This technique removes variables from Better for assessing
the model sequentially. The order is suppressor effect than the
determined by the variable’s association forward selection
with the outcome (variables with weakest method.
association leave first) after adjustment
for the variables already in the model.
Stepwise/ This is the combination of forward and Has the ability to manage
Remove backward methods. In the stepwise large number of potential
method, variables that are entered are predictor variables,
checked at each step for removal. Like- fine-tuning the model to
wise, in the removal method, variables choose the best predictor
that are excluded will be checked for variables from the
re-entry. available options.
Enter (all Enters all the variables at the same time Including all variables
variables) (does not remove any variable automati- may be problematic, if
cally from the model). there are many indepen-
dent variables and the
sample size is small
Let us see how to use the “Stepwise” method (commonly used method in mul-
tiple regression analysis) for modelling. To do this, use the following commands
(only change is in “Method”):
112
sbp, sex_1, reli_1 and reli_2” for "Independent(s)” box > Method “Stepwise”
(fig 15.4) > Statistics > Select “Estimates, Confidence interval, and Model fit”
> Continue > Ok
15.2.10.1 Outputs (only the relevant table is provided):
Table 15.16. Regression models
Coefficientsa
Error Bound Bound
1 (Constant) 19.407 2.793 6.948 .000 13.900 24.913
Systolic BP .496 .022 .847 22.961 .000 .453 .539
2 (Constant) 21.016 2.838 7.405 .000 15.421 26.611
Systolic BP .490 .022 .836 22.768 .000 .447 .532
Sex: numeric -2.180 .893 -.090 -2.441 .015 -3.941 -.419
15.2.10.2 Interpretation:
During analysis, we included 5 independent variables (age, systolic BP, sex, and
two dummy variables of religion) in the model. The SPSS has provided table 15.16
that shows the adjusted regression coefficients and models. Let us compare the
outputs in table 15.16 with those of table 15.8, where we have used the “enter”
method. In table 15.8, we can notice that the SPSS retained all the independent
variables in the model that were included, and only the “systolic BP” and “sex”
were found to be significantly associated with the dependent variable (diastolic
BP). When we used the “stepwise” method, the SPSS has provided two models –
model 1 and model 2. In the model 1, there is only one independent variable (sys-
tolic BP) and in the model 2, there are two independent variables (systolic BP and
sex; others are automatically removed). We consider the last model as the final
model.
Sometimes you may need to include certain variable(s) in the model for theo-
retical or practical reason. In such a situation, after you derive the model with
“stepwise” method, add the additional variable(s) of your choice and re-run the
model using the “enter” method.
For automatic selection method, you can specify the inclusion (entry) and
113
exclusion (removal) criteria of the variables. Usually, the inclusion and exclusion
criteria, set as default in SPSS, are 0.05 and 0.10, respectively (fig 15.5). You can,
however, change the criteria based on your requirements. Finally, for model build-
ing, the researcher should decide the variables to be included in the final model
based on theoretical understanding and empirical findings.
Figure 15.4. Model selection options
Figure 15.5. Selection criteria for entry and removal of variables
114
Section 16
Logistic Regression
Logistic regression is a commonly used statistical method for health data analysis.
Logistic regression is done when the outcome variable is a dichotomous variable,
such as diabetes (present or absent), vaccinated (yes or no) and an outcome (died
or did not die). The purposes of logistic regression analysis (and other multivari-
able analysis) are to: a) Adjust the estimate of risk for a number of confounding
factors set in the model; b) Determine the relative contribution of factors to a
single outcome; c) Predict the probability of an outcome for a number of indepen-
dent variables in the model; and d) Assess interaction of multiple variables for the
outcome. Use the data file <Data_4.sav> for practice.
16.1 Logistic regression analysis

Logistic regression is appropriate to adjust for multiple confounding factors or
model (identify) the predictors of a dichotomous categorical outcome variable
(e.g., disease present or absent). For logistic regression analysis in SPSS, the
dichotomous outcome variable should be coded as “0= disease absent” and “1=
disease present”. SPSS will consider the higher value to be the predicted outcome
and the lower value as the comparison group. Suppose, we want to predict (or
identify the factors associated with) diabetes with sex (variable name: sex_1), age,
peptic ulcer (variable name: pepticulcer) and family history of diabetes (variable
name: f_history). To perform the logistic regression analysis, recode diabetes as
“0= diabetes absent” and “1= diabetes present” (if it is not coded like this). Simi-
larly, it is better to recode the categorical predictor (independent) variables as “0
for no (comparison group)” and “1 for yes”.
Assumptions:
Logistic regression does not make any assumptions concerning the distribution of
predictor (independent) variables. However, it is sensitive to high correlation
among the independent variables (multicollinearity). The outliers may also affect
the results of logistic regression.
115
16.1.1 Commands:
Analyze > Regression > Binary logistic > Put “diabetes” in the “Dependent”
box > Put “sex_1”, age, pepticulcer, and f_history in the “Covariate” box >
Categorical > Push “sex, pepticulcer and f_history” in “Categorical covari-
ates” box > Select “f_history” > Select “first” under “Change contrast” (we are
doing this because 0 is our comparison group) > Click on “Change” > (do the
same thing for all the variables in “Categorical covariates” box) > Continue >
Options > Select “Classification plots, Hosmer-Lemeshow goodness-of-fit,
Casewise listing of residuals, Correlations of estimates, and CI for exp(B)” >
Continue > OK
16.1.2 Outputs:
SPSS provides many tables while doing the logistic regression analysis. Only the
useful tables are discussed here. After the basic tables (tables 16.1 to 16.3), the
outputs of logistic regression are provided under Block 0 and Block 1.
A. Basic tables:
Table 16.1. Case Processing Summary
Unweighted Casesa N Percent

Selected Cases Included in Analysis 210 100.0
Missing Cases 0 .0
Total 210 100.0
Unselected Cases 0 .0
Total 210 100.0
a. If weight is in effect, see classification table for the total number of cases.
Table 16.2. Dependent Variable Encoding
Original Value Internal Value

No 0
Yes 1
116
Table 16.3. Categorical Variables Codings
Frequency Parameter coding

(1)
Family history of DM No 114 .000
Yes 96 1.000
Peptic ulcer Yes 59 1.000
No 151 .000
Sex: numeric code Female 133 .000
Male 77 1.000
16.1.3 Interpretation: Basic tables

Table 16.1 (case processing summary) shows that the analysis includes all the 210
subjects and there is no missing value. If there is any missing data, this table will
show it. Note that the subjects are excluded from analysis, if there are any missing
data.
Table 16.2 (dependent variable encoding) tells us which category of the depen-
dent variable (diabetes) is the predicted outcome. The higher value is the predicted
outcome. Here, the higher value is 1 (have diabetes) and is the predicted outcome
for the dependent variable.
Table 16.3 (categorical variables codings) indicates the comparison groups of
the independent (explanatory) variables. Here, the lower value is the comparison
group. For example, for family history of diabetes, “No” is coded as 0 (.000),
while “Yes” is coded as 1 (1.000). This means that persons who do not have the
family history of diabetes is the comparison group. Similarly, not having peptic
ulcer (.000) and being female (.000) is the comparison group for peptic ulcer and
sex, respectively.
B. Outputs under Block 0:
Table 16.4. Classification Tablea,b
Observed Predicted
DIABETES MELLITUS Percentage
No Yes Correct
Step 0 DIABETES MELLITUS No 165 0 100.0
Yes 45 0 .0
Overall Percentage 78.6
a. Constant is included in the model.

b. The cut value is .500
117
16.1.4 Interpretation: Outputs under Block 0
Analysis of data without any independent variable in the model is provided under
Block 0. The results indicate the baseline information that can be compared with
the results when independent variables are put into the model (provided under
Block 1).
Look at the classification table (table 16.4). The table indicates the overall
percentage of correctly classified cases (78.6%). We will see whether this value
has increased with the introduction of independent variables in the model under
Block 1 (given in table 16.8). If the value remains the same, it means that the inde-
pendent variables in the model do not have any influence/contribution to predict
diabetes (dependent variable). In our example, the overall percentage has
increased after inclusion of the independent variables in the model (90.5%; table
16.8 under Block 1) compared to the value under Block 0 (78.6%). This means that
adding independent variables improved the ability of the model to predict the
dependent variable.
C. Outputs under Block 1:
Table 16.5. Omnibus Tests of Model Coefficients
Chi-square df Sig.
Step 1 Step 101.589 4 .000
Block 101.589 4 .000
Model 101.589 4 .000
Table 16.6. Model Summary
Step 1 -2 Log likelihood Cox & Snell R Nagelkerke R

Square Square
1 116.635a .384 .593
a. Estimation terminated at iteration number 6 because parameter estimates changed
by less than .001.
Table 16.7: Hosmer and Lemeshow Test
Step Chi-square df Sig.

1 14.663 8 .066
118
Table 16.8. Classification Tablea
Observed Predicted
Diabetes mellitus Percentage
No Yes Correct
Step 1 Diabetes mellitus No 159 6 96.4
Yes 14 31 68.9
Overall Percentage 90.5
a. The cut value is .500
Table 16.9. Variables in the Equation
B S.E. Wald df Sig. Exp(B) 95.0% C.I.for EXP(B)

Lower Upper
Step 1a age .230 .039 34.660 1 .000 1.259 1.166 1.359
sex(1) 1.587 .528 9.031 1 .003 4.891 1.737 13.774
f_history 1.119 .517 4.688 1 .030 3.062 1.112 8.430
pepticulcer 1.782 .487 13.363 1 .000 5.942 2.285 15.448
Constant -10.684 1.526 49.006 1 .000 .000
a. Variable(s) entered on step 1: age, sex, f_history, pepticulcer.
Table 16.10. Correlation Matrix
Constant age sex(1) f_history pepticulcer

Step 1 Constant 1.000 -.932 -.385 -.302 -.322
age -.932 1.000 .161 .054 .156
sex(1) -.385 .161 1.000 .388 .155
f_history -.302 .054 .388 1.000 .099
pepticulcer -.322 .156 .155 .099 1.000
Table 16.11. Casewise Listb
Case Selected Observed Predicted Predicted Temporary Variable

Statusa Diabetes mellitus Group Resid ZResid
11 S Y** .006 N .994 13.213
25 S Y** .026 N .974 6.177
38 S Y** .052 N .948 4.248
41 S Y** .062 N .938 3.883
62 S Y** .099 N .901 3.009
124 S Y** .043 N .957 4.692
137 S N** .890 Y -.890 -2.839
a. S = Selected, U = Unselected cases, and ** = Misclassified cases.
b. Cases with studentized residuals greater than 2.000 are listed.
119
16.1.5 Interpretation: Outputs under Block 1
Omnibus tests of Model Coefficients (table 16.5): This table indicates whether
the overall performance of the model is better if independent variables are includ-
ed in the model compared to the model without any independent variables (given
under Block 0). We want this test to be significant (p-value < 0.05). In this exam-
ple, the p-value of the Omnibus test is 0.000, which indicates that the proposed
model is better than the model without any predictor (independent) variables.
Model summary table (table 16.6): This table indicates usefulness of the model.
The Cox & Snell R-square and Nagelkerke R-square (called pseudo R-square)
values provide indication about the amount of variation in the outcome variable
that can be explained by the independent variables in the model. In this example,
the values of the pseudo R-square are 0.384 (Cox & Snell R-square) and 0.593
(Nagelkerke R-square), respectively. This means that between 38.4% and 59.3%
variation in the outcome variable can be explained by the independent variables set
in the model. This information is not needed if the objective of the analysis is to
adjust for the Odds Ratio.
Hosmer-lemeshow goodness-of-fit test (table 16.7): When the intention of anal-

ysis is prediction, i.e., to identify the predictors to predict the outcome, then the
question is “How good is the model for prediction”? This is judged based on the
Hosmer-Lemeshow goodness-of-fit test, and positive and negative predictive
values, given in the classification table (table 16.8).
The Hosmer-Lemeshow test indicates how well the observed and predicted
values fit with each other (i.e., 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 (non-significant). If the p-value is not significant, it means that the model
is a good fit for prediction (i.e., the observed and predicted values are close togeth-
er). In this example, the p-value is 0.066, indicating that the model is useful for
prediction of the outcome variable. If the test is significant (p<0.05), then the mod-
el is not good to predict the outcome variable by the independent variables in the
model. Note that this information is not needed if the objective of doing logistic
regression is to adjust for the confounding factors.
Classification table (table 16.8): This table indicates how well the model is able
120
to predict the correct category in each case (have or do not have the disease). This
table shows that the overall accuracy of this model to predict diabetes (with a
predicted probability of 0.5 or greater) is 90.5%. This table also shows the Sensi-
tivity and Specificity of the model as 68.9% (31 ÷ 45) and 96.4% (159 ÷ 165),
respectively. Positive and negative predictive values can also be calculated from
the table, which are 83.8% (31÷37) and 91.9% (159÷173), respectively. Interpreta-
tion of the findings of this table is a little bit complicated and needs further expla-
nation, especially to explain sensitivity, specificity, and positive and negative
predictive values.
However, the information that we need to check is the overall percentage.
Compare this value with the value under the Block 0 outputs. We expect this value
(overall percentage) to be increased, otherwise adding independent variables in the
model does not have any impact on prediction. We can see that the overall percent-
age of the model to correctly classify cases is 90.5% under Block 1 (table 16.8).
This value, compared to the value (78.6%; table 16.4) that we have seen under
Block 0, has improved. This means that adding independent variables in the model
improved the ability of the model to predict the dependent variable. This informa-
tion is needed, if the intention of this analysis is prediction. If the objective is
adjustment for confounding factors, we can ignore this information.
Variables in the equation (table 16.9): This is the most important table to look at.
This table shows the results of logistic regression analysis. This table indicates
how much each of the independent variables contributes to predict/explain the
outcome variable. This table also indicates the adjusted Odds Ratio (OR) and its
95% confidence interval (CI). The B values (column 3) indicate the logistic regres-
sion coefficients for the variables in the model. These values are used to calculate
the probability of an individual to have the outcome. The positive values indicate
the likelihood for the outcome, while the negative values indicate the less likeli-
hood for the outcome. The exponential of B [Exp(B)] is the adjusted OR.
Let us see how to interpret the results. There are 4 independent (explanatory)
variables in the model – age (as a continuous variable), sex, family history of
diabetes and peptic ulcer. The table shows the adjusted OR [Exp(B)], 95% CI for
the adjusted OR and p-value (Sig.). The adjusted OR for sex is 4.891 (95% CI:
1.737-13.774), which is statistically significant (p=0.003). Here, our comparison
group is female (see table 16.3). This indicates that males are 4.9 times more likely
121
to have diabetes compared to females after adjusting (or controlling) for age, fami-
ly history of diabetes and peptic ulcer. Similarly, persons who have the family
history of diabetes are 3.1 times more likely (OR: 3.06; 95% CI: 1.11-8.43;
p=0.03) to have diabetes compared to those who do not have the family history
after adjusting for age, sex, and peptic ulcer. Interpretation of Exp(B) for age is a
little bit different since the variable was entered as a continuous variable. Here,
Exp(B) for age is 1.259. This means that the odds of having diabetes would
increase by 25.9% [Exp(B) – 1; i.e., 1.259 – 1] (95% CI: 16.6-35.9) with each year
increase of age, which is statistically significant (p<0.001).
If we want to know which variable contributed most in the model, then look at
the Wald statistics. Higher the value (of Wald), the greater is the importance. Age
is the most important variable contributed to the model since it has the highest
Wald value (34.6).
Checking for multicollinearity (table 16.10): It is important to check for multi-

collinearity of the independent variables in the model. If there is multicollinearity,
the model becomes dubious. Multicollinearity is checked in the correlation matrix
table (table 16.10). This table shows the correlations between the independent
variables (correlation coefficients or r values). If there is multicollinearity, r values
would be higher (greater than 0.5). If we look at the correlation matrix table
(table16.10), none of the values are greater than 0.5 except for the correlation
between age and constant, which is -0.932.
Now, look at table 16.9 (variables in the equation). If multicollinearity is pres-
ent (and affects the model), the magnitude of the SEs (standard errors) would be
high or low (greater than 5.0 or less than 0.001). Existence of multicollinearity
means that the model is not statistically stable. To solve the problem (in general),
look at the SE and omit the variable(s) with large (or small) SE, until the magni-
tude of the SEs hover between 0.001 and 5.0.
If there is a high correlation between the constant and any of the predictor vari-
ables, you can omit the constant from the model. In our example, there is a high
correlation (-.932) between age and constant. However, it did not affect the results
as none of the SEs are >5.0 or <0.001. Therefore, we do not need to do anything.
If it affects the results, just omit (deselect) the constant from the model (deselect
“include constant in model” located at the bottom of the fig 16.1) from “Options
template” during analysis as shown in fig 16.1.
122
Figure 16.1. Option template for logistic regression
Casewise list (table 16.11): This table provides information about the cases for
which the model does not fit well. Look at the ZResid values (last column). The
values above 2.5 are the outliers and do not fit well in the model. The case numbers
are shown in the column 1. If present (cases that do not fit the model well), all
these cases need to be examined closely. Under the “Predicted Group” column,
you may see “Y (means yes)” or “N (means no)”. If it is “Y”, the model predicts
that the case (case no. 137, in our example) should have diabetes, but in reality (in
the data) the subject does not have diabetes (see the observed column where it is
“N”). Similarly, is if it is “N” under the “predicted group”, the model predicts that
the case should not have diabetes, but in reality the subject has diabetes.
16.1.6 ROC curve:

We can construct the ROC (Receiver Operating Characteristic) curve to assess the
model discrimination. Area under the ROC curve ranges from 0 to 1. A value of 0.5
indicates that the model is useless. Our data shows that the ROC value (area under
the curve) is 0.914 (95% CI: 0.86-0.96; p=0.000) (table 16.12). A high value of
123
ROC indicates good model for prediction. To generate the ROC curve, use the
following commands:
Analyze > Regression > Binary logistic > Put “diabetes” in the “dependent”
Categorical > Push “sex, pepticulcer and f_history” in “Categorical covari-
ates” box > Select “f-history” > Select “First” under “Change contrast” (we are
doing this because 0 is our comparison group) > Click on “Change” > (do the
same thing for all the variables in “Categorical covariates” box) > Continue >
Options > Select “Classification plots, Hosmer-Lemeshow goodness-of-fit,
Casewise listing of residuals, Correlations of estimates, and CI for exp(B)” >
Continue > Save > Select “Probabilities” under “Predicted values” > Continue
> OK
This will create a new variable, Pre_1 (predicted probability) (look at the
bottom of the variable view). Now, to get the ROC curve, use the following com-
mands:
Analyze > ROC curve > Select “Pre_1” for the “Test variable” box and “diabe-
tes” for the “State variable” box and put “1” for the value of the state variable
(since code 1 indicates individuals with diabetes) > Select “ROC curve” and
“Standard error and Confidence interval” under “Display” > OK
Table 16.12. Area Under the Curve
Area Std. Errora Asymptotic Asymptotic 95% Confidence Interval

Sig.b Lower Bound Upper Bound
.914 .027 .000 .861 .967
The test result variable(s): Predicted probability has at least one tie between the positive
actual state group and the negative actual state group. Statistics may be biased.
a. Under the nonparametric assumption
b. Null hypothesis: true area = 0.5
124
Figure 16.2. ROC Curve
16.1.7 Sample size for logistic regression:

Sample size is always a concern for analysis of data. The sample size needed for
logistic regression depends on the effect size you are trying to demonstrate and the
variability of the data. It is always better to calculate the sample size during the
design phase of the study. However, a rule-of-thumb for planning a logistic regres-
sion is that for every independent variable in the model you need to have at least
10 outcomes (some authors recommend minimum 15-25 cases for each indepen-
dent variable).
16.1.8 Variable selection for a model:

I have discussed various methods of variable selection for a model in the previous
section in detail (section 15). Like other multivariable analysis, independent vari-
ables to be selected for logistic regression should include the risk factors of interest
and potential confounders, while avoiding variables with lots of missing values.
Earlier in this section, I have used the “Enter” method for logistic regression
analysis. The “Enter” method uses all the independent variables in the model
included by the researcher. We can also use the automatic selection method for
analysis. For logistic regression, the commonly used method for automatic selec-
tion of variables is the “Backward LR” method. However, if there is multicol-
linearity, you can select the “Forward LR” method for data analysis.
125
Commands for automatic selection of independent variables (use the data file
<Data_3.sav>):
Analyze > Regression > Binary logistic > Put “diabetes” in the “Dependent”
Select “Backward LR” from “Method” (fig 16.3) > Categorical > Push “sex,
pepticulcer and f_history” in the “Categorical covariates” box > Select “f-his-
tory” > Select “First” under “Change contrast” (we are doing this because 0 is
our comparison group) > Click on “Change” > (do the same thing for all the
variables in “Categorical covariates” box) > Continue > Options > Select
“Classification plots, Hosmer-Lemeshow goodness-of-fit, Casewise listing of
residuals, Correlations of estimates, and CI for exp(B)” > Continue > OK
Figure 16.3
SPSS will give the following table (table 16.13) along with others (not shown
here as they are not relevant). We can see that analysis is done in 4 steps (Step 1 to
4; the first column of the table). In the first step, all the independent variables are
in the model. Gradually, SPSS has removed variables that are not significantly
associated with the outcome. Finally, SPSS has provided the final model (Step 4)
with a single variable (sex) in it, which is significantly associated with the
outcome. If the "Enter" method was used, SPSS would give us only the step 1 (see
table 16.9).
126
The inclusion (entry) and exclusion (removal) criteria, set as default in SPSS,
are 0.05 and 0.10, respectively. As discussed in section 15, you can change the
“Entry” and “Removal” criteria from the “Option” dialogue box (under the “Prob-
ability for Stepwise”). Finally, for model building, you should decide the variables
to be included in the final model based on theoretical understanding and empirical
findings.
B S.E. Wald df Sig. Exp(B) 95.0% C.I.for EXP(B)

Lower Upper
Step 1a age -.037 .023 2.502 1 .114 .964 .921 1.009
sex_1(1) -1.209 .388 9.707 1 .002 .299 .140 .639
f_history(1) -.425 .397 1.143 1 .285 .654 .300 1.425
pepticulcer(1) -.118 .430 .075 1 .784 .889 .383 2.064
Constant 2.997 .737 16.529 1 .000 20.028
Step 2a age -.037 .023 2.660 1 .103 .963 .921 1.008
sex_1(1) -1.212 .388 9.772 1 .002 .298 .139 .636
f_history(1) -.428 .397 1.163 1 .281 .652 .299 1.419
Constant 2.998 .737 16.530 1 .000 20.049
Step 3a age -.036 .023 2.521 1 .112 .964 .922 1.009
sex_1(1) -1.041 .347 8.980 1 .003 .353 .179 .698
Constant 2.729 .686 15.830 1 .000 15.316
Step 4a sex_1(1) -.999 .344 8.457 1 .004 .368 .188 .722
Constant 1.732 .243 50.954 1 .000 5.650
a. Variable(s) entered on step 1: age, sex_1, f_history, pepticulcer.
127
Section 17
Survival Analysis
In many situations, researchers are interested to know the progress of a patient

(with a disease) from a specific point in time (e.g., from the point of diagnosis or
from the point of initiation of treatment) until the occurrence of a certain outcome,
such as the death or recurrence of any event (such as, recurrence of cancer). Prog-
nosis is usually assessed by: a) Estimating the median survival time, and b) Cumu-
lative probability of survival after a certain time interval (e.g., 5-year, 3-year, etc.).
For example, the researchers may be interested to know what is the median surviv-
al time of colonic cancer if the patient is not treated (or treated), and what is the
estimated probability that a patient may survive for more than 5 years (5-year
cumulative survival probability) if the patient is treated (or not treated). The meth-
ods employed to answer these questions in a follow-up study are known as surviv-
al analysis (or life table analysis) methods.
Survival analysis is done in the follow-up studies. To do the survival analysis,
we need to have data (information) from each of the patients, at least on:
• Time: Length of time the patient was observed in the study (called survival
time);
• Outcome: Whether the patient developed the outcome of interest (event)
during the study period, or 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)?
The survival time is of two types – a) Censored time; and b) Event time. The
censored time is the amount of time contributed by:
a. The 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,
withdraw, etc.; or
c. Patients who developed outcome (e.g., died due to accident) due to other
reasons than the disease of interest.
128
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.
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
comparison. Such a comparison allows us to answer the question “which treatment
delays the time of occurrence of the event”. The method commonly used to
analyze the survival-time data is the Kaplan-Meier method, and SPSS can be used
to analyze such data. Use the data file <Data_survival_4.sav> for practice.
17.1 Survival analysis: Kaplan-Meier method

Suppose, a researcher has conducted a follow-up study (clinical trial) on patients
with heart failure to compare the effectiveness of a new drug (n=22) compared to
placebo (n=22). The outcome of interest in this study is death (event). The objec-
tive is to assess whether the “new treatment” delays the time to death (event) com-
pared to placebo among the patients with heart failure. Following variables are
included in the data file.
• Time: It is the amount of time each patient has spent in the study in days;
• Treatment: Which treatment did the patient receive (0= placebo; 1= new
treatment);
• Outcome (event): Whether the patient developed the event, i.e., died or not
(0= censored; 1= died)
Assumptions:
• The probability of the outcome is similar among the censored and under-ob-
servation individuals;
• There is no secular trend over the calendar period;
• The risk is uniform during the interval;
• Losses are uniform over the interval.
17.1.1 Commands:
Analyze > Survival > Kaplan Meier > Push the variable “time” to “Time” box
> Push the variable “outcome” in the “Status” box > Click “Define event” >
Select “Single value” and type “1” (here 1 is the event) in the box > Continue
> Push “treatment” in the “Factor” box > Click Options… > Select “Survival
129
table(s), Mean and Median survival” under statistics > Select “Survival” under
“Plots” > Continue > Click “Compare Factor…” > Select “Log rank” under
“Test statistics” > Continue > OK
17.1.2 Outputs:
The SPSS will give the following outputs.
Treatment Total N N of Censored

group Events N Percent
Placebo 22 16 6 27.3%
New treatment 22 11 11 50.0%
Overall 44 27 17 38.6%
Table 17.2. Means and Medians for Survival Time
Treatment Meana Median

group Estimate Std. 95% Confidence Estimate Std. 95% Confidence
Error Interval Error Interval
Lower Upper Lower Upper
Bound Bound Bound Bound
Placebo 72.545 14.839 43.462 101.629 40.000 12.899 14.719 65.281
New treatment 125.264 13.402 98.996 151.532 146.000 28.786 89.580 202.420
Overall 98.925 10.812 77.733 120.117 89.000 21.232 47.385 130.615
a. Estimation is limited to the largest survival time if it is censored.
130
Table 17.3. Survival Table
Treatment group Time Status Cumulative Proportion N of N of

Surviving at the Time Cumulative Remaining
Estimate Std. Error Events Cases
Placebo 1 2.000 Died .955 .044 1 21
2 3.000 Died .909 .061 2 20
3 4.000 Died .864 .073 3 19
4 7.000 Died .818 .082 4 18
5 10.000 Died .773 .089 5 17
6 22.000 Died .727 .095 6 16
7 28.000 Died .682 .099 7 15
8 29.000 Died .636 .103 8 14
9 32.000 Died .591 .105 9 13
10 37.000 Died .545 .106 10 12
11 40.000 Died .500 .107 11 11
12 41.000 Died .455 .106 12 10
13 54.000 Died .409 .105 13 9
14 61.000 Died .364 .103 14 8
15 63.000 Died .318 .099 15 7
16 71.000 Died .273 .095 16 6
17 127.000 Censored . . 16 5
18 140.000 Censored . . 16 4
19 146.000 Censored . . 16 3
20 158.000 Censored . . 16 2
21 167.000 Censored . . 16 1
22 182.000 Censored . . 16 0
New 1 2.000 Died .955 .044 1 21
treatment 2 6.000 Died .909 .061 2 20
3 12.000 Died .864 .073 3 19
4 54.000 Died .818 .082 4 18
5 56.000 Censored . . 4 17
6 68.000 Died .770 .090 5 16
7 89.000 Died .722 .097 6 15
8 96.000 Died . . 7 14
9 96.000 Died .626 .105 8 13
10 125.000 Censored . . 8 12
11 128.000 Censored . . 8 11
12 131.000 Censored . . 8 10
13 140.000 Censored . . 8 9
14 141.000 Censored . . 8 8
15 143.000 Died .547 .117 9 7
16 145.000 Censored . . 9 6
17 146.000 Died .456 .129 10 5
18 148.000 Censored . . 10 4
19 162.000 Censored . . 10 3
20 168.000 Died .304 .151 11 2
21 173.000 Censored . . 11 1
22 181.000 Censored . . 11 0
131
Table 17.4. Overall Comparisons
Chi-square df Sig.
Log Rank (Mantel-Cox) 4.660 1 .031
Test of equality of survival distributions for the different levels of Treatment group.
Table 17.5. Overall Comparisons
Chi-square df Sig.
Log Rank (Mantel-Cox) 4.660 1 .031
Breslow (Generalized Wilcoxon) 6.543 1 .011
Tarone-Ware 6.066 1 .014
Test of equality of survival distributions for the different levels of Treatment group.
Table 17.1 is the summary table indicating the number of study subjects in each
group (22 in the placebo and 22 in the new treatment group) and the number of
events (no. died) occurred in each group including the number censored. The table
shows that in the treatment group, 11 patients died and 11 were censored, while in
the placebo group, 16 died and 6 censored.
Table 17.2 shows the mean and median survival times for both the placebo and
new treatment groups. We do not consider the mean survival time for reporting.
We consider the median survival time. The median survival time is the time when
the cumulative survival probability is 50%. The table indicates that the median
survival time, if the patient is in the placebo group, is 40 days (95% CI:
14.71-65.28), while it is 146 days (95% CI: 89.5-202.42), if the 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 longer time to event (and placebo is
associated with shorter time to event). Thus, we conclude that the person lives
longer if s/he receives the new treatment compared to the placebo.
Table 17.3 shows the survival probability (Cumulative Proportion Surviving at
the Time) at different points of time in the placebo and treatment group. From the
table, we can see that the cumulative survival probability at the end of 71 days, in
the placebo group, is 0.273 (27.3%). Since there is no death after that, the cumula-
tive survival probability at the end of 182 days will be the same (27.3%).
On the other hand, the cumulative survival probability is 0.304 (30.4%) at the
end of 168 days, if the patient is in the new treatment group. As there is no death
after that, the cumulative survival probability at the end of 181 days will be the
132
same (30.4%). In the new treatment group, the cumulative survival probability at
the end of 71 days is about 0.722 (72.2%), which is much higher than in the place-
bo group (27.3%). This indicates that the probability of survival at the end of 71
days is higher among the patients who received new treatment compared to place-
bo. This may indicate the benefit of the new treatment (i.e., the new treatment is
better than the placebo).
However, if we consider the cumulative survival probability of patients in both
these groups at the end of 180 days, the outcome is not that different – 27.3% in
the placebo group and 30.4% in the treatment group. This information indicates
that the survival probability is still higher if the person is on the new treatment than
on the placebo.
We can also estimate the median survival time (it is the time when the cumula-
tive survival probability is 50%) in both these groups from this table. The median
survival time for placebo group is 40 days and that of the treatment group is 146
days (see table 17.3). Now, the question is whether the survival experiences of
both these groups in the population are different or not? To answer this question,
we have to use a statistical test (Log Rank test) as given in table 17.4.
Table 17.4 shows the Log Rank test results. For an objective comparison of the
survival experience of two groups, it is desirable to use some statistical methods
that would tell us whether the difference of the survival experiences in the popula-
tion is statistically significant or not. Here, the null hypothesis is “there is no
difference in the survival experience of these two groups (new treatment and
placebo) in the population”. Such a null hypothesis is tested by Log Rank test. The
Log Rank test results show that the p-value is 0.031, which is <0.05. This means
that, the survival experience of both these groups in the population is not same. In
other words, it indicates that the survival probability is better if the patient receives
the new treatment (i.e., the new treatment is more effective/better than the placebo
in improving the patients’ survival).
Note that there are alternative procedures for testing the null hypothesis that
the two survival curves are identical. They are Breslow test, Tarone-Ware test and
Peto test (table 17.5). The Log Rank test ranks all the deaths equally, while the
other tests give more weight to early deaths. The options are available in SPSS
under the “Compare Factor” tab.
Survival curve: The cumulative survival probability is usually portrayed visually

133
by a graph called survival curve (fig 17.1). The “steps” in the graph represent the
time when events (deaths or any other event of interest) occurred. The graph
allows us to represent visually the median survival time and the cumulative surviv-
al probability for any specific time period (e.g., 30-day; 6-month; 1-year, 3-year,
5-year, 10-year cumulative survival probability, etc.). In general, the line above,
indicates better survival probability. We can see that the line for the new treatment
is above the line for the placebo.
Figure 17.1. Survival Functions
134
Section 18
Cox Regression
The Cox Regression is also called Proportional Hazards Analysis. In the previous
section (section 17), I have discussed the survival analysis using the Kaplan Meier
method. Like other regression methods (e.g., multiple linear regression and logis-
tic regression), Cox Regression is a multivariable analysis technique where the
dependent measure is a mixture of time-to-event and censored time observations.
Use the data file <Data_survival_4.sav> for practice.
18.1 Cox Regression or Proportional Hazards Regression

Returning to our previous example (section 17), we analyzed data to assess the
effectiveness of a new treatment compared to the placebo. Our objective was to
determine whether the new treatment delays the time to death compared to the
placebo among patients with heart failure. We found that the new treatment signifi-
cantly delayed the time to death compared to placebo, as indicated by the median
survival time and Log Rank test. However, the effectiveness of the new treatment
might be influenced (confounded) by other factors, such as age, hypertension,
diabetes or other characteristics. All these variables, therefore, need to be
controlled during analysis for assessing 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 the new treatment.
Cox Regression gives us the Hazard Ratio, which is analogous to Relative
Risk (RR). Hazard Ratio (also called Relative Hazard) is the ratio of hazards if the
person is exposed compared to the person not exposed. In Cox Regression, the
dependent variable is the Log of hazard.
18.1.1 Commands:
Let us use the previous example and data for Cox Regression analysis along with
the variables sex and age for adjustment. Note that the variable “treatment” has
two categories – placebo (coded as “0”) and new treatment (coded as “1”).
Analyze > Survival > Cox Regression > Push “time” into the “Time” box >
135
Push “outcome” into the “Status” box > Click on “Define event” > In “Single
value” box write 1 (since 1 is the code no. of the event) > Continue > Push
“treatment, age and sex” into the “Covariate” box > Click on “Categorical” >
Push “treatment and sex” into the “Categorical Covariates” box > Select
“Last” from “Reference category” (usually the default) under “Change
contrast” > Continue > Click on “Options” > Select “CI for Exp(B) and
Correlation of estimates” > Continue > Click on “Plot” > Select “Survival and
Log minus Log” > Select the variable “treatment” from the “Covariate Values
Plotted at” and push into the “Separate Line for” box > Continue > Ok
18.1.2 Outputs:
Only relevant tables are provided below.
N Percent
Cases available in Eventa 27 61.4%
analysis Censored 17 38.6%
Total 44 100.0%
Cases dropped Cases with missing values 0 .0%
Cases with negative time 0 .0%
Censored cases before the
earliest event in a stratum 0 .0%
Total 0 .0%
Total 44 100.0%
a. Dependent Variable: Survival time in days
Table 18.2. Categorical Variable Codingsc,d
Frequency (1)b
Treatment 0=Placebo 22 1
1=New treatment 22 0
sexa 0=Male 21 1
1=Female 23 0
a. Indicator Parameter Coding
b. The (0,1) variable has been recoded, so its coefficients will not be the same as for
indicator (0,1) coding.
c. Category variable: treatment (Treatment gr)
d. Category variable: sex (Sex)
136
B SE Wald df Sig. Exp(B) 95.0% C.I.for EXP(B)

Lower Upper
treatment 1.004 .458 4.805 1 .028 2.728 1.112 6.693
age .009 .023 .150 1 .698 1.009 .965 1.055
sex .889 .436 4.169 1 .041 2.434 1.036 5.716
Figure 18.1. Survival Function of heart failure patients
Figure 18.2. LML (Log minus Log) Function
Table 18.1 shows the number of cases that are analyzed. Table 18.2 is very import-
ant for interpretation. This table indicates which category of the categorical vari-
137
ables is the comparison group. Look at the last column [(1)b]. The value “0” in this
column indicates the comparison group. In the table “New treatment” is indicated
as “0” in the last column. Therefore, in the analysis, “new treatment” is the com-
parison group (though “new treatment” is actually coded as “1”). Similarly,
“females” are the comparison group in this analysis since the value of being
“female” is “0” in the last column. We shall, therefore, get the Hazard Ratio for the
“placebo” group compared to the “new treatment” group and for the “males” com-
pared to the “females”, as shown in table 18.3 (Variables in the Equation).
Our main interest is in table 18.3 (Variables in the Equation). The table indi-
cates the Hazard Ratio [Exp(B)], p-value (Sig.) and 95% confidence interval (CI)
for the Hazard Ratio [95% CI for Exp(B)]. The Hazard Ratio for the variable
“treatment” is 2.72 (95% CI: 1.11-6.69) and the p-value is 0.028. This indicates
that compared to the “new treatment”, patients in the “placebo” group are 2.72
times more likely to have shorter time to event after controlling for “age” and
“sex”, which is statistically significant (p=0.028). On the other hand, males are
more likely (2.43 times) to have shorter time to event compared to the females
after controlling for the variables “treatment” and “age” (p=0.041). Age, inde-
pendently, does not have any significant effect on the survival time, since the
p-value is 0.698.
Figure 18.1 shows the survival plot of the heart failure patients by treatment
group. The upper line is for the new treatment group and the lower one is for the
placebo group. The figure shows the outcome difference between the new treat-
ment and placebo. The group represented by the upper line has the better survival
probability.
However, before we conclude the results, we have to check if: a) there is multi-
collinearity among the independent variables; and b) relative hazards over the time
are proportional (also called the proportionality assumption of the proportional
hazards analysis). Look at the SE of the variables in the model (table 18.3). There
is no value which is very small (<0.001) or very large (>5.0) (refer to the logistic
regression analysis in section 16), indicating that there is no problem of multicol-
linearity in the model.
For the second assumption, we need to check the log-minus-log survival plot
(fig 18.2). If there is a constant vertical difference between the two curves (i.e.,
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
138
points in time and much further apart at other points in time, then the assumption
is violated. In our example, two lines are more or less parallel indicating that the
assumption is not violated. When the proportional hazard assumption is violated,
it is recommended to use the Cox regression with time dependent covariate to
analyze the data.
139
Section 19
Non-parametric Methods
Non-parametric tests, in general, are done when the quantitative dependent vari-
able is not normally distributed. Non-parametric tests are also used when the data
are measured in nominal and ordinal scales. Table 19.1 shows the types of
non-parametric methods recommended against the parametric tests, when the
dependent variable is not normally distributed in the population. Note that
non-parametric tests are less sensitive compared to the parametric tests and may,
therefore, fail to detect differences between groups that actually exist. Use the data
file <Data_3.sav> for practice.
Table 19.1. Types of non-parametric techniques against the alternative para-

metric methods
Non-parametric 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 for goodness-of-fit None
Chi-square test for independence None
(discussed earlier)
Spearmen’s correlation (discussed Pearson correlation
earlier)
19.1 Mann-Whitney U test

This test is the alternative test for Independent Samples t-test, when the dependent
variable is not normally distributed. This test compares the differences between
two groups on a continuous measure (variable). This test is based on ranks of
observations and is better than the median test. This test, tests the null hypothesis
that the two populations have equal medians. For example, we may want to know
whether the median systolic BP (where the distribution of systolic BP is non-nor-
mal) of males and females is same.
140
19.1.1 Commands:
Analyze > Nonparametric tests > 2 Independent samples > Select “sbp” and
push into the “Test Variable List” box > Select “sex_1” and push into the
“Grouping Variable” box > Click on “Define Groups” > Write 0 in “Group1”
box and 1 in “Group 2” box (note: our code nos. are 0 for female and 1 for
male) > Continue > Select “Mann-Whitney” under “Test Type” > Ok
19.1.2 Outputs:
Table 19.2. Mann-Whitney: Ranks
Sex: numeric N Mean Rank Sum of Ranks

Systolic BP Female 133 109.90 14616.50
Male 77 97.90 7538.50
Total 210
Table 19.3. Test Statisticsa
Systolic BP
Mann-Whitney U 4535.500
Wilcoxon W 7538.500
Z -1.379
Asymp. Sig. (2-tailed) .168
a. Grouping Variable: Sex: numeric
Our interest is in table 19.3. Just look at the p-value of the test. Here, the p-value
is 0.168, which is >0.05. This indicates that the distribution of systolic BP among
males and females is not different (or median of systolic BP of males and females
is not different). However, with this test result, the median systolic BP of females
and males should be reported. To get the medians, use the following commands.
Analyze > Compare means > Means > Select “sbp” and push into the “Depen-
dent List” box > Select “sex_1” and push it into the “Independent List” box >
Remove “Mean, Number of cases and Standard deviation” from the “Cell
Statistics” box > Select “Median” from “Statistics” box and push it into the
“Cell Statistics” box > Continue > OK
You will get the following tables (19.4 & 19.5). Tables 19.5 shows the median
systolic BP of males and females.
141
Cases
Included Excluded Total
N Percent N Percent N Percent
Systolic BP * Sex: numeric 210 100.0% 0 .0% 210 100.0%
Table 19.5. Report
Median
Sex: numeric Systolic BP
Female 124.00
Male 122.00
Total 123.00
19.2 Wilcoxon Signed Ranks test

This test is the non-parametric alternative of the paired samples t-test. This test
compares the distribution of two related samples (e.g., pre-test and post-test).
Wilcoxon test converts the scores into ranks and then compares. Suppose, to evalu-
ate the impact of a training, you have taken the pre- and post-tests, before and after
the training. You want to assess if there is any change in the post-test score com-
pared to the pre-test score due to the training.
19.2.1 Commands:
Analyze > Nonparametric tests > 2 Related Samples > Select “post-test and
pre-test” together and push into the “Test Pairs” box > Options > Select “De-
scriptive” and “Quartile” > Continue > Select “Wilcoxon” under “Test Type”
> Ok
19.2.2 Outputs:
N Mean Std. Minimum Maximum Percentiles

Deviation 25th 50th 75th
(Median)
Post test score 32 90.9844 8.44096 62.00 100.00 87.0000 92.5000 98.1250
Pre test score 32 53.5781 15.42835 23.50 87.50 41.8750 52.0000 64.5000
142
Table 19.7. Test Statisticsb
Pre test score - Post test score

Z -4.938a
Asymp. Sig. (2-tailed) .000
a. Based on positive ranks.
b. Wilcoxon Signed Ranks Test
Table 19.6 shows the descriptive statistics of pre- and post-test scores. The median
(50th percentile) score of pre-test is 52.0, while the median score is 92.5 for the
post-test. The difference between these scores is quite big. Look at table 19.7. The
p-value of the Wilcoxon Signed Ranks test is 0.000, which is highly significant.
This indicates that the pre- and post-test scores (medians) are significantly differ-
ent. We, therefore, conclude that the training has significantly improved the
knowledge of the participants (since the median of the post-test score is signifi-
cantly higher than that of the pre-test score).
19.3 Kruskal-Wallis test

It is the non-parametric 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,
we want to test the hypothesis whether the systolic BP is different among religious
groups (Muslim, Hindu and Christian) [the null hypothesis is systolic BP is not
different across the religious groups].
19.3.1 Commands:
Analyze > Nonparametric Tests > K Independent Samples > Select “sbp” and
push into “Test Variable List” box > Select “religion” and push it into “Group-
ing Variable” box > Click “Define Range” > Write 1 in “Minimum” box and
3 in “Maximum” box (the religion has code numbers from 1 to 3) > Continue
> Options > Select “Quartile” > Continue > Select “Kruskal Wallis H” > Ok
143
19.3.2 Outputs:
Table 19.8. Ranks
Religion N Mean Rank

Systolic BP MUSLIM 126 105.54
HINDU 58 106.47
Christian 26 103.13
Total 210
Table 19.9. Test Statisticsa,b
Systolic BP
Chi-Square .054
df 2
Asymp. Sig. .973
a. Kruskal Wallis Test
b. Grouping Variable: religion
Table 19.9 shows the Kruskal Wallis test results (dependent variable is systolic BP
and grouping variable is religion with 3 levels – Muslim, Hindu and Christian as
shown in table 19.8). The p-value (Asymp. Sig.) of the Chi-square test is 0.973,
which is >0.05. Therefore, we are unable to reject the null hypothesis. We
conclude that the median of the systolic BP among religious groups is not signifi-
cantly different. You can get the median of the systolic BP using the commands as
mentioned under Mann-Whitney U test. The medians of systolic BP in different
religious groups are provided in table 19.10.
Table 19.10. Report
Median
Religion Systolic BP
MUSLIM 122.00
HINDU 126.00
Christian 121.50
Total 123.00
19.4 Friedman test

The Friedman test is the non-parametric alternative of the one-way repeated mea-
sures ANOVA test. Suppose, we are interested to evaluate the changes in blood
sugar levels (if they are different or not) at four different time intervals (e.g., at
144
hour 0, hour 7, hour 14 and hour 24) after administration of a drug. To conduct this
study, we have selected 15 individuals randomly from a population and measured
their blood sugar levels at the baseline (hour 0). All the individuals were then given
the drug, and their blood sugar levels were measured again at hour 7, hour 14 and
hour 24. The blood sugar levels at hour 0, hour 7, hour 14, and hour 24 are named
in SPSS as Sugar_0, Sugar_7, Sugar_14 and Sugar_24, respectively. Use the data
file <Data_Repeat_anova_2.sav> for exercise.
19.4.1 Commands:
Analyze > Nonparametric Tests > K Related Samples > Select “sugar_0,
sugar_7, sugar_14 and sugar_24” and push them into “Test Variables” box >
Statistics > Select “Quartile” > Continue > Select “Friedman” > Ok
19.4.2 Outputs:
N Percentiles
25th 50th (Median) 75th
Blood sugar at hour 0 15 106.0000 110.0000 115.0000
Table 19.12. Ranks
Mean Rank
Blood sugar at hour 0 3.80
Table 19.13. Test Statisticsa
N 15
Chi-Square 27.562
df 3
Asymp. Sig. .000
a. Friedman Test
Outputs are provided in tables 19.11-19.13. Table 19.11 shows the median blood
145
sugar levels at 4 different time periods. Look at the Friedman test results as provid-
ed in table 19.13. The Chi-square value is 27.56 and the p-value (Asymp. Sig.) is
0.000, which is <0.05. This indicates that there is a significant difference in blood
sugar levels across the 4 time periods (p<0.001). The findings indicate that the
drug is effective in reducing the blood sugar levels.
19.5 Chi-square test for goodness-of-fit

The Chi-square test of independence, which is most frequently used to determine
the association between two categorical variables, has been discussed in section
13. The Chi-square test for goodness-of-fit is also referred to as one-sample
chi-square test. It is often used to compare the proportion of cases with a hypothet-
ical proportion. Suppose, we have conducted a survey taking a random sample
from a population, and the data show that the prevalence of diabetes is 21.4%.
Now, we want to test the hypothesis, whether the prevalence of diabetes in the
population is 18% or not (the null hypothesis is “the prevalence of diabetes in the
population is 18%”). To have the answer, we shall do the Chi-square test for good-
ness-of-fit (seldom we test such a hypothesis).
19.5.1 Commands:
Analyze > Nonparametric tests > Chi-square > Move the variable “diabetes”
into the “Test Variable List” box > Select “Values” under “Expected Values” >
Write “0.18” in the box > Add > Again write “0.72” (1 minus 0.18) in the box
> Add > Click on “Options” > Select “Descriptive” > Continue > Ok
19.5.2 Outputs:
Table 19.14. Diabetes mellitus
Observed N Expected N Residual

Yes 45 42.0 3.0
No 165 168.0 -3.0
Total 210
146
19.15. Test Statistics
Diabetes mellitus
Chi-Square .268a
df 1
Asymp. Sig. .605
a. 0 cells (.0%) have expected frequencies less than 5.
The minimum expected cell frequency is 42.0.
Table 19.14 provides the observed and expected frequencies for those who have
diabetes (as “Yes”) and those who do not have diabetes (as “No”). These are the
descriptive information, and you do not need to report them. Table 19.15 is the
main table to interpret the results. Our interest is at the p-value. The Chi-square
value is 0.268 and the p-value is 0.605. Since the p-value is >0.05, we cannot reject
the null hypothesis. This means that the prevalence of diabetes in the population
may not be different from 18%.
147
Section 20
Checking Reliability of Scale: Cronbach’s Alpha
When the researchers select a scale (e.g., a scale to measure depression) in their
study, it is important to check that the scale is reliable. One of the ways to check
the internal consistency (reliability) of the scale is to calculate the Cronbach’s
alpha coefficient. Cronbach’s alpha indicates the degree to which the items in the
scale correlate with each other in the group.
Ideally, the Cronbach’s alpha value should be above 0.7. However, this value
is sensitive to the number of items in the scale. If the number of items in the scale
is less than 10, the Cronbach’s alpha coefficient tends to be low. In such a situation,
it is more appropriate to use the “mean inter-item correlations”. The optimum
range of the mean inter-item correlation is between 0.2 and 0.4. Use the data file
<Data_cronb.sav> for practice.
20.1 Cronbach’s alpha

Before doing the procedure to get the Cronbach’s alpha coefficient, be sure that all
the negatively worded values are “reversed” by recoding. If this is not done, it
would produce very low (or negative) value of the Cronbach’s alpha coefficient.
Suppose, we have used a scale to measure the depression. The scale has 4 ques-
tions, q1, q2, q3 and q4. To get the Cronbach’s alpha coefficient, use the following
commands:
Analyze > Scale > Reliability analysis > Select all the items (q1, q2, q3 & q4)
that construct the scale and push them into the “Items” box > Make sure that
“Alpha” is selected in “Model” section (it is usually the default) > Type name
of the scale (e.g., depression or any other name suitable for the data) in the
“Scale label” box > Statistics > Select “Item, Scale, & Scale if item deleted”
under “Descriptives for” section > Select “Correlations” under “Inter-item”
section > Select “Correlation” under “Summaries” section > Continue > Ok
148
20.1.1 Outputs:
N %
Cases Valid 60 100.0
Excludeda 0 .0
Total 60 100.0
a. Listwise deletion based on all variables in the procedure.
Table 20.2. Reliability Statistics
Cronbach's Alpha Cronbach's Alpha Based N of Items

on Standardized Items
.839 .840 4
Table 20.3. Item Statistics

q1 3.1333 1.08091 60
q2 3.2667 1.00620 60
q3 3.0167 1.08130 60
q4 3.2833 1.13633 60
Table 20.4. Inter-Item Correlation Matrix
q1 q2 q3 q4
q1 1.000 .512 .491 .548
q2 .512 1.000 .635 .630
q3 .491 .635 1.000 .589
q4 .548 .630 .589 1.000
Table 20.5. Summary Item Statistics
Mean Minimum Maximum Range Maximum Variance N of Items

/ Minimum
Inter-Item
.567 .491 .635 .143 1.292 .003 4
Correlations
Table 20.6. Item-Total Statistics
Scale Mean if Scale Variance if Corrected Squared Cronbach's Alpha if

Item Deleted Item Deleted Item-Total Multiple Item Deleted
Correlation Correlation
q1 9.5667 7.741 .600 .364 .827
q2 9.4333 7.572 .711 .518 .781
q3 9.6833 7.373 .677 .476 .794
q4 9.4167 6.993 .705 .500 .781
149
Table 20.7. Scale Statistics
Mean Variance Std. Deviation N of Items

12.7000 12.519 3.53817 4
The Reliability Statistics table (table 20.2) shows the Cronbach’s alpha value. In
this example, the value is 0.839. This indicates very good correlation among the
items in the scale (scale is reliable).
However, before looking at the value of Cronbach’s alpha, look at the table
“Inter-item Correlation Matrix” (table 20.4). All the values in the table should be
positive (all are positive in our example). One or more negative values (if there is
any) indicate that some of the items have not been “reverse scored” correctly. This
information is also provided in the table “Item-total Statistics (table 20.6)”. All the
values under “Corrected item - Total Correlation” should be positive (there should
not have any negative value).
The corrected item-total correlation (table 20.6) indicates the degree to which
each item correlates with the total score. In our example, the values are 0.60, 0.71,
0.67 and 0.70. Low value for any item (<0.3) could be a problem. If the Cronbach’s
alpha value (<0.7) (table 20.2) and the corrected item-total correlation value (<0.3)
is low, one may consider omitting the item from the scale with low value. In our
example, there is no such problem.
However, if the number of items is small in the scale (fewer than 10), it may be
difficult to get a reasonable Cronbach’s alpha value. In such a situation, report the
Mean Inter-item Correlation value (Summary-item Statistics table; table 20.5). In
this example, the Inter-item Correlation values range from 0.491 to 0.635, and the
mean is 0.567 (optimum range of the mean is 0.2 to 0.4). This indicates a strong
relationship among the items.
150
Section 21
Analysis of Covariance (ANCOVA): One-way ANCOVA
ANCOVA stands for Analysis of Covariance, which is done to statistically control

the extraneous variable(s) [called covariate] for comparison of the mean 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 the categorical and quantitative variables in
the model, including the interaction between the categorical and quantitative inde-
pendent variables. The ANCOVA can be performed as One-way, Two-way and
Multivariate ANCOVA techniques. Use the data file <Data_3.sav> for practice.
21.1 One-way ANCOVA

The purpose of doing the one-way ANCOVA test is to understand the differences
of the means of the dependent variable (e.g., systolic BP) against a categorical
variable (e.g., sex, or effect of drugs, etc.) after controlling for the quantitative
variable(s) [called covariates, such as age, diastolic BP, etc.] in the model. The
one-way ANCOVA test involves at least three variables:
• One quantitative dependent variable (e.g., systolic BP, post-test score, blood
sugar level, etc.);
• Only one categorical independent variable with two or more levels (e.g.,
sex, type of intervention, or type of drug, etc.); and
• One (or more) covariate (continuous quantitative variable), e.g., diastolic
BP, age, pre-test score, baseline blood sugar level, etc.
The covariates to be selected for the model should be one or more continuous
variable(s) and they should significantly correlate with the dependent variable.
One can also include categorical variables as covariate in the model.
Suppose, the researcher is interested to compare the effectiveness of 3 drugs
(drug A, drug B and drug C) in reducing the systolic BP. To conduct the study, the
researcher has randomly selected three groups of people and assigned these drugs,
one in each group. In such a situation, one-way ANOVA could be done. However,
it was observed that the mean age and pre-treatment systolic BP of these three
groups are not same. Since age and pre-treatment systolic BP can influence the
151
effectiveness of the drugs in reducing the systolic BP, it requires adjustment for
these variables (here, age and pre-treatment systolic BP are the covariates) to
conclude the results. In such a situation, one-way ANCOVA can be used. Note that
for ANCOVA, the independent variable must be a categorical variable (here it is
“type of drug”). ANCOVA can adjust for more than one covariate, either continu-
ous or categorical.
Another example, suppose you have organized a training. To evaluate the
effectiveness of the training, you have taken the pre- and post-tests of the partici-
pants. Now, you want to conclude if males and females (independent variable)
have similar performance in the post-test (dependent variable), after controlling
for age and pre-test results (covariates). One-way ANCOVA is the appropriate test
for both these situations, if the assumptions are met.
Hypothesis:
Suppose, you want to assess if the mean systolic BP (dependent variable) is same
among males and females (independent variable) after controlling for diastolic BP
(covariate).
H0: There is no difference of the 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.
Assumptions:
1. The dependent variable is normally distributed at each level of the indepen-
dent variable;
2. The variances of the dependent variable for each level of the independent
variable are same (homogeneity of variance);
3. The covariates (if more than one) are not strongly correlated with each other
(r<0.8);
4. There is a linear relationship between the dependent variable and the covari-
ates at each level of the independent variable;
5. There is no interaction between the covariate (diastolic BP) and the inde-
pendent variable (sex) [called homogeneity of regression slopes].
152
21.1.1 Commands:
A. Homogeneity of regression slopes (Assumption 5): First, we shall have to
check the homogeneity of regression slopes, using the following commands. Note
that, the SPSS variable names for sex is “sex_1 (0= female; 1= male)”, Systolic BP
is “sbp” and Diastolic BP is “dbp”.
Analyze > General linear model > Univariate > Push “sbp” into the “Depen-
dent Variables” box > Push “sex_1” into the “Fixed Factor” box > Push “dbp”
in the “Covariate box” > Click Model > Select “Custom” under “Specify mod-
el” > Confirm that interaction option is showing in the “Build Terms” box >
Push “sex_1” and “dbp” into the “Model” box > Click on “sex_1” in “Factors
& Covariates” box > Pressing the control button click on the “dbp” in “Factors
& Covariates” box > Push them into the “Model” box (you will see “dbp*-
sex_1” in the Model box) > Continue > Ok
21.1.2 Outputs: Homogeneity of regression slopes

Table 21.1. Between-Subjects Factors
Value Label N
Sex_1 0 Female 133
1 Male 77
Table 21.2. Tests of Between-Subjects Effects
Dependent Variable:SYSTOLIC BP
Source Type III Sum of df Mean Square F Sig.
Squares
Corrected Model 61902.160a 3 20634.053 194.296 .000
Intercept 385.931 1 385.931 3.634 .058
Sex_1 7.019 1 7.019 .066 .797
Dbp 41714.418 1 41714.418 392.795 .000
Sex_1 * dbp 17.964 1 17.964 .169 .681
Error 21877.006 206 106.199
Total 3515465.000 210
a. R Squared = .739 (Adjusted R Squared = .735)
21.1.3 Interpretation: Homogeneity of regression slopes

Only look at table 21.2 (tests of between-subjects effects). Our interest is on the
significance (Sig.) of the interaction (Sex_1*dbp). We can see that the p-value for
153
interaction is 0.681, which is >0.05. This indicates that the homogeneity of regres-
sion slopes assumption is not violated. A p-value of <0.05 indicates that the regres-
sion slopes are not homogeneous and the ANCOVA test is inappropriate.
B. One-way ANCOVA:
To perform the one-way ANCOVA, use the following commands:
Analyze > General linear model > Univariate > Push “sbp” into the “Depen-
dent Variables” box > Push “sex_1” into the “Fixed Factor” box > Push “dbp”
in the “Covariate” box > Click “Model” > Select “Full Factorial” > Continue
> Options > Select “sex_1” and push it into the “Display Means for” box (this
would provide the adjusted means) > Select “Compare main effects” > Select
“Bonferroni” from “Confidence interval adjustment” > Select “Descriptive
Statistics, Estimates of effect size, and Homogeneity” tests under “Display” >
Continue > Ok
21.1.4 Outputs: One-way ANCOVA

Value Label N
Sex 0 Female 133
1 Male 77

Sex Mean Std. Deviation N
Female 129.73 21.309 133
Male 124.56 17.221 77
Total 127.83 20.021 210
Table 21.5. Levene's Test of Equality of Error Variancesa

F df1 df2 Sig.
.365 1 208 .546
Tests the null hypothesis that the error variance of the
dependent variable is equal across groups.
a. Design: Intercept + dbp + sex_1
154
Dependent Variable:SYSTOLIC BP
Source Type III Sum of df Mean Square F Sig. Partial Eta
Squares Squared
Corrected Model 61884.196a 2 30942.098 292.534 .000 .739
Intercept 688.574 1 688.574 6.510 .011 .030
Dbp (diastolic BP) 60580.272 1 60580.272 572.740 .000 .735
Sex_1 143.945 1 143.945 1.361 .245 .007
Error 21894.971 207 105.773
Total 3515465.000 210
a. R Squared = .739 (Adjusted R Squared = .736)
Table 21.7. Estimated Marginal
Sex
Sex Mean Std. Error 95% Confidence Interval
Female 127.191a .898 125.421 128.962
Male 128.942a 1.186 126.604 131.281
a. Covariates appearing in the model are evaluated at the following values:
DIASTOLIC BP = 83.04.
Table 21.8. Pairwise Comparisons
Dependent Variable:Systolic BP
(I) Sex: (J) Sex: Mean Std. Error Sig.a 95% Confidence Interval
numeric numeric Difference (I-J) Differencea
Female Male -1.751 1.559 .263 -4.825 1.322
Male Female 1.751 1.559 .263 -1.322 4.825

155
Table 21.9. Pairwise Comparisons
Dependent Variable:Systolic BP
(I) Religion (J) Religion Mean Std. Error Sig.a 95% Confidence Interval
Difference (I-J) for Differencea
Lower Upper
Bound Bound
MUSLIM HINDU -.448 1.705 1.000 -4.562 3.666
CHRISTIAN .948 2.313 1.000 -4.635 6.532
HINDU MUSLIM .448 1.705 1.000 -3.666 4.562
CHRISTIAN 1.397 2.535 1.000 -4.721 7.514
CHRISTIAN MUSLIM -.948 2.313 1.000 -6.532 4.635
HINDU -1.397 2.535 1.000 -7.514 4.721
21.1.5 Interpretation: One-way ANCOVA

Table 21.3 and 21.4 displayed the descriptive statistics. Table 21.4 shows the unad-
justed means of the systolic BP by sex (female: 129.7 and male: 124.5).
Table 21.5 shows the Levene’s test of Equality of Error Variances. This is the
test for assumption 2. We expect the p-value (Sig.) to be >0.05 to meet the assump-
tion. In this example, the p-value is 0.54, which is more than 0.05. This means that
the variances of the dependent variable (systolic BP) are same at each level of the
independent variable (sex).
Table 21.6 (tests of between-subjects effects) is the main table showing the
results of one-way ANCOVA test. We tested the hypothesis whether the population
mean of the systolic BP of males and females is same after controlling for diastolic
BP. Look at the p-value for sex in the table, and it is 0.245. Since the p-value is
>0.05, we cannot reject the null hypothesis. This indicates that the mean systolic
BP (in the population) of males and females is not different after controlling for
diastolic BP. Also look at the value for Partial Eta Squared. Eta indicates the
amount of variance (also called effect size) in the dependent variable that is
explained by the independent variable (sex). We can see that the effect size is very
small (0.007 or 0.7%).
We can also have information about the influence 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. The value of the Partial
Eta Squared for diastolic BP is 0.735 (73.5%). This means that 73.5% variance of
156
systolic BP can be explained by the diastolic BP, after controlling for sex.
Table 21.7 (estimated marginal) shows the adjusted (adjusted for diastolic BP)
means of the dependent variable (systolic BP) at different levels of the indepen-
dent variable (sex). We can see that the mean systolic BP of females is 127.19
mmHg and that of males is 128.94 mmHg, after adjusting for diastolic BP (note
that the adjusted means are different from the unadjusted means as shown in table
21.4).
Table 21.8 is the table for pairwise comparison. This table is not necessary in
this example, since the independent variable (sex) has two levels. If the indepen-
dent variable has more than two levels, then the table for pairwise comparison is
important to look at, especially if there is a significant association between the
dependent and independent variable. Look at table 21.9 [this is an additional table
I have provided where the independent variable (religion) has three categories],
which shows the pairwise comparison of mean systolic BP by religious groups.
The results indicate that there is no significant difference of the mean systolic BP
among different religious groups after controlling for diastolic BP, since all the
p-values are >0.05.
157
Section 22
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. Therefore, in two-way ANCOVA,
four variables are involved. They are:
• One continuous dependent variable (e.g., diastolic BP, blood sugar, post-test
score, etc.);
• Two categorical independent variables (with two or more levels) [e.g., occu-
pation, diabetes, type of drug, etc.]; and
• One or more continuous covariates (e.g., age, systolic BP, income, etc.).
Use the data file <Data_3.sav> for practice.
22.1 Two-way ANCOVA

The two-way ANCOVA provides information, after controlling for the covari-
ate(s), on:
1. Whether there is a significant main effect on the dependent variable for the
first independent variable (e.g., occupation);
2. Whether there is a significant main effect on the dependent variable for the
second independent variable (e.g., diabetes);
3. Whether there is an interaction between the independent variables (e.g.,
occupation and diabetes).
Suppose, we want to assess, after controlling for age (covariate):
• Whether, occupation influences the diastolic BP (i.e., is mean diastolic BP in
different occupation groups same);
• Whether, diabetes influences the diastolic BP (i.e., is the mean diastolic BP
same for diabetics and non-diabetics); and
• Does the influence of occupation on diastolic BP depend on the presence of
diabetes (i.e., is there interaction between occupation and diabetes)?
The question numbers 1 and 2 refer to the main effect, while question 3
158
explains the interaction of two independent variables (occupation and diabetes) on
the dependent variable (diastolic BP). For analysis, we shall use the data file
<Data_3.sav>. Note that the SPSS variable names for diastolic BP is “dbp”, occu-
pation is “occupation”, diabetes is “diabetes” and age is “age”.
Assumptions:
All the assumptions mentioned under one-way ANCOVA are applicable for
two-way ANCOVA. Look at one-way ANCOVA for the assumptions and how to
check them.
22.1.1 Commands:
To perform the two-way ANCOVA, use the following commands:
Analyze > General linear model > Univariate > Push “dbp” into the “Depen-
dent Variables” box > Push “occupation” and “diabetes” into the “Fixed
Factor” box > Push “age” into the “Covariate” box > Click “Model” > Select
“Full Factorial” > Continue > Options > Push “occupation, diabetes and occu-
pation*diabetes” into the “Display Means for” box (this would provide the
adjusted means of the diastolic BP for occupation and diabetes) > Select
“Compare main effects” > Select “Bonferroni” from “Confidence interval
adjustment” > Select “Descriptive Statistics, Estimates of effect size, and
Homogeneity tests” > Continue > Plots > Select “occupation” and push into
the “Horizontal” box > Select “diabetes and push it into the “Separate Lines”
box > Click “Add” > Continue > Ok
22.1.2 Outputs:
Value Label N
OCCUPATION 1 GOVT JOB 60
2 PRIVATE JOB 49
3 BUSINESS 49
4 OTHERS 52
DIABETES MELLITUS 0 No 165
1 yes 45
159
Table 22.2. Descriptive Statistics (unadjusted)
Dependent Variable: DIASTOLIC BP

OCCUPATION DIABETES MELLITUS Mean Std. Deviation N
GOVT JOB No 84.58 12.862 50
yes 82.60 11.559 10
Total 84.25 12.583 60
PRIVATE JOB No 82.80 14.128 41
yes 79.75 9.468 8
Total 82.31 13.443 49
BUSINESS No 83.97 12.940 36
yes 84.69 12.977 13
Total 84.16 12.818 49
OTHERS No 80.87 12.036 38
yes 82.43 7.822 14
Total 81.29 11.009 52
Total No 83.15 12.982 165
yes 82.64 10.410 45
Total 83.04 12.454 210
Table 22.3. Levene's Test of Equality of Error Variancesa

F df1 df2 Sig.
.963 4 202 .459
Tests the null hypothesis that the error variance of the
dependent variable is equal across groups.
a. Design: Intercept + age + occupation + diabetes + occupation * diabetes
Dependent Variable:DIASTOLIC BP
Source Type III Sum of df Mean Square F Sig. Partial Eta
Squares Squared
Corrected Model 495.245a 8 61.906 .390 .925 .015
Intercept 101042.759 1 101042.759 636.198 .000 .760
age 34.661 1 34.661 .218 .641 .001
Occupation 229.135 3 76.378 .481 .696 .007
diabetes 12.293 1 12.293 .077 .781 .000
occupation * diabetes 124.301 3 41.434 .261 .854 .004
Error 31923.370 201 158.823
Total 1480603.000 210
a. R Squared = .015 (Adjusted R Squared = -.024)
160
Table 22.5. Estimated Marginal for occupation (adjusted)
1. OCCUPATION
OCCUPATION Mean Std. Error 95% Confidence Interval
GOVT JOB 83.605a 2.183 79.300 87.910
PRIVATE JOB 81.253a 2.436 76.449 86.056
BUSINESS 84.374a 2.041 80.350 88.399
OTHERS 81.708a 1.974 77.815 85.601
a. Covariates appearing in the model are evaluated at the
following values: age = 26.5143.
Table 22.6. Estimated Marginal for diabetes (adjusted)
2. DIABETES MELLITUS
DIABETES Mean Std. Error 95% Confidence Interval
MELLITUS Lower Bound Upper Bound
No 83.037a .990 81.086 84.988
yes 82.433a 1.930 78.627 86.239
a. Covariates appearing in the model are evaluated at the

following values: age = 26.5143.
Table 22.7. Estimated Marginal
3. OCCUPATION * DIABETES MELLITUS

OCCUPATION DIABETES Mean Std. Error 95% Confidence Interval
MELLITUS Lower Bound Upper Bound
GOVT JOB No 84.545a 1.784 81.028 88.063
yes 82.665a 3.988 74.802 90.528
PRIVATE JOB No 82.777a 1.969 78.894 86.659
yes 79.729a 4.456 70.942 88.515
BUSINESS No 83.979a 2.100 79.837 88.121
yes 84.769a 3.499 77.870 91.669
OTHERS No 80.847a 2.045 76.815 84.880
yes 82.569a 3.381 75.901 89.236
a. Covariates appearing in the model are evaluated at the following values: age = 26.5143.
161
22.8. Pairwise Comparisons
Dependent Variable:Diastolic BP
(I) Occupation (J) Occupation Mean Std. Error Sig.a 95% Confidence Interval
Difference (I-J) for Differencea
Lower Upper
Bound Bound
GOVT JOB PRIVATE JOB 2.139 3.089 1.000 -6.093 10.371
BUSINESS -.379 2.821 1.000 -7.896 7.138
OTHERS 1.778 2.778 1.000 -5.625 9.180
PRIVATE JOB GOVT JOB -2.139 3.089 1.000 -10.371 6.093
BUSINESS -2.518 3.002 1.000 -10.519 5.482
OTHERS -.361 2.963 1.000 -8.257 7.534
BUSINESS GOVT JOB .379 2.821 1.000 -7.138 7.896
PRIVATE JOB 2.518 3.002 1.000 -5.482 10.519
OTHERS 2.157 2.678 1.000 -4.978 9.291
OTHERS GOVT JOB -1.778 2.778 1.000 -9.180 5.625
PRIVATE JOB .361 2.963 1.000 -7.534 8.257
BUSINESS -2.157 2.678 1.000 -9.291 4.978
Figure 22.1 Mean diastolic BP of different occupation groups by diabetes after adjustment for age
Table 22.1 and 22.2 shows the descriptive statistics. All the means provided in
table 22.2 are the crude (unadjusted) means, i.e., without adjusting for age.
162
Table 22.3 shows the results of Levene’s test of Equality of Error Variances. This
is the test for homogeneity of variances. We expect the p-value (sig.) to be >0.05
to meet the assumption. In this example, the p-value is 0.459, which is more than
0.05. This means that the variances of the dependent variable (diastolic BP) are
same for each level of the independent variables (occupation and diabetes).
Table 22.4 (tests of between-subjects effects) is the main table showing the
results of the two-way ANCOVA test. We tested the hypothesis whether:
• Mean diastolic BP (in the population) in different occupation groups is same
after controlling for age;
• Mean diastolic BP (in the population) among diabetics and non-diabetics is
same after controlling for age; and
• Is there any interaction between occupation and diabetes after controlling for
age?
Look at the p-values for occupation, diabetes and occupation*diabetes in table
22.4. They are 0.696, 0.781 and 0.854, respectively, indicating that none of them
are statistically significant. This means that occupation and diabetes do not have
any influence on the diastolic BP after controlling for age. There is also no interac-
tion between occupation and diabetes after controlling for age. However, we
should always check the p-value of interaction first. If the interaction is significant
(p-value <0.05), then the main effects (of occupation and diabetes) are not import-
ant, because effect of one independent variable is dependent on the level of the
other independent variable.
The effect size is indicated by the value of Partial Eta Squared. Eta indicates
the amount of variance in the dependent variable that is explained by the indepen-
dent variable (also called effect size). We can see that the effect sizes are very
small both for occupation (0.007) and diabetes (0.000) (table 22.4).
We can also have information about the influence of the covariate (age) on the
dependent variable (diastolic BP). We can see (table 22.4) that the p-value for age
is 0.641, which is not statistically significant. This indicates that there is no signifi-
cant association between age and diastolic BP after controlling for occupation and
diabetes. The value of Partial Eta Squared for age is 0.001 (0.1%). This means that
less than 1% variance in diastolic BP can be explained by age, after controlling for
occupation and diabetes.
Tables 22.5 and 22.6 (estimated marginal) show the adjusted means of the
163
diastolic BP (dependent variable) at different levels of the independent variables
(occupation and diabetes) after controlling for age. In this example, the adjusted
mean of diastolic BP of government job holders is 83.6 mmHg and that of the
diabetics (diabetes mellitus: yes) is 82.4 mmHg, after controlling for age. Similar-
ly, the last table (table 22.7) shows the adjusted mean of diastolic BP of different
occupation groups by diabetes.
Table 22.8 is the table of pairwise comparison of the mean diastolic BP in
different occupation groups. This table is necessary when the independent variable
has more than two levels, and there is a significant association between the depen-
dent and independent variable. Look at the p-values (Sig.) in table 22.8. Since all
the p-values are >0.05, there is no significant difference of the mean diastolic BP
in the population between different occupation groups after controlling for age.
Figure 22.1 plotted the mean diastolic BP of different occupation groups disag-
gregated by diabetes. Finally, from the data, we conclude that the diastolic BP is
not influenced (there is no association) by occupation and diabetes after
controlling for age.
164
Annex
Table A.1. Codebook of data file <Data_3.sav>
SPSS 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 Actual value
Tk.
sbp Systolic blood pressure in Actual value
mmHg
dbp Diastolic blood pressure in Actual value
mmHg
f_history Family history of diabetes 0= No
1= Yes
pepticulcer Have peptic ulcer 1= Yes
2= No
diabetes Have diabetes mellitus 1= Yes
2= No
post_test Post-test score Actual value
pre_test Pre-test score Actual value
date_ad Date of hospital admission Actual value
date_dis Date of discharge Actual value
165
References
1. Daniel WW. (1999). Biostatistics: A Foundation for Analysis in the Health
Science. 7th edition. John Wiley & Sons, Inc.
2. Altman DG. (1992). Practical Statistics for Medical Research. 1st Edition.
Chapman & Hill.
3. Katz MH. (2011). Multivariable Analysis: A Practical Guide for Clinicians and
Public Health Researchers. 3rd Edition. London, Cambridge University Press.
4. Katz MH. (2009). Study Design and Statistical Analysis: A Practical Guide for
Clinicians. Cambridge University Press.
5. Gordis L. (2014). Epidemiology. 5th Edution. ELSEVIER Sounders.
6. Szklo M, Nieto FJ. (2007). Epidemiology: Beyond the Basics. 2nd Edition.
Jones and Bartlett Publishers.
7. Field A. (2002). Discovering Statistics Using SPSS for Windows. SAGE Publi
cations: London, California, New Delhi.
8. 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.
9. Katz MH. Multivariable Analysis: A Primer for Readers of Medical Research.
Ann Intern Med 2003; 138:644–650.
10. Pallant J. (2007). SPSS Survival Manual. 3rd edition. Open University Press.
11. Chan YH. Biostatistics 103: Qualitative Data – Tests of Independence. Singa-
pore Med J 2003; Vol 44(10):498-503.
12. Chan YH. Biostatistics 104: Correlational Analysis. Singapore Med J 2003;
Vol 44(12):614-619.
13. Chan YH. Biostatistics 201: Linear Regression Analysis. Singapore Med J
2004; Vol 45(2):55-61.
14. Chan YH. Biostatistics 202: Logistic regression analysis. Singapore Med J
2004; Vol 45(4):149-153.
15. Chan YH. Biostatistics 203. Survival analysis. Singapore Med J 2004; Vol
45(6):249-256.
16. Amderson M, Nelson A. Data analysis: Simple statistical tests. FOCUS on
Field Epidemiology; Vol 3(6).
166

SPSS Without Pain

Uploaded by

Copyright:

Available Formats

SPSS Without Pain

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

SPSS Without Pain

Uploaded by

Copyright:

Available Formats

!"#$!!$%&'$()*&+,*$!

Learning SPSS without Pain

Mohammad Tajul Islam

Professor (Adjunct), Department of Public Health

Copyright © 2020 by Author

Price: Tk. 250

Links for the e-manual and data files:

I am particularly indebted and grateful to Harun Or Rashid, Monjura Khatun Nisha

Section 2 Generating Data File 3

Section 3 Data Cleaning and Data Screening 11

Section 4 Data Analysis: Descriptive Statistics 14

Section 5 Checking Data for Normality 25

Section 6 Data Management 30

Section 7 Testing of Hypothesis 46

Section 8 Student’s t-test for Hypothesis Testing 51

Section 9 Analysis of Variance (ANOVA): One-way ANOVA 58

Section 10 Two-way ANOVA 64

Section 12 Repeated Measures ANOVA: Within and Between-

Section 13 Association between Two Categorical Variables:

Section 14 Association between Two Continuous Variables:

Section 15 Linear Regression 96

Section 16 Logistic Regression 115

Section 17 Survival Analysis 128

Section 18 Cox Regression 135

Section 19 Non-parametric Methods 140

Section 20 Checking Reliability of Scale: Cronbach’s Alpha 148

Section 21 Analysis of Covariance (ANCOVA): One-way ANCOVA 151

Section 22 Two-way ANCOVA 158

SPSS stands for Statistical Package for Social Sciences. It is a powerful

1.1 Steps of data analysis

2.1 Generating data file

2.1.1 Defining variables:

Align: You do not need to do anything for this.

2.2 Data used in this manual

Links for the data files:

3.1 Checking for out-of-range errors

3.2 Checking for outliers

Figure 3.1. Box and Plot chart of systolic blood pressure

Systolic BP Statistics Std. Error

3.3 Assessing normality of a dataset

4.1 Frequency distribution

Figure 4.1. Commands for frequency distribution of variables

Table 4.1. Frequency distribution of sex with no missing value

Table 4.2. Frequency distribution of sex with 4 missing values

Figure 4.3. Commands for obtaining central tendency and dispersion

Figure 4.5. Selection of statistics for the variable(s)

4.3 Alternative method of getting measures of central tendency and dispersion

Statistics Std. Error

Figure 4.6. Histogram of age

Figure 4.7. Stem and leaf chart of age

Stem width: 10.00

Figure 4.9. Box and plot chart of diastolic BP

4.4 Descriptive statistics and histogram disaggregated by Sex

Table 4.5. Descriptive statistics of age by sex

4.5 Checking for outliers

Table 4.6. Extreme values of the variable age

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

Figure 5.2. Q-Q plot of Systolic BP

Table 5.1 Descriptive statistics of Systolic BP

Table 5.2. Statistical tests for normality (of Systolic BP)

!"#$!!$%&'$()&+,$!