University of Nebraska - Lincoln

DigitalCommons@University of Nebraska - Lincoln

Zea E-Books Zea E-Books

3-12-2018

Essentials of Structural Equation Modeling

Mustafa Emre Civelek
Istanbul Commerce University, [email protected]

Follow this and additional works at: https://digitalcommons.unl.edu/zeabook

Part of the Applied Statistics Commons, Multivariate Analysis Commons, Other Applied
Mathematics Commons, Other Mathematics Commons, Quantitative, Qualitative, Comparative,
and Historical Methodologies Commons, and the Statistical Models Commons

Recommended Citation
Civelek, Mustafa Emre, "Essentials of Structural Equation Modeling" (2018). Zea E-Books. 64.
https://digitalcommons.unl.edu/zeabook/64

This Book is brought to you for free and open access by the Zea E-Books at DigitalCommons@University of Nebraska - Lincoln. It has been accepted
for inclusion in Zea E-Books by an authorized administrator of DigitalCommons@University of Nebraska - Lincoln.
Essentials of
Structural
Equation
Modeling

Dr. Mustafa Emre Civelek

Essentials of Structural Equation Modeling

Dr. Mustafa Emre Civelek

Structural Equation Modeling is a statistical method increasingly used in

scientific studies in the fields of Social Sciences. It is currently a preferred
analysis method, especially in doctoral dissertations and academic
researches. However, since many universities do not include this method
in the curriculum of undergraduate and graduate courses, students and
scholars try to solve the problems they encounter by using various books
and internet resources.

This book aims to guide the researcher who wants to use this method in
a way that is free from math expressions. It teaches the steps of a research
program using structured equality modeling practically. For students
writing theses and scholars preparing academic articles, this book aims
to analyze systematically the methodology of scientific studies conducted
using structural equation modeling methods in the social sciences.

This book is prepared in as simple language as possible so as to convey

basic information. It consists of two parts: the first gives basic concepts of
structural equation modeling, and the second gives examples of
applications.

ISBN: 978-1-60962-129-2

doi:10.13014/K2SJ1HR5

This work is licensed under a

Creative Commons Attribution 4.0 International License.
ESSENTIALS OF STRUCTURAL
EQUATION MODELING

Dr. Mustafa Emre Civelek

Zea Books
Lincoln, Nebraska 2018
1
 ISBN: 978-1-60962-129-2
doi: 10.13014/K2SJ1HR5

Copyright © 2018 Mustafa Emre Civelek

This work is licensed under a

Creative Commons Attribution 4.0 International License.

Zea Books are published by the University of Nebraska–Lincoln Libraries.

Electronic (pdf ) editions published online at
https://digitalcommons.unl.edu/zeabook/
Print editions sold at
ttp://www.lulu.com/spotlight/unllib

UNL does not discriminate based upon any protected status.

Please go to unl.edu/nondiscrimination
PREFACE
Structural Equation Modeling is a statistical method
increasingly used in scientific studies in the field of Social
Sciences. Nowadays, it is a preferred analysis method
especially in doctoral dissertations and academic researches.
However, since most of the universities do not include this
method in the curriculum of undergraduate and graduate
courses, students and scholars try to solve the problems they
encounter by using various books and internet resources. The
aim of this book is to guide the researcher who wants to use
this method in a way that is free from math expressions and to
teach the steps of a research using structured equality
modeling practically. For the students who write thesis and the
scholars who prepare academic articles, this book aims to
systematically analyze the methodology of the scientific studies
which is conducted by using structural equation modeling
method in social sciences. This book is prepared in a simple
language as possible so as to convey basic information. This
book consists of two parts. In the first part, basic concepts of
structural equation modeling are given. In the second part,
examples of applications are given.

1
CONTENT
PREFACE .......................................................................................................................... 1
CONTENT ........................................................................................................................ 2
1. INTRODUCTION ....................................................................................................... 4
I. CHAPTER...................................................................................................... 6
CONCEPTS AND METHODOLOGY ...................................................... 6
2. BASIC CONCEPTS OF STRUCTURAL EQUATION
MODELING ..................................................................................................................... 6
2.1. Definitions and Features .................................................................. 6
2.2. Latent and Observed Variables .....................................................10
2.3. Endogeneous and Exogeneous Variables ...................................13
2.4. Parameters ........................................................................................16
2.5. Fit Indices.........................................................................................17
2.6. Assumptions of Structural Equation Modeling ..........................22
2.7. Types of Structural Equation Models ..........................................26
2.7.1. Path Analysis Models ..............................................................26
2.7.2. Confirmatory Factor Analysis Models .................................27
2.7.3. Structural Regression Models ................................................28
2.7.4. Latent Change Models ............................................................28
3. VALIDITY AND RELIABILITY ANALYSIS ..................................................... 31
3.1. Determination of Convergent Validity ........................................32
3.2. Determination of Discriminant Validity ......................................41
3.3. Determination of Reliability ..........................................................42
4. STRUCTURAL REGRESSION MODEL ............................................................. 44
5. PATH ANALYSIS ...................................................................................................... 48
6. MODEL MODIFICATION .................................................................................... 50
7. MEDIATOR VARIABLE ANALYSIS METHODOLOGY ............................. 52
II. CHAPTER ..................................................................................................61
APPLICATIONS ............................................................................................61

2
8. USE OF AMOS Program .......................................................................................... 61
9. SAMPLE OF MEDIATOR VARIABLE ANALYSIS ......................................... 73
9.1. Title of the Research.......................................................................73
9.2. Purpose .............................................................................................74
9.3. Conceptual Model and Scales........................................................74
9.4. Determination of Validity and Reliability ....................................75
9.5. Analysis Results ...............................................................................77
9.6. Conclusion .......................................................................................79
10. MULTITRAIT-MULTIMETHOD MODEL ...................................................... 80
11. METHODS TO BE APPLIED IN CASE OF DATA
INADEQUACY ............................................................................................................ 100
11.1. Bootstrap Method ...................................................................... 100
11.2. Partial Least Square Structural Equation Modeling ( PLS-
SEM) ..................................................................................................... 102
Bibliography .................................................................................................................... 111

3
1. INTRODUCTION
There are various computer programs that are used when the
structural equation modeling method is applied. The most
common ones are LISREL (Linear Structural Relations),
AMOS (Analysis of Moment Structures), MPlus, EQS
(Equation Modeling Software) (Taşkın & Akat, 2010). The
samples and illustrations in this book were made according to
the AMOS program. The AMOS program is a visual program
that is easier to use than other programs. This is why it is
preferred in this book. There will also be plenty of videos
related to AMOS on YouTube and similar sites. That will
make the learning process easier for readers. There are
different versions of the AMOS program. AMOS 22 version
was used for the analysis examples in the book. All the
narrations in the book have been tried to be supported by
visuals. In this book, it is assumed that the readers have basic
statistical information. The main aim is to provide a guide to
the readers on structural equation modeling.

Although the structural equation modeling method is similar

to linear regression analysis, it has many advantages. Some of
the features that outperform the structural equation modeling
are summarized below. These superior features distinguish
structural equation modeling from other classical linear
modeling approaches (Çelik & Yılmaz, 2013).

1. It reveals the relationship among hidden structures that

are not directly measured.

4
 2. Possible mistakes in the measurements of the observed
variables are taken into consideration. The classic regression
approach assumes no measurement error.
3. It is a very useful method to analyze highly complex
multiple variable models and to reveal direct and indirect
relationships between variables.

In this book, firstly the basic concepts related to structural

equation modeling are discussed. Basic concepts such as
measurement model versus structural model, latent versus
observed variables, and endogeneous versus exogeneous
variables are explained. Once the goodness of fit indices have
been defined, all the processes of a research have been dealt
with up to hypothesis testing, from determining the validity
and reliability of the scale. The first chapter of the book
consists of a description of the topics related to the structural
equation modeling. In the second chapter, there are sample
applications. Samples consist of summaries of actual research
done on the field. This is intended to give readers a template
that they can follow in their own research. It is recommended
that the readers who will be using AMOS for the first time
have started this book after looking at the tutorial videos on
YouTube and similar sites. Also, to use AMOS, it is absolutely
necessary to be able to use SPSS and to have basic statistical
information. Assuming that the readers have this information,
the subject explanations have been prepared.

5
 I. CHAPTER
CONCEPTS AND METHODOLOGY

2. BASIC CONCEPTS OF STRUCTURAL

EQUATION MODELING

2.1. Definitions and Features

Structural equation modeling is a statistical method
increasingly used in scientific studies in the field of social
sciences in recent years. The most important reason of the
spread of this statistical technique is that the direct and
indirect relationships among causal variables can be measured
with a single model (Meydan & Şen, 2011). Structural equation
modeling is a statistical method used to test the relationships
between observed and latent variables. Observed variables are
the measured variables in the data collection process and
latent variables are the variables measured by connecting to
the observed variables because they can not be directly
measured. Structural Equation Modeling consists of two basic
components as structural model and measurement model.

Another reason for the widespread adoption of this method is

the ability of taking in to the account of the measurement
errors and the relationships between errors in the observed

6
variables. In this way measurement errors can be minimized.
In traditional regression analysis, potential measurement errors
are neglected. Another difference from the regression models
of structural equality models is that they are based on the
covariance matrix. For this reason, in some sources, it is
named as covariance structure modeling or analysis of
covariance structure (Bayram, 2013). On the other hand, , the
correlation matrix is the basis of the regression. Covariance is
a nonstandardized measure of the relationship between two
variables, so it can take values between - ∞ and + ∞.
Correlation, however, can take values between -1 and +1,
since it is standardized (Gujarati, 1999). The covariance which
is the basic statistic of the structural equation model can be
shown for two observed and continuous variables as follows:

Figure 1. Demarcation between Measurement Model and

Structural Model
Source: Byrne, B. M. (2010). Structural Equation Modeling with AMOS. New York:
Routledge Taylor & Francis Group.

7
 Covxy = rxy SDx SDy

In the formula shown above, rxy indicates the Pearson

correlation coefficient, and also, SDx and SDy indicate the
standard deviation of each variable (Kline, 2011). Structural
equation modeling differs from some other multivariate
statistical methods in that it is a confirmatory approach. In
Table 1, it can be seen that the explanatory and confirmatory
ones of the multivariate methods are (Hair, Hult, Ringle, &
Sarstedt, 2017).

Tablo 1. Explanatory and Confirmatory Multivariate

Methods

Explanatory Confirmatory

x CulusterAnalysis x Analysis of Variance

x Explanatory Factor x Logistic Regression

Analysis
x Multiple Regression
x Multidimensional
Scaling x Confirmatory Factor

Analysis
x Partial Least Squares
Structural Equation x Covariance Based
Modeling (PLS-SEM) Structural Equation
Modeling (CB-SEM)
Source: Hair, J., Hult, G., Ringle, C., & Sarstedt, M. (2017). A primer on partial least
squares structural equation modeling PLS-SEM. Los Angeles: SAGE.

Most of the statistical methods other than structural equation

modeling try to discover relationships through the data set.

8
However, structural equation modeling confirms the
correspondence of the data of the relations in the theoretical
model. For this reason, it can be said that structural equation
modeling is more suitable for testing the hypothesis than other
methods (Karagöz, 2016). Structural equation modeling
consists of a system of linear equations. The key in the
regression analysis is to determine how much of the change in
the dependent variable is explained by the independent
variable or variables. Although multiple regression analysis can
only be applied to observed variables, the basic principles can
be applied to structural equation modeling (Kline, 2011).
Differently from the regression, structural equation modeling,
as a new statistical analysis technique, allows to test research
hypotheses in a single process by modeling complex
relationships among many observed and latent variables. In
traditional regression analysis, only direct effects can be
detected. However, in the method of structural equation
modeling, direct and indirect effects are put together.

In order to test the accuracy of the conceptual model, the

most common method encountered in the literature on
structural equation modeling is a two-stage method consisting
of measurement model and structural model. In the first stage,
the measurement model is tested; in the second stage the
structural model is tested. The measurement model measures
how well hidden variables are represented by the observed
variables. It is mainly confirmatory factor analysis (CFA) and
indicates the contruct validity of scales. Therefore, if the
measurement model fit indices are low, it will not make sense
to test the structural model (Dursun & Kocagöz, 2010). As
seen in Figure 1, structural equation modeling is a compound
of factor analysis and regression analysis. The measurement

9
model and the structural model are interwoven. But the
structural equation modeling is based on the confirmatory
approach. It is based on the statistical confirmation of the
theoretical model. For this reason, the measurement model is
confirmatory factor analysis.

2.2. Latent and Observed Variables

The obseved variable (manifest variable) is the measured

variable in the data collection process; latent variables are
variables that are measured by connecting to observed
variables because they can not be directly measured. Latent
variables must be represented by more than one observed
variable since they represent abstract concepts. It is
recommended that the number of observed variables
connected to a latent variable in structural equality models be
at least three. Latent variables represent hypothetical
constructs in a research model (Raykov & Marcoulides, 2006).

Observed variables in the structural equation modeling can be

categorical, discrete or continuous variable, but latent variables
must always be continuous variable. There are other statistical
techniques that allow analysis with categorical latent variables,
but in structural equation modeling only continuous variables
can be analysed (Kline, 2011). Continuous variable is a
variable that takes any random value from the set of real
numbers. However discrete variable can only take value from
the set of even number. Research questionnaires used to
collect data in social sciences are generally prepared by using
the Likert type ordinal scale. Values in this scale type take
integer values ordered by importance. Therefore, the
indicators used to describe a concept are discrete variables.

10
Latent variables are always continuous variables due to they
are linked to more than one indicator.

Figure 2. Endogeneous and Exogeneous Variables

As shown in Figure 1, the first latent variable (Latent Variable

1) is linked to the observed variables (Question 1, Question 2,
and Question 3) which are composed of the questions
answered in the research questionnaire. As seen in Figure 1,
the observed variables and latent variables are connected to
each other in a reflective way.

In the same way, the observed variables of the second

dimension (Question 4, Question 5 ve Question 6) in the
conceptual model of the research are connected to another
latent variable (Latent Variable 2). The direction of the arrows

11
connecting the observed variables and the hidden variables is
important.

Figure 3. Symbols in Structural Equation Models

Again referring to the examples in Figure 1 and Figure 2, the

quadrangular symbols indicate the observed variables, while
the elliptical symbols represent the predicted latent variables
(this book is based on the notation used in the AMOS
program). Figure 3 shows the meanings of the most
commonly used symbols in structural equation models.

12
 2.3. Endogeneous and Exogeneous Variables
Variables in the structural equation modeling, except for the
distinction between latent and observed, are dealt with in two
groups: endogeneous and exogeneous. In structural equation
modeling, the Endogeneous Exogeneous distinction is used as
a more accurate distinction because a variable can assume the
role of both the dependent variable and the independent
variable at the same time. Endogeneous variables are
dependent variables explained by other variables. In Figure 2,
the variables Z, W and T are endogeneous variables.
Exogeneous variables are independent variables that are not
explained by any variables. In Figure 2, the variables X and Y
are external variables. If there are more than one exogeneous
variable, covariance between these variables is required. As
shown in Figure 2, a bi-directional arrow is placed between the
X and Y latent variables. It should not be forgotten to add an
error term to the endogeneous variables. The terms Res1,
Res2 and Res3, which appear in Figure 2, represent the
residuals of each endogeneous variable. These residuals are
also called as error terms of the structural model. The error
terms in the measurement model are shown in Figure 2 by the
notation 'e'. The error terms in both groups are usually marked
with a separate notation in this way. Unlike regression models,
structural equality models are based on the covariance matrix.
But it mainly consists of the system of linear equations.
Therefore, the linear equations of the model in Figure 2 can
be written as:
(1. Equation) Z= β1.X+β2.Y+1.Res1

(2. Equation) W= β3.X+1.Res2

(3. Equation) T= β4.Z+β5.Y+ β6.W+1.Res3

13
As shown in Figure 2, there are 3 endogeneous (dependent)
variables, and the above three equations are written. The
residual term of the first equation named as Res1, the residual
term of the second and third equation named as Res2 and
Res3 respectively. The coefficients of residual terms are fixed
as '1'. β coefficients are free parameters. The variables to the
left of the above equations are dependent variables, while the
ones to the right are independent variables. There is another
point of interest in equations. W and Z variables are
dependent variables in the first and second equations while
they play an independent role in the third equation. In the
group of these equations only the variables X and Y are always
independent. This is because X and Y variables are
exogeneous variables as seen in Figure 2. And the variables T,
W and Z are endogeneous variables. Again as seen in Figure
2., residual error of each latent variable correspond to the
residual term of each regression equation in which the latent
variable has the dependent variable role. Residual terms mainly
represent variance that can not be explained by factor (Kline,
2011). The square root of the variance equals to the standard
deviation. This value represents how far or near the
distribution of the values in a serial from the mean. If the
standard deviation is small, the values are scattered close to
the serial average.

Measurement errors of the observed variables that form the

latent variable are symbolized by "e" notation. Once again, the
superiorities of structural equation modelling are that more
than one regression analysis can be conducted at the same
time and take into account the measurement errors in the
observed variables when doing so. For this reason, it can also
be regarded as a simultaneous equation system. The three

14
regression equations described above are equations for the
structural model. However, as already mentioned, the
structural equation model consists of two parts, the
measurement model and the structural model. In the example
in Figure 2, there are 15 more regression equations in the
measurement model as seen below:

(1. Equation) Question1= λ 1.X+e1

(2. Equation) Question 2= λ 2.X+e2

(3. Equation) Question 3= λ 3.X+e3

(4. Equation) Question 4= λ4.Y+e4

(5. Equation) Question 5= λ5.Y+e5

(6. Equation) Question 6= λ 6.Y+e6

(7. Equation) Question 7= λ7.Z+e7

(8. Equation) Question 8= λ8.Z+e8

(9. Equation) Question 9= λ9.Z+e9

(10. Equation) Question 10= λ10.W+e10

(11. Equation) Question 11= λ11.W+e11

(12. Equation) Question 12= λ12.W+e12

(13. Equation) Question 13= λ13.T+e13

(14. Equation) Question 14= λ14.T+e14

(15. Equation) Question 15= λ15.T+e15

15
In these above equations, each observed variable is a
dependent variable. Latent variables are independent variables.
This is due to the fact that the observed variables and the
latent variables are connected in a reflective way in the
measurement model.

2.4. Parameters

There are three kinds of parameters in structural equation

models. These are called free, fixed and constrained
parameters. Free parameters are parameters for which no
value is assigned and whose value is to be estimated. Fixed
parameters are parameters with a specific value, such as 0 or 1.
Constrained parameters are estimated parameters depending
on value of other parameters (Raykov & Marcoulides, 2006).
In Figure 2, the estimated total number of parameters in
equations in both the structural model and the measurement
model is 21. Of these, 6 are the regression coefficients
forming the structural model and indicated by the β symbol,
15 are the factor loads of the measurement model and are
indicated by the λ symbol. In order to estimate the parameters
included in the structural equation model given in Figure 2, it
is seen that a total of 21 regression models are working
together at the same time. The assumption of regression,
which is a prerequisite for each regression, needs to be
provided. Therefore, the assumptions that apply to the
regression models apply within the structural equation models.
These are the assumptions of linearity, normal distribution of
error terms (normality), lack of multicollinearity, constant
variance of error terms (homoscedasticity), and no relation
among error terms (authocorrelation). The assumptions about
structural equation modeling will be explained in section 2.6.

16
 2.5. Fit Indices
In the method of structural equation modeling, the measures
that assess the compliance of the models with the data are
called fit indices or fit statistics. There are many fit indices in
the literature. Below there are definitions of the most
commonly used of these fit indices. The size of the sample
should be considered in the analyzes to be done by the
structural equation modeling. Because many of the fit indices
are affected by sample size. The minimum sample size that
must be used in the structural equation modeling method is at
least 10 times the number of parameters that can be estimated
in the model (Jayaram, Kannan, & Tan, 2004). In addition,
the minimum sample size for structural equation modeling is
suggested as 150 (Bentler & Chou, 1987). Some researchers
suggest that the sample size for Structural Equation Models
should be 200-500, at least 200 (Çelik & Yılmaz, 2013).

CMIN is the likelihood ratio chi-square test. This test shows

the correspondence between the proposed model and the
actual model and it is most commonly used fit indice. As a
result of this test, it is evaluated whether the covariance matrix
of the sample with which the model is tested is equal to the
population covariance matrix. Furthermore, since this test is a
difference test, it is not desirable that chi-square value is
significant. The fact that the CMIN / DF ratio is less than 3
and the chi-square value is insignificant indicates that the
model's overall fit is within acceptable limits (Meydan & Şen,
2011). The DF, ie the degree of freedom, is calculated from
the number of observations in a model and number of the
parameters needed estimation. Assumed that the number of
observed variable in a models equals p. In this model, the
estimated number of parameters can not be more than

17
p(p+1)/2. In this case, the degree of freedom is found as
follows (Raykov & Marcoulides, 2006):

DF= p(p+1)/2-(number of parameters in the model)

Models with zero degree of freedom in the structural equation

model are called saturated models. Saturated model shows
perfect fit with data. Negative degree of freedom indicates that
the model can not be defined. The model can be defined if the
degree of freedom is not negative, ie zero or positive.

Table 2. Good Fit Values

Fit Indices Goodness of Fit Values

CMIN/DF 0 <CMIN/DF< 2

CFI 0,97 < CFI < 1

AGFI 0,90 < AGFI < 1

GFI 0,95 < GFI < 1
NFI 0,95 <NFI< 1

RMSEA 0 < RMSEA < 0,05

Source: Bayram, N. (2013). Yapısal Eşitlik Modellemesine Giriş. Bursa: Ezgi
Kitapevi.

CFI (Comparative Fit Index) is a fit indice that compares the

saturated model with the independent model. In the
independent model, there is no relationship among the

18
dimensions that form the research model. CFI values can
range from 0 to 1, values above 0,90 and close to 1 show good
fit (Schermelleh-Engel, Moosbrugger, & Müller, 2003). CFI is
in the group of fit indices based on independent models.
Table 2 summarizes the goodness of fit values according to
the literature. Table 3 summarizes acceptable fit values
according to the literature. The AGFI fit indice is calculated
using the degree of freedom. It is affected by sample size.
When the sample size increases, the value of the AGFI indice
also increases. AGFI takes a value between 0 and 1. Values
over 0.90 indicate that the fit is good (Bayram, 2013).

Table 3. Acceptable Fit Values

Fit Indices Goodness of Fit Values

CMIN/DF 2 <CMIN/DF< 3

CFI 0,95 < CFI < 0,97

AGFI 0,85 < AGFI < 0,90

GFI 0,90 < GFI < 0,95

NFI 0,90 <NFI< 0,95
RMSEA 0,05 < RMSEA < 0,08
Source: Bayram, N. (2013). Yapısal Eşitlik Modellemesine Giriş. Bursa: Ezgi
Kitapevi.

The GFI fit indice is a measure of the degree of variance and

covariance that is explained by the model. The value of the
GFI fit indice rises as the sample size increases. This feature
can prevent accurate results when sample size is low. The GFI

19
value ranges from 0 to 1. Values above 0.90 are considered
acceptable model indices. Values above 0.90 indicate that
covariance is calculated among the observed variables. GFI
and AGFI fit indices are based on the residuals (Bayram,
2013).

RMSEA is a measure of fit that compares the mean

differences of each expected degree of freedom that can occur
in the population with each other. This scale is adversely
affected by sample size. A value of 0.05 or less for the
RMSEA fit indice indicates good fit (Bayram, 2013). Values
between 0.05 and 0.08 indicate acceptable fit (Byrne, 2010).

NFI (Normed Fit Index) takes values between 0 and 1. Higher

values show better fit. Values greater than 0.90 are acceptable,
while values greater than 0.95 are good fit. It is in the group of
the fit indices based on independent model.

In the "output model fit" section of the AMOS program, the

model's fit indices are displayed as follows. In the following
example, the values of CMIN/DF and RMSEA are
acceptable. However other fit indices show a problem. In this
case, indices may come up to normal values after model
modifications. Fit indice values should be read from the
"default model" line. Default model refers to the model being
tested. On the bottom line there is the "saturated model".
Saturated model is the mode where the degree of freedom is
zero and the data is perfectly matched to the model. For this
reason, the indice values of the best model are in this line. On
the bottom line there is independence model. It's the worst

20
possible model. This line contains the worst possible indice
values. The CMIN (chi-square likelihood ratio) value appears
to be significant when the P value is 0,000 in the example
below. However, this test is required to be insignificant since it
is a difference test. However, in most cases this value is
significant. This may be due to the neglect of some of the
assumptions of structural equation modeling described in the
next section as general practice during analyzes.

Model NPAR CMIN DF P CMIN/DF

Default
111 1529,801 555 ,000 2,756
model

Saturated
666 ,000 0
model

Independence
36 9511,261 630 ,000 15,097
model

Model RMR GFI AGFI PGFI

Default model ,045 ,827 ,793 ,689

Saturated model ,000 1,000

Independence model ,248 ,211 ,166 ,199

21
 NFI RFI IFI TLI
Model CFI
Delta1 rho1 Delta2 rho2

Default model ,839 ,817 ,891 ,875 ,890

Saturated model 1,000 1,000 1,000

Independence model ,000 ,000 ,000 ,000 ,000

Model RMSEA LO 90 HI 90 PCLOSE

Default model ,066 ,062 ,070 ,000

Independence model ,186 ,183 ,190 ,000

In addition to the fit indices described above, there are also

indices used for model comparison. These are called model
comparison adaptation indices. AIC (Akaike Information
Criterion) is one of them. In the compared models the one
which has lowest AIC value is considered as the closest model
to reality (Karagöz, 2016).

2.6. Assumptions of Structural Equation Modeling

Similar to regression analysis, structural equation modeling has
its assumptions. But in structural equality models, many
regression equations work together, whether in the structural
model part or in the measurement model part. Therefore, the
assumptions that apply to the regression models are valid for
the structural equation models. As these assumptions are
known, linearity, that is, the relationship between dependent

22
and independent variables is linear, normal distribution of
error terms (normality), no multicollinearity which means
independent variables are not related to each other, the
variance of error terms is fixed (homoscedasticity) or in other
words there is no relationship between independent variables
and error terms ve no authocorrelation that means that there
is no relationship between error terms (Wooldridge, 2003). If
these assumptions are met, it should be considered whether
the assumptions required for the structural equation models
are also met. These assumptions can be summarized as
follows (Bayram, 2013).

x Observed variables have multivariate normality:

The multivariate normal distribution is the most important
assumption of the maximum likelihood estimation method
used in structural equation modeling. This rule is often
violated when ordinal and discrete scales are used. Neglecting
the assumption of multivariate normal distribution of
observed variables leads to a high CMIN / DF value and a
significant test outcome. In case of violation of this
assumption, it is recommended to use the estimation methods
such as weighted least squares (WLS) instead of the maximum
likelihood estimation method. This method can be used if the
data is continuous but does not meet the normal distribution
requirement. Other prediction methods that may be preferred
in this case are ADF (asymptotically distribution free), MLM
(Robust Maximum Likelihood) and GLS (generalized least
squares) (Tabachnick & Fidell, 2001). As the complexity level
of the model tested in the structural equation modeling
method increases, the number of sample observations must
also be increased. However, as the distribution of the data

23
becomes farther away from the normal distribution, it is
necessary to increase the number of data (Kline, 2011).

The skewness and kurtosis values are examined to determine

whether the variables in the data set are normally distributed.
These values are calculated on the basis of moments. In
general, the packaged softwares calculate these values to be 0
as base value. In this case values between -2 and +2 are
considered normal. In addition, Kolmogorov-Smirnov and
Shapiro-Wilk tests can be conducted to test whether the data
set is normally distributed (Sarstedt & Mooi, 2014).

In cases where the data set does not fit the normal
distribution, the outliers (extreme values) should be cleared
first. In AMOS program in analysis properties window in
output tab, normality and outliers can be tested. It is sufficient
to mark the “test for normality” and “outliers” options so that
these test values can be obtained in a tabular form.
Additionally in SPSS, outliers can be determined by examining
the Mahalonobis distance value. If the dataset is not normally
distributed, what can be done is covered in the last section of
this book.

x Latent variables have multivariate normal

distribution:
It refers to the endogeneous latent variables have normal
distribution. In practice, it is a violated assuption.

24
 x Linearity:
As stated at the beginning of the book, structural equation
modeling is a component of factor and regression analysis.
Therefore, linearity, which is the most important assumption
of regression analysis, also applies to structural equation
modeling. In the structural equation model, it is assumed that
there are linear relationships between latent variables and also
between observed and latent variables.

x Absence of outliers:
The outlier affects the significance of the existence model
negatively.

x Multiple measurements:
In the structural equation model, three or more observed
variables must be used to measure each latent variable.

x No multicollinearity:
It is assumed that there is no relation between the
independent variables in the structural equation model.

x Sample size:
In the structural equation modeling, many of the fit indices are
influenced by sample size. In some sources, a minimum

25
sample size of 150 is recommended for structural equation
models (Bentler & Chou, 1987). The minimum sample size
that should be used in the structural equation modeling
method is at least 10 times the number of parameters that can
be estimated in the model. (Jayaram, Kannan, & Tan, 2004).
According to some researchers, the sample size required for
structural equation modeling should be at least 200 and 200-
500 (Çelik & Yılmaz, 2013).

x No correlation between error terms:

It is assumed that there is no correlation between error terms
in the structural equation modeling method. However, if it is
explicitly stated by the researcher in the conceptual model, a
correlation can be made between the error terms (Doğan,
2015).

2.7. Types of Structural Equation Models

There are four basic types of structural equation models.
These are explained below:

2.7.1. Path Analysis Models

In the method of structural equation modeling, the models
established with only observed variables are called path
analysis models. The basis of the structural equation modeling
depends upon path analysis. Path analyzes first started to be
implemented in the 1920s. Developed by biologist Sewall
Wright (Taşkın & Akat, 2010). The path analysis is similar to
multiple regression as it is done with observed variables. But it
is superior than multiple regression. Because there is one

26
dependent variable in the multiple regression. However, there
may be more than one dependent variable in the path analysis,
and a variable can be both a dependent variable and an
independent variable. In path analyzes, more than one
regression model can be analyzed at the same time, and
indirect and direct effects can be measured at the same time.
Direct effect is the effect of one variable on another variable
without any mediation. However, the indirect effect arises
from the intervention of a variable which is playing mediator
role between independent and dependent variables. This
variable is named as the mediator variable. The sum of the
direct effect and the indirect effect of a variable on another
variable is called the total effect (Raykov & Marcoulides,
2006). Since path analyzes do not contain latent variables, they
can not be saved from measurement errors (Meydan & Şen,
2011). For this reason, structural regression models generated
by latent variables give more accurate results. Because
structural regression models include measurement model. Path
analysis models and examples will be discussed in more detail
in chapter 5.

2.7.2. Confirmatory Factor Analysis Models

Factor analysis is divided into two types as exploratory and
confirmatory. In explanatory factor analysis, factors are
revealed from relations among variables. In explanatory factor
analysis, the observed variables can be loaded on any factor or
on multiple factors. However, in the confirmatory factor
analysis, the theoretically predetermined factor structure is
confirmed by the current data. In other words, in the
confirmatory factor analysis, which factor will be loaded on an
observed variable is predetermined. By means of the
explanatory factor analysis, the latent variables are revealed

27
from the observed variables. However, in the confirmatory
factor analysis, previously discovered scales are confirmed
again with the collected data. In Section 3, examples of first-
level one-factor, first-level multi-factor and second-level multi-
factor confirmatory factor analysis are included.

2.7.3. Structural Regression Models

It is regression models formed between latent variables in

structural equation models. It consists of a combination of
measurement model and structural model. Incorporating the
measurement model and the structural model allows the
inclusion of measurement errors so that more accurate results
can be obtained. In other words, confirmatory factor analysis
and multiple regression analysis coexist. Structural regression
models will be discussed in more detail in Section 5.

2.7.4. Latent Change Models

They are also named as “latent change models”, “latent
growth curve models” or “latent curve analysis”. Models that
describe longitudinal variation in time series (Raykov &
Marcoulides, 2006). These models are the models used to
explain the growth and decay of an event over time,
similarities or differences within and between units. (Doğan,
2015). In Figure 4, two factorial growth models are observed
for two time points (T1, T2). Structural equation modeling is a
very useful method for analyzing changes in time. Repeated
measurements over time are needed to use the latent change
models. Such data are called longitudinal data (vertical cross-
section data). In AMOS program, under the "Plugins" menu,
latent change models can be drawn from the "Growth Curve
Model" window. According to Baltes and Nesselroade, this

28
model can be used for the following purposes. (Baltes &
Nesselroade, 1979):

Figure 4. Latent Change Model Example

Source: Raykov, T., & Marcoulides, G. (2006). A First Course in Structural Equation
Modeling. Mahwah: Lawrence Erlbaum Associates.

(1) Describe observed and unobserved vertical section data.

29
(2) Characterize the development of individuals and groups.
(3) To predict individual and group differences in
developmental forms.
(4) To examine the dynamic determinants among variables in
time.
(5) To reveal the group differences of the dynamic
determinants between variables in time.

30
3. VALIDITY AND RELIABILITY ANALYSIS
Scale is the method used to find the numerical values of the
dimensions that constitute a concept. Since concepts can not
be directly measured in social sciences, questionnaires are
formed to define these concepts. Reliability means that a scale
is always measure the same value under the same conditions
consistently. For example, a questionnaire form is reliable if
the same group is given the same result when applied two
different times. So if we ask the same questions about the
same people, if the conditions are not changed, they are
expected to give the same answers. Otherwise, this means that
the persons in the sample either they did not understand the
questions on the questionnaire or they did not read them.
Validity is a measure of what we really want to measure. For
example, if a questionnaire actually measures a different
concept than the dimension we want to measure, it is not
valid. If the questions we ask about the concept A are
confused with the questions about the concept B, then it
means that the concepts we consider to measure are not
perceived or perceived as different from those in the sample.
In this case, the scale we use is not a valid measurement tool
for this sample. For this reason, it is necessary to test the
validity and reliability of the scale before any analysis is started.
As a result of these tests, verification of unidimensionality is
generally provided. Unidimensionality means that the
observed variables used to measure each dimension must
measure only one dimension (Avcılar & Varinli, 2013).
Construct validity and reliability must be determined in order
to confirm unidimensionality. In a theoretically determined
model, construct validity refers to convergence of observed

31
variables that are connected to the same latent variable
(convergent validity) and dissociation of observed variables
from other observed variables that are connected to other
latent variables (discriminant validity). The construct validity
indicates that the observed variables do not measure any latent
variable other than they connected in the conceptual model.
But in this case it would not be correct to say that the validity
of the construct is fully realized without confirming the
reliability of the scale (Gerbing & Anderson, 1988).

3.1. Determination of Convergent Validity

Convergent validity indicates that the correlations between
questions constituting a construct are high. In structural
equation modeling method, it is necessary to look at the
results of confirmatory factor analysis to determine the
convergent validity of the scales used to measure the
dimensions constituting the conceptual model of the research.
The measurement model part of structural equation models
correspond to confirmatory factor analysis (Confirmatory
Factor Analysis - CFA). Therefore, if the measurement model
fit indices are low, there is no need to test the structural model
(See also Figure 1. Demarcation between Measurement Model
and Structural Model). Because the scales used to measure the
dimensions that make up the conceptual model will not be
validated. Therefore, if the measurement model is insufficient,
the fit indices of the structural model will be low. The t test
results of all the coefficients in the measurement model should
indicate that the coefficient values are different from zero.
The standard value of each coefficient in the measurement
model is the factor loadings of the confirmatory factor
analysis. Each factor load should be higher than 0.50.
Otherwise, the fit indices of the general model will be

32
adversely affected. The fact that the factor loads are above 0,5
is evidence of convergent validity. If the critical rate value of a
question in CFA results is greater than 2 as an absolute value
this means that this item is loaded to the factor it is connected.

In structural equation modeling method, while the CFA model

is set, some parameters are freed while some are fixed. As
shown in Figure 5, Parameter of question 1 and parameters of
error terms are fixed by assigning 1 value. Others are fixed.
The freed parameters will be estimated by the program.

Before applying confirmatory factor analysis (CFA), it is first

necessary to look at the results of explanatory factor analysis
(EFA) in practice. Even though scales generally accepted in
the literature are used, to see if the survey fillers correctly
perceive the questions principle component analysis should be
conducted in SPSS before set up CFA model in AMOS. And
how many different dimensions the questions are perceived by
those who solve the questionnaire should be clarified.

At this stage, the necessary questions should be eliminated.

This step is also called the purification stage. Principle
component analysis is a type of analysis that assigns the
variables in the data set into groups so that the relationship
between the variables in the group is maximized. Main
purpose of this analysis is to obtain the least number of
factors to represent the relationship among items at the
highest level.

33
 Figure 5. Single Factor CFA Model

Bartlett test of Sphericity and Kaiser-Meyer-Olkin test were

performed to determine the suitability of the dataset to the
principle component analysis. If the null hypothesis is not
rejected as the result of the Bartlett test of Sphericity, the
analysis is not continued. This test detects whether the
correlation matrix indicating the inter-variable relation is a unit
matrix. There is no relation between variables in case of unit
matrix. The Kaiser-Meyer-Olkin test is affected by sample
size. This test compares the values of the observed correlation
coefficients with the values of the partial correlation
coefficients. In this way, it tests whether the sample size is
sufficient to perform principle component analysis. Values
above 0.7 are considered good. If it is below 0.5, factor
analysis can not be continued (Karagöz, 2016).

Essentially, principle component analysis is done to determine

scale validity, but it also fulfills the purpose of making the data

34
set analyzable. It fulfills the following functions for this (Aksu,
Eser, & Güzeller, 2017):

- To remove the dependency between variables.

- To obtain fewer new variables those are not related to each
other.

Table 4. Results of Confirmatory Factor Analysis

t-Value
Standardized Unstandar (Critical
Conceptual Standard
Items
Variable dized Error Ratio)
Factor Loads
Factor
Loads

Qestion1 0,818 1

Qestion2 X 0,906 1,104 0,049 22,523

Qestion3 0,907 1,111 0,049 22,570

Qestion4 0,825 1

Qestion5 Y 0,732 0,882 0,057 15,549

Qestion6 0,718 0,885 0,058 15,187

Qestion7 0,757 1

Qestion8 Z 0,835 1,102 0,062 17,785

Qestion9 0,939 1,255 0,062 20,176

Qestion10 0,676 1

Qestion11 W 0,799 1,131 0,083 13,555

Qestion12 0,785 1,158 0,087 13,379

Note: For all values P<0.01

35
After conducting explanatory factor analysis and purification
by principle component analysis, the remaining indicators
(questions) are linked under the structures they belong to and
the CFA model is created in the AMOS program as shown in
Figure 5 (The principle component analysis performed in the
SPSS program is not explained because it is out of scope of
this book). The number of factors is freed in EFA. In CFA,
the number of factors and which indicators are connected is
determined in advance.

Table 4 shows the way in which confirmatory factor analysis

results are given. What is important here is that the standard
factor loads of the questions under each conceptual variable
are over 0.50. By looking at this table, questions with a
standard factor load of less than 0.50 are discarded.

In this case factor analysis is done again. Figure 8 shows a

second-order CFA model. It is important to note here that
residual terms are to be placed in the first-level latent variables
(Res1, Res2 and Res3). Figure 9 shows another second-level
multi-factor CFA model. The difference of this model is that
there are more than one factor in the second level. Since there
are more than one exogeneous variables, covariance is placed
between the second level latent variables (between A and B).
The fit indices of the CFA model are then tabulated. Figure 6
shows the analysis features that must be marked in the
Analysis Properties window before running the CFA model in
the AMOS program. Figure 7 shows the first-order multi-
factor CFA model. In this case it must be considered to place
covariance between the latent variables. The rule of placing
covariance between exogeneous variables is also valid here. In

36
Figure 8, covariance is not placed when there is only one
exogeneous variable. However Figure 9 shows covariance
between two exogeneous variables.

Figure 6. Analysis Properties Window for CFA Models in

AMOS program.

37
Figure 7. Multi-Factor First-Order CFA Model

38
Figure 8. Second-Order CFA Model

39
 Figure 9. Multi-Factor Second-Order CFA Model

Another indicator of convergent validity is the Average

Variance Extracted (AVE) value. To be able to confirm the
convergent validity, it must be more than 0.50 or 0.50.
(Fornell & Larcker, 1981).

40
 3.2. Determination of Discriminant Validity
Discriminant validity is the measure of the level at which a
structure in a measurement model differs from other
structures. It is an indicator of a low correlation between the
questions that form a construct and other questions that form
other construct. To find the discriminant validity for each
dimension, we first need to calculate the Average Variance
Extracted (AVE) value for each dimension. The acceptable
AVE value must be greater than 0.50 or 0.50. However, as
noted in the previous section, this value confirms convergent
validity when examined alone (Fornell & Larcker, 1981). In
order to determine discriminant validity, it is also desirable
that the values of the AVE for each construct in the data set
are larger than the correlation coefficients of that construct
with the other constructs. In this case, it can be determined
that the scales used have discriminant validity for each
dimension. AVE value alone does not indicate discriminant
validity but the square root of the AVE value of each
construct is larger than the inter-dimensional correlation value
it can be said that there is discriminant validity (Fornell &
Larcker, 1981). In Table 5, the values shown in parentheses as
crosswise is the square root of the AVE values. For the 9
dimensions in the table, the square root of the AVE values in
each column is higher than the correlation coefficients in that
column. In addition, the AVE values are above the 0.50
threshold. In this case, it can be said that the scales used for
this example have discriminant validity. The AVE value is not
calculated by the AMOS package program. However, it is easy
to find ready-made excel files that provide this value
calculation on the internet.

41
 3.3. Determination of Reliability
After determination of the validity of the scales by means of
CFA reliability analysis must be conducted for each construct.
First of all, Cronbach's α value is calculated for each
dimension separately. Values greater than 0.7 threshold
indicate that the internal reliability of the scale used is
sufficient. Cronbach’s α is a measure based on correlations
between items in a consruct. It is obtained by dividing the sum
of the variances of the items constituting a scale by the general
variance. It takes a value between 0 and 1. Values beyond 0.7
threshold indicate that the scale is reliable. If it is below 0.6,
the reliability of the scale is low. (Karagöz, 2016).

Another value that is used to calculate the reliability of the

scale for each dimension is the composite reliability value. The
composite reliability value is calculated from the factor loads
found in the confirmatory factor analysis. After CR values
beyond 0.7 threshold or equals to 0.7 it can be said that there
is composite reliability (Raykov, 1997).

Table 5 shows a sample table showing Cronbach's α, AVE and

CR values calculated for each construct and the correlation
values between constructs. Cronbach's α value can be
calculated from the scale reliability menu in the SPSS program.
The AVE and CR values are found by placing the results of
the CFA factor loadings in to the formulas. There are ready-
made calculation tools on the Internet.

42
 Table 5. Descriptive Statistics, Correlation Coefficient, Reliability
Results and Discriminant Validity
Std.
Avr. 1 2 3 4 5 6 7 8 9
Dev.

1. Construct 3,25 0,81 (0,842)

2. Construct 3,28 0,71 ,216* (0,711)

3. Construct 3,63 0,72 ,427* ,383* (0,840)

4. Construct 3,72 0,68 ,228* ,533* ,457* (0,718)

5. Construct 3,62 0,70 ,449* ,192* ,378* ,298* (0,769)

6. Construct 3,76 0,68 ,430* ,394* ,551* ,450* ,499* (0,734)

7. Construct 3,23 0,87 ,585* ,166* ,452* ,174* ,449* ,479* (0,800)

8. Construct 3,68 0,67 ,394* ,496* ,672* ,508* ,350* ,508* ,358* (0,722)

9. Construct 3,02 0,77 ,340* ,374* ,353* ,335* ,209* ,302* ,219* ,410* (0,754)

Cronbach Alfa Reliability

0,927 0,861 0,901 0,851 0,781 0,771 0,828 0,808 0,721
Coefficient

Composite Reliability Coefficient

0,924 0,854 0,905 0,841 0,791 0,777 0,840 0,813 0,725
(CR)

Average Variance Extracted (AVE) 0,710 0,506 0,706 0,516 0,592 0,539 0,640 0,522 0,570

* P<0,05, Note: the values written in brackets indicate the square root of the AVE values.

There are statistically significant relationships among the

constructs in the sample in Table 5. Correlation is the
coefficient that indicates the power of linear relationship
between variables. This coefficient must be statistically
significant in order to be able to say that there is a relationship
between variables. The correlation coefficient takes a value
between -1 and +1 (Sipahi, Yurtkoru, & Çinko, 2010).

43
4. STRUCTURAL REGRESSION MODEL
As a result of the processes described above, convergence
validity and discriminant validity are determined. Then the
phase for forming the structural model begins. The structural
model is based on measurement model. In this phase it should
be noted that the structure and order of CFA model is
preserved in the model that will be established with latent
variables. Depending on the causal relationships between
latent variables, the directions of the arrows are determined in
accordance with the developed hypotheses and a structural
model is constructed. Two-way arrows indicate the covariance
between two variables without specifying the direction of
causality. While constructing the structural model, the
conceptual model leads the way. The conceptual model is
mainly based on the relations found in the literature. Finally,
the conceptual model is tested using real data. Before the
hypothesis tests, the fit indices of the model are examined.

If the fit indices of the model are not within the limits
recommended in the literature, the modifications are made
and the fit indices are improved provided that they are
compatible with the literature. If the fit indices are at an
adequate level, the predicted values of the model parameters
are checked first. Then the hypothesis test results of the
research are given in a table. Figure 10 shows the properties
to be marked in the analysis properties menu in the AMOS
program when starting the path analysis. Table 6 shows an
example of the hypothesis test results. The values in this table
are in the estimates section of the output screen of the AMOS

44
program. The notation *** in AMOS output means that P is
equal to zero.

Table 6. Hypothesis Test Results Table Example

Unstandar
Standard d
Relations
Coefficients Coefficien
ts

X→Z 0.533* 0.594*

Y→Z 0.437* 0.638*

Z→W 0.493* 0.377*

*p < 0.05

Figure 11 and Figure 12 show examples of structural models

based on the first-level multi-factor DFA model given in
Figure 7.

45
Figure 10. Analysis Properties Window for the Structural
Model in the AMOS program.

These examples are called structural regression models.

Although the models in Figure 11 and Figure 12 are based on
the same measurement model, the path analysis created is
different. In Figure 11, there are more than one exogenous
variable and therefore covariance is placed between them. In
Figure 12 there is only one exogeneous variable. In both
models residual terms are linked to the endogeneous variables.

46
Careful attention should be paid to these rules when
constructing structural models. Otherwise the model will not
work.

Figure 11. Structural Regression Model Example

Figure 12. Structural Regression Model Example

47
5. PATH ANALYSIS
Structural model can be established by directly observed
variables in cases factor analysis is done before and the
average of the questions that make up the constructs or when
working with the secondary data. Such models are called as
path analysis. Figures 13 and 14 show the models constructed
in Figures 11 and 12, respectively, with the observed variables.
As seen in the figures, covariance and residuals remained the
same in the models. This is because the rules are the same in
the path analyzes made with the observed variables. But the
measurement models are gone. Therefore, it is assumed that
there are no measurement errors in the path analysis.
Predictive error terms are included in dependent variables.
Before running the path analysis in the AMOS program, the
options in Figure 10 should be checked in the analysis
properties window.

Figure 13. Example of Path Analysis

48
Figure 14. Example of Path Analysis

49
6. MODEL MODIFICATION
If the fit indices of the structural model do not come to an
adequate level, the AMOS program suggests modifications to
the user. These modifications improve fit indices. For
example, Figure 15 shows the modifications suggested under
the Modification Indices menu in the AMOS program. In the
case of placing covariance among those with the highest value
among the proposed changes in Figure 13, the model's fit
indices will improve. Because the modification index value
corresponds to the decrease in the chi-square value of the
model. However, these covariances are not to be considered
as unrelated concepts in the literature that should be taken
into account when setting. Because each modification changes
the conceptual model which is first introduced. It should be
noted that the changes made by this reason do not contradict
the purpose of the research and the relations in the literature.
To reach the modification indices, the modification indices
box in the analysis properties window in the AMOS program
must be marked as shown in Figure 10. In the example shown
in Figure 15, first the model is run again by adding covariance
between the error terms e24 and e22 and compared with the
previous situation. Because the highest modification value in
the table is 17.457, which is between e24 and e22. After the
modifications are made, the model is retested, and if sufficient
compliance values can not be obtained, the proposed
modifications can be repeated. Again, modifications made
should be consistent with the literature. Unrepeatable and
inappropriate modifications made are scientifically
unacceptable and result in inconsistent situations with the
population. For this reason, it is necessary to apply it very

50
carefully. Just in order to raise the fit indices, covariance
should not be placed between concepts that are not related to
each other.

Figure 15. AMOS Modification Indices

51
7. MEDIATOR VARIABLE ANALYSIS
METHODOLOGY
The variable starting the causality relation between the
independent and the dependent variable is called as mediator
variable (Wu & Zumbo , 2008). It is also called as intervening
variable (MacKinnon, Lockwood, Hoffman, West, & Sheets,
2002). It can also be defined as the variable that transfers the
indirect effect of an independent variable to the dependent
variable. Analysis of mediator variable is based on the
hierarchical regression method introduced by Baron and
Kenny in 1986. In order to apply this method, the following
conditions must first be met (Baron & Kenny, 1986) :

a. The changes that occur in the independent variable cause a

change in the mediator variable,
b. Changes in the mediator variable cause changes in the
dependent,
c. If the mediator and independent variables are together
included in the regression analysis, the effect of the
independent variable on the dependent variable either falls or
completely ceases.

In Figure 16, there is a sample mediator variable model. After

this model is created, it is first checked whether there is a
correlation between all variables. It is thus tested whether the

52
model meets the first two preconditions put forward by Baron
and Kenny. Table 7 shows an example of correlation table.

Figure 16. Mediator Variable Model and Hypotheses

Table 7. Example of Correlation Coefficient Table

Constructs A B C

A - - -

B 0,492* - -

C 0,575* 0,672* -
* P< 0.01

53
When an mediator analysis is performed, three different
models are run and the coefficients of the models are
compared with one another. Models and hypotheses tested are
as follows:

H1: Variable A affects variable B in the positive direction.

H2: Variable B affects variable C in the positive direction.

H3: Variable A affects variable C in the positive direction.

H4: Variable B plays mediator role in the relationship
between Variable A and Variable C.

Model 1: ‫ܥ‬௜ ൌ  ߚ଴ ൅ ߚଵ ‫ܣ‬௜ ൅ ߝ௜ (H3)

Model 2: ‫ܤ‬௜ ൌ  ߚ଴ ൅ ߚଵ ‫ܣ‬௜ ൅ ߝ௜ (H1)

Model 3: ‫ܥ‬௜ ൌ  ߚ଴ ൅ ߚଵ ‫ܣ‬௜ ൅ ߚଶ ‫ܤ‬௜ ൅ ߝ௜ (H2 ve H4)

Once the three regression models given above are run

separately, the results found are compared as seen in Table 8.
As shown in Table 8, when the B variable is added in model 3,
the coefficient of the relationship between A and C is lowered
and turns to be insignificant. This indicates that the variable B
has mediator role in the relationship between variable A and
variable C. As a result, hypotheses H1, H2, H3 and H4 are
accepted. In this way, the Baron and Kenny method can be
easily applied in the SPSS program in the presence of a third
variable that plays a role of mediator variable between the two
variables.

54
 Table 8. Example of Regression Coefficients Table
Relations Model 1 Model 2 Model 3

A→C 0,492* - 0,07

A→B - 0,575* -

B→C - - 0,582*

R2 0,242 0,331 0,468

Adjusted R2 0,240 0,329 0,466

F 129,216* 200,205* 177,925*

* P<0.001

Structural equation modeling can be used in mediator variable

analysis when testing multiple mediator variables or multiple
independent or dependent variables. For example, in the
conceptual model shown in Figure 17, there are two mediator
variables.

In the same way, The Baron and Kenny method is applied

while the analysis of the mediator variables is done in the
structural equality models. For this reason, once the model is
constructed, first, it is checked whether there is a correlation
among all variables. It is thus tested whether the model meets
the first two preconditions put forward by Baron and Kenny.
Table 9 shows an example of correlation table. When an
mediator analysis is performed, the three different models
shown in Figure 18, Figure 19, and Figure 20 below are run in
AMOS and the coefficients of the models are compared
against each other. Table 10 shows an example of a table

55
comparing the β coefficients found after running models in
AMOS.

As shown in Table 10, the coefficient β for the relationship

between A and D, which is significant and high in Model 1,
has fallen and become insignificant with the inclusion of the
variables B and C in Model 3. In this case, the mediator role of
B and C are found statistically significant. In addition, fit
indices of each model tested should be given. Therefore Table
10 shows the fit indices of each model.

Figure 17. Two Mediator Conceptual Model and Its

Hypotheses

56
For scientific validation of mediator variable roles, these fit
indices must also be within acceptable limits. This method can
be applied to more variables. The basic logic of this analysis is
the comparison of the coefficients of the model to which the
mediator variables are not included and the coefficients of the
next model to which the mediator variables are included.

Table 9. Correlation Coefficients Sample Table

Variables 1 2 3 4
 A –
 B 0.883* –
 C 0.861* 0.928* –
 D 0.430* 0.653* 0.703* –
*p < 0.01.

Figure 18. Model 1

57
If the statistically significant coefficients become insignificant,
the role of intermediate variable is confirmed. If the p value of
the coefficient remains significant but there is a serious
decrease in the coefficient, it can be said that there is semi-
mediator effect. In the conceptual model in Figure 17, the
mediator roles of B and C can be tested together. But if
needed, these variables can be included to analyses one by
one.

Figure 19. Model 2

58
 Tablo 10. Analysis Results Sample Table

Relations Model 1 Model 2 Model 3

A → D (H1) 0.414* 0.035

A → B (H2) 0.863* 0.865*

A → C (H3) 0.279* 0.870*

C → D (H4) 0.532*

B → D (H5) 0.935*
χ2/df=2.554 χ2/df=2.921 χ2/df=2.856
CFI=0.947
Model Fit CFI=0.985 CFI=0.954
Indices IFI=0.985 IFI=0.947
RMSEA=0.14 IFI=0.954 RMSEA=0.14
RMSEA=0.14

Note: Path analysis coefficients are standardized.

*p<0.01

59
Figure 20. Model 3

60
 II. CHAPTER
APPLICATIONS

8. USE OF AMOS PROGRAM

This section provides general information about the menus
and use of the AMOS program. The AMOS program runs on
the Windows operating system. It is a graph based program.
In this way, users can easily analyze by drawing conceptual
models they create. The results of the analysis are given both
on the drawn model and also on the tables.

As the AMOS program carries copy and paste features of the

Windows operating system, the results can easily be
transferred to programs such as Word and Excel. To open the
AMOS program in Windows 10 operating system, the Start →
IBM SPSS Statistics → AMOS Graphics menus must be
followed. After the AMOS program is turned on, the screen
shown in Figure 21 is displayed. There are various icons on
the left side of this screen. The functions of these icons are
summarized in Figure 22. When you wait for a while on these
icons with the mouse, the function of the icon is displayed on
the screen. At the top of the screen there are File, Edit, View,
Diagram, Analyze, Tools, Plugins and Help menus.

61
The AMOS program is a program that works together with
the SPSS program. For this reason, it is first necessary to
contact the SPSS file containing the data to be worked on
before starting the model drawing. For this, File → Data Files
menus are followed and the window that appears in Figure 23
opens. In this screen, click the File Name menu to select the
relevant file and click OK to close the window. In this way,
the AMOS file is linked to the data to be run.

When starting to draw the model after this step, firstly the
indicator symbol is pressed to draw the hidden variable
and the three observed variables (indicator) which are shown
in Figure 24.

Figure 21. AMOS Startup Screen

62
Each click creates an indicator. Therefore, it is necessary to
click three times to draw three indicators. Then change
direction of the indicators by clicking on the direction
change icon. One more latent variable is drawn by following
the same sequence and the model shown in Figure 25 is
obtained. In this model, relations between two latent variables
are drawn by using the regression symbol and covariance
is drawn between two external variables using the covariance
symbol .

63
Figure 22. Drawing Tools in the AMOS Program

64
Figure 22. Continued

65
 Figure 23. Data Files Window

Figure 24. Drawing Latent Variable

66
 Figure 25. Sample Model Drawing

The data set icon is pressed and the observed variables in

the data set are listed. The variables found in this list can be
linked to the indicators by means of drag and drop method.
Select each latent variable and click on the right mouse button
and select the object property from the pop-up menu and
write the names of the F1, F2 and F3 factors appearing in
Figure 25.

After that, by pressing error symbol to add residual to

endogeneous variable. Finally, the plugins → Name
Unobserved Variables menus are followed and the variable
name is assigned to the error term of each indicator and the
residue term of the endogeneous variable. Therefore, in the
model shown in Figure 25, error names from e1 to e10 are
automatically assigned by the program. Right click on the
variables on the screen, and when the object properties is
selected in the drop-down menu, the window shown in Figure

67
26 opens. In this window in text tab, variable names can be
assigned.

In Figure 25, F1, F2, F3 and R1 variable names are assigned in

this way. In this tab, the font size can be adjusted. In this
window in the parameters tab, it is possible to assign the
values of parameters which is desired to be kept constant in
the model. F1 and F2 are exogeneous variables. Due to the
fact that between them covariance is inserted by using the
symbol . F3 is endogeneous variable. Therefore residual
term (R1) is added by using the symbol . After the model is
drawn, the Analysis Properties menu opens in order to select
the analysis that you want to perform.

Click on the icon to open this menu. When this icon is

clicked, the window shown in Figure 27 opens. The options
marked on the Output tab in Figure 27 must be noted. On the
Estimation tab, the Maximum likelihood option is usually
selected. After the selections are made, the program is run by
clicking on the icon. The special cases in which other
estimation methods can be used are mentioned in previous
chapters.

The items to be marked on the Output tab vary according to

the properties of the analysis to be performed. For example,
when performing confirmatory factor analysis, the factor score
weights option should be marked. After running the analyzes,

68
click the icon to view the results on the model. Click on
the view text icon to view the results in tabular form.

Figure 26. Object Properties Window Text Tab

When this icon is clicked, the screen shown in Figure 28

opens. For this window to be opened separately, the results
can be viewed both in text format and on the model at the
same time as shown in Figure 29. If you click on the Model Fit
tab in the menu on the left side of this screen, the values of
the fit indices of the model are reached. Regression weights
are reached from the Estimate tab. Modification suggestions
can be obtained from the modification indices tab.

69
Figure 27. Analysis Properties Window Estimation Tab

70
Figure 27. Analysis Properties Window Output Tab
(Continued)

71
 Figure 28. Text View Window

Figure 29. Showing Results on the Model

72
9. SAMPLE OF MEDIATOR VARIABLE
ANALYSIS

In Section 7, we focused on the methodology of mediator

variable analysis. Mediator analysis methodology is based on
the hierarchical regression method introduced by Baron and
Kenny in 1986 (Baron & Kenny, 1986). The structural
equation modeling method allows analysis of direct and
indirect relations together, so the application of the Baron and
Kenny method in the structural equation model provides an
advantage especially when there are more than one mediator
variable in the conceptual model.

In Section 7, how the mediator analysis methodology is

applied is described both by regression analysis method and by
structural equation modeling method. In the case of mediator
variable analysis in the structural equation modeling, a real
research example is summarized below in order that the reader
can better understand the way the tables are given (Civelek,
İnce, & Karabulut, 2016). The literature section of the
research is not given and only the parts enough to understand
the method are given.

9.1. Title of the Research

Mediator role of attitude towards site in the relationship

between system quality and net benefit.

73
 9.2. Purpose
The aim of this research is to determine the role of users'
attitudes towards the site and satisfaction levels in system
quality which was found to have a positive impact on net
benefit in previous research.

9.3. Conceptual Model and Scales

While constructing the conceptual model, the models
developed in previous researches measuring the success of
information systems were used.

Figure 30. Conceptual Model

The scale developed by Wu and Wang in 2006 was used to

measure the net benefit dimension (Wu & Wang, 2006). Scales
developed by Chen et al. In 2013 were used to measure other

74
dimensions (Chen, Rungruengsamrit, Rajkumar, & Yen, 2013).
The scales were measured according to the five-point Likert
scale. The conceptual model of the research is shown in
Figure 30.

9.4. Determination of Validity and Reliability

Confirmatory factor analysis (CFA) was performed to
determine the construct validity of the scales used. The fit
indice values of the CFA model were found satisfactory:
χ2/DF =3.790, CFI=0.920, IFI=0.921, RMSEA= 0.083.
Table 10 shows the standard factor loadings of the
confirmatory factor model. The standard factor loads are
above 0,50 and the fit indices are close to the threshold values.
Therefore the convergent validity of the scales is determined.

The AVE (Avarage Variance Extracted) values given in Table

12 are above the 0.50 threshold and the square root of the
AVE values is greater than the correlation values in that
column for each dimension. Therefore the discriminant
validity of the scales is determined.

In addition, Cronbach α and Composite Reliability values are

above the threshold value of 0.70, indicating the reliability of
the scales used.

75
 Table 11. Confirmatory Factor Analysis Results
Standardized Unstandardize
Constructs Items
Factor Loads d Factor Loads

Use1 0.760 1.00

Usability Use2 0.786 1.10

Use3 0.647 0.80

Sa1 0.912 1.00

System Quality Sa2 0.807 0.91

Sa3 0.663 0.62

Ats1 0.717 1.00

Attitude Ats2 0.747 1.13

Toward the
Site Ats3 0.847 1.11

Ats4 0.763 1.10

Us1 0.761 1.00

User Us2 0.831 1.09

Satisfaction Us2 0.941 1.25

Us4 0.818 1.06

Nf1 0.772 1.00

Nf2 0.630 0.99

Net Benefit
Nf3 0.818 1.15

Nf4 0.646 0.95

Note: For all p<0.01

76
 9.5. Analysis Results
Three separate models have been analyzed as described in
Chapter 7. The analysis results of three different models are
compared in Table 13. In Figure 31, the results of the analysis
of model 3 are given as an image.

Table 12. Correlation, AVE and Reliability Values

Variables 1 2 3 4

(0.767)
1. System Quality

2. User Satisfaction 0.575* (0.840)

3. Attitude Toward Site

0.566* 0.677* (0.770)

4. Net Benefit 0.492* 0.672* 0.709* (0.721)

Cronbach α 0.821 0.901 0.848 0.808

Composite Reliability 0.895 0.905 0.853 0.810
(CR)
Avarage Variance
Extracted (AVE) 0.589 0.706 0.593 0.520

*p < 0.01

Note: The values in brackets indicate the square root of the AVE
values.

77
 Table 13. Test Results
Relations Model 1 Model 2 Model 3

System Quality
0.701* -0.03**
→Net Benefit

System Quality →
0.733* 0.839*
User Satisfaction

System Quality →
0.487* 0.880*
Attitude Toward Site

User Satisfaction
0.331*
→Net Benefit

Attitude Toward Site

0.651*
→ Net Benefit

χ2/df=4.039 χ2/df=3.678 χ2/df=3.750

CFI=0.944
CFI=0.974 CFI=0.967
Model Fit Indices
IFI=0.974,RMS IFI=0.944
EA=0.08 IFI=0.967 RMSEA=0.08
RMSEA=0.08

Note: Regression coefficients are standard values.

*p<0.01
**Insignificant

As shown in Table 13, in the first model, the regression

coefficient of the relationship between system quality and net
benefit is statistically significant and quite high. However, in
the third model, when the relationship includes user
satisfaction and attitude toward site dimension, the coefficient
of the relationship decreases and it turns out to be statistically
insignificant.

78
 Note: χ2/DF = 3.750, CFI = 0.944, IFI = 0.944, RMSEA=
0.082

Figure 31. Model 3

9.6. Conclusion

As a result, this research shows that the quality of the system

has no direct effect on the net benefit perceptions of the
users. The system quality indirectly affects the net benefit
perception. In other words, the relation between the system
quality and the net benefit perception user satisfaction and
attitudes of the users towards the site play mediator role. This
means: Increasing the quality of the system primarily increases
the satisfaction of users and improves their attitude towards
the site. In the end, user satisfaction and attitude, which
turned into positive, increase the net benefit perception.

79
10. MULTITRAIT-MULTIMETHOD MODEL
In some cases, nested structures can be found. In other words,
some scale questions can be designed to measure nested
concepts at the same time. As an example, assume that a scale
to measure teachers' competence in a school has been
developed. On this scale, there are questions that measure the
social, academic and English competence of the teachers. The
questions were asked to the same number of groups of
students and colleagues. In this example, there is a structure
consisting of three features and two separate methods as
shown in Figure 32, the best method that can be used to
determine the construct validity of such models is Multitrait-
Multimethod Model. As shown in Figure 32, there are two
separate questions asked to two different groups. In the model
shown in Figure 32, social, academic and English competences
are the traits and students and peers are the methods.

In the model there are as many questions as the multiplication

of the numbers of traits and methods. Multitrait-Multimethod
Model was first proposed by Campbell and Fiske in a paper
published in 1959.

Although different alternatives have been proposed over

time models based on covariance structure became important
(Byrne, 2010).

80
 Figure 32. Multitrait-Multimethod Model
Source: Byrne, B. M. (2010). Structural Equation Modeling with AMOS. New York:
Routledge Taylor & Francis Group.

Construct validity, according to Campbell and Fiske, focused

on the determination of the convergent validity, which
expresses the correlation of the components that make up, the
discriminant validity, which expresses low degree of
correlation with the components constituting other constructs
and method effect, which is an extension of discriminant
validity (Campbell & Fiske, 1959). The method effect can be
defined as the bias that results from using the same method to
evaluate different properties (Byrne, 2010).

81
In Figure 33, there is a larger example that will provide a
better understanding of Multitrait-Multimethod Model. In the
figure, there is an assumption model of this example. In
Figure 33, there are 7 traits as X, Y, V, W, Z, T, U and 5
methods as A, B, C, D, E. Therefore, a scale consisting of 35
questions was used. The notation of “,1” on the latent
variables forming traits and methods indicates that the factor
variance is fixed at 1 (The another notation “1,” indicates that
the factor averages are fixed at 1). This is a symbolic notation,
and comma notation does not appear in the AMOS program,
it looks as seen in Figure 35. To fix the variance of a hidden
variable in the AMOS program, enter the value in the variance
box under the "Parameter" tab in the "Object Properties"
dialog box which opened by right mouse button on selected
variable. When we look at the parameter summaries in Table
14, it is seen that the variances of 12 variables are kept
constant.

It is also seen that the variance of 35 error terms obtained as a

result of estimation is released. 70 regression coefficients were
estimated and 35 regression coefficients belongs to the error
terms were fixed. Therefore, there are a total of 105 regression
weights. In this case, when the fixed regression weights and
latent variable variances are evaluated together, it is seen that
47 parameters are kept constant. There are totally 183
parameters in the whole model.

82
Figure 33. Hypothesized Multitrait-Multimethod Model

83
 Table 14. Parameter Summary
Weights Covariances Variances Total

Fixed 35 0 12 47

Labeled 0 0 0 0

Unlabeled 70 31 35 136

Total 105 31 47 183

Figure 34. Multitrait-Multimethod Matrix

84
Figure 35. Model 1 (freely correlated traits; freely
correlated methods)

85
 Figure 36. AMOS Warnings Window

The matrix structure in the background of the example model

in Figure 33 is as seen in Figure34. Due to the difficulty in
displaying the figure, only a limited drawing including methods
A, B and C was made.

86
 Table 15. Summary of Fit Indices
Fit Indices

Models x2 df CFI RMSEA 90% C.I. PCLOSE

1. Freely
correlated
traits; freely 86.62 78 .897 .015 .000 .048 .897
correlated
methods

2. No traits;
freely 459.1
98 .693 .204 .122 .157 .000
correlated 2
methods

3. Perfectly
correlated
317.1
traits; freely 85 .795 .086 .081 .110 .000
2
correlated
methods

4. Freely
correlated
123.3
traits; 81 .964 .058 .037 .065 .000
9
uncorrelated
methods

In figurel 34, Heterotrait-Heteromethod Triangles,

Heterotrait-Monomethod Triangles, Heterometdod Blocks
and Monomethod Blocks in Multitrait-Multimethod Matrix
are shown. When the Multitrait-Multimethod models are run,
the warning shown in Figure 36 appears. The reason for this
warning is that there are no covariances among all the
exogeneous variables. In this case, "proceed with the analysis"
should be selected and continued.

87
Figure 37. Post Hoc Model 1

88
Figure 38. Model 2 (no traits; freely correlated methods)

89
Figure 39. Model 3 (perfectly correlated traits; freely
correlated methods)

90
Figure 40. Model 4 (freely correlated traits; uncorrelated
methods)

91
In some cases, the error terms get negative values when the
model is run. In this case, a post hoc model is created as
shown in Figure 37.

In this model, Var (a) value is assigned in order to fix the

negative error term. The assignment of value Var(a) is also
made in alternative models created for comparison with
Model 1. Note that the same value assignment is shown in
Figure 38.-39.-40.

Table 16. Comparisons of Nested Models

Difference

Models x2 df CFI

Test of Convergent
Validity

Model 1 and Model 2

372.506 20 .204
(traits)

Test Discriminant Validity

Model 1 and Model 3 (traits) 230.502 7 .102

Model 1 and Model 4

74.230 3 .067
(methods)

Then, 3 alternative models are created. These models are as

seen in Figures 38-39-40. Model 2 consists only of methods.
The latent variables that make up the traits are not included in
the model. The parameters of the covariances between the

92
methods are released. In Model 3, the features and methods
are included in the same model but this time the covariance
parameters between the latent variables forming the traits are
fixed to 1.

Table 17. Model 1 Trait and Method Loadings

X Y V W Z T U A B C D E

X .920 .008

Y .901 .601

V .898 .007

W .794 .405

Z .854 .506

T .426 .522

U .324 .714

X .401 .302

Y .306 .852

V .384 .701

W .399 .628

Z .424 .701

T .789 .574

U .698 .358

X .640 .406

Y .501 .720

93
 V .654 .506

W .701 .525

Z .689 .603

T .745

U .754 .289

.597 .374

X .256 .356

Y .406 .902

V .250 .604

W .274 .712

Z .372 .586

T .755 .457

U .743 .771

X .595 .411

Y .489 .398

V .525 .549

W .424 .375

Z .445 .601

T .379 .832

U .408 .566

Note: Estimates are standardized

For this reason this model is called as perfectly correlated

traits; freely correlated methods. In Model 4, there is no
covariance between methods. In Model 1, which is the model

94
where the correlations between traits and methods are free,
calculated correlation values are shown in Table 18. In the
Multitrait-Multimethod models approach, analyzes are
performed at the matrix level and the parameter level to test
the construct validity (both convergent and discriminant
validity).

In matrix level analysis, comparison of fit indices of alternative

models with model 1 is made. Table 15 shows the values of fit
indices of each model. Table 16 compares each alternative
model with model 1. The comparison between Model 1 and
Model 2 shows convergent validity, comparisons between
Model 1 and Model 3 and Model 4 show the discriminant
validity. The significance of difference between X2 values of
Model 1 and Model 2 indicating the convergent validity is
sufficient.

As shown in Table 16, ΔX2 (372.506, p<0.01) and ΔCFI

(0.204, p<0.01) are significant. In addition, the difference
between X2 values of Model 1 and Model 3 and Model 4 are
the basis for confirming the discriminant validity. As shown in
Table 16, between Model 1 and Model 3, ΔX2 (230.502,
p<0.01) and ΔCFI (0.102, p<0.01) are significant, between
Model 1 and Model 4, ΔX2 (74.230, p<0.01) and ΔCFI (0.067,
p<0.01) are significant. Another indicator of construct validity
is parameter level analysis. For comparison at the parameter
level, the factor loadings and factor correlations in Table 17
are compared.

95
 Table 18. Traits and Methods Correlations in Model 1
Traits Methods

Construct X Y V W Z T U A B C D E

X 1

Y .345 1

V .302 .789 1

W .220 .720 .487 1

Z .351 .698 .501 .607 1

T .455 .521 .201 .421 .584 1

U .248 .836 .478 .225 .160 .370 1

A 1

B .197 1

C .201 .421 1

D .254 .428 .222 1

E .218 .648 .168 .334 1

Factor loads are expected to be significant. the Multitrait-

Multimethod models, four models are formed as described
above. There is a need for a model that is free from bias in
these models. In this case, correlated uniqueness model can be
established. In Figure 41, correlated uniqueness model is
shown.

First, it is checked whether the fit indices of the model are

within the acceptable range. In this model, the factor loads
must be significant in order to confirm the construct validity.

96
Figure 41. Model 5 (Correlated Uniqueness Model)

97
Table 19. Factor Loading in Model 5
X Y V W Z T U

X .810

Y .801

V .797

W .644

Z .751

T .800

U .721

X .501

Y .416

V .353

W .489

Z .518

T .468

U .499

X .741

Y .428

V .555

W .721

98
Z .489

T .706

U .742

X .368

Y .517

V .398

W .277

Z .365

T .479

U .398

X .498

Y .489

V .520

W .328

Z .531

T .544

U .298
Note: Estimates are standardized

99
11. METHODS TO BE APPLIED IN CASE OF
DATA INADEQUACY
Sometimes, there may be cases where the assumptions of the
estimation methods used are not met by the existing data set.
In this case, there are methods that can be applied if it is
necessary to be satisfied with the dataset available. Leading
methods among them are bootstrap partial least sequare.

11.1. Bootstrap Method

The bootstrap technique is applied when one of the
assuptions of normal distribution or being continuous variable
is not met. This method was developed by B. Efron in 1979
(Efron, 1979). In many studies in the literature the condition
of normal distribution obligation is neglected. It is also seen in
many studies in the literature that X2 value is derived by
maximum likelihood and generalized least squares methods.
Estimation methods which are frequently used in the
structural equation model are these two methods. In
particular, with the non-normal distribution, the number of
observations is also low cause X2 value to increase. At the
same time, irreversible and inadequate modifications made
during the analysis of such data are not scientifically
acceptable and result in inconsistent estimations about the
population. In the bootstrap method, a different data set is
obtained from the existing observations (Sacchi, 1998). This
method is basically the derivation of the sample from the
sample.

100
There are advantages and limitations of the bootstrap process.
The main advantage of the bootstrap technique is the ability to
evaluate the accuracy of the predicted parameters. The idea
underlying the bootstrap technique is to create sub-samples of
the current data and look at the distribution of the parameters
computed from each sub-sample.

Figure 42. Bootstrap Properties Window

101
There are various criticisms in the literature about the
correctness of the results obtained with this technique. (Kline,
2011) (Ichikawa, 1995) (Yung & Bentler, 1994) (Hancock &
Nevitt, 1999). It's wrong to see this technique as a magical
method. This method should not be used especially in small
samples with low representation ability and in extreme non-
normal distribution (Kline, 2011). Because especially in small
samples, there is the posibility to further enhance its
properties that do not match the population (Rodgers, 1999).
The AMOS program includes bootstrap analysis. Figure 42
shows the bootstrap tab under the analysis properties window.
In this window it is initially marked that how much the
bootstrap technique can be applied in the sample. The
window is closed after the confidence interval and prediction
method are selected.

11.2. Partial Least Square Structural Equation

Modeling ( PLS-SEM)
It is also called covariance-based structural equation modeling
since the structural equation model that has been examined in
the previous sections is based on the covariance matrix.
However partial least square structural equation modeling is
based on variance. For this reason, it is also called as the
variance-based structural equation modeling. Partial least
square structural equation modeling (PLS-SEM) is an
advantageous method when the assumptions of least squares
are not met. It is an alternative of covariance-based structural
equation modeling (CB-SEM). It is a second generation
multivariate analysis method that enables measurement model
and structural model to be analyzed together like covariance-
based structural equation modeling.

102
 Table 20. Classification of Multivariate Methods

Exploratory Confirmatory

x Cluster Analysis x Analysis of

Variance
x Exploratory Factor
Analysis x Logistic
First Regression
Generation x Multidimensional
Techniques Scaling x Multiple
Regression

x Confirmatory
Factor Analysis

x Partial Least Square x Covariance-

Structural Equation Based
Second
Modeling (PLS- Structural
Generation
SEM) Equation
Techniques
Modeling (CB-
SEM)
Source: Hair, J., Hult, G., Ringle, C., & Sarstedt, M. (2017). A primer on partial least
squares structural equation modeling PLS-SEM. Los Angeles: SAGE.

But it is not a confirmatory analysis technique like Covariance-

Based Structural Equation Modeling. Table 20 shows the
classification of multivariate methods (Hair, Hult, Ringle, &
Sarstedt, 2017). As shown in Table 20, it is an exploratory
analysis technic.

103
According to some sources in the literature, covariance-based
structural equation modeling is a more powerful and reliable
method. For this reason, the partial least square structural
equation modeling method is generally preferred in cases
where the conditions listed below are found:

• If the sample is small.

• If the data do not distribute normally.

• If the number of indicators connected to the latent variable
is less than three.

• If there is a multicollinearity.
• There is missing value.

• If the number of observations is less than the number of

explanatory variables.

If the above listed conditions are found, method PLS-SEM

method is far superior to method CB-SEM. Because, in these
cases, it reduces the unexplained variance to the lowest level.
As the model is complex, such as in the CB-SEM method, no
larger sampling is required in PLS-SEM. However, some
researchers who have done research on the sampling
sensitivity of the PLS-SEM method have raised the ten-fold
rule. According to this rule, there is a necessity to have 10
times observation of the number of indicators used to
measure a construct in the measurement model and 10 times
observation the number of the path in a structural model
(Barclay, Higgins, & Thompson, 1995).

104
However, the PLS-SEM method is a non-parametric method
because it does not have any distributional assumption (Hair,
Hult, Ringle, & Sarstedt, 2017). It is also an explanatory
approach, which is why it is preferred in exploratory research.
In other words, when the theory is underdeveloped, it can be
said that researchers prefer to use partial least squares
structural equation modeling. This judgment is partially
correct in cases where the structure need to be predicted and
relations need to be explained (Rigdon, 2012).

When the theory needs to be tested and verified, in case of

there is cycles in the structural model and if the model needs
to be verified in general with fit indices it is more accurate to
use CB-SEM method. Because the PLS-SEM method can not
explain loop-related relations. In addition, it does not give
general fit indices of the model.

Partial least squares method can be easily implemented by

means of a packet program called SmartPLS. SmartPLS is a
packet program that allows the creation of partial least squares
based structural equation models. Structural equation
modeling programs outside of SmartPLS makes the maximum
likelihood estimation method the default choice. Because
covariance based structural equation modeling is defined by
this estimation method in the literature. However, as explained
in previous chapters, different estimation methods can be used
in case of necessity. Therefore, the method that is most likely
to be used in the covariance-based structural equation
modeling should be based on a valid reason (Kline, 2011).
This is the most important reason for criticism of PLS-SEM
method. For this reason, it is necessary to make sure that all

105
solution alternatives are exhausted if the partial least squares
method is used in a research according to some sources. But
there are also resources that assess them as blind and
misleading criticisms, and that reveal the advantages of the
partial least squares method (Henseler, Dijkstra, Sarstedt,
Ringle, Diamantopoulos, & Straub, 2014). Despite all this
criticism and hesitation, the PLS-SEM method has become an
increasingly used method in scientific studies (Hair, Hult,
Ringle, & Sarstedt, 2017).

As mentioned above, a packet program called SmartPLS is

used to construct the least squares based structural equation
models. The following steps describe how to build a structural
equation model with SmartPLS. These descriptions are based
on SmartPLS 3 version. When SmartPLS is first turned on, the
screen shown in Figure 43 'opens. In order to create a new
project, in the top menu "File" is selected, and after that
"Create New Project" command is clicked on.

Figure 43. SmartPLS Start Screen

106
When the command is clicked, the "Create Project" window
shown in Figure 44 opens. Enter the project name in the
Name field of this window and press the OK button to close
the window. As a result, the project name is displayed in the
Project Explorer section of the main screen. Double click on
"Double-click to import data" on the project name and the
data set to be worked on is connected to the project. The
SmartPLS program is not a SPSS-compatible program like
AMOS, so it only accepts Excel files. If the data set to be used
is in the SPSS program, data can be easily transferred from the
SPSS to the Excel file.

Figure 44. Project Creation Screen

After the data set has been connected, in the indicator section,
indicators are displayed as listed. Once you click on the project
name in the Project Explorer section, the white screen opens
again. When the desired indications in this segment are
selected in groups and are dragged in to the middle of screen,

107
it can be seen that the latent variable and its indicators are
automatically drawn as shown in Figure 45.

Figure 45. Drawing Latent Variable

The hidden variable plotted in Figure 45 can be easily dragged

with the mouse and the variable can be renamed by right
clicking on it. After all the variables are drawn in this way,
paths are drawn between the hidden variables using the
Connect button at the top of the screen. After the model
drawing process is completed, the Calculate → PLS Algorithm
command, which is located at the top of the screen, is
executed. The predicted values of the path coefficients in the
opened window can be obtained in matrix form or graphically.
Figure 46 shows the estimated values in a matrix form.

108
 Figure 46. Coefficient Estimation Results

If you click on the project name again in the Project Explorer

pane, all the coefficient values can be displayed on the main
model, including the measurement model, as shown in Figure
47. In this Figure, the values in the middle of each latent
variable indicate R2 values. As a result, analysis reports can be
written by interpreting these values.

109
Figure 47. Estimation Results

110
BIBLIOGRAPHY
Aksu, G., Eser, M. T., & Güzeller, C. O. (2017). Açımlayıcı ve
Doğrulayıcı Faktör Analizi ile Yapısal Eşitlik Modeli
Uygulamaları. Ankara: Detay Yayıncılık.

Anderson, J., & Gerbing, D. (1988). Structural Equation

Modelling in Practice: A Review and Recommended
Two-Step Approach. Psychological Bulletin.
Avcılar, M., & Varinli, İ. (2013). Perakende Marka Değerinin
Ölçümü ve Yapısal Eşitlik Modeli Uygulaması (1 b.). Ankara:
Detay Yayıncılık.
Bagozzi, R. P., & Yi, Y. (1990). Assessing Method Variance in
Multitrait-Multimethod Matrices: The Case of Self-
reported Affect and Perceptions at Work. Journal of
Applied Psychology, 75(1), 547-560.
Baltes, P., & Nesselroade, J. (1979). History and rationale of
longitudinal research. New York: Academic Press.
Barclay, D., Higgins, C., & Thompson, R. (1995). The partial
least sequares approach to causal modeling: Personal
computer adoption and use as illustration. Technology
Studies, 2, 285-309.
Baron, R., & Kenny, D. (1986). The Moderator-Mediator
Variable Distinction in Social Phychological Research:
Conceptual, Strategic and Statistical Considerations.
Journal of Personality and Social Phychology, 6(51), 1173-
1182.

111
Bayram, N. (2013). Yapısal Eşitlik Modellemesine Giriş. Bursa:
Ezgi Kitapevi.
Bentler, P. M., & Chou, C.-P. (1987). Practical Issues in
Structural Modeling. Sociological Methods Research, 16(1),
78-117.
Byrne, B. M. (2010). Structural Equation Modeling with AMOS.
New York: Routledge Taylor & Francis Group.

Campbell, D., & Fiske, D. (1959). Covergent and

Discriminant Validation by the Multitrait-Multimethod
Matrix. Psychological Bulletin, 56, 81-105.
Chen, J., Rungruengsamrit, D., Rajkumar, T., & Yen, D.
(2013). Success of Electronic Web Sites: A Comparative
Study in Two Countries. Information & Management,
50(6), 344-355.
Civelek, M., İnce, H., & Karabulut, T. (2016). The Mediator
Roles Of Attitude Toward The Web Site And User
Satisfaction On The Effect Of System Quality On Net
Benefit: A Structural Equation Model On Web Site
Success. European Scientific Journal, 61-73.
Çelik, H. E., & Yılmaz, V. (2013). Lısrel 9.1 ile Yapısal Eşitlik
Modellemesi. Ankara: Anı Yayıncılık.
Doğan, İ. (2015). Farkli Veri Yapisi ve Örneklem Büyüklüklerinde
Yapisal Eşitlik Modellerinin Geçerliği ve Güvenirliğinin
Değerlendirilmesi. Eskişehir: Eskİşehİr Osmangazİ
Ünİversİtesİ Doktora Tezi.
Dursun, Y., & Kocagöz, E. (2010). Yapısal Eşitlik
Modellemesi ve Regresyon: Karşılaştırmalı Bir Analiz.
Erciyes Üniversitesi İ.İ.B.F. Dergisi(35), 1-17.

112
Efron, B. (1979). Bootstrap method: Another look at the
jackknife. Annals of Statistics, 7, 1-26.
Fornell, C., & Larcker, D. (1981). Evaluating Structural
Equation Models with Unobservable Variables and
Measurement Error. Journal of Marketing Research, 18(1),
39-50.
Gerbing, D., & Anderson, J. (1988). An Updated Paradigm for
Scale Development Incorporating Unidimensionality
and Its. Joournal of Marketing Research, 25(2), 186-192.
Gujarati, D. (1999). Essentials of Econometrics (2 b.). Singapore:
McGRAW-HILL.
Hair, J., Hult, G., Ringle, C., & Sarstedt, M. (2017). A primer on
partial least squares structural equation modeling PLS-SEM.
Los Angeles: SAGE.
Hancock, G., & Nevitt, J. (1999). Bootstrapping and the
identifcation of exogenous latent variables within
structural equation models. Structural Equation Modeling,
6, 394-399.
Henseler, J., Dijkstra, T., Sarstedt, M., Ringle, C.,
Diamantopoulos, A., & Straub, D. (2014). Common
beliefs and reality about partial least squares: Comments
on Rönkkö & Evermann (2013). Organizational Research
Methods, 17, 182-209.
Ichikawa, M. (1995). Application of the bootstrap methods in
factor analysis. Psychometrika, 60, 77-93.
Jayaram, J., Kannan, V., & Tan, K. (2004). Influence of
initiators on supply chain value creation. International
Journal of Production Research, 42(20), 4377-4399.

113
Karagöz, Y. (2016). SPSS ve AMOS 23 Uygulamalı İstatistiksel
Analizler. Ankara: Nobel.
Kline, R. (2011). Principles and practice of structural equation
modeling (3nd Ed. b.). New York: Guilford.
MacKinnon, D., Lockwood, C., Hoffman, J., West, S., &
Sheets, V. (2002). A comparison of methods to test
mediation and other intervening variable effects.
Psychological Methods, 7, 83-104.

Meydan, C. H., & Şen, H. (2011). Yapısal Eşitlik Modellemesi

AMOS Uygulamaları. Ankara: Detay Yayıncılık.
Raykov, T. (1997). Estimation of composite reliability for
congeneric measures. Applied Psychological Measurement,
21(2), 173-184.
Raykov, T., & Marcoulides, G. (2006). A First Course in
Structural Equation Modeling. Mahwah: Lawrence
Erlbaum Associates.
Rigdon, E. (2012). Rethinking partial least squares path
modelling: In praise of simple methods. Long Range
Planning, 45, 341-358.
Rodgers, J. (1999). The bootstrap, the jackknife, and the
randomization test: A sampling taxonomy. Multivariate
Behavioral Research, 34, 441-456.
Sacchi, M. (1998). A bootstrap procedure for high-resolution
velocity analysis. Geophysics, 63(5).

Sarstedt, M., & Mooi, E. (2014). A concise guide to market research:

The process, data, and methods using IBM SPSS statistics (2nd
b.). Berlin: Springer.

114
Schermelleh-Engel, K., Moosbrugger, H., & Müller, H. (2003).
Evaluating the Fit of Structural Equation Models: Tests
of Significance and Descriptive Goodness-of-Fit
Measures. Methods of Psychological Research Online, 8(2), 23-
74.
Sipahi, B., Yurtkoru, E., & Çinko, M. (2010). Sosyal bilimlerde
SPSS'le Veri Analizi. İstanbul: Beta Basım A.Ş.
Tabachnick, B., & Fidell, L. (2001). Using multivariate statistics.
Boston: Ally and Bacon.
Taşkın, Ç., & Akat, Ö. (2010). Araştırma Yöntemlerinde Yapısal
Eşitlik Modelleme. Bursa: Ekin Basım Yayın.

Widaman. (1985). Hierarchically Tested Covariance Structure

Models for Multitrait-Multimethod Data. Applied
Psychological Measurement, 9, 1-26.
Wooldridge, J. (2003). Introductory Econometrics: A Modern
Approach . Mason: Thomson.
Wu, A., & Zumbo , B. (2008). Understanding and Using
Mediators and Moderators. Social Indicators Research, 87,
367-392.
Wu, J.-H. W., & Wang, Y.-M. (2006). Measuring KMS
Success: A Respecification of the DeLone and
McLean’s Model. Information & Management, 43, 728-
739.
Yung, Y., & Bentler, P. (1994). Bootstrap-corrected ADF test
statistics in covariance structure analysis. British Journal
of Mathematical and Statistical Psychology, 47, 63-84.

115