Structural Equation Modeling Using AMOS

Preface
In the era of advanced technology, vast amount of data concerning science, social
science, and economy are gathered and transmitted efficiently worldwide. Thus,
academicians as well as practitioners need to analyze those data so that it will be
meaningful for decision makers, policy makers, and the like. As everybody knows,
making quick and accurate decision is crucial for the growth and survival of an
organization, especially in the borderless world where competitions are fierce. Thus,
employing the advanced research methodology in data analysis often differentiates
between success and failure of an organization. Structural Equation Modeling, or
popularly known as SEM, is one of the newest methods of multivariate data analysis
developed specifically to overcome the limitations experienced in the previous
methodology namely the Ordinary Least Square (OLS) regressions.

SEM, which is often termed as the Second Generation Method, could entertain latent
constructs with multiple indicators as well as the usual observed variables into the model
concurrently and, more importantly, the inter-relationship among them are analyzed
simultaneously. This book is suitable for the beginners since it explains the basic concept
of SEM and how it works together with easy to understand examples. The book also
introduces the application of the most popular SEM software namely AMOS Graphic. As
the chapter moves on, it provides practical examples where SEM with AMOS Graphic is
employed to model and solve the problems in postgraduate studies. Among the statistical
analysis procedure explained in detail include the testing of hypothesis for Path Model
(direct effect), testing the mediation effect of a Mediator (direct as well as indirect effect),
and testing the moderation effect of a Moderator (interference effect) in the model.

Zainudin Awang, PhD

 Table of Contents Page

Overview: Structural Equation Modeling (SEM) using AMOS 1

Why One Needs to Employ SEM? 4

Examples Where AMOS Graphic is Employed in Research 8

Chapter 1

1.0 Introduction to Structural Equation Modeling (SEM) 17

1.1 The Concept of SEM and How it Works 17

1.2 The Advantages of SEM Compared to OLS 18

1.3 Converting Regression Models into AMOS Graphic 20

1.4 The Concept of Latent Construct in Research 22

1. 5 The Minimum Sample Size Required in using SEM 24

1.6 Introduction to AMOS Software 24

1.7 The Variable Terms in SEM using AMOS 25

1.8 Modeling the Observed and Unobserved variables in AMOS Graphic 27

1.9 Modeling Multiple Variables in AMOS Graphic 32

1.10 Modeling the Mediator Variable in AMOS Graphic 34

1.11 Modeling the Moderator Variable in AMOS Graphic 40

i
 Chapter 2

2.0 The Models Involved in Structural Equation Modeling 46

2.1 The Role of Theory in Structural Equation Modeling 46

2.2 The Measurement Model for a Construct 47

2.3 The Structural Model in AMOS Graphic 50

2.3.1 The Structural Model: Modeling the Correlational Relationship 50

2.3.2 The Structural Model: Modeling the Causal Relationship 52

2.3.3 The Structural Model: Modeling the Mediator 53

2.4 The Types of Constructs in Structural equation Modeling 55

2.4.1 The Reflective Construct 55

2.4.2 The Formative Construct 56

2.4.3 The Second Order Construct 59

Chapter 3

3.0 Validating the Measurement Model using CFA 61

3.1 Evaluating the Fitness a Model: The Measurement and Structural 63

Model

3.2 The Steps Involved in Validating the Measurement Model 68

3.3 The Confirmatory Factor Analysis (CFA) 69

3.31 Validating the Pooled Measurement Model 72

3.32 The Measurement Model for Pooled Latent Constructs 73

ii
3.33 Assessing the Validity and Reliability for a Pooled Measurement 78
Model

3.4 The Assessment of Normality Distribution for the Data 80

Chapter 4

4.0 Analyzing the SEM Structural Model in AMOS Graphic 84

4.1 The Steps Involved in Performing Structural Equation Modeling 84

4.2 Performing the Structural Equation Modeling Analysis 85

4.2.1 Interpreting the AMOS Text Output: The Standardized Regression 91

Weights

4.2.2 Interpreting the AMOS Test Output: The Regression Weights 92

4.2.3 Interpreting the AMOS Test Output: The Standardized Residual 94

Covariance

4.3 An Example of More Complicated Amos Output 95

Chapter 5

5.0 Analyzing the Relationships between Variables in a Model 99

5.1 Analyzing the Relationships between Observed Variables 99

5.2 Analyzing the Relationships between Latent Constructs 101

5.3 Analyzing the Causal Effects for Observed Variables 104

5.4 Analyzing the Causal Effects for Simple Latent Constructs 105

iii
5.5 Analyzing the Causal Effects for Multiple Latent Constructs 107

5.6 Path Analysis: Analyzing the Mediation Effects in a Model 112

Chapter 6

6.0 Analyzing the Mediating Variable in a Model 118

6.1 Analyzing the Mediating Effects for Observed Variables 120

6.2 Analyzing the Mediating Effects for Latent Constructs 124

6.3 Analyzing the Mediator in a Complex Model 129

6.4 Mediation Test: Confirming the Test Results through Bootstrapping 134

6.5 Computing the Effect Size in a Mediation Test 136

Chapter 7

7.0 Analyzing the Moderating Variable in a Model 144

7.1 The Schematic Diagram for Moderator Variable in a Model 145

7.2 Modeling the Interaction effects for Observed variables 147

7.3 Scale of Measurement for the Moderator Variable 148

7.4 Modeling the Moderating Effects for Observed Variable 149

7.5 Analyzing the Moderating Effects for Observed Variable 151

7.6 Modeling the Moderating Effects for Latent Constructs 154

iv
7.7 Analyzing the Moderator for Latent Constructs: The Multi-Group 155
CFA

7.8 Comparing the Group Effect for Moderator Variable 172

Chapter 8

8.0 The Second Order Confirmatory factor Analysis 176

8.1 The Steps Involved in Performing Second Order CFA 176

8.2 Performing Second Order CFA for Training Transfer Constructs 177

8.3 Performing Second Order CFA for Job Satisfaction Constructs 181

8.4 Performing Second Order CFA for Technological Innovation 191

Constructs

Chapter 9

9.0 The Application of AMOS Graphic in Postgraduate Researches 199

9.1 Modeling the Loyalty of banking Customers 199

9.2 Modeling the Loyalty of Outgoing Undergraduates towards 201

University for Postgraduates

9.3 Modeling Customer Loyalty towards their Service Provider 209

9.4 Modeling Investor Loyalty towards Listed Companies 212

9.5 Modeling the Effects of Entrepreneurial Orientation on Business 213

performance

v
9.6 Modeling the Effects of University reputation on Students Loyalty 214

9.7 Modeling the Effects of Individual Factors, Transfer Climate, and 215
Training Design on Training Transfer in an Organization

9.8 Modeling Motivation to Transfer as a Mediator in The Transfer 221

Training Model

Bibliography 224

Index 228

vi
 Overview of Structural Equation Modeling (SEM)

Academicians, researchers, as well as postgraduate students, are debating theories

concerning the relationships among certain hypothetical constructs. They are modeling
this theorized relationship with the intention to test their theoretical model with the
empirical data from the field. The example of a Theoretical Framework is given in Figure
1.

Figure 1: The Schematic Diagram Showing the Theoretical Framework of a Study

The schematic diagram in Figure A is converted into AMOS Graphic and

analyzed using empirical data. In AMOS Graphic, the rectangle represents the directly

1
observed variables, while the ellipse represents the unobserved variable or latent
constructs. The schematic diagram of theoretical framework in Figure 1 is converted into
AMOS Graphic as shown in Figure 2.

In Figure 2, the researcher needs to determine the types of variables and

constructs involved in the model and how the study is going to measure them. The types
of latent construct and variables involved in Figure 2 are identified as follows:

1. Financial position is a latent construct. The type of this latent construct is called
Formative Construct. This Formative Construct is formed by four variables namely
Leverage, Growth, Bankruptcy, and Tax Consideration. The data for these variables
are obtained directly from the annual reports. The type of variables here are measured
or observed variables.

2. Risk Taking Propensity is another latent construct. The type of latent construct in
this case is called Second Order Construct. This Second Order Construct consists of
three first order constructs or dimensions namely Comp1, Comp2, and Comp3. The
variables for Comp1 (R11, R12, CR13), Comp2 (R21, R22, R23), and Comp3 (R31,
R32, R33) are measured using items in a questionnaire.

3. Attitude towards Risk is another Second Order construct. This construct consists of
two first order constructs or dimensions namely Att1 and Att2. The latent construct
Att1 and Att2 are measured using four items in a questionnaire (St11, St12, St13,
St14) and (St21, St22, St23, St24), respectively.

4. Demand for Insurance is an observed variable. Sometimes, this variable is termed as

a directly measured variable. The value is obtained from secondary data.

For the sake of providing the overall picture of what to expect from the book, the
author would demonstrate the process from the beginning (theoretical framework)
until the output is obtained for analysis through Figure 1, Figure 2, and Figure 3.

2
 Figure 2: The Schematic Diagram is Converted into AMOS Graphic

Figure 2 shows the measuring items for all constructs and variables involved in a
model. In AMOS Graphic, ellipses represent latent constructs, while rectangles represent
input variables. Once the data are obtained, the researcher would input these data into
their respective variable indicated by rectangles. The data input is made through “click
and drag” procedure.

Figure 3 shows the output produced by AMOS Graphic after data input was made
and the model was executed. This output is ready for interpretation and hypothesis
testing.

3
 Figure 3: The AMOS Graphic Showing the Output after the Model is Executed

Why One Needs to Employ SEM?

SEM is a Second Generation multivariate analysis technique developed due to some

limitations in the traditional Ordinary Least Squares (OLS), especially when dealing with
latent constructs.

In the above example (Figure B), Financial Positions, Risk-Taking Propensity,

Attitude towards Risk, and Intention to Protect are latent constructs. Latent constructs are
unobserved variables. They are measured using multiple items in a questionnaire. The

4
researchers should switch from employing the traditional OLS to SEM so as to keep pace
with the latest development in research methodology.

AMOS is an acronym for Analysis of Moments Structures. This is one of the

newest software developed for Structural Equation Modeling (SEM). The researchers
could employ AMOS Graphic to model and analyze the inter-relationships among latent
constructs effectively, accurately, and efficiently. More importantly, the multiple
equations of inter-relationships in a model are computed simultaneously.

Why One Should Employ AMOS Graphic?

As has been said earlier, researchers could convert their theoretical framework directly
into AMOS graphic for analysis. More importantly, using AMOS Graphic interface,
researchers can create path diagrams using drawing tools, rather than by writing
equations or by typing commands. Even the researchers could validate the measurement
model of a latent construct using Confirmatory Factor Analysis (CFA). Once CFA is
completed, the researcher could move into modeling the Structural Equation Modeling
(SEM). Thus, analyzing and testing the theory using AMOS is fast, efficient, and user
friendly.

The conversion of a research framework into AMOS Graphic will be illustrated

using the following example. The theoretical framework in Figure 4 consists of four
constructs namely Entrepreneurial Orientation (EO), Marketing Orientation (MO),
Organizational Innovation (OI), and Business Performance (BP). Construct EO and MO
have two sub-constructs each. The sub-constructs are measured using certain number of
items. Meanwhile, Constructs OI and BP are measured using five items and four items,
respectively. An example of a researcher’s theoretical framework is given in Figure 4.

5
Figure 4: The Theorized Framework Showing the Hypotheses to be Tested
Empirically in the Study

The framework in Figure 4 indicates several hypotheses to be tested in the study.

Hypotheses H1, H2, H3, and H5 are testing for causal effects. Hypothesis H4 is intended to
test the mediation effects, while another hypothesis namely H6 is testing the moderation
effect in the model. The procedure for hypothesis testing would be easy in SEM using
AMOS Graphic. The researcher only needs to convert the theoretical model into AMOS
Graphic. Once converted, the inter-relationship among constructs could be analyzed.

Now, let’s see how the theoretical framework in Figure 4 is converted into AMOS
Graphic in Figure 5.

6
 Figure 5: The Research Framework Converted into AMOS Graphic

AMOS Graphic could find models that best fit data at hand

The researcher should take advantage from the results of previous researches by
specifying constraints on a certain parameter in the model. Researchers can fit multiple
models in a single analysis. AMOS Graphic examines every pair of the models, where
one model can be obtained by placing parameter restrictions on the other. AMOS could
also identify the pair of redundant items in a measurement model that jeopardize the
fitness of the model. The researcher could either constrain a pair of redundant items in a
measurement model, or delete the item altogether from the model in order to improve the
fitness of the model.

7
Examples where AMOS Graphic is employed in research:

Now let’s go through some practical examples to familiarize how AMOS Graphic is
being employed to model and analyze research problems. Do not worry if you found the
applications to be too difficult for you to understand at this stage because we will go back
to basic and begin our journey from ground zero, from beginning Chapter 1.

1) In the field of psychological research — AMOS Graphic could be employed to

model and evaluate the role of medical counseling in helping the healing process of
patients undergoing treatment in a hospital.

The above objective could be modeled using AMOS Graphic as shown in Figure
6.

In Figure 6, the Amount of Medication administered is an independent variable.

Note: The independent variable is the variable that can be manipulated by the researcher.
The Degree of Wellness is a dependent variable since it is the outcome of independent
variable. The medical counseling is a moderator variable. The study is intended to
evaluate the moderation effects of medical counseling in the relationship between
Amount of Medication administered by the doctor and the Degree of Wellness
experienced by the patients. Both variables are directly observed.

In the model (Figure 6), the Degree of Wellness could also be latent construct if it
is measured through a set of questionnaire given to the patients to measure their wellness.

8
 All variables involved in the
model are directly observed

Figure 6: Modeling the Moderator Variable Namely “Counseling Intervention”

2) In the field of medical and healthcare research — AMOS Graphic could be

employed to determine the influence of Corporate Image of drugs manufacturers and
Cost of Medicine on the doctors’ Willingness to Prescribe generic drugs to their
patients.

The above problem is modeled in AMOS Graphic as shown in Figure 7. In the study,
the Manufacturer Image and Willingness to Prescribe are latent constructs measured
through a set of questionnaire, while the Cost of Medicine is a directly measured variable.
The directly observed variable is also termed as an observed variable.

9
 Manufacturer Image and
Willingness to Prescribe are
latent constructs

Cost of Medicine is a
directly measured variable

Figure 7: Modeling the Causal Effects of Manufacturer Image and Cost of Medicine
on Doctors’ Willingness to Prescribe

10
3) In the field of social science — AMOS Graphic could be employed to determine the
effects of respondents’ Socio-Economic Status (SES) on their stress and health
condition.

The above problem could be modeled in AMOS Graphic as shown in Figure 8. In the
model (Figure 8), SES is an independent latent construct measured using three variables
namely Education, Occupation, and Income. Another latent construct namely Stress is
measured using eight items in a questionnaire. The other latent construct namely Health
is measured using two sub-constructs; Mental Health (5 items) and Physical Health (5
items).

SES is a
formative Stress is a
construct reflective construct

Health is a second order

construct measured by
mental and physical health

Figure 8: Modeling the Effects of SES on Health with Stress as a Mediator

Construct
11
4) In the field of educational research — AMOS Graphic could be employed to
evaluate how Infrastructure Facilities, Academic Facilities, Academic Instructors, and
Program Schedules are influencing students’ Academic Performance in a university.

The above problem is modeled in Figure 9. In the model, the variable Student
Academic Performance is a latent construct. This construct is measured using three
variables namely Individual Performance (observed score), Group Performance
(observed score), and Final Examination (observed score). In this case, the Academic
Performance is called a formative construct since it is formed by the observed score
instead of the items in a questionnaire.

Academic Performance is
a Formative Construct

Figure 9: Modeling the Multiple Regression Equations with Latent Constructs

12
5) In the field of academic research – AMOS Graphic could be employed to assess
how students’ satisfaction mediates the relationship between University Reputation
and the Loyalty of outgoing undergraduates to continue into postgraduate study.
Figure 10 illustrates how the Students’ Satisfaction construct is modeled as a
mediator in the relationship between University Reputation and Students’ Loyalty.

University Reputation and

Students Loyalty are
Second Order constructs

Main construct is University Reputation. Sub-

constructs are Faculties, Programs, and
Collaboration.

Main construct is Students Loyalty. Sub-

constructs are Repurchase and Recommend.

Figure 10: Modeling the Students’ Loyalty towards a University

13
6) In the field of market research — AMOS Graphic could be employed to model the
effects of firm’s Corporate Reputation on the Competitiveness of its products in the
market.

Corporate Reputation is a second order construct. It is measured through three first

order constructs namely Company Reputation, Product Reputation, and Personnel
Reputation.

In the above model – the main domain is Corporate Reputation. Sub-Domains are
Company, Product, and Personnel. All sub-domains are measured using questionnaires.

Corporate Reputation is a second order

construct. Its first order constructs are
Company, Product, and Personnel

Figure 11: Modeling the Effect of Corporate Reputation on the Competitiveness of

its Product in the Market
14
7) In the field of institutional research — AMOS Graphic is employed to study the
significance of Organizational Climate in a workplace as a moderator in the
relationship between employees’ Job Satisfaction and their Work Commitment.

Independent Construct Dependent Construct

Moderator

Figure 12: Modeling the Organizational Climate as a Moderator in the Relationship

between Job Satisfaction and Work Commitment

In Figure 12, the construct Organizational Climate is a moderator in the Job

Satisfaction and Work Commitment relationship.

However, in Figure 13, the same construct Organizational Climate could also be used
as a mediator in the same study (refer to Figure 13).

15
 Mediator

Independent Construct Dependent Construct

Figure 13: Modeling the Organizational Climate as a Mediator Linking Employees’

Job Satisfaction and their Work Commitment

As opposed to the model in Figure 12, the Organizational Climate in Figure 13 is

modeled as a mediator that links the relationship between Job Satisfaction and Work
Commitment.

After going through some research examples where AMOS Graphic is employed to
model the stated research framework, now we begin our learning process on SEM using
AMOS Graphic in Chapter 1.

16
 CHAPTER 1

INTRODUCTION TO STRUCTURAL
EQUATION MODELING
The Structural Equation Modeling or popularly known as SEM is a second generation
statistical analysis technique developed for analyzing the inter-relationships among
multiple variables in a model. The inter-relationships among variables could be expressed
in a series of single and multiple regression equations. The Structural Equation Modeling
technique employs the combination of quantitative data and the correlational or causal
assumptions into the model.

SEM is a more powerful statistical technique to solve the following requirements:

1) Running the Confirmatory Factor Analysis (CFA).

2) Analyzing multiple regression models simultaneously.
3) Analyzing regressions with multi-collinearity problem.
4) Analyzing the path analysis with multiple dependents.
5) Estimating the correlation and covariance in a model.
6) Modeling the inter-relationships among variables in a model.

1.1 THE CONCEPT OF SEM AND HOW IT WORKS

SEM begins with a theory where the researcher intends to test the relationship among
constructs of interest in the study. The relationships are modeled into a theoretical
framework represented by a schematic diagram. The schematic diagram presents the
hypotheses of interest to be tested in the study. The constructs of interest involved are
measured using a set of items in a questionnaire. The measurement scale for each item
should be either interval or ratio. The ideal measurement should be in the interval from 1
to 10 so that the data is more independent and meet the requirement for parametric

17
analysis. The researcher should develop at least four items to measure each latent
construct.

Throughout the chapter, the readers would find the terms variable and construct
are used interchangeably. A variable is the directly measured score, while the construct is
meant for an indirectly measured score. In fact the construct is only a hypothetical
concept of something, or the respondents’ perception concerning certain issue. A
construct is measured through the respondent’s response towards a set of items in a
questionnaire.

1.2 THE ADVANTAGES OF SEM COMPARED TO OLS

SEM is capable of estimating a series of inter-relationships among latent constructs

simultaneously in a model. In fact, SEM is the most efficient method to handle the
Confirmatory Factor Analysis (CFA) for measurement models, analyze the causal
relationships among latent constructs in a structural model, estimating their variance and
covariance, and test the hypotheses for mediators and moderators in a model.

As has been said earlier, latent constructs could not be measured directly since it
is only a hypothetical concept of something. Thus, the researcher could not model them
using the Ordinary Least Squares (OLS) regression. The examples of latent constructs
measured through a set of items in a questionnaire are:

1) Service Quality.
2) Customer Satisfaction.
3) Job Satisfaction.
4) Corporate Image.
5) Product Image.
6) Customer Loyalty.
7) Purchase Intention.
8) Consumer Behaviour.

18
9) Employee Soft Skills.
10) Perceived Usefulness.
11) Relational Bond.
12) Financial Bond.
13) Structural Bond.
14) Relationship Quality.
15) Attitudinal Loyalty.
16) Behavioural Loyalty.

Those constructs cannot be measured directly like counting the number of kids in
a family, total income of a household, monthly phone bills, daily production, weekly
price of chicken, etc. The variable which could be measured directly is called the
observed variable, while the variable which could not be measured directly is called
latent construct. These latent constructs could only be measured indirectly using a set of
items in a questionnaire.

Example of items in a questionnaire to measure student satisfaction as a latent

construct: In this example, the construct Students Satisfaction is measured using eight
items in a questionnaire.

As a student of this university, I….. Strongly Strongly

Disagree Agree
1 am satisfied with the lecture schedules 1 2 3 4 5 6 7 8 9 10
2 am satisfied with the learning process 1 2 3 4 5 6 7 8 9 10
3 am satisfied with the academic system 1 2 3 4 5 6 7 8 9 10
4 am satisfied with the continuous 1 2 3 4 5 6 7 8 9 10
evaluation
5 am satisfied with academic regulations 1 2 3 4 5 6 7 8 9 10
6 am satisfied with the library references 1 2 3 4 5 6 7 8 9 10
7 am satisfied with classroom facilities 1 2 3 4 5 6 7 8 9 10
8 am satisfied with the campus security 1 2 3 4 5 6 7 8 9 10
Source: Research Methodology and Data Analysis 2nd Edition by Zainudin Awang (2012)

19
Other advantages of Structural Equation Modeling (SEM) include:

1) Could run the Confirmatory Factor Analysis (CFA) to reduce measurement errors.
2) Could deal with the problem of multicollinearity among independent constructs.
3) Could assess the fitness of measurement model, as well as the structural model.
4) Could analyze the model with multiple independents, as well as multiple dependents.
5) Could include the mediating variable in a model and analyze its effects (mediator).
6) Could analyze the effects of moderating variable in certain path of a model
(moderator).
7) Could model the error terms and handle the correlated errors among response items.
8) Could analyze both First Order and Second Order Constructs in the structural model.
9) Could include both observed variables and latent constructs in the structural model.

1.3 CONVERTING REGRESSION MODELS INTO AMOS

GRAPHIC

Modeling the Simple Linear Regression Y = Bo + B1X1 + e1 in AMOS

Graphic

Usually, the researchers could model the above equation using Ordinary Least
Squares (OLS) regression and analyze the model using ANOVA since X1 and X2 are
observed variables. However, the researchers could also employ AMOS graphic software
to model and analyze the regression equation as shown in Figure 1.

20
 Observed
Residual e1
Variable

1
β1

X1 Y

Figure 1: The Simple Regression Model in AMOS Graphic

Key: X1 = Independent variable (observed), Y = dependent variable (observed), e1= error in the
equation or residuals (unobserved).

Note: In Figure 1, the researcher is interested to estimate the causal effect of X1

on Y and subsequently test the hypothesis to prove its significance. In Figure 1, both X1
and Y are observed variables. In AMOS, the observed variables are represented using
rectangles.

Modeling the Multiple Linear Regression Y = Bo + B1X1 + B2X2 + B3X3 + e1

The researchers could model the above equation using Ordinary Least Squares (OLS)
regression and analyzed the model using ANOVA. However, the researchers could also
employ AMOS to model the equation as shown in Figure 2 below.

21
 X1 e1
1

X2 Y

X3

Figure 2: The Multiple Regression Models in AMOS Graphic

Note: X1, X2, X3, and Y are represented by rectangles since they are directly observed variables

1.4 THE CONCEPT OF LATENT CONSTRUCTS IN

RESEARCH

The Simple Regression with Multiple Indicators to Analyze Latent

Constructs

In science and social science researches, most of the times the researchers are dealing
with latent constructs. As has been said earlier, these constructs are measured using a set
of items in a questionnaire. Since the OLS procedures could not entertain latent
constructs, the researchers need to employ SEM for the analysis. Using SEM, the
researcher could model the relationship among these constructs together with their
respective items in the model and analyze them simultaneously.

In this case, at least two measurement models are involved – one for independent
construct and the other one is for dependent construct. The theorized link between
measurement model for independent construct and measurement model for dependent
construct is called a structural model. Thus, instead of modeling the Ordinary Least

22
Squares regression (OLS) and analyzed using ANOVA, the researcher is working with
the Structure Equation Modeling (SEM) and analyzed using AMOS as shown in Figure 3.

Figure 3: The Structural Equation Model for Analyzing Latent Constructs in a Model

Note: X1 and Y are latent constructs. In AMOS syntax, latent constructs are represented
by the ellipses. The latent construct X1 is measured using items X11 to X15, while latent
construct Y is measured using items Y1 to Y5. These measured items are represented by
rectangles in the model.

Key: X1 = Exogenous construct, while X11 to X15 is a set of 5 items to measure latent construct
X1

e1 to e5 are errors in measurement for items X11 to X15

Y = Endogenous construct, while Y1 to Y5 is a set of 5 items to measure latent construct Y

e6 to e10 are errors in measurement for items Y1 to Y5

e11 is an error in the equation or the residual.

23
 1.5 THE MINIMUM SAMPLE SIZE REQUIRED FOR SEM

There are endless debates in the literatures as to how many respondents should be
obtained in order to employ SEM. However, there is no clear-cut answers to it since
every research differs (among other things) in terms of the population characteristics, and
the number of constructs employed in a model. Hair et al. (2010) offer the following
suggestion for minimum sample size depending on the model complexity and basic
measurement model characteristics.

Model Characteristics Minimum Sample

(Number of latent constructs and items) Required
1. Five or less latent constructs. Each latent construct has more than three items. 100 sample
2. Seven or less latent constructs. Each construct has more than three items. 150 sample
3. Seven or less latent constructs. Some constructs have less than three items 300 sample
(just identified model).
4. More than seven latent constructs. Some constructs have less than three items 500 sample
(just identified model).

1.6 INTRODUCTION TO AMOS SOFTWARE

AMOS is an acronym for Analysis of Moments Structure – the software developed for
analyzing the Structure Equation Modeling (SEM). SEM is also known as Covariance
Structure Analysis or Covariance Structure Modeling. Other software available to
analyze SEM includes LISREL, SEPATH, PRELIS, SIMPLIS, MPLUS, EQS, and SAS.
The advantage of AMOS compared to other software in its class is its graphics
representation of the model. So, instead of writing instructions through computer
program, researchers only need to draw the AMOS graphic identical to the schematic
diagram of a model in the study.

AMOS software could be utilized to explore statistical relationships among the

items of each construct and also between constructs. Using AMOS, the researcher can
specify, estimate, assess, and present the model in a causal path diagram to show the

24
hypothesized relationships among constructs of interest. The empirical model can be
tested against the hypothesized model for goodness of fit. If the researchers found any
path that does not fit with the original model, they could either modify the path to
improve the fitness of the model or remove that particular path completely from the
hypothesized model.

1.7 THE VARIABLE TERMS IN SEM USING AMOS

GRAPHIC

The explanation below refers to Figure 4.

1. Exogenous construct is the independent variable in the Ordinary Least Squares

(OLS) regression. In AMOS, the independent variable is drawn as an upstream
variable with the causal arrow pointing out to its corresponding dependent variable.

In Figure 4: X1 and X2 are exogenous construct with five response items. The arrows
from exogenous constructs X1 and X2 are pointing out to their endogenous construct
Y to indicate that X1 and X2 are theorized to have some causal effects on Y.

2. Endogenous construct is the dependent variable in the Ordinary Least Squares

regression. In AMOS, the dependent variable is drawn as a downstream variable with
the arrow pointing in from its corresponding independent variable.

In Figure 4: Y is an endogenous latent construct with five response items.

3. Mediating variable is the variable which has a double role. This variable acts as a
dependent variable in the first equation, and acts as an independent variable in the
second equation. In theory, the mediator variable mediates the relationship between
an independent variable and a dependent variable. In Figure 4, M is the mediating
construct with four response items.

25
4. Moderating variable is the variable that moderates the effects of independent
variable on its dependent variable. In the case of latent constructs, the moderating
variable is the variable that moderates the effects of exogenous construct on the
endogenous construct. The representation for moderating variable is shown in Figure
4. Unlike the mediating variable, the moderating variable is not in the model.

Exogenous

Construct

Endogenous Residual

Mediator Variable Construct

Measurement Error

Figure 4: The Sequence of Constructs Assembled in a Model in AMOS Graphic

Note: X1 and X2 are exogenous constructs while Y is an endogenous construct. All constructs are
latent.

26
5. Error in measurement - an error depicted from each measuring item of a variable.

In Figure 4 – we can see that e1 to e5 are the measurement errors for construct X1, e11
to e15 are the measurement errors for construct X2, while e6 to e10 are the
measurement errors for construct Y.

6. Error in equation – a residual in the respective regression equation. In Figure 4 – we

can see that e20 is the residual for the equation Y = f(X1, X2) or Y = Bo + B1X1 + B2X2
+ e1.

Remember: The numbering for measurement errors, as well as residuals in the model,
are assigned randomly by AMOS Graphic.

1.8 MODELING THE OBSERVED AND UNOBSERVED

VARIABLES IN AMOS GRAPHIC

Let X1 and X2 be independent variables while Y is a dependent variable in a multiple

regression model. Both variables are directly observed. The researcher could model this
multiple regression in AMOS Graphic as shown in Figure 5.

X1
1
Y e1

X2

Figure 5: Modeling the Observed Variables in AMOS Graphic for Multiple Regression
Models

27
The model in Figure 5 is equivalent to the following model in a multiple regression
equation:

Y = Bo + B1X1 + B2X2 + e1

This model in Figure 5 is valid and workable only if the independent variables X1
and X2 do not have a multi-collinearity problem between them. Remember, one of the
main assumptions in the Ordinary Least Squares (OLS) is no significant multi-
collinearity exists between the independent variables. The Structural Equation Modeling
(SEM) technique could deal with the multi-collinearity problem. In fact, AMOS requires
the researcher to estimate the correlation between independent variables as well as
between exogenous constructs.

28
 e1 e2 e3 e4 e5
Residual
1 1 1 1 1

X11 X12 X13 X14 X15

1
e15

X1 1 1
1 Y1 e11

1
Y Y2 e12

1
Y3 e13
X2
1

X21 X22 X23 X24 X25 Response Item

1 1 1 1 1

e10 e9 e8 e7 e6 Error in Measurement

Figure 6: Modeling the Latent Constructs in the Multiple Regression Models

As shown in Figure 6, the latent constructs X1 and X2 are measured using five
questionnaire items respectively, while the latent variable Y is measured using three
questionnaire items. However, in reality, each latent construct could be measured using
as many as ten to twenty questionnaire items.

The model in Figure 6 is valid only if the latent variables X1 and X2 do not have
significant multi-collinearity problem between them. Remember, the main assumption for
Ordinary Least Squares Regression (OLS) is no significant multi-collinearity exists
among the independent variables or exogenous constructs in a model.

29
 Taking multi-collinearity problem into perspective, AMOS software requires the
researcher to estimate the covariance between independent variables or between
exogenous constructs in a model. The program would not run until the researchers
employ the double-headed arrow to link the pair of exogenous constructs in a model to
set the pair as “free parameter estimates” concerning the multi-collinearity effects
between them. The application of double headed arrow linking two independent variables
is shown in Figure 7.

However, if the correlation between X1 and X2 is greater than 0.85, then the
assumption of discriminant validity has failed. It means, one variable is like a mirror of
the other. Thus, the researcher needs to drop one of the two variables from the model and
continue the analysis using a single variable.

X1

1
Y e1

X2

Figure 7: Modeling the Multiple Regressions and Estimating the Correlation between
Independent Variables in AMOS Graphic

30
 If the reader could still recall, the double-headed arrow is used to estimate the
correlational relationship while the single-headed arrow is used to estimate the causal
relationship. In the model shown in Figure 7, the researchers could test the significance of
covariance between X1 and X2. At the same time, the researchers could also test the
significance of causal effect of X1 on Y, and also the causal effect of X2 on Y. All tests
are carried out simultaneously.

e1 e2 e3 e4 e5
1 1 1 1 1

X11 X12 X13 X14 X15

1
e15
Exogenous
X1 Construct 1 1
1 Y1 e11

1
Y Y2 e12

1
Y3 e13
X2
1
Endogenous
Construct

X21 X22 X23 X24 X25

1 1 1 1 1

e10 e9 e8 e7 e6

Figure 8: Modeling the Multiple Regressions and Estimating the Correlation between
Exogenous Constructs in AMOS Graphic

31
 The analysis of correlational and causal relationship for the model in Figure 8 is
equivalent to the analysis stated in Figure 7. The advantage of analysis as stated in Figure
8 is the researcher could assess the importance of each item in measuring their underlying
latent construct. In short, the researcher could assess which item contributes more
information in measuring their respective construct. In SEM, the researchers could even
test the significance of each response item on its respective construct.

1.9 MODELING MULTIPLE VARIABLES IN AMOS GRAPHIC:

THE MULTIPLE REGRESSION MODEL

The Multiple Regression Analysis Y = Bo + B1X1 + B2X2 + B3X3 + e1

Again, the researchers could model the observed variables using Ordinary Least Squares
(OLS) regression and analyze using ANOVA. For the same problem, the researchers
could model the equation in AMOS software as shown in Figure 9. The result of two
methods would be identical. However, the output from AMOS is much more informative
and friendly.

Figure 9: The Multiple Regression Models for the Observed Variables in AMOS Graphic

Key: X1, X2, X3 = Independent variables, Y = dependent variable, e1= residual

32
The Multiple Regression Models for Latent Constructs

AMOS Graphic can model the relationship among latent constructs with multiple items.
In this case, more than one measurement models are involved. The researchers need to
validate each of these measurement models prior to running structural model. Thus,
instead of modeling the OLS, the researchers are modeling SEM as shown in Figure 10.

Figure 10: The Structural Equation Modeling for the Latent Constructs in AMOS Graphic

33
Key: X1 = Exogenous latent construct, X11 to X15 = a set of 5 items to measure X1

e1 to e5 = error in measurement for items X11 to X15

X2 = Exogenous latent construct, X21 to X25 = a set of 5 items to measure X2

e6 to e10 = error in measurement for items X21 to X25

X3 = Exogenous latent construct, X31 to X35 = a set of 5 items to measure X3

e11 to e15 = error in measurement for items X31 to X35

Y = Endogenous latent construct, Y1 to Y5 = a set of 5 items to measure Y

e16 to e20 = error in measurement for items Y1 to Y5

e21= residual

1.10 MODELING THE MEDIATOR VARIABLE IN AMOS

GRAPHIC

Once the regression relation exists and the direct effect of X1 on Y is significant, the
researchers could determine a variable that mediates the relationship between X1 and Y.
This variable is called a mediator. The role of a mediator is providing an indirect effect
of X1 on Y. Thus, the researcher needs to test the significance of a mediator in the X1
and Y relationship. The method of path analysis using OLS is quite tedious. However, the
testing procedure of path analysis is much easier in SEM. Let X1, Y and M be an
independent variable, a dependent variable, and a moderator variable respectively. Refer
to Figure 11.

34
 To begin with, the simple effect of X1
on Y has to be significant

Mediator

Mediator M enters the model

Figure 11: Modeling the Mediator Variable M in AMOS Graphic

In Figure 11, X1 is an independent variable, Y is a dependent variable, and M is a

mediating variable. All variables in the model are directly observed.

The regression equations involved:

Y = Bo + B1X1 + B2M + e2 ...(1)

Y = Bo + B1X1 + e2 ...(2)

Y = Bo + B2M + e2 ...(3)

M = Bo + B3X1 + e3 ...(4)

35
 Using OLS, the researcher needs to analyze all four regression equations
separately in determining the mediating effect of M. The analysis would be quite tedious.

However, in SEM the researcher could include those four regression equations
simultaneously in one model. The researcher could also convert the schematic diagram
into a model in AMOS. Furthermore, the output from AMOS and the subsequent analysis
is simple, informative, and presentable. Now let’s discuss in detail the process involved
in testing the effect of mediating variable. Our discussion centers on the schematic
diagram showing the mediating variable in a model, as shown in Figure 12. In the
diagram, the researcher is interested to assess the effects of mediator variable M in
linking the relationship between X1 and Y.

Key: The coefficient B1 would reduce when the mediator M enters into the
model. If it reduces and become non- significant, then the full mediation occurs.
However, if it reduces but still significant, then the partial mediation occurs. As
for B2 and B3, both of them must be significant for a mediation to occur.

Figure 12: The Diagram Showing B1, B2, and B3 in the Analysis for Mediator Variable

36
The schematic diagram in Figure 12 reveals the following regression equations:
Y = Bo + B1X1 + e is the path from X1 to Y (represented by B1).

Y = Bo + B2M + e is the path from X2 to Y (represented by B2).

M = Bo + B3X1 + e is the path from X1 to M (represented by B3).

The path analysis to assess the effect of M in mediating the relationship between
X1 and Y could result in one of the three following possibilities:

1. M plays a complete mediation role in the relationship between X1 and Y.

2. M plays a partial mediation role in the relationship between X1 and Y.
3. M plays no mediation role in the relationship between X1 and Y.

The complete mediation role of M occurs only if these conditions are met (refer to
Figure 12):

1. The hypothesis testing for regression coefficient B1 is not significant.

2. The hypothesis testing for regression coefficient B3 is significant.
3. The hypothesis testing for regression coefficient B2 is significant.

The partial mediation role of M occurs only if these conditions are met (refer to Figure
12):

1. The hypothesis testing for regression coefficient B1 is still significant

2. The hypothesis testing for regression coefficient B3 is significant.
3. The hypothesis testing for regression coefficient of B2 is significant.
4. The absolute value of B3 x B2 is higher than the absolute value of B1.

The no mediation role of M occurs if at least one of these three conditions is met (refer
to Figure 12):

1. The hypothesis testing for regression coefficient B3 is not significant.

2. The hypothesis testing for regression coefficient B2 is not significant.
3. Both regression coefficients B2 and B3 are not significant.

37
Question: What if both coefficients B3 and B2 are significant but B3*B2is lower than B1?

In this case, one needs to compare the value of B1 in the single model (X1 alone)
with its value when the mediator M enters the model. If its value is reduced when the
mediator is included, then the partial mediation occurred.

AMOS could also analyze the mediating effects of latent construct in a model.
The theoretical model is illustrated in Figure 13. In the model, construct X1 has five
items, and the mediator M also has five response items, while Y has three response items.

In Figure 13, the researcher models the mediating effect of construct M in linking
the relationship between construct X1 and construct Y. So, in this diagram, X1 is an
exogenous construct (arrow pointing out), and Y is an endogenous construct (arrow
pointing in), while M is a mediating construct (two arrows involved - one is pointing in
and another one is pointing out). In this model, the researcher is interested to assess
whether construct M is a really a significant mediator in the X1 to Y relationship.

38
 Figure 13: Modeling the Mediating Effect of Latent Construct M in AMOS Graphic

The hypothesis testing involved in determining whether construct M has full

mediation, partial mediation, or no mediation role is similar to the explanation given for
Figure 12. In AMOS Graphic, one can model more than mediator in a model. The model
with more than one mediator is shown in Figure 14. In Figure 14, the independent
variable (Leverage) and dependent variable (Demand) are observed directly, while the
two mediators are latent constructs.

39
 Figure 14: The Model Contains More than One Mediators Namely RTP and Attitude

1.11 Modeling the Moderating Variable in AMOS Graphic

Sometimes, the researcher is also interested to assess the moderating effects of certain
variable in the model, normally demographic characteristics of the respondents. As its
name implies, the role of a moderator variable is to moderate the relationship between the
independent and its corresponding dependent variable. The position of a moderating
variable in a schematic diagram of a model is presented in Figure 15:

40
 Figure 15: The Moderating Variable M in a Schematic Diagram of a Model

If you could recall from the earlier explanation, the single headed arrow
originating from the independent variable and pointing to its dependent variable indicates
the causal effects of X on Y that is being estimated. Now, the existence of variable M in
the path could play a significant role in altering the effects of independent variable X on
its corresponding dependent variable Y.

For example, the effectiveness of certain teaching method in improving the

academic performance of school children could depend on the background of the
respondents under study. Here, teaching method applied is an independent variable,
academic performance is a dependent variable, while background or demographic
characteristics of the children is a moderating variable. Let’s see how the above theory is
presented in a schematic diagram of a model as shown in Figure 16.

41
 Figure 16: The Moderating Variable M in a Schematic Diagram of a Model

Analyzing the moderation effects using the traditional Ordinary Least Squares
(OLS) is quite tedious and sometimes can be misleading. However, AMOS could handle
this job quite easily. The researcher needs to draw the AMOS graphic as shown in Figure
18 and Figure 19, execute the software, analyze the output and interpret the results. Let’s
assume that variable T is Teaching Method, variable M is Family Background of
Children, and variable P is their Academic Performance. The schematic diagram is given
in Figure 17:

42
 Figure 17: The Model Showing the Independent, Dependent and Moderating Variables

The AMOS Graphic model representing the schematic diagram in Figure 17 is

presented in Figure 18 and Figure 19.

Modeling the Moderator in the Model of Observed Variables

First of all, the study needs to prove that the regression effect of variable T on variable P
is significant. Refer to Figure 18.

Figure 18: Modeling the Effect of Variable T on Variable P in AMOS Graphic

43
 In analyzing the effect of moderator variable B in the observed model, the
researcher needs to compute the interaction effect between independent variable T and
moderator variable M. The product of T multiply by M is termed as TM. Now the model
will estimate the effect of T, the effect of M, and the interaction effect between T and M,
termed as TM as shown in Figure 19.

Figure 19: Modeling the Effects of Moderating Variable M using AMOS Graphic

In order to prove that the child’s Family Background (M) is a significant

moderator in the relationship between Teaching Method (T) and child’s Academic
Performance (P), the study needs to prove the hypothesis of causal effects as follows:

1) The effect of T on P is reduced when moderator variable M enters the model.

2) At the same time, the effect of interaction TM on P is statistically significant.

44
Modeling the Moderator in the Model of Latent Constructs

Modeling the moderator in the model consisting latent constructs is not as easy as
modeling it in the observed variables. First of all, the researcher needs to determine the
path where the moderator effect is to be analyzed. In the first place, the effect of
exogenous construct on the endogenous construct in that particular path must be
significant. The modeling of moderator for latent constructs is shown in Figure 19.

Figure 19: Modeling the Moderator M in a Model Consisting of Latent Constructs

Analyzing the moderator in the latent constructs model is not as straight forward
as in the observed model. The method used in the analysis is called the Multi-Group
CFA. The procedure for analyzing a moderator for latent constructs is explained in
Chapter 7.

45
 CHAPTER 2

THE MODELS INVOLVED IN STRUCTURAL

EQUATION MODELING
In dealing with SEM, there are two models involved in the analysis namely:

i) The measurement model – the model that demonstrates the relationship between
response items and their underlying latent construct. The researcher needs to assess
this model for unidimensionality, validity, and reliability prior to modeling the
structural model.
ii) The structural model – the model that demonstrates the inter-relationships among
constructs in the study. The constructs are assembled into the structural model
based on the hypothesis stated in the theoretical framework.

2.1 THE ROLE OF THEORY IN STRUCTURAL EQUATION

MODELING
The researcher begins by specifying a theoretical model on the basis of theory. Every
construct in the model is conceptualized as a latent, and the researcher is measuring these
constructs using multiple items in a questionnaire. These items can be developed by
researchers, or adopted from literatures customized to suit the field of present study. The
researcher analyzes the measurement model for each construct to assess how well the
generated items represent their underlying construct. The analysis is carried out using
Confirmatory Factor Analysis (CFA). Through the process of CFA, the issues of
construct validity and reliability would be addressed.

The researcher examines the factor loading for each item and the fitness indexes
for the construct. The item with low factor loading that causes poor fitness indexes for the
construct should be deleted from measurement model. After deletion the model is re-
46
specified, and the fitness indexes would improve. If low factor loadings items are deleted
but the fitness indexes for the construct is still low, then the researcher needs to obtain the
Modification Indices (MI). MI indicates the correlation between a pair of items in a
measurement model. High MI means the items are redundant. The researcher needs to
delete one of the two items in order to improve the model fit.

2.2 THE MEASUREMENT MODEL FOR A LATENT

CONSTRUCT

An example of a measurement model is given in Figure 1. Here, the latent construct is

“Service Quality”, which cannot be measured directly. Instead, the Service Quality
construct is measured using a set of ten items in a questionnaire. The ten measuring items
are enclosed in Table 1. Caution: The number of items employed to measure a latent
construct should be a minimum of four in order to avoid the “model identification
problem” during the analysis part.

47
 Service Quality
is a latent
construct

Figure 1: The Measurement Model for Service Quality Construct using Ten Measuring
Items

Table 1: The Measuring Items for Service Quality Construct

This university......... Strongly Strongly

Disagree Agree
Q1 has an adequate infrastructure facilities 1 2 3 4 5 6 7 8 9 10
Q2 has a beautiful campus environment 1 2 3 4 5 6 7 8 9 10
Q3 has a good campus transportation system 1 2 3 4 5 6 7 8 9 10
Q4 has adequate academic facilities 1 2 3 4 5 6 7 8 9 10
Q5 has comfortable cafeteria or food courts 1 2 3 4 5 6 7 8 9 10
Q6 has good sports facilities 1 2 3 4 5 6 7 8 9 10
Q7 has conducive rooms in hostels 1 2 3 4 5 6 7 8 9 10
Q8 has library with adequate materials 1 2 3 4 5 6 7 8 9 10
Q9 offers a flexible class schedule 1 2 3 4 5 6 7 8 9 10
Q10 has an effective information delivery system 1 2 3 4 5 6 7 8 9 10
Source: Research Methodology and Data Analysis 2nd Edition by Zainudin Awang (2012)

48
 Another example of a measurement model is given in Figure 2. Here, the latent
construct is “customer satisfaction”, which cannot be measured directly. Instead, the
customer satisfaction construct (presented in Table 2) is measured using a set of five
items in a questionnaire.

e1 e2 e3 e4 e5
1 1 1 1 1

Q1 Q2 Q3 Q4 Q5

Measuring Item
Customer
Latent Construct
Satisfaction

Figure 2: The Measurement Model for Customer Satisfaction Construct

Table 2: The Measuring Items for Customer Satisfaction Construct

As a student of this university, I am….. Strongly Strongly

Disagree Agree
Q1 satisfied with the system at this university 1 2 3 4 5 6 7 8 9 10
Q2 satisfied with academic rules and regulations 1 2 3 4 5 6 7 8 9 10
Q3 satisfied with the process in obtaining 1 2 3 4 5 6 7 8 9 10
references
Q4 satisfied with academic facilities provided 1 2 3 4 5 6 7 8 9 10
Q5 satisfied with the security system on campus 1 2 3 4 5 6 7 8 9 10
Source: Research Methodology and Data Analysis 2nd Edition by Zainudin Awang (2012)

49
 In the measurement model, the researcher models a latent construct and its
measuring items. As presented in Figure 2, the latent construct is Customer Satisfaction
and it is measured using item Q1 to Q5. The researcher analyzes a measurement model
using CFA to assess the meaningfulness of its items in measuring the construct. In
literatures, one would find the measuring items for latent construct are also called latent
indicators, indicator variables, or manifesting variables.

2.3 THE STRUCTURAL MODEL IN AMOS GRAPHIC

In the structural model, the researchers assemble the constructs involved in the study. The
positions of these constructs in the structural model would depend on how these
constructs are theorized to relate to each other – or in other words, depending on the
schematic diagram of the model. The beauty of AMOS is its ability to draw the schematic
diagram of a model directly into its graphical interface. Furthermore, the analysis and its
corresponding output are presented exactly on the graphic diagram. To put it in a more
appealing perspective – with AMOS, what you see is what you get.

2.3.1 The Structural Model: Modeling the Correlational

Relationship

An example of a structural model is given in Figure 3. This AMOS graphic presents the
correlational relationship between Service Quality construct (with ten response items)
and Customer Satisfaction construct (with five response items).

In this structural model (Figure 3), the researcher intends to estimate the
correlational relationship between Service Quality and Customer Satisfaction. The
Service Quality construct is measured using ten response items, while the Customer
Satisfaction construct is measured using five response items.

50
 Remember, the correlational relationship is represented by a double-headed arrow
in AMOS graphic interface. This arrow should be drawn in the clockwise direction to
link one latent construct to another latent construct.

Correlational Path

Figure 3: The Structural Model Linking Service Quality to Customer Satisfaction. In

AMOS, the Double-headed Arrow is Used to Estimate the Correlation between the Two
Latent Constructs

51
2.3.2 The Structural Model: Modeling the Causal
Relationship

The example of a structural model for a causal relationship is given in Figure 4. This
AMOS graphic measures the causal effects of service quality construct (with ten
measuring items) on customer satisfaction construct (with five measuring items). Both
constructs are latent. The measuring items for service quality construct and customer
satisfaction construct are presented in Table 1 and Table 2, respectively.

Causal

Path

Figure 4: The Structural Model Linking Service Quality to Customer Satisfaction. The
Single Headed Arrow is Used to Estimate the Causal Effects of Service Quality on Customer
Satisfaction

52
 In the above structural model (Figure 4), the researchers are interested to
determine the causal effect of service quality construct on customer satisfaction construct.
Remember: A causal effect is represented by a single-headed arrow. The arrow is drawn
from the latent exogenous construct pointing to the latent endogenous construct. The
position of the arrow itself indicates the exogenous construct is hypothesized to have
significant causal effect on the endogenous construct. Service Quality is a latent
exogenous construct, while Customer Satisfaction is a latent endogenous construct.

2.3.3 The Structural Model: Modeling the Mediator

The example of a structural model with a mediator construct is given in Figure 5. This
AMOS graphic presents the mediating role of Customer Satisfaction construct. The
causal arrow from Service Quality construct pointing towards Customer Satisfaction
construct indicates this particular study believes that Service Quality performance has a
causal effect on Customer Satisfaction.

In the meantime, another arrow originates from Customer Satisfaction construct

pointing towards the Customer Loyalty construct indicate that the level of Customer
Satisfaction is believed to have a causal effect on Customer Loyalty.

53
 Mediator

Construct

Endogenous
Construct
Exogenous
Construct

Figure 5: The Structural Model Linking Service Quality, Customer Satisfaction and
Customer Loyalty

In Figure 5, the researcher is interested to assess the mediating role of Customer

Satisfaction in linking Service Quality to Customer Loyalty. Here, Customer Satisfaction
is a mediator construct.

The researcher models the mediating construct of Customer Satisfaction (Figure

5). The regression equations involved this structural model are:

1. Customer Loyalty is a function of Service Quality.

2. Customer Loyalty is a function of Customer Satisfaction.
3. Customer Satisfaction is a function of Service Quality.

54
2.4 The Types of Construct in Structural Equation Modeling

Apart from the observed variable, there are a few types of latent construct involved in
SEM. Among the types of construct that researchers need to know are Reflective
Construct, Formative Construct, and Second Order Construct.

2.4.1 The Reflective Construct

As its name implies, the influence of this construct “reflects” on its response items. Thus
the single-headed arrow flows from the construct onto its items. This construct is
measured using certain number of items in a questionnaire. Most of the times, researchers
are dealing with the Reflective Constructs in the structural model. Among the examples
of Reflective Constructs are:

i) Customers’ Perceived Quality concerning their service experience – measured using

certain number of items in a questionnaire.
ii) Customers’ Perceived Usefulness towards certain cosmetic product– measured using
certain number of items in questionnaire.
iii) Employee’s Knowledge – the knowledge obtained from undergoing certain training
programs.
iv) Employee’s Skills – the perceived skills acquired after undergoing certain training
programs.
v) Employee’s Attitude – the attitude towards task after undergoing certain training
programs.

55
 A Reflective Construct:
Arrows flow from
Construct to the items

Figure 6: The Measurement Model of the Reflective Constructs in AMOS Graphic

2.4.2 The Formative Construct

Formative Construct is a latent construct which is “formed” by certain number of

variables. The variables forming the formative construct are directly measured, thus no
measurement error is required.

56
 Among the examples of latent formative constructs are:

i) Socio-Economic Status of respondents, which is formed by three factors namely

Education, Occupation, and Income.
ii) Student’s Academic Performance, which is formed by three assessments made on the
student namely Individual Assessment, Group Assessment and Final Exam.
iii) Company Performance, which is formed by three financial measures namely Return on
Investment (ROI), Return of Assets (ROA) and Return on Equity (ROE).

The Formative Construct could also become exogenous, as well as endogenous

construct in the structural model. The illustration of Formative Constructs in AMOS
Graphic is given in the following figure.

57
 Variables formed the construct.
No measurement error since the
variables are directly observed. Formative Construct: The arrow
flows from variables to form the
construct.

One of the variables must

have a reference point 1.

Figure 7: Modeling the Formative Construct in AMOS Graphic

58
2.4.3 The Second Order Construct

The Second Order Construct is a construct which consists of a few First Order
Constructs. The First Order Construct is measured using certain number of items in a
questionnaire. Sometimes the First Order Construct is called the Main Construct, while
the First Order Constructs are the Components of the main construct. Among the
examples of Second Order Construct are:

Measurement error of
the sub-construct

Measurement error
of the item

The Second Order Construct

- Arrow flows from main
construct to sub-constructs

First Order
Construct or
Sub-Construct

Figure 8: Modeling the Second Order Construct in AMOS Graphic

59
Example: AMOS Graphic can take all types of construct in the model

The model in Figure 9 consists of all three types of construct that we have just learned.

First order construct

STRESS has eight items Second order construct
HEALTH has
two sub-constructs

Formative
construct SES is
measured using
three variables

Figure 9: The Model in AMOS Graphic which Consists of Three Types of

Constructs

60
 CHAPTER 3

VALIDATING THE MEASUREMENT MODEL: CFA

The measurement model of a latent construct:

SEM is a confirmatory method providing a comprehensive means for assessing and
modifying the measurement model of a latent construct. The procedure is called CFA.
The method has the ability to assess the unidimensionality, validity and reliability of the
measurement model (construct).

Remember - the researchers need to confirm the unidimensionality, validity, and

reliability for all latent constructs involved in the study before modeling their inter-
relationship in a structural model (SEM). However, the unidimensionality assessment
should be made first prior to assessing validity and reliability.

With Confirmatory Factor Analysis (CFA), any item that does not fit the
measurement model due to low factor loading should be removed from the model. The
researchers need to perform the CFA for all latent constructs involved in a model. The
researchers could run the CFA for every measurement model individually or using a
pooled CFA. However, the pooled CFA procedure is more preferred. Thus, this
procedure for model assessment will be demonstrated.

The assessment for each element is done as follows:

1. Unidimensionality

Unidimensionality is achieved when the measuring items have acceptable factor loadings
for the respective latent construct. In order to ensure unidimensionality of a measurement
model, any item with a low factor loading should be deleted.

61
How low is considered to be “too low” and has to be deleted from the measurement
model?

a. For a newly developed item, the factor loading for an item should be 0.5 or higher,
or
b. For an already established item, the factor loading for an item should be 0.6 or
higher.

The deletion should be made one item at a time with the lowest factor loading item
to be deleted first. After an item is deleted, the researcher needs to run the new
measurement model. The process continues until the unidimensionality requirement is
achieved.

2. Validity

Validity is the ability of instrument to measure what it supposed to be measured for a

construct. Three types of validity required for each measurement model are:

a. Convergent validity. This validity is achieved when all items in a measurement

model are statistically significant. The convergent validity could also be verified by
computing the Average Variance Extracted (AVE) for every construct. The value of
AVE should be 0.5 or higher to achieve the Convergent Validity.

b. Construct validity. This validity is achieved when the Fitness Indexes for a
construct achieved the required level. The fitness indexes and the level of
requirement are presented in Table 1.

c. Discriminant validity. This validity is achieved when the measurement model is

free from redundant items. AMOS will identify the pair of redundant items in the
model in term of high Modification Indices (MI). The researcher could delete one
of the items and run the model. The researcher could also set the correlated pair as
“free parameter estimate”. Another requirement for discriminant validity is the
correlation between exogenous constructs should be less than 0.85.

62
3. Reliability

Reliability is the extent of how reliable is the said measurement model in measuring the
intended latent construct. The assessment for reliability for a measurement model could
be made using the following criteria.

a. Internal reliability – This reliability is achieved when the Cronbach’s Alpha

coefficient is greater than 0.7 (calculated in SPSS).
b. Composite Reliability – The measure of reliability and internal consistency for a
latent construct. A value of CR > 0.6 is required in order to achieve composite
reliability for a construct. (CR is calculated using the given formula).
c. Average Variance Extracted – The average percentage of variation as explained by
the measuring items for a construct. An AVE > 0.5 is required (AVE is calculated
using the given formula).

AVE= ∑ Қ2 / n Қ = Factor loading of every item

CR = (∑Қ)2 / [(∑Қ)2 + (∑1- Қ2)] n = Number of items in a model

3.1 EVALUATING THE FITNESS OF A MODEL

(MEASUREMENT MODEL AND STRUCTURAL MODEL)

In SEM, there is several Fitness Indexes that reflect how fit is the model to the data at
hand. However there is no agreement among researchers which fitness indexes to use.
Hair et al. (1995, 2010) and Holmes-Smith (2006) recommend the use of at least one
fitness index from each category of model fit. There are three model fit categories namely
absolute fit, incremental fit, and parsimonious fit.

The choice of index to choose from each category to report depends on which
literature is being referred. The information concerning the model fit category, their level
of acceptance, and comments are presented in Table 1.
63
 Table 1: Index Category and the Level of Acceptance for Every Index

Name of category Name of index Level of acceptance Comments

1. Absolute fit Chisq P > 0.05 Sensitive to sample size >200

RMSEA RMSEA < 0.08 Range 0.05 to 0.1 is

acceptable

GFI GFI > 0.90 GFI = 0.95 is a good fit

2. Incremental fit AGFI AGFI > 0.90 AGFI = 0.95 is a good fit

CFI CFI > 0.90 CFI = 0.95 is a good fit

TLI TLI > 0.90 TLI = 0.95 is a good fit

NFI NFI > 0.90 NFI = 0.95 is a good fit

3. Parsimonious fit Chisq/df Chi square/ df < 5.0 The value should be less than
5.0.

***The indexes in bold are recommended since they are frequently reported in literatures

The acceptable cut-off values reported by researchers may vary depending on

literatures support they are referring to. However, the following table presents the
literature support for the widely employed fitness indexes.

64
 Table 2: The Literature Support for the Respective Fitness Index

Name of category Name of index Index full name Literature

1. Absolute fit Chisq Discrepancy Chi Square Wheaton et al. (1977)

RMSEA Root Mean Square of Browne and Cudeck

Error Approximation (1993)

GFI Goodness of Fit Index Joreskog and Sorbom

(1984)

2. Incremental fit AGFI Adjusted Goodness of Tanaka and Huba (1985)

Fit

CFI Comparative Fit Index Bentler (1990)

TLI Tucker-Lewis Index Bentler and Bonett

(1980)

NFI Normed Fit Index Bollen (1989b)

3. Parsimonious fit Chisq/df Chi Square/Degrees of Marsh and Hocevar

Freedom (1985)

*** One could ignore the absolute fit index of minimum discrepancy chi-square if the
sample size obtained for the study is greater than 200 (Hair et al., 1996; Joreskog and
Sorbom, 1996).

An example a measurement model of a latent construct is presented in Figure 1.

65
e1 e2 e3 e4 e5 e6 e7 e8 e9
1 1 1 1 1 1 1 1 1

Wl 1 Wl 2 Wl 3 Wl 4 Wl 5 Wl 6 Wl 7 Wl 8 Wl 9

1
Response Item
Work
Load
Reference

Point
Figure 1: The Measurement Model for Measuring Employee Workload Consists of Nine
Items

Note: Workload is a latent construct (represented in ellipse). This latent construct

is measured using nine questionnaire items namely wl 1 to wl 9 (represented by
rectangles since they are observed score). In Figure 1, wl 1 to wl 9 are the response items
for the construct while e1 to e9 are their respective measurement errors. The data (wl 1 to
wl 9) are inserted into the model through the “click-and-drag” procedure.

Through CFA, the researcher instructs AMOS to calculate the standardized

estimate and squared multiple correlation by clicking the respective box in the analysis
properties menu. The standardized estimates indicate the factor loading for each item in a
measurement model.

The output showing the factor loading and squared multiple correlation for every
item in a measurement model are presented in Figure 2.

66
e1 e2 e3 e4 e5 e6 e7 e8 e9

.50 .53 .64 .67 .50 .68 .60 .42 .56

Wl 1 Wl 2 Wl 3 Wl 4 Wl 5 Wl 6 Wl 7 Wl 8 Wl 9

.71 .77
.73 .80 .82 .83 .65
.70 .75
Response
Item
Response
Work R2

Factor Loading Load Latent Construct

Figure 2: The Factor Loading for Every Item in the Measurement Model for Employee’s
Workload

Figure 2 presents the factor loading for each item in a measurement model to
measure the latent construct namely workload. The factor loading for a particular item is
shown near the arrow pointing to the respective item, while the value shown above for
each response item is the squared multiple correlation or R2 for that particular item.

Any item having a factor loading less than 0.6 and an R2 less than 0.4 should be
deleted from the measurement model of a construct. However, the researcher may not do
so if the Fitness Indexes for that measurement model has already achieved the required
level as shown in Table 1. An item having low factor loading simply means that
particular item is deemed useless to measure that particular construct. Keeping useless
item in a model will affect the Fitness Index of the model.

67
3.2 THE STEPS INVOLVED IN CFA FOR THE
MEASUREMENT MODEL OF A LATENT CONSTRUCT

1. Run Confirmatory Factor Analysis (CFA) for the measurement model.

2. Examine the required Fitness Indexes for the measurement model (refer to Table 1.
If the Fitness Index does not meet the required level, examine factor loading).
3. Delete an item with factor loading less than 0.6 (suggested by the literatures).
4. Delete one item at a time (select the lowest factor loading to delete first).
5. Run this new measurement model (the model after an item is deleted).
6. Examine the Fitness Indexes – repeat step 3-5 until the Fitness Indexes are
achieved.
7. If the Fitness Index is still not achieved, look at the Modification Indices (MI).
8. High value of MI (above 15) indicates there are redundant items in the model.

(The MI indicate a pair of items which is redundant in the model)

To solve the redundant items, the researcher could choose one of the following:

Choice 1:

a. Delete one of the item (choose the lower factor loading).

b. Run the measurement model and repeat the above steps.

Choice 2:

a. Set the pair of redundant item as “free parameter estimate”.

b. Run the measurement model and repeat the above steps.

9. Obtain the Cronbach’s Alpha, CR, and AVE for the final measurement model.
10. Report the normality assessment for all measurement models involved.

68
 Note: The acceptable value of factor loading, the level of fitness indexes, and also
the method of modification to the measurement model varies among the literatures.

3.3 THE CONFIRMATORY FACTOR ANALYSIS (CFA)

Confirmatory Factor Analysis (CFA) is a special form of factor analysis. It is employed

to test whether the measures of a construct are consistent with the researcher’s
understanding of the nature of that construct. The CFA procedure replaced the older
methods to determine construct validity. Every measurement model of a latent construct
needs to undergo CFA before modeling in SEM.

Example: Let’s assume that the model has two exogenous constructs; X1 and X2,
and one endogenous construct; Y. The representation of the constructs in a schematic
diagram is shown in Figure 3. In this study, the researcher is interested to determine the
causal effect of two exogenous constructs (X1 and X2) on one endogenous construct (Y).

X1

X2

Figure 3: A Schematic Diagram of a Model

Key: X1 and X2 are latent exogenous constructs (represented by an ellipse).

Y is the latent endogenous construct (represented by an ellipse).

69
 In this example, the exogenous latent construct X1 has five items, and X2 has four
items, respectively. The measurement model for X1 is shown in Figure 4, while the
measurement model for X2 is shown in Figure 5. Figure 6 presents the measurement
model for endogenous construct Y. The structural model is presented in Figure 7.

e1 e2 e3 e4 e5

1 1 1 1 1

X11 X12 X13 X14 X15

X1

Figure 4: A Latent Exogenous Construct X1 has Five Items

70
 e1 e2 e3 e4

1 1 1 1

X21 X22 X23 X24

X2

Figure 5: A Latent Exogenous Construct X2 has Four Items

e1 e2 e3 e4

1 1 1 1

Y1 Y2 Y3 Y4

Figure 6: A Latent Endogenous Construct Y has Four Items

71
 1
e5 X15

1
e4 X14

1
e3 X13 X1 e14
1
Y1 e10
1 1 1
e2 X12
1
1 Y2 e11
1
e1 X11 Y
1
Y3 e12
1
e9 X24
1
Y4 e13
1
e8 X23

X2
1
e7 X22

1 1
e6 X21

Figure 7: The Structural Model in AMOS for the Schematic Diagram Shown in Figure 3

3.31 Validating the Pooled Measurement Model

The researcher can assess and validate the measurement model of more than one
latent constructs at a time. Actually, the measurement model for all constructs involved in
the study should be assessed together at once if this is possible. This method is called
pooled CFA. However, in many studies, this is not possible when the model has too many
latent constructs. In the Pooled CFA, the item-deletion process is made for every
construct by selecting the item having the lowest factor loading in each construct to be
deleted.

72
3.32 The Measurement Model for Pooled Constructs

Figure 8: The Measurement Model combining All Latent Constructs Simultaneously

73
 The CFA results showing fitness indexes and factor loading for every item
together with its R2 are presented in Figure 9. Using this method, the correlations
between constructs are computed simultaneously. If one has too many constructs and thus
cannot pool them together into one measurement model, he can always pool the
constructs into two separate measurement models.

Factor Loading R2

Correlation
between Constructs

Figure 9: The Factor Loading for All Items of the Respective Construct

As shown in Figure 9, certain fitness indexes for the pooled constructs do not
achieve the required level. When we examine the factor loading, we found the factor
loading for item X12 (from construct X1), and item X24 (from construct X2) are below 0.6.
These two “useless items” has caused the measurement model for the constructs to be
poorly fit. Therefore, one has to delete these two items and run the new measurement

74
model. Now the two items namely X12 and X24 were deleted and the new measurement
model is run and shown in Figure 10.

.36

X11 e1
.60

.83
.91
X1 X13 e3
.88 .77

X14 e4
.93
.86

X15 e5
.86
.60
X21 e6
.93
.93

.97 X22 e7
.56
X2 .65
.42

X23 e8

.46 Certain index is

.91 still not
Y1 e10 achieved the
.96
.95
required level
.98 Fitness Indexes
Y2 e11 1.ChiSq = 122.335
2.df = 41
Y .94
.89
3.ChiSq/df = 2.984
Y3 e12 4.GFI = .935
5.AGFI = .895
.63 6.CFI = .977
.80
7.RMSEA = .079
Y4 e13

Figure 10: The New Factor Loading after the Two Items were Deleted

75
 The fitness index for AGFI (Figure 10) is still below the required level even
though the factor loading for all items are above 0.6. Thus, one might suspect that certain
items are redundant of each other in the measurement model. The items redundancy can
be examined through inspecting the Modification Indexes (MI). Table 3 presents the MI
for a pair of correlated errors which reflect redundant items exist in the model. The MI
value of 26.79 is considered high since it is greater than 15.0. The correlated
measurement error here is between e12 and e13. If we look at the items, the redundant
item is between Y3 and Y4. These items have caused the measurement model to have a
poor fit.

Table 3: The Modification Indices Presents the Covariance between each Pair of Items
(Redundant Items is Shown through The Correlated Measurement Error of The Respective
Item)

M.I. Par Change Comment

MI > 15 indicates item 3 and item 4 are
e12 e13 26.789 0.127
redundant
e10 e13 4.643 -0.044
e5 e7 4.568 -0.032
e1 e7 10.890 0.085

The researcher needs to modify the measurement model since the fitness index
AGFI does not meet the required level 0.9 due to redundant items. In dealing with
redundant items in the model, the researcher has two choices:

Choice 1: to delete one of the two redundant items and run the new measurement model.

Choice 2: to set these two correlated measurement errors of redundant items as a “free
parameter” and run the new measurement model.

Let’s say, the researcher decides to employ choice 2 for this particular case. The
result is presented in the following diagram (Figure 11).

76
 The fitness
indexes have
improved after
modification.

Figure 11: The New Measurement Model after e12 and e13 are Set as “Free Estimate”

The fitness indexes for the model are assessed in the following table (Table 4).

77
 Table 4: The Fitness Indexes for new Measurement Model

Name of category Name of index Index value Comments

1. Absolute fit RMSEA 0.065 The required level is achieved
GFI 0.949 The required level is achieved

2. Incremental fit CFI 0.985 The required level is achieved

3. Parsimonious fit Chisq/df 2.344 The required level is achieved

Note: Look how the fitness indexes have improved after the two redundant items are
constrained in the model.

3.33 ASSESSING THE VALIDITY AND RELIABILITY FOR A

POOLED MEASUREMENT MODEL

Once the CFA procedure for every measurement model is completed, the researcher
needs to compute certain measures which indicate the validity and reliability of the
construct and summarize them in a table. As has been said earlier, the assessment for
unidimensionality, validity, and reliability for measurement models are required prior to
modeling the structural model. The author would like to suggest the following format for
reporting the CFA results:

Unidimensionality: This requirement was achieved through the item-deletion

process for low factor loading item. The new model is run and the item deletion process
is repeated until the fitness indexes achieved the required level. All processes have been
demonstrated.

78
 Validity: This requirement was achieved through the following processes:

1. 1. Convergent Validity: AVE > 0.50: Refer to the following table (refer Table 5)

2. Construct Validity: All fitness indexes for the models meet the required level

3. Discriminant Validity: The redundant items are either deleted or constrained as

“free parameter”, also the correlation between exogenous
construct X1 and X2 is lower than 0.85

Reliability: This requirement was achieved through the following process

1. Internal Reliability: Cronbach’s Alpha > 0.70: Refer to Table 5 (use

SPSS)
2. Composite Reliability: CR > 0.60: Refer to Table 5 (using a formula)
3. Average Variance Extracted: AVE > 0.50: Refer to Table 5 (using a formula)

Table 5: The CFA Results for the Measurement Model

Construct Item Factor Loading Cronbach’s CR AVE

Alpha (Above (Above
(Above 0.7) 0.6) 0.5)
X1 X11 0.60 0.893 0.904 0.707
X12 This item was deleted due to low factor loading
X13 0.91
X14 0.88
X15 0.93
X2 X21 0.93 0.869 0.894 0.743
X22 0.97
X23 0.65
X24 This item was deleted due to low factor loading

79
 Y Y1 0.96 0.946 0.955 0.844
Y2 0.98
Y3 0.94
Y4 0.78

Table 6: The Discriminant Validity Index Summary

Construct X1 X2 Y
X1 0.83
X2 0.60 0.85
Y 0.56 0.46 0.92

The diagonal values (in bold) is the square root of AVE while other values are
the correlation between the respective constructs. The discriminant validity for all
constructs is achieved when a diagonal value (in bold) is higher than the values in its row
and column. Referring to Table 6, one can conclude that the discriminant validity for all
three constructs is achieved.

3.4 THE ASSESSMENT OF NORMALITY FOR THE DATA

After the fitness indexes have been achieved, the researcher needs to examine the
normality assessment for the data at hand before proceeding to modeling the structural
model. Using the final measurement model, select the output box named “test for
normality and outliers” in order to assess the distribution for every variable in a dataset.

The following table (Table 7) presents the resulted output from the
abovementioned procedure. The table presents the normality assessment for every items
involved in the measurement model.

80
 Table 7: The Assessment of Normality Distribution for Items in the Respective Construct

Variable min max skewness c.r. kurtosis c.r.

Y4 1.000 7.000 -0.561 -2.077 -0.230 -0.837
Y3 1.000 7.000 -0.786 -2.713 0.297 1.078
Y2 2.000 7.000 -0.566 -2.116 -0.274 -0.995
Y1 2.000 7.000 -0.505 -1.672 -0.429 -1.559
X23 1.000 7.000 -0.690 -2.015 0.124 0.450
X22 2.000 7.000 -0.362 -2.632 -0.165 -0.600
X21 2.000 7.000 -0.400 -2.909 -0.220 -0.799
X15 2.000 7.000 -0.507 -2.683 -0.386 -1.403
X14 2.000 7.000 -0.535 -2.891 -0.022 -0.078
X13 2.000 7.000 -0.475 -1.453 -0.437 -1.590
X11 1.000 7.000 -0.109 -0.792 -0.158 -0.574
Multivariate 22.887 12.048

The normality assessment can be made by assessing the measure of skewness for
every item. The absolute value of skewness 1.0 or lower indicates the data is normally
distributed. However, SEM using the Maximum Likelihood Estimator (MLE) like Amos
is fairly robust to skewness greater than 1.0 in absolute value if the sample size is large.
Meaning, the researcher could proceed into further analysis (SEM) since the estimator
used is MLE. Normally the sample size greater than 200 is considered large enough in
MLE even though the data distribution is slightly non-normal.

Another method for normality assessment is by looking at the multivariate

kurtosis statistic. However, SEM using Maximum Likelihood Estimator (MLE) is also
robust to kurtotic violations of multivariate normality as long the sample size is large.

81
 If the distribution is found to depart from normality, the researcher could assess
the Mahalanobis distance to identify for the potential outliers in dataset. Amos computes
the distance for every observation in dataset from the centroid. The centroid is the center
of all data distribution. It tabulates the distance of potential outliers from the centroid
together with the probability for an observation suspected to be an outlier in the first
column and the probability that an observation of similar extremity would occur given a
multivariate normal population (the second column).

The outlier occurs when the distance of certain observation is too far compared to
the majority other observations in a dataset. The deletion of few extreme outliers in the
model might improve the multivariate normality. Once the outlier is identified, the
researcher could go back to dataset and get them deleted (based on the observation
number). The new measurement model is re-specified using the cleaned dataset. The
process could be repeated. However, there is no necessity to examine Mahanolobis
Distance if the non-normality issue does not arise.

As a summary, in the case when the normality assumption is not fulfilled, the
researchers still have many options to take. One of them is to remove the non-normal
items from the measurement model (based on the measure of skewness) and continue
with the analysis. Another option is to remove the farthest observation from the center
(outlier) of distribution. However, the most popular method lately is to continue with the
analysis with MLE (without deleting any item and also without removing any
observation) and re-confirm the result of analysis through Bootstrapping.

82
 Bootstrapping is the re-sampling process on the existing dataset using the method
of sampling with replacement. The statistical procedure would compute the mean and
standard deviation for every sample of size n to create the new sampling distribution. The
researcher could instruct Amos to collect 1000 random sample from the dataset and re-do
the analysis. Since the sample size is large (1000), the new sampling distribution would
be closer to normal distribution. Amos would analyze the Bootstrapping data and produce
the confidence intervals as well as the significance for every parameter involved in the
analysis. The researcher could compare the actual results with the bootstrapped results to
confirm the analysis. If the results differ, the bootstrapped result will be acceptable.

83
 CHAPTER 4

ANALYZING THE SEM STRUCTURAL MODEL

After the researchers have addressed the issues of unidimensionality, validity, and
reliability of the latent constructs involved in the study, the next step is to model these
constructs into structural model for analysis using SEM. The normal practice is to
assemble the constructs from left to right. Starting with the exogenous constructs on the
left, followed by the mediating construct (if any) in the middle, and finally the
endogenous construct on the far right.

The arrow to link the constructs is determined by the direction of hypotheses. The
single-headed arrow is used to test the causal effects, while the double-headed arrow is
used to test the correlational effects among constructs. In SEM, the researcher could
model and analyze the multiple relationships among the constructs simultaneously.

4.1 STEPS INVOLVED IN PERFORMING THE STRUCTURAL

EQUATION MODELING (SEM)

There are six steps involved in analyzing the Structural Equation Modeling.

1. Specify the schematic diagram of the research model (based on theory).

2. Convert the schematic diagram of research into AMOS structural model.

3. Collect data: Items for latent constructs are measured using interval scale. The items
could be in the form of statement in a questionnaire.

4. Perform the CFA for measurement models of the latent constructs. This has been
discussed in Chapter 3. Remember: No need to assess CFA for observed variables.

84
5. Assemble the AMOS structural model. Make sure the residual terms and reference
point are put accordingly. Input data and execute the model.

6. Obtain the AMOS Graphic output for the standardized and unstandardized estimate.
Verify the direction of relationship between constructs conform to theory.

Lastly, obtain the AMOS text output of the estimates. Assess the significance of the
path analysis. Interpret the results according to the stated research hypotheses.

4.2 PERFORMING THE STRUCTURAL EQUATION

MODELING (SEM)

For the sake of continuation, we will be using the same example that was used in the
previous chapter. The step by step explanation is made for every procedure. Using the
example, let’s begin the steps in performing SEM.

Step 1: Specify the schematic diagram of the research model (based on the
theory)

Suppose the schematic diagram for the model in the study is presented in Figure
1.

X1

X2

Figure 1: The Schematic Diagram of the Model in a Study

85
Note: X1 and X2 are the latent exogenous constructs, while Y is the latent endogenous construct
in the model. There is no mediating construct involved in this particular study. The oval shapes
are used to model the latent constructs of X1, X2, and Y.

Step 2: Draw the structural model based on the specified diagram. Include
all items for every construct in the structural model

The representation of schematic diagram in AMOS syntax is given in Figure 2.

1
e5 X15

1
e4 X14

1
e3 X13 X1 e14
1
Y1 e10
1 1 1
e2 X12
1
1 Y2 e11
1
e1 X11 Y
1
Y3 e12
1
e9 X24
1
Y4 e13
1
e8 X23

X2
1
e7 X22

1 1
e6 X21

Figure 2: The Representation of a Schematic Diagram of the Model in AMOS Graphic

86
Note: Latent construct X1 is measured using five items namely X11, X12, X13, X14, and X15.

Latent construct X2 is measured using four items namely X21, X22, X23, and X24.

Latent construct Y is measured using four items namely Y1, Y2, Y3, and Y4.

The residual term is needed for construct Y since it is an endogenous construct.

Step 3: Perform the CFA for the measurement model of latent constructs

The CFA procedures for measurement model of latent constructs have been explained in
the previous chapter. The resulted measurement models after CFA are assembled and
presented in Figure 3.

1
e1 X11
1 Reference point
Residual term

1
e3 X13 X1 e14
1
1
Y1 e10
e4 X14 1 1

1
1
Y2 e11
e5 X15
Y
1
1 Y3 e12
e6 X21
1
1
1 Y4 e13
e7 X22

X2
1
e8 X23

Figure 3: The Constructs are Assembled into Structural Model for Path Analysis

87
Step 4: Execute the structural model. Choose the required output for the
structural model in the “output” menu. Obtain the standardized
regression weights and observe the fitness indexes.

The researcher could still modify the structural model if certain Fitness Index is not
achieved. The structural model in Figure 3 is executed after inserting data for the
respective items (click and drag procedure) and selecting the required analysis procedure
to perform.

The standardized estimate for the model is presented in Figure 4, while the
unstandardized estimate is presented in Figure 5. Their corresponding Fitness Indexes
and assessment is presented in Table 1.

.36
Standardized
e1 X11 Factor Loading
Beta
.60
Item R2

.83
.91
e3 X13 X1 e14
.88 .91
.77 R2 Y1 e10
e4 X14 .45 .96
.93
.98 .96
.86 Y2 e11
e5 X15
.60 .34 Y
.88
.86
.19 .94 Y3 e12
e6 X21
.93 .33
.78 .61
.93 Fitness Indexes Y4 e13
e7 X22 .97
1 ChiSq = 93.747
2 df = 40
.42 X2 3 ChiSq/df = 2.344
e8 X23 4 GFI = .949
.65
5 AGFI = .917
6 CFI = .985
7 RMSEA = .065

Figure 4: The Standardized Regression Weights (All Units are in Standard Deviation)

88
 Note: As shown in Figure 4, the standardized beta estimate for effect of X1 on Y is 0.45,
while the standardized beta estimate for X2 is 0.19.

The measure of correlation between exogenous constructs X1 and X2 is 0.60. This

value is important since it indicates that the discriminant validity between exogenous
constructs is achieved and the two constructs are not redundant (correlation is below
0.85).

The value of coefficient of determination R2 is 0.34. The figure indicates the

contribution of exogenous constructs X1 and X2 in estimating the endogenous construct Y
is 34%.

Table 1: The Fitness Indexes Assessment for the Structural Model in Figure 4

Name of category Name of index Index value Comments

1. Absolute fit RMSEA 0.065 The required level is achieved
Absolute fit GFI 0.949 The required level is achieved
2. Incremental fit CFI 0.985 The required level is achieved
3. Parsimonious fit Chisq/df 2.944 The required level is achieved

89
 .87
1
e1 X11 Residual Variance in
Error variance measurement
1.00
Variance error
Actual beta
.23 .76
1 1.51
e3 X13 X1 .49
e14 .11
.28 1.40 1
1
Y1 e10
e4 X14 .68 1 1.00
.05
1.59
.20 1.03 1
1
Y2 e11
e5 X15
.45 Y .19
.20 1
1 .19 1.09 Y3 e12
e6 X21
1.00 .93 .14
1.12
.09 1
1 Fitness Indexes Y4 e13
e7 X22 1.00
1 ChiSq = 93.747
2 df = 40
1.07 X2 1.19 3 ChiSq/df = 2.344
1
e8 X23 4 GFI = .949
.81 5 AGFI = .917
6 CFI = .985
7 RMSEA = .065

Figure 5: The Regression Weights (The Measures of Beta Estimate in its Actual Unit)

Step 5: Obtain the AMOS text output: Examine the standardized regression
weights and regression weights. Interpret the results and make a
decision concerning the hypothesized relationships.

AMOS gives two sets of text output namely the standardized regression weights and
regression weights. The text output explaining the results in Figure 4 are presented in
Table 2, Table 3 and Table 4. Meanwhile, the text output explaining the results in Figure
5 are presented in Table 5, Table 6 and Table 7.

90
4.2.1 Interpreting the AMOS Text Output: The
Standardized Regression Weights

Table 2: The Standardized Regression Weight as Shown in Figure 4

Path Path Standardized Beta Estimate Comment

Y X2 0.190 Standardized beta
Y X1 0.450 Standardized beta

The definition for the standardized regression weight is given below.

Extracted from Table 2:

Path Path Standardized Beta Estimate
Y X2 0.190

When X2 goes up by 1 standard deviation, Y goes up by 0.19 standard deviations.

Extracted from Table 2:

Path Path Standardized Beta Estimate
Y X1 0.450

When X1 goes up by 1 standard deviation, Y goes up by 0.45 standard deviations.

Table 3: The Correlation Estimate for Each Pair of Exogenous Construct

Construct Construct Estimate
X1 X2 0.60

The correlation between latent construct X1 and X2 is estimated to be 0.60.

*The measure of correlation indicates that the strength of the relationship

between the two latent exogenous constructs (X1 and X2) is not strong. In this case, the
researchers could continue with further analysis.

***Caution: If the measure of correlation is higher than above 0.85, the two
latent exogenous constructs (X1 and X2) are highly correlated, and the discriminant
validity failed. Thus the researcher could not treat the two exogenous constructs
separately since they are redundant. In order to avoid redundancy, the researcher
should drop one of these two exogenous constructs (X1 and X2) for further analysis.

91
 Table 4: The Squared Multiple Correlations (R2)

Variable Estimate (R2)

Y 0.341

The definition for the squared multiple correlations are given below.

It is estimated that the predictors of Y explain 34.1 percent of its variance. In

other words, the error variance of Y is approximately 65.9 percent of the variance of
Y itself.

4.2.2 Interpreting the AMOS Text Output: The Regression

Weights

Table 5: The Regression Weight for X1 and X2 in Predicting Y

The Actual Beta
S.E. C.R. P-value
Values
Y X1 0.682 0.106 6.433 0.001

Y X2 0.188 0.062 3.032 0.002

***Indicate highly significant at < 0.001

Extracted from Table 5: Estimate of Regression Weight for X1

The Actual Beta Value S.E. C.R. P-value

Y X1 0.682 0.106 6.433 0.001

Definition for actual beta 0.682: When X1 goes up by 1 unit, Y goes up by 0.682 unit.

Definition for Standard Error of regression weight 0.106: The regression weight estimate,
0.662, has a standard error of about 0.106.

Definition for Critical Ratio for regression weight: Dividing the regression weight
estimate by the estimate of its standard error gives z = 0.662/0.106 = 6.433. In other
words, the regression weight estimate is 6.433 standard errors above zero.

92
Definition for Level of Significance for regression weight: The probability of getting a
critical ratio 6.433 in absolute value is less than 0.001. In other words, the regression
weight for X1 in the prediction of Y is significantly different from zero at the 0.001
level (two-tailed test).

Extracted from Table 5: Estimate of Regression Weight for X2

The Actual beta value S.E. C.R. P-value

Y X2 0.188 0.062 3.032 0.002

Definition for actual beta 0.188: When X2 goes up by 1unit, Y goes up by 0.188 unit.

Definition for Standard Error of regression weight 0.062: The regression weight estimate
of 0.188 has a standard error of about 0.062.

Definition for Critical Ratio for regression weight: Dividing the regression weight
estimate by the estimate of its standard error gives z = 0.188/0.062 = 3.032. In other
words, the regression weight estimate is 3.032 standard errors above zero.

Definition for Level of Significance for regression weight: The probability of getting a
critical ratio 3.032 in absolute value is 0.002. In other words, the regression weight for X2
in the prediction of Y is significantly different from zero at the 0.05 level (two-tailed
test).

93
 4.2.3 Interpreting the AMOS Text Output: The Standardized
Residual Covariance

The researcher could examine the output of Standardized Residual Covariance as a guide
to examine the “wellness” of the structural model. The residual covariance is the
difference between the sample covariance and the model-implied covariance. The model
is correctly specified if most of the standardized residuals are less than two in absolute
value. Table 6 presents the standardized residual estimate for the above structural model.

Table 6: The Standardized Residual Covariance: The Standardized Difference between

Sample Covariance and Model-Implied Covariance
Y4 Y3 Y2 Y1 X15 X14 X13 X12 X11 X24 X23 X22 X21
Y4 .000
Y3 .740 .000
Y2 -.144 -.020 .000
Y1 -.265 -.106 .055 .000
X15 -.478 .034 -.401 -.106 .000
X14 .183 .931 .821 1.08 -.045 .000
X13 -.866 -.342 -.762 -.283 .281 -.092 .000
X12 .346 .394 .139 .700 -.568 .167 -.574 .000
X11 .523 .751 .392 .753 -.479 .095 -.625 4.91 .000
X24 3.14 3.29 3.54 3.94 1.68 1.78 2.03 1.78 3.20 .000
X23 2.37 2.98 2.79 3.39 .518 .193 .146 1.36 1.76 3.77 .000
X22 -.487 -.116 -.251 .459 -.508 .025 -.223 1.20 2.31 -.425 -.161 .000
X21 -1.04 -.702 -1.04 -.172 -.427 -.233 -.230 .994 1.43 -.300 -.271 .070 .000

The researcher could employ the standardized residual estimates as an alternative

option to delete the “poor item” from the construct in the model. The “poor item” is the
item that does not contribute in estimating the respective latent construct. These items
caused the measurement model to have a poor fit. Based on the standardized residual
covariance matrix, the researcher needs to delete item X23 and item X24 from the
structural model.

94
4.3 An Example of More Complicated AMOS Output

Sometimes, the results of the structural model are quite complicated to read and interpret,
especially for the beginners. As an example, the result in Figure 6 shows the Standardized
Regression Weights for the Event Loyalty Model.

The most important output in the Standardized Regression Weight is the value of
R2 for the model. In Figure 6, one could find that the value of R2 for the whole model is
0.89 (this value is located at the endogenous construct). By looking at this value, one
could conclude that the model is good, since it could capture 89% of the estimate on
endogenous construct by including certain exogenous constructs in the model. At the
same time, the Fitness Indexes are good, and the factor loading for all items are also good
(above the required 0.6).

95
 Figure 6: The Standardized Regression Weights for the Model

The output in Figure 6 indicates that 89% of the performance in Event Loyalty
could be estimated by using three exogenous constructs into the model namely Service
Performance, Perceived Quality, and Customer Emotion into the model. At the same
time, 47% of the Customer Emotion could be measured by using Service Performance
and Service Quality.

The actual Regression Weights are presented in Figure 7. The value of regression
weight indicates the effect of an exogenous construct on its corresponding endogenous
construct. As has been explained earlier, one could differentiate exogenous construct
from endogenous based on the one-sided arrow. Remember, one-sided arrow represents
the causal effect of an exogenous construct (arrow pointing out) on its corresponding

96
endogenous construct (arrow pointing in). AMOS Graphic also produced the text output
for the researcher to test the required hypothesis.

Figure 7: The Regression Weights for the Model

Table 7 presents the causal effects of all exogenous constructs on their corresponding
endogenous constructs in the model.

97
Table 7: The Regression Weights and the Probability Value which Indicates its Significance

Estimate S.E. C.R. P Result

Emotion Service_Performance 0.371 .086 4.292 *** Significant
Emotion Perceived_Quality 0.388 .069 5.608 *** Significant
Event_Loyalty Emotion 0.272 .060 4.561 *** Significant
Event_Loyalty Service_Performance 0.154 .057 2.694 .007 Significant
Event_Loyalty Perceived_Quality 0.132 .047 2.783 .005 Significant
Q11 ACC 1.098 .085 12.898 *** Significant
Q12 ACC 1.115 .101 11.047 *** Significant

98
 CHAPTER 5

ANALYZING THE RELATIONSHIPS AMONG

VARIABLES IN A MODEL
The advantage of SEM using AMOS graphic is its effectiveness in presenting the
correlational relationships and causal effects among the variables. The presentation of
output is visually informative since the researchers could include the mean and variance
for each variable involved in the model.

5.1 ANALYZING THE RELATIONSHIP BETWEEN

OBSERVED VARIABLES IN A MODEL

Suppose the researchers have two measured variables namely X1 and Y. In AMOS, the
correlational relationship between these two variables could be modeled using a double-
headed arrow as shown in Figure 1:

X1 Y

Figure 1: Analyzing Correlational Effect for the Observed Variables

99
 The researchers could present valuable information in the diagram such as mean,
variance, and covariance by selecting the appropriate box in the output menu. This
information (output) is presented in Figure 2.

The Mean and The Mean and

Variance of data X1 Variance of data Y

7.35, 2.48 8.52, 1.26

X1 Y

The Covariance between

X1 and Y is 0.66
.66

Figure 2: The Mean, Variance and Covariance Estimates for X1 and Y

Referring to Figure 2, the covariance between X1 and Y is 0.66. The result shows
the mean score for variable X1 is 7.35 with variance of 2.48. Meanwhile, the mean score
for variable Y is 8.52 with variance 1.26. The estimated covariance, its standard error,
critical region, and probability value are presented in Table 1.

The information obtained would be adequate for the researchers to test the
following research hypothesis regarding the relationship between X1 and Y.

H1: The covariance between X1 and Y is significant.

100
 Table 1: The Hypothesis Testing for the Relationship between X1 and Y

Estimate S.E. C.R. Probability

X1 Y 0.664 0.107 6.205 0.001

***Indicate a highly significant at < 0.001

Level of significance for covariance (refer to AMOS output in Table 1)

The probability of getting a critical ratio 6.205 in absolute value is less than
0.001. In other words, the covariance between X1 and Y is significantly different from
zero at the 0.001 level (two-tailed). Thus, the above research hypothesis is supported.

5.2 ANALYZING THE RELATIONSHIP BETWEEN LATENT

CONSTRUCTS IN A MODEL

Suppose the researchers have an exogenous construct namely X and an endogenous

construct namely Y, where both constructs have five measuring items respectively. The
researchers’ intention here is to analyze the correlational relationship between these two
latent constructs. The AMOS graphic representation is given in Figure 3.

101
e1 e2 e3 e4 e5 e6 e7 e8 e9 e10
1 1 1 1 1 1 1 1 1 1

X11 X12 X13 X14 X15 Y1 Y2 Y3 Y4 Y5

1 1

X Y

Figure 3: Analyzing the Relationships between Two Latent Constructs

102
The results of analysis for the model in Figure 3 are presented in Figure 4.

.38 .32 .51 .63 1.31 .28 .19 .40 .13 .57

e1 e2 e3 e4 e5 e6 e7 e8 e9 e10
1 1 1 1 1 1 1 1 1 1

X11 X12 X13 X14 X15 Y1 Y2 Y3 Y4 Y5

.97 1.0 1.08
.94 1.23 1.10
1.00 5 .79 1.00 1.11

X Y .97
2.61

Variance in X Variance in Y

Covariance between
X and Y
.60

Figure 4: The Variances and Covariance Estimate for Latent Constructs in the Model

The estimated covariance, its standard error, critical region and probability value
are shown in Table 2. The information given would be adequate for the researchers to test
the following research hypothesis regarding the relationship between the two constructs.

H1: The covariance between X and Y is significant.

Table 2: The Hypothesis Testing for the Covariance between X and Y

Estimate S.E. C.R. Probability

X Y 0.603 0.102 5.911 0.001

***Indicate a highly significant at < 0.001

103
Level of significance for covariance (Table 2)

The probability of getting a critical ratio 5.911 in absolute value is less than
0.001. In other words, the covariance between X and Y is significantly different from
zero at the 0.001 level (two-tailed). Thus, the above research hypothesis is supported.

5.3 ANALYZING THE CAUSAL EFFECTS FOR OBSERVED

VARIABLES IN A MODEL

Suppose the researchers are interested to analyze the causal effect of X1 on Y. The causal
effect is drawn using a single-headed arrow as shown in Figure 5:

1
X1 Y e1

Figure 5: Analyzing Causal Effect for Observed Variables

The causal effect or beta coefficient is given in Figure 6.

2.48
1.08
.27 1
X1 Y e1

Figure 6: The Regression Coefficient 0.27 Reflects the Amount of Causal Effect of X1 on Y

The estimated beta, its standard error, critical region, and probability value is
given in Table 3. The information given would be adequate for the researcher to test the
hypothesis for beta.

104
 Table 3: The Hypothesis Testing for the Causal Effect of X1 on Y

Estimate S.E. C.R. Probability

Y X1 0.270 0.038 7.105 0.001

***Indicate a highly significant at < 0.001

H1: X1 has a significant and direct effect on Y

Level of significance for regression weight (Table 3)

The probability of getting a critical ratio 7.105 in absolute value is less than
0.001. In other words, the regression weight for X1 in the prediction of Y is significantly
different from zero at the 0.001 level (two-tailed). Thus, the above research hypothesis is
supported.

5.4 ANALYZING THE CAUSAL EFFECTS FOR SIMPLE

LATENT CONSTRUCTS IN A MODEL

Suppose the researchers have two latent constructs namely X and Y. Let X is a latent
exogenous construct while Y is a latent endogenous construct. Both X and Y has five
indicators. The causal effect of X on Y is drawn using a single-headed arrow as shown in
Figure 7.

105
e1 e2 e3 e4 e5 e6 e7 e8 e9 e10
1 1 1 1 1 1 1 1 1 1

X11 X12 X13 X14 X15 Y1 Y2 Y3 Y4 Y5

1
1
1
X Y e11

Figure 7: Analyzing the Causal Effect for Latent Constructs

The causal effect of X on Y is presented in Figure 8.

.38 .32 .51 .63 .31 .28 .19 .40 .13 .57

e1 e2 e3 e4 e5 e6 e7 e8 e9 e10
1 1 1 1 1 1 1 1 1 1

X11 X12 X13 X14 X15 Y1 Y2 Y3 Y4 Y5

.97 1.05 1.08 1.10
.94 1.00 1.23 1.11
1.00 .79
.23 1
X Y e11

1.61 .83

Figure 8: The Regression Coefficient Reflecting the Causal Effect of X on Y

The estimated coefficient beta, its standard error, critical region and probability
value is given in Table 4. The information given would be adequate for the researcher to
test the hypothesis for regression coefficient beta.
106
 H1: X has a significant and direct effect on Y

Table 4: The Hypothesis Testing for the Causal Effect of X and Y

Estimate S.E. C.R. Probability

Y X 0.230 0.035 6.571 0.001

Level of significance for regression weight (Table 4)

The probability of getting a critical ratio 6.571 in an absolute value is less than
0.001. In other words, the regression weight for X in the prediction of Y is significantly
different from zero at the 0.001 level (two-tailed). Thus, the above research hypothesis is
supported.

5.5 ANALYZING THE CAUSAL EFFECTS FOR MULTIPLE

LATENT CONSTRUCTS IN A MODEL

Suppose the study has two independent constructs, one mediator, and one dependent
construct in the model. All four constructs are latent and they are measured through a set
of items in a questionnaire. First and foremost, the researcher needs to run the CFA for
the measurement model of the respective constructs. The CFA procedure would assess
the three requirements for modeling SEM namely unidimensionality, validity, and
reliability. The CFA procedure for measurement model has been discussed in Chapter 3.

Let’s say the CFA procedure for measurement models have been completed. The
next stage is to assemble the construct in the structural model as shown in Figure 9.
Observe how the mediator construct namely Customer Satisfaction is being modeled in
SEM for further analysis.

107
 The standardized estimate for structural model showing the factor loading for
every item and the correlation between exogenous constructs is presented in Figure 9.

Mediator

Figure 9: The Standardized Regression Weights for the Model

The important figures to highlight from standardized estimate are the correlation
between exogenous constructs, the factor loading for every item, the value of R2, and the
fitness indexes for the model.

However, one should observe the fitness indexes for the model which reflect how
fit is the hypothesize model with the data at hand. If the fitness indexes do not achieve the
required level for each fitness category, then one could suspect many problems with the
model. Among the problems with the model are low factor loading items, redundant

108
items, and the correlation between exogenous constructs is high. Remember, high
correlation between exogenous constructs (correlation above 0.85) indicates the model is
having a multi-collinearity problem.

The Regression Weights are presented in Figure 10.

Figure 10: The Regression Weights for the Model

The regression weights indicate the estimate of beta coefficient which measures
the effects of every exogenous construct on the endogenous construct.

109
 Table 5: The Regression Weights for Every Path Estimate in Figure 10

Construct Path Construct Estimate S.E. C.R. P-value Result

Customer_
Service_Quality 0.503 0.062 8.064 0.001 Significant
Satisfaction
Customer_
Corporate_Image 0.202 0.043 4.670 0.001 Significant
Satisfaction
Customer_
Customer_Satisfaction 0.609 0.076 8.046 0.001 Significant
Loyalty
Customer_
Service_Quality 0.223 0.063 3.519 0.001 Significant
Loyalty
Customer_ Not
Corporate_Image 0.065 0.040 1.616 0.106
Loyalty Significant

Figure 5 presents the path (arrow) and its coefficients in bold, which indicates
how much the effects of every exogenous construct on the respective endogenous
construct. In the path analysis, the researcher could verify the significance of every path
coefficient.

As an example, the path coefficient of Service Quality to Customer Satisfaction

is 0.503. This value indicates that for every one unit increase in Service Quality, its
effects would contribute 0.503 unit increase in Customer Satisfaction. And more
importantly, the effects of Service Quality on Customer Satisfaction is significant
(p<0.001). Thus, the hypothesis that Service Quality has significant and positive effects
on Customer Satisfaction is supported.

The result of every hypothesis is presented in Table 5A.

110
 Table 5A: The Results of Hypothesis Testing for the Respected Path

P- Results on
Hypothesis Statement of Path Analysis Estimate
value Hypothesis
H1: Service Quality has significant effect on customer Supported
0.503 0.001
satisfaction
H2: Corporate Image has significant effect on Customer Supported
0.202 0.001
Satisfaction
H3: Customer Satisfaction has significant effect on Supported
0.609 0.001
Customer Loyalty
H4: Service Quality has significant effect on Customer Supported
0.223 0.001
Loyalty
H5: Corporate Image has significant effect on Customer Not
0.065 0.106
Loyalty Supported

111
5.6 THE PATH ANALYSIS: ANALYZING THE MEDIATION
EFFECTS IN A MODEL

The path analysis is shown in Figure 11.

Figure 11: The Schematic Diagram of the Model

The schematic diagram of the model is converted into AMOS Graphic as shown
in Figure 11. Observed that the model contains two mediators namely Customer Emotion
and Relationsip Quality.

112
 Figure 12: The Schematic Diagram of the Model CConverted into AMOS Graphic

The CFA was performed for the measurement models of latent construct to
establish their unidimensionality, validity and reliability prior to modeling the structural
model for analysis using SEM. The output of SEM is presented starting from Figure 13.

113
Figure 13: The Standardized Regression Weights for Every Path in the Model

Table 6: The Standardized Regression Weights (Factor Loading)

Construct Path Construct Estimate

Customer_Emotion  Financial_bond 0.436
Customer_Emotion  Social_Bond 0.289
Relationship_Quality  Social_Bond 0.364
Relationship_Quality  Financial_bond 0.044
Relationship_Quality  Customer_Emotion 0.509
Customer Loyalty  Customer_Emotion -0.021
Customer Loyalty  Relationship_Quality 0.591
Customer Loyalty  Financial_bond 0.087
Customer Loyalty  Social_Bond 0.215

114
The Regression Weights for every path is presented in Figure 14.

Figure 14: The Regression Weights for Every Path in the Model

115
 Table 7: The Regression Weights for Every Path and Its Significance

Construct Path Construct Estimate S.E. C.R. P-value Result

Customer_
 Social_Bond 0.314 0.066 4.727 0.001 Significant
Emotion
Relationship_
 Social_Bond 0.358 0.055 6.515 0.001 Significant
Quality
Customer_
 Social_Bond 0.223 0.063 3.510 0.001 Significant
Loyalty
Customer_
 Financial_bond 0.367 0.053 6.915 0.001 Significant
Emotion
Customer_ Not
 Financial_bond 0.070 0.044 1.595 0.111
Loyalty Significant
Relationship_ Not
 Financial_bond 0.033 0.041 0.803 0.422
Quality Significant
Relationship_
 Customer_Emotion 0.460 0.058 7.900 0.001 Significant
Quality
Customer_ Not
 Customer_Emotion -0.020 0.066 -0.297 0.767
Loyalty Significant
Customer_
 Relationship_Quality 0.623 0.089 7.038 0.001 Significant
Loyalty

The result of every hypothesis is presented in Table 7A.

Table 7A: The Results of Hypothesis Testing for the Respected Path

Result on
Hypothesis Statement Estimate P-value
Hypothesis
H1: Social Bond has significant effect on
0.314 0.001 Supported
Customer Emotion
H2: Social Bond has significant effect on
0.358 0.001 Supported
Relationship Quality
H3: Social Bond has significant effect on
Customer Loyalty 0.223 0.001 Supported

116
 Result on
Hypothesis Statement Estimate P-value
Hypothesis
H4: Financial bond has significant effect on
0.367 0.001 Supported
Customer Emotion
H5: Financial bond has significant effect on
0.033 0.422 Not Supported
Relationship Quality
H6: Financial bond has significant effect on
0.070 0.111 Not Supported
Customer Loyalty
H7: Customer Emotion has significant effect on
0.460 0.001 Supported
Relationship Quality
H8: Customer Emotion has significant effect on
-0.020 0.767 Not Supported
Customer Loyalty
H9: Relationship Quality has significant effect on
0.623 0.001 Supported
Customer Loyalty

Quick Exercises:

Based on the above results, decide on the following hypothesis regarding the mediators:

1) Customer Emotion mediates the relationship between Social Bond and Relationship
Quality.

2) Customer Emotion mediates the relationship between Financial Bond and

Relationship Quality.

3) Relationship Quality mediates the relationship between Social Bond and Customer
Loyalty.

4) Relationship Quality mediates the relationship between Financial Bond and Customer
Loyalty.

5) Relationship Quality mediates the relationship between Customer Emotion and

Customer Loyalty.

117
 CHAPTER 6
ANALYZING THE MEDIATING VARIABLE
IN A MODEL
Sometimes, research questions intend to address the effect of a mediating variable in the
relationship between an independent variable and its corresponding dependent variable in
a model. The diagram below illustrates the position of a mediator in the relationship
between independent variable and its corresponding dependent variable.

Direct Effect

from IV to
DV

Indirect Effect

Through
Mediator
First of all, the direct effect of independent variable on dependent variable is
significant. When the mediator variable M enters the model, the direct effect would be
reduced since some of the effect has shifted through the mediator. If it is reduced but still
significant, the mediation effect here is called “partial mediation”. However, if the direct
effect is reduced and no longer significant, then the mediation is called “complete
mediation”.

118
 The modeling of a mediator variable in AMOS graphic is illustrated in Figure 1.

When analyzing the mediator, there are two effects involved namely direct effect
and indirect effect. The direct effect is the effect from independent variable directly to
dependent variable, while the indirect effect is the effect from independent variable to
dependent variable that goes indirectly through the mediating variable.

Mediating Indirect Effect

Variable: M
Indirect Effect
Direct Effect

Independent Dependent
Variable: X1 Variable: Y

Figure 1: Modeling the Mediator in the Structural Model

Figure 1: Let X1 be an independent variable, Y as a dependent variable, and M as

a mediator in the model. The single-headed arrow indicates a causal effect of X1 on Y, X1
on M, and M on Y. In this case, M is modeled as a mediator variable in AMOS graphic.
In the diagram, the researchers can examine the direct effect of X1 on Y, the direct effect
of X1 on M, and the direct effect of M on Y.

In this model, the researchers will examine the direct effect and indirect effect of
X1 on Y. If the direct effect of X1 on Y is reduced, and the indirect effect (through M) is
significant, then M is said to play a mediating role in linking X1 to Y indirectly.

119
6.1 ANALYZING THE MEDIATING EFFECTS USING AMOS
GRAPHIC: THE OBSERVED VARIABLES

Firstly, we begin by modeling the simple effect of X1 on Y as shown in Figure 2. We test

the direct effect of X1 on Y as shown in Figure 3. The output in Table 1 shows B1 is
0.361 and it has a significant effect on Y (p-value < 0.001).

Secondly, we enter the mediator variable M into the model as shown in Figure 4.
Now we test the direct effect of X1 on Y as shown in Figure 5. The output in Table 1a
shows B1 is reduced from 0.361to 0.103 when M entered the model. Now the direct
effect on Y is no longer significant (p-value = 0.062). Here, the requirement for complete
mediation is met. Finally, we need to test the hypothesis for B2 and B3 to confirm the
complete mediation.

e1
1

X1 Y

Figure 2: Modeling the Direct Effect of X1 on Y

120
 e1
1
4.90, 1.36 3.87
.36
X1 Y

Figure 3: The Result of Direct Effect of X1 (Beta Coefficient is 0.36)

Table 1: The Direct Effect of X1 on Y is significant (Beta Coefficient is 0.361)

Estimate S.E. C.R. P-Value Result

Y X1 0.361 0.050 7.235 0.001 Significant

Mediator M
1
enters the M e2
model

e1
1

X1 Y

Figure 4: Modeling the Mediator M in AMOS Graphic for Observed Variable

121
 3.10

1
M e2

.53 .31
e1
1
4.90, 1.36 3.01
.10
X1 Y

Figure 5: The AMOS Output showing Regression Coefficient between Variables

P- Result
Estimate S.E. C.R.
Value
Y X1 0.103 0.066 1.547 0.062 Not significant
M X1 0.527 0.041 12.991 0.001 Significant
Y M 0.312 0.050 6.282 0.001 Significant
***Observe that the direct effects of X1 is reduced from 0.36 to 0.1 after M entered the model

The Hypotheses Required in Determining the Mediation Effects

Hypothesis 1: X1 has significant and direct effects on Y

Table 1a: The Hypothesis Testing for a Direct Effect of X1 on Y

Estimate S.E. C.R. P-Value Result

Y X1 0.103 0.066 1.547 0.062 Not significant
(Beta coefficient is reduced from 0.361 in Table 1 to 0.103 in Table 1a)

Result: Hypothesis 1 is not supported.

122
 Hypothesis 2: X1 has significant and direct effects on M

Table 2: The hypothesis testing for the causal effect of X1 on X2

Estimate S.E. C.R. P-Value Result

M X1 0.527 0.041 12.991 0.001 Significant

Result: Hypothesis 2 is supported

Hypothesis 3: M has significant and direct effects on Y

Table 3: The Hypothesis Testing for the Causal Effect of X2 on Y

Estimate S.E. C.R. P-Value Result

Y M 0.312 0.050 6.282 0.001 Significant

Result: Hypothesis 3 is supported.

Conclusion: M is a mediating variable in the relationship between X1 and Y.

The type of mediation here is a called a “complete mediation” since the direct
effect of X1 on Y is no longer significant after M entered the model (Hypothesis 1).
Instead, the indirect effect is significant. Thus, X1 has an indirect effect on Y through the
mediator variable M.

123
6.2 ANALYZING THE MEDIATING EFFECTS FOR LATENT
CONSTRUCTS

The concept of mediation in latent constructs is equivalent to that of the observed

variables as discussed above. First of all, we need to show that the direct effect of X1 on
Y is significant.

Figure 6: Modeling the Direct Effect of X1 on Y for Latent Constructs

124
 The output when the model in Figure 6 is run is shown in Figure 7.

Figure 7: The Result shows the Direct Effect of X1 on Y (Beta Coefficient = 0.87)

Table 4: The direct effect of X1 on Y is significant (Beta Coefficient 0.868)

Beta Estimate S.E. C.R. P-Value Result

Y X1 0.868 0.104 8.365 0.001 Significant

The analysis for mediation begins by showing that the direct effect of X1 on Y is
significant. The direct effect is measured through beta coefficient. In this case, B1 is
significant.

When the mediating variable M enters the model, the value of beta coefficient for
X1 is expected to reduce, or in other words the direct effect of X1 on Y would be reduced
when the mediator enters the model.

The new model when M is entered as a mediator is presented in Figure 8.

125
 e6 e7 e8 e9
1 1 1 1

X21 X22 X23 X24

1
e15 M
e14
1
Y1 e10
1 1

1
Y2 e11
Y
X1 1
Y3 e12
1
1
Y4 e13
X15 X14 X13 X11
1 1 1 1

e5 e4 e3 e1

Figure 8: Modeling the Mediator for Latent Construct M in AMOS Graphic

In the above diagram, X1, M, and Y are latent constructs. The study is interested
to prove that the construct M mediates the relationship between construct X1 and
construct Y.

The Regression Weight estimates for the model are presented in Figure 9, while
the text output is presented in Table 5.

126
 .45
.19 .09 1.39

e6 e7 e8 e9
1 1 1 1

X21 X22 X23 X24

1.00 .81
1.00
.64

1 .75
.77 e15 M
e14 .11
1
.94 .79 Y1 e10
1 1.00
.06
.48 1
1.03
Y2 e11
.68 Y .17
1.10
X1 1
Y3 e12
1.59 1.15
1.00 .87
1.40 1.51
1
Y4 e13
Fitness Indexes
X15 X14 X13 X11 1.ChiSq = 151.309
2.df = 50
1 1 1 1 3.ChiSq/df = 3.026
.19 .28 .23 .87 4.GFI = .929
5.AGFI = .919
6.CFI = .972
e5 e4 e3 e1
7.RMSEA = .079

Figure 9: The AMOS Output showing Regression Weights between Constructs

Note: All fitness indexes achieved the required level

***Observed that the value of beta coefficient linking X1 to Y is reduced from 0.87 to
0.68.

127
 Table 5: The Multiple Regression Weights

Beta Estimate S.E. C.R. P-Value Result

Y X1 0.682 0.110 6.202 0.001 Significant

M X1 0.941 0.108 8.712 0.001 Significant
Y M 0.787 0.161 4.888 0.001 Significant

The Hypotheses Required in Determining the Mediation Effects

Hypothesis 1: X1 has significant and direct effects on Y

Table 6: The Hypothesis Testing for the Causal Effect of X1 on Y

Beta Estimate S.E. C.R. P-Value Result

Y X1 0.682 0.110 6.202 0.001 Significant

(The direct effects has reduced after mediator entered the model from 0.868 in Table 4 to
0.682 in Table 6)

Hypothesis 2: X1 has significant and direct effects on M

Table 7: The Hypothesis Testing for the Causal Effect of X1 on X2

Beta Estimate S.E. C.R. P-Value Result

M X1 0.941 0.108 8.712 0.001 Significant

Hypothesis 3: M has significant and direct effects on Y

Table 8: The Hypothesis Testing for the Causal Effect of M on Y

Beta Estimate S.E. C.R. P-Value Result

Y M 0.787 0.161 4.888 0.001 Significant

Conclusion: The construct M does mediate the relationship between X1 and Y.

128
 The type of mediation here is a called a “partial mediation” since the direct effect
of X1 on Y is still significant after M entered the model even though the beta coefficient
for X1 is reduced from 0.87 (in Figure 7) to 0.68 (in Figure 9). In this case, X1 is both
significant direct effect on Y and also significant indirect effect on Y through the
mediator variable namely M.

6.3 ANALYZING THE MEDIATOR IN A COMPLEX MODEL

In practice, researchers are working with complex models, and their theoretical
framework in AMOS Graphic is also complicated. Thus, analyzing the mediator in a
complex model is quite challenging for new researchers. At all times, the researcher
needs to focus on the concept of triangle as stated in Figure 1, which indicates the direct
effect and indirect effect when analyzing the mediator.

The direct effect is the effect that goes directly from exogenous construct to
endogenous construct, while the indirect effect is the effect from exogenous construct to
endogenous construct that goes indirectly through the mediator in the model.

The following example is based on Structural Model in Figure 10.

Figure 10 illustrates the position of mediator construct namely Customer

Satisfaction in the model. In this example, the researcher is interested to assess the role of
Customer Satisfaction as a mediator in the following relationships:

i) The relationship between Service Quality and Customer Loyalty, and

ii) The relationship between Corporate Image and Customer Loyalty.

129
 The structural model shows there are two
mediating roles of Customer Satisfaction:
i) Service Quality to Customer Loyalty
ii) Corporate Image to Customer Loyalty

Figure 10: The Standardized Regression Weights for Every Path in the Model

TESTING THE MEDIATION EFFECT OF CUSTOMER SATISFACTION IN THE

RELATIONSHIP BETWEEN SERVICE QUALITY AND CUSTOMER LOYALTY

First of all, obtain the standardized regression weights and the probability values, which
indicate the significance for the respective path (Figure 10). The required information is
given in Table 9. Draw the triangle as shown in Figure 11.

130
Table 9: The Standardized Regression Weights and Its Significance for Each Path

Construct Path Construct Estimate P-Value Result

Customer_Satisfaction Service_Quality 0.51 0.001 Significant
Customer_ Loyalty Customer_Satisfaction 0.58 0.001 Significant
Customer_ Loyalty Service_Quality 0.22 0.001 Significant

THE PROCEDURE FOR TESTING

MEDIATION

1) The Indirect Effect = 0.51 x 0.58 = 0.2958

2) The Direct Effect = 0.22
3) Both indirect path (SQ to CS and CS to CL) are significant
4) Since Indirect Effect > Direct Effect, the mediation occurs
5) Type of mediation here is Partial Mediation since the Direct Effect
is still significant after mediator enters the model

Figure 11: The Procedure for Testing Mediation in a Complex Model

The hypothesis statement:

Ha: Customer Satisfaction mediates the relationship between Service Quality and
Customer loyalty

131
The Results of Mediation Test

P- Results on
Hypothesis Statement of Path Analysis Estimate
Value Hypothesis
Ha1: Service Quality has significant effect on Customer Supported
0.51 0.001
Satisfaction
Ha2: Customer Satisfaction has significant effect on Supported
0.58 0.001
Customer Loyalty
Ha3: Service Quality has significant effect on Customer Supported
0.22 0.001
Loyalty

TESTING THE MEDIATION EFFECT OF CUSTOMER SATISFACTION IN THE

RELATIONSHIP BETWEEN CORPORATE IMAGE AND CUSTOMER LOYALTY

First of all, obtain the standardized regression weights and the probability values which
indicate the significance for the respective path (Figure 10). The required information is
given in Table 10. Draw the triangle as shown in Figure 12.

Table 10: The Standardized Regression Weights and Its Significance for Each Path

P-
Construct Path Construct Estimate Result
Value

Customer_Satisfaction Corporate_Image 0.27 0.001 Significant

Customer_ Loyalty Customer_Satisfaction 0.58 0.001 Significant

Not
Customer_ Loyalty Corporate_Image 0.08 0.106
Significant

132
 THE PROCEDURE FOR TESTING
MEDIATION

1) The Indirect Effect = 0.27 x 0.58 = 0.1566

2) The Direct Effect = 0.08
3) Both indirect path (CI to CS and CS to CL) are significant
4) Since Indirect Effect > Direct Effect, the mediation occurs
5) Type of mediation here is Complete Mediation since the Direct
Effect is not significant after mediator enters the model

Figure 12: The Standardized Regression Weights for the Model

The hypothesis statement:

Hb: Customer Satisfaction mediates the relationship between Corporate Image and
Customer loyalty

Table 11: The Results of Mediation Test

Results on
Hypothesis Statement of Path Analysis Estimate P-Value
Hypothesis
Hb1: Corporate Image has significant effect on Supported
0.202 0.001
Customer Satisfaction
Hb2: Customer Satisfaction has significant Supported
0.609 0.001
effect on Customer Loyalty
Hb3: Corporate Image has significant effect on Not Supported
0.065 0.106
Customer Loyalty

133
 The results of hypothesis testing in Table 11 indicate that Customer Satisfaction
does mediate the relationship between Corporate Image and Customer Loyalty. Thus, the
type of mediation here is full mediation since the direct effect is no longer significant
after the mediator enters the model.

6.4 MEDIATION TEST: CONFIRMING THE RESULT

THROUGH BOOTSTRAPPING

Lately, there are demands from many quarters, including the examiners that researchers
re-confirm the results of their mediation tests using the resampling procedure called
Bootstrapping. This is especially for testing the indirect effect. Bootstrapping is the
method of sampling with replacement, whereby one instructs the algorithm to take the
sample of size n from the existing dataset. The number of re-sampling could be between
500 to 1,000 times. The algorithm would compute the mean and standard error for every
sample. From re-sampling process, the algorithm develops sampling distribution for the
estimates.

From the sampling distribution, the total effect, the direct effect, and also the
indirect effect between constructs are estimated. Finally, the 95% confidence interval
values for total effect, direct effect, and indirect effect will be tabulated. The algorithm
would tabulate the lower limit and the upper limit, as well as the two-tailed significant
values for the effects. Using these values, the researcher could compare the mediation test
results with the bootstrapping results. Most of the times, the results are equivalent.
However, for any contradictory, the bootstrapping result will be applicable.

134
 Figure 13: Testing the Mediator MRPI

Table 12: The Regression Weights between Constructs

Std.
Estimate S.E. C.R. P Result
Estimate
MRPI PERSONALITI 0.020 0.58 .006 3.440 *** Significant
BEHAVIOR MRPI 1.337 0.43 .290 4.606 *** Significant
BEHAVIOR PERSONALITI 0.025 0.29 .010 2.470 .014 Significant

The bootstrapping results are shown in Table 13 and 14.

135
 Table 13: The Significance of Indirect Effects - Two Tailed Significance (BC)

PERSONALITI MRPI BEHAVIOR

PERSONALITI - .000 .027
MRPI .000 - .000
BEHAVIOR .027 .000 -

Table 14: The Significance of Direct Effects - Two Tailed Significance (BC)

PERSONALITI MRPI BEHAVIOR

PERSONALITI - .000 .001
MRPI .010 - .001
BEHAVIOR .038 .002 -

Based on the results in Table 13 and Table 14, one can conclude that the results of
mediation test is consistent with Table 12.

6.5 CCOMPUTING THE EFFECT SIZES IN A MEDIATION TEST

Supposed we are working with the following model (Figure 14). In this model, the
researcher is interested to carry out the following analysis:

1) To determine whether Job Attitude mediates the relationship between Skills &
Training and Career Advancement.
2) To measure the effect size for every single path in the model.
3) To measure the mediated effect of the mediator in the model.

136
 The effect size of Skills and
Training on Job Attitude = 0.78

r2XMY= 0.67

The total effect size of Skills &

Training and Job Attitude on
Career Advancement = 0.67

Figure 14: The Model Containing All Constructs namely Skill & Training (X), Job Attitude
(M), and Career Advancement (Y)

The coefficients and their probability values are summarized in Table 15. As
discussed earlier, these values can be used to determine the significance of a mediator in
the mediation model.

137
 Table 15: The Regression Weight and Its Probability Value

Estimate P-Value Results

Job Attitude Skills & Training 0.88 0.001 Significant
Not
Career Advancement Skills & Training 0.22 0.064
Significant
Career Advancement Job Attitude 0.61 0.001 Significant

From the results in Figure 14, the indirect effect is 0.537 (0.88 * 0.61), higher
than the direct effect of 0.22. Thus, we can conclude that the construct Job Attitude is a
mediator in the relationship between Skills & Training and Career Advancement. The
type of mediation here is full mediation since the direct effect is not significant.

Now, we focus on computing the two effect sizes (Objectives 2 and 3). The
computation will be based on figures obtained in Figure 14, Figure 15, and Figure 16.

138
 The effect size of Skills and Training
on Career Advancement = 0.59

r2XY = 0.59

Figure 15: The Model Containing Construct Skills and Training (X) and Career
Advancement (Y)

Table 16: The Regression Weights when Skills & Training is a Sole Predictor

Estimate P-value Result

Attitude Knowledge 0.770 0.001 Significant

139
 The effect size of Job Attitude
on Career Advancement = 0.66

r2MY = 0.66

Figure 16: The Model containing Job Attitude (M) and Career Advancement (Y)

THE EFFECT SIZES IN A MEDIATION MODEL

Effect size is the amount of variance explained in the mediation model contributed by
every single path. The researcher might be interested to know the relative contribution of
individual paths in the mediation model, especially the effect of a mediator on the
dependent variable. The researcher could assess the effect size for every path in the
model, and also the mediated effect of the mediator variable itself. Specifically, the two
effect sizes are:

(1) The r2 measures the effect size of individual path in the model (3 paths), and

(2) The r2 measures the size of mediated effect of a mediator in the model.

140
 The range of values of r2 and its relative effect size based on definition by Cohen
(1988) is given in Table 17.

Table 17: The Cohen's (1988) Benchmark Range of Effect Sizes

Range of R2 The Effect Size

Below 0.13 Small Range
Between 0.13-0.26 Medium Range
Above 0.26 Large Range

1) r2 measures for the individual path in the mediation model is defined as follows:

i) r2XM represents the squared partial correlation between the X and M variables in
the model,
ii) r2XY denotes the squared partial correlation between the X and Y variables in the
model, and
iii) r2MY.X corresponds to the squared partial correlation between the M and Y
variables with the influence of X removed.

For the above example - the following measures are obtained:

i) The effect size of Skills & Training (X) on Job Attitude (M) is 0.78 (Figure 14).
ii) The effect size of Skills & Training (X) on Career Advancement (Y) is 0.59 (Figure
15).
iii) The effect size of Job Attitude (M) on Career Advancement (Y), controlling for Skills
& Training. This is equivalent to r2XMY - r2XY = 0.67 – 0.59 = 0.08.
(Note: r2XMY=0.67 - Figure 14)
Conclusion: The effect size of a mediator on DV in small range

141
2) R2 Measures for the mediated effect of the mediator (Job Attitude).
This value is computed using the following formula:

r2MY – (r2XMY - r2XY)

Where:

r2XY is the R2 for the model containing X and Y only (Figure 15).

r2MY is the R2 for the model containing M and Y only (Figure 16).

r2 XMY is the R2 for the model containing all X, M, and Y (Figure 14).

r2MY – (r2XMY - r2XY)

0.66 – (0.67-0.59) = 0.58

Thus, based on Cohen (1988), the mediated effect size of the mediator (Job Attitude)
in the relationship between Skills & Training and Career Advancement is large.

In Class Exercises:

a) Refer to the model in Figure 10 above. Obtain the following values:

1) The effect size of Service Quality on Customer Satisfaction.
2) The effect size of Service Quality on Customer Loyalty.
3) The effect size of Customer Satisfaction on Customer Loyalty.
4) The mediated effect of Customer Satisfaction in the model linking Service
Quality to Customer Loyalty.

b) Refer to the model in Figure 10 above. Obtain the following values:

1) The effect size of Corporate Image on Customer Satisfaction.
2) The effect size of Corporate Image on Customer Loyalty.
3) The effect size of Customer Satisfaction on Customer Loyalty.
4) The mediated effect of Customer Satisfaction in the model linking Corporate
Image to Customer Loyalty.

142
Power in the Mediation Models

The power to detect the significance of mediation effects is always lower than the power
to detect the significance of the main effects because the magnitude of the mediated
effect is bounded by the individual coefficients from which it is formed. Recent research
has shown that causal steps tests for mediation and normal theory point estimators of the
mediated effect are underpowered (MacKinnon et al., 2002).

Methodological work has shown that asymmetric confidence limits based on the
distribution of the product and re-sampling methods such as the percentile bootstrap and
the bias-corrected bootstrap give the best combination of low Type 1 error rates and
power to detect effects in mediation models (MacKinnon et al., 2002). Consequently,
these methods should be implemented to test mediation hypotheses in substantive
research.

143
 CHAPTER 7

ANALYZING THE MODERATING VARIABLE

IN A MODEL
Researches in business, social sciences and other disciplines involve theories concerning
moderating variables. Thus, researchers in these areas should know how to model the
moderators and analyse them in their work.

Moderating variable is the variable that “moderates the effects” of an independent

variable on its dependent variable. The social science researchers, in particular, define
moderator as the variable that “interfere” in the relationship between an independent
variable and its corresponding dependent variable. For illustration, let M be the
moderator variable in the X-Y relationship. Then the moderation role of M is “to alter”
the effects of X on Y.

Before introducing a moderator into the model, the effects of independent variable
X on its dependent variable Y must exist and significant. Thus, when a moderator M
enters the model, the causal effects would change due to some “interaction effect”
between independent variable X and moderator variable M that has just entered. As a
result, the “effects” of X on Y could either increase or decrease. In other words, the effect
of independent variable on its dependent variable would depend on the level of moderator
variable.

7.1 THE SCHEMATIC DIAGRAM FOR MODERATING

144
 VARIABLE IN A MODEL

Example 1:

Let X = the amount of environmental news in the media educating the public concerning
the safe and clean environment. The campaign intends to make the public aware of
environmental degradation and that they should help the environment by switching to
environmental friendly products. Let Y = the respondents’ intention to purchase green
products, and let M = their level of education as a moderator. If the effect of
environmental campaign (X) in influencing the public to purchase green products (Y) is
more visible among higher educated consumers compared to lower educated consumers,
then we can say that education (M) is the moderating variable that moderates the
relationship between Environmental Awareness Campaign and Intention to Purchase
Green Products by the public.

Figure 1 illustrates the position of moderating variable M in the in the X-Y

relationship.

Environmental Intention to
Awareness Purchase
Campaign Green Products

Independent Variable
Level of
Moderator Variable Education Dependent Variable

Figure 1: The Schematic Diagram showing the Independent Variable, Dependent Variable
and a Moderator in a Model

Example 2:

145
Let X = Monetary Incentives, Y = Work Motivation, and M = Age of Workers. If the
effects of Monetary Incentives (X) on Work Motivation (Y) are more visible on certain
age groups (M), then one could claim that Age of Workers (M) moderates the
relationship between Monetary Incentives (X) and their Work Motivation (Y). Figure 2
illustrates the position of variable M in the X-Y relationship.

Monetary Motivation
Incentives to Work

Age Group

Figure 2: The Representation of Age as Moderating Variable in the Relationship between

Monetary Incentives and Work Motivation

Example 3:

Let X = the corporate reputation of manufacturers, Y = customers’ brand loyalty, and M

= the customers’ socio-economic status. If the effects of manufacturer’s corporate
reputation (X) on customers’ brand loyalty (Y) are depending on their level of
socioeconomic status (M), then one could claim that respondents’ socioeconomic status
moderates the effects of the firm’s corporate reputation (X) on consumers’ brand loyalty
(Y). Figure 3 illustrates the position of socioeconomic status (M) in the corporate
reputation and brand loyalty relationship.

146
 Manufacturer's Customer
Corporate Loyalty
Reputation on Product

Socio
Economic
Status

Figure 3: The Representation of Socio-Economic Status as Moderating Variable in the

Corporate Reputation – Brand Loyalty Relationship

7.2 MODELING THE INTERACTION EFFECTS OF A

MODERATOR FOR THE OBSERVED VARIABLE

As has been said earlier, although moderation implies a weakening of a causal effect, a
moderator can also enhance the causal effect. Remember: The term interaction and
moderation carries the same meaning. The interaction between independent variable and
moderator in the model could decrease or increase the effects on dependent variable.

A key part of moderation is the measurement of causal effect of independent

variable X on dependent variable Y for different level of moderator variable M. In
statistics, the effect of X on Y for a fixed value of M is referred as the “simple effect” of
independent variable on its dependent variable. Let’s X be an independent variable and Y
is a dependent variable. The simple regression equation will be:

Y = β0 + β1X + e

147
 Let’s assume that the above regression relation does exists and statistically
significant. When the moderator variable M enters the model, the moderation effect of M
is modeled in the regression equation as follows:

Y = β0 + β1X + β2M + β3XM + e

The regression coefficient β3 measures the interaction effect between

independent variable X and moderating variable M. Note that the regression coefficient
β1 measures the simple effects of X when the value of M = 0 (no interaction effects
involved). Then, the test of moderation is operationalized by the product term XM (the
multiplication between independent variable X and moderator variable M).

In order to test the moderation in a model, one needs to test β3 (the coefficient of
interaction term XM). If β3 is significant, then one could conclude that moderator variable
M moderates the relationship between X and Y.

Testing moderator for observed variables

Testing moderation for observed variables involve the Ordinary Least Squares
(OLS) regression, in which the dependent variable, Y, is regressed on the interaction term
XM and the main effects X and M. If both variables X and M are continuous, the
researcher needs to create the mean-centred value for X and M, where Xi’ = (Xi – mean
of X) and Mi’ = (Mi – mean of M). Thus, the new variable X and M has a mean of zero.
Now XM = (Xi’) * (Mi’). Variable Y does not have to be cantered.

7.3 SCALE OF MEASUREMENT FOR A MODERATING

VARIABLE

The researcher should employ the interval or ratio scale for measuring both independent
and dependent variables since the analysis involves parametric methods. As for the
moderator variable, it can be measured using any scale (nominal, ordinal, interval, and
ratio). Among the popularly used moderating variables in research are the respondent’s
148
demographic characteristics (nominal) and the level of treatment variable applied
(ordinal).

Both Ordinary Least Square regression (OLS) and Structural Equation Modeling
(SEM) could be employed if the dependent variable (Y) is measured using the interval or
ratio scale. However, if the dependent variable is measured using a dichotomous scale
(outcome is either yes or no), then the logistic regression should be employed.

7.4 MODELING THE MODERATING EFFECTS FOR

OBSERVED VARIABLES

Having all variables and data in hand, the next thing the researcher needs to know is how
to analyze the moderator and prove that M is actually moderating the relationship
between X and Y. In addition to the variable X, M, and Y, the researcher needs to create
a new variable namely XM from the product of X multiply M. Thus, the variables
involve will be X, Y, M, and XM. The information can be modeled in the following
regression equation:

Y = β0 + β1X + β2M + β3XM + e1

Figure 4 illustrates how the regression equation is modeled in AMOS graphic.

149
 X
e1
1

M Y

XM

Figure 4: The Modeling of Moderating Variable M in AMOS for Observed Variables

As shown in Figure 4, three hypotheses testing for path analysis is required:

1. The X-Y relationship (testing for β1) – we indicate as Hypothesis 1.

2. The M-Y relationship (testing for β2) – we indicate as Hypothesis 2.
3. The XM-Y relationship (testing for β3) – we indicate as Hypothesis 3.

The moderation effects of moderator variable M in the model occurs if

Hypothesis 3 (β3) is significant and Hypothesis 2 (β2) is not significant. As for
Hypothesis 1 (β1), there are two possibilities that could occur:

1. If Hypothesis 1 is not significant – then the “complete moderation” occurs, or

2. If Hypothesis 1 is significant – then the “partial moderation” occurs.

150
7.5 ANALYZING THE MODERATING EFFECTS FOR
OBSERVED VARIABLES

We shall go through some practical examples in order to enhance our understanding

concerning the concept of moderation.

Suppose the researcher is interested to assess the moderation effects of age of

workers (M) in the relationship between the monetary incentives given to them (X) and
their monthly productivity (Y). The AMOS model illustrating the researcher’s theoretical
argument is given in Figure 5. All variables in the model are directly observed, thus the
rectangles are employed instead of ellipses.

Incentives

e1
1

Age Productivity

Incentive*Age

Figure 5: The Model in AMOS Graphic for Testing Age as Moderator

The measurement of variables involved in the model.

Independent variable = Monetary Incentives (can be interval or ratio scale)

Dependent variable = Monthly Productivity (can be interval or ratio scale)

Moderating variable = Age of Workers (can be interval or ratio scale)

151
 The corresponding AMOS output for the above model where all variables
involved are interval or ratio is given in Figure 6.

.48

Incentives
.60

.59 e1
.24 1
.49

Age
.89 Productivity
.66

.68 -.88
.20

Incentive*Age

Figure 6: The AMOS Output showing the Regression Coefficients, Variance and
Covariance

The hypotheses testing required in analyzing a moderator in Figure 6.

Hypothesis 1: The Monetary Incentives given to workers has significant effects on

their monthly productivity.

Table 1: Testing the Causal Effects of Incentives on Productivity

Estimate S.E. C.R. P Result

Productivity Incentives 0.59 0.068 8.636 0.001 Significant

152
 In this case, Hypothesis 1: the hypothesis that the causal effects of incentives on
productivity are significant is supported.

Hypothesis 2: The Workers’ Age Level has significant effects on their productivity.

Table 2: Testing the Causal Effects of Age on Productivity

Estimate S.E. C.R. P Result

Productivity Age 0.89 0.61 1.451 0.072 Not Significant

In this case, the hypothesis that the effects of age on productivity are significant is
not supported.

Hypothesis 3: The Workers’ Age moderates the relationship between incentives and
productivity.

Table 3: Testing the Moderating Effects of Incentives*Age on Productivity

Estimate S.E. C.R. P Result

Incentive
Productivity -0.88 0.186 -4.742 0.001 Significant
*Age

In this case, the hypothesis that the moderating effects of workers’ age (M) on
relationship between incentives (X) and their productivity (Y) are significant is
supported.

The type of moderation that occurs in this case is partial moderation since the
hypothesis for the main effect is still significant after the moderator enters the model.

Note: The regression coefficient of product term (incentive*age) on productivity

is negative, which indicates that the moderating variable (age) weakens the causal effects
of monetary incentives (X) on monthly productivity (Y). In other words, the increase in
workers age would give negative effects on the firm’s productivity.

153
7.6 MODELING THE MODERATING EFFECTS FOR LATENT
CONSTRUCTS

Analyzing the moderating effects for the model with latent constructs is very
complicated. The normal modeling procedure using interaction terms is not practical with
latent constructs since it would cause problems with model convergence, as well as
distortion of standard errors. In the end, it resulted in model misfit and the procedure
stops.

Figure 7 illustrates how the moderator is modeled when analyzing the model
consisting latent constructs.

e1 e2 e3 e4 e5 e6 e7 e8 e9 e10
1 1 1 1 1 1 1 1 1 1

X11 X12 X13 X14 X15 Y1 Y2 Y3 Y4 Y5

1 1
1
X1 Y e11

Moderator

Figure 7: Modeling the Moderator Variable in the Path between X1 and Y

Alternatively, the Multi-Group CFA has been suggested as an alternative method

for assessing the effect of moderator variable in the model. The researcher only needs to
identify the path of interest where the moderator variable is to be assessed. This particular
path would be constrained with parameter = 1 and the model is termed as the constrained
model. The procedure will estimate two models separately. One is the constrained
model while the other one is the unconstrained model. The step-by-step process for
Multi-Group CFA is discussed.

154
7.7 ANALYZING THE MODERATOR FOR LATENT
CONSTRUCTS: THE MULTI-GROUP CFA

There are few steps involved in performing multi-group CFA:

1) Split data into two groups based on the moderator variable to be tested.
2) Save data into two separate files: Name the files as dataset 1 and dataset 2.
3) Select the path of interest in the model to test the moderator variable.
4) Develop two separate AMOS models: Rename as model 1 and model 2.
5) In Model 1, constraint the parameter in the path of interest to be equal to 1.
6) Name model 1 as the constrained model.
7) In model 2, do not constrain the relationship in the path of interest.
8) Name model 2 as the unconstrained model.
9) Use dataset 1: Estimate the constrained model.
10) Use the same dataset 1: Estimate the unconstrained model.
11) Obtain the difference in Chi-Square value between the constrained and the
unconstrained model. If the value differs by more than 3.84, then the moderation
occurs in that path.
12) Repeat the same procedure using dataset 2.
13) Use dataset 2: Estimate the constrained model.
14) Use the same dataset 2: Estimate the unconstrained model.
15) Obtain the difference in Chi-Square value between the constrained and the
unconstrained model. If the value differs by more than 3.84, then the moderation
occurs in that path.

Example: Suppose we are modeling the effect of X1 and X2 on Y (Figure 8). One of the
objectives for this research is to examine the moderation effect of a variable namely
education in the relationship path between X1 and Y.

155
 e1 e3 e4 e5

1 1 1 1

X11 X13 X14 X15 Path of interest to

test the
1 Moderator

e14
X1 1
Y1 e10
1 1

1
Y2 e11

Y
1
Y3 e12

1
Y4 e13
X2

X24 X23 X22 X21

1 1 1 1

e9 e8 e7 e6

Figure 8: The path where the Moderator Education is to be Examined

The path of interest where the moderation tests is to be carried out is shown in
Figure 8. First of all, the data is sorted in ascending order based on respondents’ level of
education. Then the data is split and save into two separate data files. Data 1 is renamed as
low education group, while data 2 is renamed as high education group.

Secondly, put a parameter constraint on the selected path to be equal to “1” as

shown in Figure 9. This model is renamed as the constrained model.

156
 e1 e3 e4 e5

1 1 1 1

X11 X13 X14 X15

Constrained
1

e14
X1 1
constrained Y1 e10
1 1
1

1
Y2 e11

Y
1
Y3 e12

1
Y4 e13
X2
1

X24 X23 X22 X21

1 1 1 1

e9 e8 e7 e6

Figure 9: The Constrained Model: The Parameter in the Path of Interest (X1 to Y) is
Constrained to 1

Thirdly, using the same model, remove the parameter constraint in the path as
shown in Figure 10. This model is renamed as the unconstrained model. Now the
researcher has two models to be assessed namely the constrained and the unconstrained
model.

157
 e1 e3 e4 e5

1 1 1 1

X11 X13 X14 X15

Unconstrained
1

e14
X1 1
Y1 e10
1 1

1
Y2 e11

Y
1
Y3 e12

1
Y4 e13
X2

X24 X23 X22 X21

1 1 1 1

e9 e8 e7 e6

Figure 10: The Unconstrained Model: The Coefficient in the Path (X1 to Y) is Not
Constrained

Next, obtain the estimate for both the constrained model and also the
unconstrained model using the first dataset (low education group). The output is
presented in Figure 11 for the constrained model and in Figure 12 for the unconstrained
model.

The procedure for testing moderation is carried out as shown in Table 1a.

158
 .87 .19 .24 .25

e1 e3 e4 e5

1 1 1 1

X11 X13 X14 X15

1.78
1.00 1.65 .73
1.76

e14 .09
X1 1
Y1 e10
1 1.00
1.00 .04
.39
1
1.02 Y2 e11

.43 Y 1.09
.20
1
Y3 e12
.07 1.05
1.31 .72 .12
1
Y4 e13
X2 Fitness Indexes
1.ChiSq = 107.979
1.00 2.df = 50
.62 .83 1.01 3.ChiSq/df = 2.160
4.GFI = .905
5.AGFI = .851
X24 X23 X22 X21 6.CFI = .970
7.RMSEA = .086
1 1 1 1
1.41 .96 .04 .18

e9 e8 e7 e6

.51

Figure 11: Low Education Group: The Output for the Constrained Model

The Chi-Square Value and DF for the constrained model

Model NPAR CMIN DF P CMIN/DF
Default model 28 107.979 50 0.000 2.160
Saturated model 78 0.000 0
Independence model 12 2000.617 66 0.000 30.312

159
 .86 .19 .24 .25

e1 e3 e4 e5

1 1 1 1

X11 X13 X14 X15

1.55
1.00 1.44 .72
1.54

e14 .09
X1 1
Y1 e10
1 1.00
.69 .04
.51
1
1.03 Y2 e11

.49 Y 1.09
.20
1
Y3 e12
.14 1.06
1.31 .72 .12
1
Y4 e13
X2 Fitness Indexes
1.ChiSq = 104.520
1.00 2.df = 49
.62 .83 1.01 3.ChiSq/df = 2.133
4.GFI = .906
5.AGFI = .851
X24 X23 X22 X21 6.CFI = .971
7.RMSEA = .085
1 1 1 1
1.41 .96 .05 .18

e9 e8 e7 e6

.51

Figure 12: Low Education Group: The Output of the Unconstrained Model

The Chi-Square Value and DF for the unconstrained model

Model NPAR CMIN DF P CMIN/DF

Default model 29 104.520 49 0.000 2.133
Saturated model 78 0.000 0
Independence model 12 2000.617 66 0.000 30.312

160
 Table 1a: The Moderation Test for Low Education Group Data

Constrained Unconstrained Chi-Square Result on Result on

Model Model Difference Moderation Hypothesis
Chi-Square 107.979 104.520 3.459 Not Significant Not Supported

DF 50 49 1
GFI 0.905 0.906
AGFI 0.851 0.851
CFI 0.970 0.971
RMSEA 0.086 0.085
CMIN/DF 2.160 2.133
The hypothesis statement: Not
Ha: Respondent’s education moderates the relationship between X1 and Y Supported

***The moderation test is not significant since the difference in Chi-Square value
between the constrained and unconstrained model is less than 3.84.

The difference in Chi-Square value is 3.459 (107.979 - 104.520), while the

difference in Degrees of Freedom is 50 – 49 = 1. For the test to be significant, the
difference in Chi-Square value must be higher than the value of Chi-Square with 1 degree
of Freedom, which is 3.84.

The procedure for performing the test of moderation for the same variable
(education) using another dataset (high education group) is carried out in Table 1b. The
test of hypothesis should the produce the same result.

If the result differs, then go back to the original data and redefine the levels of
education. Regroup the data based on the new definition for low education level and high
education level. Repeat the same procedure again.

161
 .92 .27 .29 .14

e1 e3 e4 e5

1 1 1 1

X11 X13 X14 X15

1.73 1.61
.79
1.00 1.93

e14
X1 1
Y1 e10 .12
1 1.00
1.00
.32
1
1.04 Y2 e11 .06

.34 Y 1.09
.17
1
Y3 e12
.17
1.08 1.21 .15
1
Y4 e13
X2 1.07
Fitness Indexes
1.ChiSq = 75.987
.66 2.df = 50
.79 .97 1.00 3.ChiSq/df = 1.520
4.GFI = .928
5.AGFI = .888
X24 X23 X22 X21 6.CFI = .985
7.RMSEA = .058
1 1 1 1
1.38 1.17 .12 .20

e9 e8 e7 e6

.40

Figure 13: High Education Group: The Output for the Constrained Model

The Chi-Square value and DF for the constrained model

Model NPAR CMIN DF P CMIN/DF

Default model 28 75.987 50 0.010 1.520
Saturated model 78 0.000 0
Independence model 12 1760.721 66 0.000 26.678

162
 .91 .27 .29 .14

e1 e3 e4 e5

1 1 1 1

X11 X13 X14 X15

1.53 1.41
.78
1.00 1.70

e14
X1 1
Y1 e10 .12
1 1.00
.70
.42
1
1.04 Y2 e11 .06

.39 Y 1.10
.17
1
Y3 e12
.23
1.08 1.21 .15
1
Y4 e13
X2 1.07
Fitness Indexes
1.ChiSq = 73.575
.66 2.df = 49
.79 .98 1.00 3.ChiSq/df = 1.502
4.GFI = .930
5.AGFI = .888
X24 X23 X22 X21 6.CFI = .985
7.RMSEA = .057
1 1 1 1
1.38 1.17 .12 .20

e9 e8 e7 e6

.40

Figure 14: High Education Group: The Output for the Unconstrained Model

The Chi-Square Value and DF for the unconstrained model

Model NPAR CMIN DF P CMIN/DF
Default model 29 73.575 49 0.013 1.502
Saturated model 78 0.000 0
Independence model 12 1760.721 66 0.000 26.678

163
 Table 1b: The Moderation Test for High Education level

Constrained Unconstrained Chi-Square Result on Result on

Model Model Difference Moderation Hypothesis
Chi-Square 75.987 73.575 2.412 Not Significant Not Supported

DF 50 49 1
GFI 0.928 0.930
AGFI 0.888 0.888
CFI 0.985 0.985
RMSEA 0.058 0.057
Chisq/df 1.520 1.502
The hypothesis statement: Not
Ha: Respondent’s education moderates the relationship between X1 and Y Supported

***The moderation test is not significant since the difference in Chi-Square value
between the constrained and unconstrained model is less than 3.84. The difference in
Chi-Square value is 2.412 (75.987 - 73.575), while the difference in Degrees of Freedom
is 50 – 49 = 1. For the test to be significant, the difference in Chi-Square value must be
higher than the value of Chi-Square with 1 degree of Freedom, which is 3.84.

The test of hypothesis for moderation that has been carried out found that
the moderator variable “respondents’ education” does not moderate the causal
effects of X1 on Y.

Suppose that the researcher has another objective - to determine whether the same
moderator variable (respondents’ education) moderates another causal path namely X2 to
Y. Now the selected path has changed to the new path (X2 to Y). To test the moderation
effect of respondents’ education for the new path, the researcher needs to repeat the same
procedure that has been explained earlier.

The analysis and moderation test for the new path is explained in the following
example.

164
 e1 e3 e4 e5

1 1 1 1

X11 X13 X14 X15

e14
X1 1
Y1 e10
1 1

1
Y2 e11

Y
1
Y3 e12
1

1
Y4 e13
X2
1

X24 X23 X22 X21 Constrained

1 1 1 1

e9 e8 e7 e6
constrained

Figure 15: The Constrained Model: The Parameter in the Selected Path (X2 to Y) is
Constrained to 1

Recall: The parameter constraint is fixed in the path where the moderation effect
will be examined, and the data is split based on the moderator variable of interest.

In the above example, the path of interest is X2 to Y and the moderator variable to
be tested is respondents’ education. Let’s begin the analysis using the low education
group.

165
 The output for the constrained and unconstrained models is presented in Figure 16
and Figure 17, respectively. The test of moderation is carried out in Table 2a.

.87 .18 .26 .24

e1 e3 e4 e5

1 1 1 1

X11 X13 X14 X15

1.57
1.00 1.44 1.21
1.56

e14 .09
X1 1
Y1 e10
1 1.00
-.10 .05
.51
1
.99 Y2 e11

.46 Y 1.06
.20
1
Y3 e12
1.00 1.02
1.12 .73 .12
1
Y4 e13
X2 Fitness Indexes
1.ChiSq = 185.363
1.00 2.df = 50
.69 .91 1.08 3.ChiSq/df = 3.707
4.GFI = .848
5.AGFI = .763
X24 X23 X22 X21 6.CFI = .930
7.RMSEA = .131
1 1 1 1
1.39 .94 .06 .20

e9 e8 e7 e6

.49

Figure 16: Low Education Group: The Output for Constrained Model

The Chi-Square value and DF for the constrained model

Model NPAR CMIN DF P CMIN/DF
Default model 28 185.363 50 0.000 3.707
Saturated model 78 0.000 0
Independence model 12 2000.617 66 0.000 30.312

166
 .86 .19 .24 .25

e1 e3 e4 e5

1 1 1 1

X11 X13 X14 X15

1.55
1.00 1.44 .72
1.54

e14 .09
X1 1
Y1 e10
1 1.00
.69 .04
.51
1
1.03 Y2 e11

.49 Y 1.09
.20
1
Y3 e12
.14 1.06
1.31 .72 .12
1
Y4 e13
X2 Fitness Indexes
1.ChiSq = 104.520
1.00 2.df = 49
.62 .83 1.01 3.ChiSq/df = 2.133
4.GFI = .906
5.AGFI = .851
X24 X23 X22 X21 6.CFI = .971
7.RMSEA = .085
1 1 1 1
1.41 .96 .05 .18

e9 e8 e7 e6

.51

Figure 17: Low Education Group: The Output for Unconstrained Model

The Chi-Square Value and DF for the Unconstrained Model

Model NPAR CMIN DF P CMIN/DF

Default model 29 104.520 49 0.000 2.133
Saturated model 78 0.000 0
Independence model 12 2000.617 66 0.000 30.312

167
 Table 2a: The Moderation Test for Low Education Group Data

Constrained Unconstrained Chi-Square Result on Result on

Model Model Difference Moderation Hypothesis
Chi-Square 185.363 104.520 80.843 Significant Supported

DF 50 49 1
GFI 0.848 0.906
AGFI 0.763 0.857
CFI 0.930 0.971
RMSEA 0.131 0.085
Chisq/df 3.707 2.133
The hypothesis statement:
Ha: Respondent’s education moderates the relationship between X2 and Y Supported

***The moderation is significant since the difference in Chi-Square value between the
constrained and unconstrained model is more than 3.84. The difference in Chi-Square
value is (185.363 – 104.520) = 80.843, while the difference in Degrees of Freedom is 50
– 49 = 1. For the test to be significant, the difference in Chi-Square value must be higher
than the value of Chi-Square with 1 degree of Freedom, which is 3.84

The test of hypothesis for moderation that has been carried out found that the
moderator variable “respondents’ education” does moderate the causal effects of X2 on
Y.
The procedure for performing the test of moderation for the same variable
(education) using another data-set (data 2) is carried out in Table 2b. The test of
hypothesis is expected to produce the same result that the respondents’ level of education
does moderate the causal effects of X2 on Y.

168
 .91 .26 .30 .14

e1 e3 e4 e5

1 1 1 1

X11 X13 X14 X15

1.53 1.40
1.04
1.00 1.69

e14
X1 1
Y1 e10 .12
1 1.00
.06
.42
1
1.00 Y2 e11 .06

.36 Y 1.06
.18
1
Y3 e12
1.00
.85 1.18 .15
1
Y4 e13
X2 1.06
Fitness Indexes
1.ChiSq = 123.410
.75 2.df = 50
.90 1.10 1.00 3.ChiSq/df = 2.468
4.GFI = .884
5.AGFI = .819
X24 X23 X22 X21 6.CFI = .957
7.RMSEA = .097
1 1 1 1
1.36 1.15 .11 .26

e9 e8 e7 e6

.39

Figure 18: High Education Group: The Output for Constrained Model

The Chi-Square value and DF for the constrained model

Model NPAR CMIN DF P CMIN/DF

Default model 28 123.410 50 0.000 2.468
Saturated model 78 0.000 0
Independence model 12 1760.721 66 0.000 26.678

169
 .91 .27 .29 .14

e1 e3 e4 e5

1 1 1 1

X11 X13 X14 X15

1.53 1.41
.78
1.00 1.70

e14
X1 1
Y1 e10 .12
1 1.00
.70
.42
1
1.04 Y2 e11 .06

.39 Y 1.10
.17
1
Y3 e12
.23
1.08 1.21 .15
1
Y4 e13
X2 1.07
Fitness Indexes
1.ChiSq = 73.575
.66 2.df = 49
.79 .98 1.00 3.ChiSq/df = 1.502
4.GFI = .930
5.AGFI = .888
X24 X23 X22 X21 6.CFI = .985
7.RMSEA = .057
1 1 1 1
1.38 1.17 .12 .20

e9 e8 e7 e6

.40

Figure 19: High Education Group: The output for Unconstrained Model

The Chi-Square Value and DF for the Unconstrained Model

Model NPAR CMIN DF P CMIN/DF

Default model 29 73.575 49 0.013 1.502
Saturated model 78 0.000 0
Independence model 12 1760.721 66 0.000 26.678

170
 Table 2b: The Moderation Test for High Education Group Data

Constrained Unconstrained Chi-Square Result on Result on

Model Model Difference Moderation Hypothesis
Chi-Square 123.410 73.575 49.835 Significant Supported
DF 50 49 1
GFI 0.884 0.930
AGFI 0.819 0.888
CFI 0.957 0.985
RMSEA 0.097 0.057
Chisq/df 2.468 1.502
The hypothesis statement:
Ha: Respondent’s education moderates the relationship between X2 and Y Supported

***The moderation test is significant since the Chi-Square difference between the
constrained and unconstrained model is greater than 3.84. Recall: The Chi-Square value
with 1 degree of freedom is 3.84.

Referring to Table 2b: All fitness indexes for the unconstrained model is
significantly better (smaller Chi-Square) than the constrained model, indicating that the
two group’s coefficient differ.

The results show support for the hypothesis that education moderates the
relationship between latent exogenous construct X2 and its corresponding latent
endogenous construct Y.

Once the moderation effect is established, the study might be interested to

determine in which group (low education or high education) the relationship between X2
on Y is more pronounced?

171
 To address this particular research question, the researcher needs to run the
unconstrained model separately using both datasets (Low Education and High
Education). Compare the standardized parameter estimates and its significance for both
datasets. The result is presented in Figure 21 for dataset 1 (low education), and Figure 22
for dataset 2 (high education).

7.8 COMPARING THE GROUP EFFECTS FOR A MODERATOR

VARIABLE

Suppose one has the following research question to address: In which group (Low
Education or Higher Education), the effect of moderator variable (Education) is more
pronounced? To address this RQ, the researcher needs to obtain the standardized estimate
for the path of interest for both datasets. The procedure is demonstrated in Figure 20 and
Figure 21.

172
 e1 e3 e4 e5

.37 .87 .81 .83

X11 X13 X14 X15

.93
.61 .90
.91

e14 .92
X1
Y1 e10
.48 .96
.97

.98 Y2 e11

.60 Y .93
.86

Y3 e12
.15 .34 .79
.63 .31

Y4 e13
X2 Fitness Indexes
1.ChiSq = 104.520
.94 2.df = 49 Low Education
.51 .70 .98 3.ChiSq/df = 2.133 Group
.26 .49 .97 .88 4.GFI = .906
5.AGFI = .851
X24 X23 X22 X21 6.CFI = .971 Slope = 0.15, Not
7.RMSEA = .085 Significant

e9 e8 e7 e6

.44
Figure 20: The Standardized Beta Estimate for Low Education Group in Path X 2 to
Y

The Effect of X2 on Y is Not Significant for “Low Education” Group

Standardized beta Result

P
Estimate
Not Significant at
Y X2 0.15 0.077
0.05

173
 e1 e3 e4 e5

.32 .78 .74 .90

X11 X13 X14 X15

.88 .86
.56 .95

e14 .90
X1
Y1 e10
.42 .95
.96

.98 Y2 e11

.58 Y .94
.89

Y3 e12
.22 .34 .62
.79 .35

Y4 e13
X2 Fitness Indexes
1.ChiSq = 73.575
.50 2.df = 49
.61 .95 .92 3.ChiSq/df = 1.502 High Education
.25 .37 .89 .84 4.GFI = .930
5.AGFI = .888 Group
X24 X23 X22 X21 6.CFI = .985
7.RMSEA = .057
Slope = 0.22,
Significant

e9 e8 e7 e6

.32
Figure 21: The Standardized Beta Estimate for High Education Group in Path X 2 to
Y

The Effect of X2 on Y is Significant for “High Education” Group

Standardized beta Result

P
Estimate
Y X2 0.22 0.011 Significant at 0.05

***The standardized parameter estimate for “High Education Group” is 0.22, while the
same estimate for “Low Education Group” is 0.15. Thus, one can conclude that the effect
of X2 on Y is more pronounced in “Higher Education Group” compared to “Low
Education Group”.
174
 Now the researcher wants to determine the type of moderation that occurs in the
X2 and Y relationship. The results show that the type of moderation is full moderation
since the standardized estimate for High Education is significant, while the standardize
estimate for Low Education is not significant. If both estimates are significant, then
partial moderation occurs.

175
 CHAPTER 8

THE SECOND ORDER CONFIRMATORY

FACTOR ANALYSIS
The Second Order CFA is a statistical method employed by the researcher to confirm that
the theorized construct in a study loads into certain number of underlying sub-constructs.
For example, the theory posits that service quality construct consists of five underlying
sub-constructs and each sub-construct is measured using certain number of items. The
researcher might want to estimate the effect of main construct on its sub-constructs. Here,
the main construct has become second order construct, while the sub-constructs have
become the first order construct.

8.1 THE STEPS IN PERFORMING SECOND ORDER CFA

The procedure of performing the second order factor analysis is as follows:

Step 1: In AMOS Graphic, draw the Main Construct followed by the Sub-Constructs or
components.

Using the one sided arrow, link the Main Construct to its Sub-Constructs. Put the
residual for every Sub-Construct since the Sub-Construct has an arrow pointing to it from
the Main Construct. Put a parameter ‘1’ on one of the arrows pointing to Sub-Construct.
If the Main Construct has only two Sub-Constructs, then both Sub-Constructs must have
parameter ‘1’.

176
Step 2: Run the Second Order CFA for the main construct on its sub-constructs

In this step, the researcher estimates the causal effects from the main construct to
all its sub-constructs. The objective here is to estimate the factor loading of main
construct on its sub-constructs in order to confirm that the theorized second order
construct loads into the respective sub-constructs.

8.2 PERFORMING SECOND ORDER CFA FOR TRAINING

TRANSFER CONSTRUCTS

In this study, the researcher is estimating the Training Transfer construct (main construct)
on its three sub-constructs namely Knowledge, Skills, and Attitude. The three latent sub-
constructs are measured using certain number of items.

As usual, the researcher runs the CFA to validate the measurement models of the
latent constructs for unidimensionality, validity, and reliability. The Second Order
Construct is shown in Figure 1.

177
 The model is estimating the effects of
Training Transfer on its sub-constructs.
Thus, the residual is required.

Sub-Construct

Main Construct

Figure 1: Estimating the Factor Loading for the Main Construct (Second Order) Training
Transfer

In the above diagram, Training Transfer is the main construct while Knowledge,
Skills, and Attitude are three sub-constructs. In second order CFA, the main construct
Training Transfer will become second order construct, and the three sub-constructs will
become the first order constructs.

178
 The Second Order CFA results are presented in Figure 2.

R2 for second order construct R2 for first order

construct

Factor loading for

second order
construct

Factor loading for first order construct

Figure 2: The Factor Loading for Second Order as well as the First Order Constructs

First of all, observe that all fitness indexes for the second order have achieved the
required level. Thus, no item deletion and modification is required. The results showed
that Training Transfer construct loads well on its three sub-constructs. The factor loading
of Training Transfer on Knowledge, Skills, and Attitude are 0.91, 0.97, and 0.84,
respectively. Furthermore, the R2 for all sub-constructs are high (0.83, 0.93, and 0.7),
which reflect the contribution of Training Transfer on its three constructs is good. In
other word, the theory that Training Transfer consists of three sub-constructs is well
supported.

179
 Figure 3: The Regression Weight of Training Transfer Construct on its Sub-Constructs

Table 1: The Path Analysis and Its Significance

Estimate S.E. C.R. P Results

TRAINING_
Attitude 0.754 0.067 11.264 0.001 Significant
TRANSFER
TRAINING_
Knowledge 0.909 0.063 14.341 0.001 Significant
TRANSFER
TRAINING_
Skills 1.000
TRANSFER

Table 1 shows the effect of Training Transfer on all sub-constructs are highly
significant.

180
8.3 PERFORMING SECOND ORDER CFA FOR JOB
SATISFACTION CONSTRUCTS

In this example, the study wants to confirm that the main construct Job Satisfaction
consists of six sub-constructs as given below:

1) The chance of getting a promotion (Promotion).

2) The pay or remuneration received (Remuneration).
3) The relationship with top management (Management).
4) The working environment or atmosphere (Environment).
5) The relationship with colleagues (Relationship).
6) The workload assigned to an employee (Workload).

As shown in Figure 4, each of the six sub-constructs has its own measurement
model. Firstly, the researcher performs the first order CFA for all measurement models of
the sub-constructs. Secondly, the researcher links the Job Satisfaction construct to all
measurement models of sub-constructs using single headed arrow. Since the single-
headed arrow indicates a causal effect of Job Satisfaction on its sub-constructs, the model
needs to include the residual estimate accordingly (Figure 4).

181
 e1 e2 e3
First Order
1 1 1
Residual Estimate
Construct
Prm1 Prm2 Prm3

1
1 1
e20 wl11 1 promo Rem1 e4
e42
tion 1
1 1
e19 wl10 Rem2 e5

1 work remu 1
e18 wl9 Rem3 e6
load neration
1 1
e17 wl8 1 1 Rem4 e7

1 1 1
e16 wl7 e37 e38 Rem5 e8
1 1
e34 Rel8 Mgt1 e9
Job
1 Satisfaction 1
e33 Rel7 Mgt2 e10
1
1 1
e32 Rel6 e39 e40 Mgt3 e11

1 1 1 1
e31 Rel5 Mgt4 e12
1
1 relation 1
e30 Rel4 mgmnt Mgt5 e13
ship
1 1
e29 Rel3 Mgt6 e14
1
1 1
e28 Rel2 Mgt7 e15
environ 1
1 e41 1
e36 Rel1 ment Mgt8 e35
1

Env1 Env2 Env3 Env4 Env5 Env6 Env7

Reference 1 1 1 1 1 1 1

Point e27 e26 e25 e24 e23 e22 e21

Figure 4: Modeling the Second Order CFA for Job Satisfaction on Its Sub-Constructs

Note: The parameter “1” is required at one of the arrows pointing towards the sub-constructs as a
reference point in the analysis. In the above diagram, the arrow pointing towards sub-construct
Environment is selected as a reference point.

182
 e1 e2 e3
Factor Loading for Factor Loading for
.73 .64 .61
First Order Construct Second Order Construct
Prm1 Prm2 Prm3

.80
.86 .78
.49 .87
e20 wl11 promo Rem1 e4
e42
.70 tion .93
.81 .88
e19 wl10 .90 .94 Rem2 e5
.41
.51 .71 .85
work remu
e18 wl9 Rem3 e6
.94 load .60 .44 neration .92
.88 .64
.79
e17 wl8 .89 Rem4 e7
.77 .66
.90 .74 .55
.81
e16 wl7 e37 e38 Rem5 e8
.76 .72
e34 Rel8 Mgt1 e9
Job
.84 .79
e33 Rel7
Satisfaction Mgt2 e10
.85
.87
.77 .89 .68
e32 Rel6 .92 e39 e40 Mgt3 e11
.78 .82
.88 .82
.88 .72
e31 Rel5 .94 .85 Mgt4 e12

.80
.84 .86
.90 relation .93
e30 Rel4 .61 .68 mgmnt Mgt5 e13
.94 ship
.79
.89 .62
e29 Rel3 Mgt6 e14
.90
.80 .71 .80 .64
e28 Rel2 .84 Mgt7 e15
.77
environ
.71 e41 .60
e36 Rel1 ment Mgt8 e35
.87 .89
.89 .90 .75 .61 .87

Env1 Env2 Env3 Env4 Env5 Env6 Env7

.76 .79 .79 .82 .56 .38 .75 Item to

e27 e26 e25 e24 e23 e22 e21 delete: Mgt8

Figure 5: The Standardized Factor Loading for Each Component of Job Satisfaction
Construct

183
 Figure 5 presents the factor loading of Job Satisfaction towards every component.
Since all factor loadings are higher than 0.6, the study managed to prove that Job
Satisfaction consists of those six components. The component and their respective factor
loading are Workload (0.77), Promotion (0.64), Remuneration (0.66), Management
(0.82), Environment (0.84), and Relationship (0.78).

Now, the researcher needs to examine the fitness indexes for the model.

Table 2: The Fitness Indexes for model in Figure 6: The Baseline Comparisons

Model NFI RFI IFI TLI CFI

Default model 0.769 0.753 0.802 0.788 0.802

The baseline comparison indexes indicate that the model is not a good fit to the
data at hand. Thus, the researcher needs to modify the model (as suggested by the
modification indices) in order to improve the fit. The ideal fitness index for the model is
above 0.90.

Table 3: The Fitness Indexes for Model in Figure 6: The RMSEA

Model RMSEA LO 90 HI 90

Default model 0.123 0.118 0.127

The RMSEA (Root Mean Square Error Approximation) of 0.123 also indicates
that the model does not provide a good fit to the data at hand. The value indicates that the
modification to the model is required. The ideal value for RMSEA for the model should
be 0.08 or lower.

As explained earlier, the researcher needs to modify the model based on the
modification indices (MI) proposed in the AMOS output. The MI can be obtained by
ticking the “modification indices” in the output box prior to running the model.

184
 Table 4 presents the list of MI for the above model. The modification is suggested
for every pair of error terms when the correlation is high. The high correlation indicates
that these two items are redundant or correlated. Remember: The items measuring the
same construct should be independent of each other.

In order to implement the modification as listed in Table 4, firstly select the pair
which has the highest MI. Secondly, delete one item from the pair. The procedure is to
delete the item with lower factor loading and to re-specify the model.

In Table 4, the highest MI is 122.652, and the correlated items are between e15
and e35. In terms of item, it is between mgt7 and mgt8. In this case, one should delete
mgt8 since it has a lower factor loading between the two (0.77 vs 0.80). The re-specified
model is shown in Figure 6, and the new output is presented in Figure 7

Table 4: The List of Modification Indices Produced by AMOS

M.I. Par Change

e30 ↔ e36 36.644 -0.171

e28 ↔ e36 41.023 0.152
e26 ↔ e27 58.394 0.260
e22 ↔ e26 37.557 -0.372
e21 ↔ e22 31.987 0.291
e19 ↔ e33 44.329 0.158
e18 ↔ e36 37.123 0.182
e18 ↔ e33 36.393 -0.184
e18 ↔ e28 43.122 0.194
e18 ↔ e20 43.063 0.164

e15 ↔ e35 122.652 0.583

e11 ↔ e15 32.021 -0.319

e9 ↔ e10 36.384 0.181

185
 e1 e2 e3
1 1 1

Prm1 Prm2 Prm3

1
1 1
e20 wl11 1 promo Rem1 e4
e42
tion 1
1 1
e19 wl10 Rem2 e5

1 work remu 1
e18 wl9 Rem3 e6
load neration
1 1
e17 wl8 1 1 Rem4 e7

1 1 1
e16 wl7 e37 e38 Rem5 e8
1 1
e34 Rel8 Mgt1 e9
Job
1 Satisfaction 1
e33 Rel7 1 Mgt2 e10

1 1
e32 Rel6 e39 e40 Mgt3 e11

1 1 1 1
e31 Rel5 Mgt4 e12
1
1 relation 1
e30 Rel4 mgmnt Mgt5 e13
ship
1 1
e29 Rel3 Mgt6 e14
1
1 1
e28 Rel2 Mgt7 e15
environ 1
1 e41
e36 Rel1 ment
1

Env1 Env2 Env3 Env4 Env5 Env6 Env7 Item Mgt8

1 1 1 1 1 1 1
has been
e27 e26 e25 e24 e23 e22 e21 deleted

Figure 6: The Diagram shows Item e35 (mgt8) has Been Deleted as Suggested by the MI in
Table 4

186
 e1 e2 e3

.73 .64 .61

Prm1 Prm2 Prm3

.80
.86 .78
.49 .87
e20 wl11 promo Rem1 e4
e42
.70 tion .93
.82 .88
e19 wl10 .90 .94 Rem2 e5
.41
.51 .71 .85
work remu
e18 wl9 Rem3 e6
.94 load .60 .44 neration .92
.88 .64
.79
e17 wl8 .89 Rem4 e7
.77 .66
.90 .74 .55
.81
e16 wl7 e37 e38 Rem5 e8
.76 .74
e34 Rel8 Mgt1 e9
Job
.84 .82
e33 Rel7
Satisfaction Mgt2 e10
.86
.87
.77 .90 .69
e32 Rel6 .92 e39 e40 Mgt3 e11
.79 .83
.88 .83
.88 .70
e31 Rel5 .94 .84 Mgt4 e12

.80
.84 .86
.90 relation .93
e30 Rel4 .62 .68 mgmnt Mgt5 e13
.94 ship
.78
.89 .60
e29 Rel3 Mgt6 e14
.90
.80 .71 .78 .60
e28 Rel2 .84 Mgt7 e15
environ
.71 e41
e36 Rel1 ment
.87 .89
.89 .90 .75 .61 .87

Env1 Env2 Env3 Env4 Env5 Env6 Env7

.76 .80 .79 .82 .56 .38 .75

e27 e26 e25 e24 e23 e22 e21

Figure 7: The New Factor Loadings after the Required Modification is Made to the Model
(Mgt8 is Deleted)

187
 Table 5: The Fitness Indexes for Model in Figure 8: The Baseline Comparisons

Model NFI RFI IFI TLI CFI

Default model 0.929 0.910 0.913 0.907 0.912

All fitness indexes in Table 5 are satisfactory since they are well above 0.9. The
fitness indexes indicate that the proposed model is a good fit to the data at hand.

Table 6: The Fitness Indexes for Model in Figure 8: The RMSEA

Model RMSEA LO 90 HI 90
Default model 0.077 0.100 0.108

The RMSEA presented in Table 6 is lower than 0.08. This value also indicates
that the proposed model is a good fit to the data at hand.

Note: All fitness indexes have improved to the acceptable level after the
suggested modifications are implemented to the model. Thus, no other modification is
required. Now the researcher could proceed with further analysis (Table 7).

Table 7: The Standardized Factor Loading for Each Component

Component Construct Standardized Factor Loading

Promotion ← Job Satisfaction 0.64
Remuneration ← Job Satisfaction 0.66
Management ← Job Satisfaction 0.83
Environment ← Job Satisfaction 0.84
Relationship ← Job Satisfaction 0.79
Workload ← Job Satisfaction 0.77

The standardized factor loadings in Table 7 are above 0.6 for all components. In
other words, the Job Satisfaction construct consists of these six components. In this case,
no component should be dropped from the model. Figure 8 presents the unstandardized
regression weight.

188
 .61 .63 1.34

e1 e2 e3
1 1 1

Prm1 Prm2 Prm3

.82
.99 1.00 1.12
1 1
.92 e20 wl11 1 promo Rem1 e4 .38
e42
.79 tion 1.00
1 1
.34 e19 wl10 .97 Rem2 e5 .33
1.04

1 .71 work remu 1

.69 e18 wl9 Rem3 e6 .50
1.17 load neration 1.06
.69
1 1
.26 e17 wl8 1 1 .94 Rem4 e7 .63
.76 .89
1 1.00 .78 1
.33
e16 wl7 e37 e38 Rem5 e8 1.31
1 1
.52 e34 Rel8 .56 1.47 Mgt1 e9 .48
Job
1 1.44 1
.35
e33 Rel7
Satisfaction Mgt2 e10 .40
.52 1.00
1.10 .43
1 1.15 1
.64 e32 Rel6 1.14e39 e40 Mgt3 e11 .99
.77 .81
1.24 1.25
1 1 1 1
.23 e31 Rel5 1.09 1.16 Mgt4 e12 .78
1.00
1 1.20 relation 1.17 1
.48 e30 Rel4 mgmnt Mgt5 e13 .30
1.11 ship
.93
1 1
.21 e29 Rel3
1.00
Mgt6 e14 .78

1 1.01 1
.33 e28 .59
Rel2 .87 Mgt7 e15 .92
environ 1
1 e41
.43 e36 Rel1 ment
.92 .97 .75
.88 1.00
1.03 1.07

Env1 Env2 Env3 Env4 Env5 Env6 Env7

1 1 1 1 1 1 1

e27 e26 e25 e24 e23 e22 e21

.54 .49 .41 .48 1.84 1.90 .68

Figure 8: The Regression Weights of Job Satisfaction on Its Sub-Constructs

189
 Note: The effects of Job Satisfaction on its respective components.

Table 8: The Regression Weights for Every Construct in the Model (in Figure 8)

Sub- Results
Construct Estimate S.E. C.R. P
Construct
Promotion ← Job Satisfaction 0.691 0.072 9.638 *** Significant

Environment ← Job Satisfaction 1.000 Reference point

Management ← Job Satisfaction 0.761 0.062 12.192 *** Significant

Relationship ← Job Satisfaction 0.748 0.059 12.576 *** Significant

Workload ← Job Satisfaction 0.766 0.062 12.409 *** Significant

Remuneration ← Job Satisfaction 0.883 0.082 10.758 *** Significant

*** Indicate highly significant, p-value < 0.001

The probability values in Table 8 indicate that Job Satisfaction has significant
effects on all constructs. Meaning: The second order construct job satisfaction consists of
the respective six first order constructs.

Once the second order CFA is completed, the researcher could proceed for further
analysis (SEM) using the second order construct (with the measurement models just
confirmed) – in this example the second order construct (new construct) is Job
Satisfaction.

In Figure 9, Job Satisfaction now consists of six components. Each component

comes from the respective measurement model. All items in the respective components
shown in Figure 8 have been condensed into one score as presented if Figure 9.

190
 e1 e2 e3 e4 e5 e6

1 1 1 1 1 1

Promotion remnrtion Mngmnt Envirmnt Reltnship Workload

Job
Satisfaction

Figure 9: The Measurement Model for Job Satisfaction

8.4 PERFORMING SECOND ORDER CFA FOR

TECHNOLOGICAL INNOVATION CONSTRUCTS

The study is interested to perform Second Order CFA in order to confirm that the
construct namely Technological Innovation consists of six sub-constructs. The theorized
six sub-constructs are:

1) Marketing and Planning (3 items).

2) Idea and Creativity (5 items).
3) Commercialization (3 items).
4) Collaborations (2 items).
5) Internationalization (3 items).
6) Research and Development (3 items).

191
 e14
e13 e15
1
1 1

Mktg2 e4 e5 e6
Mktg1 Mktg3
e1 e2 e3
1 1 1 1
1 1 1

RdF1 RdF2 RdF3

Marketing 1
IdeaF1 IdeaF2 IdeaF3 e22
Planning 1
1

Research 1
e21
1 Idea Development
e18
Creativity

1
Technological
Innovation

1 International 1
Commercial e20
e19 ization
ization

Collaborations
1
1

1
ComF1 ComF2 1 IntnF1 IntnF2 IntnF3
ComF3

CollbF1 CollbF2
1 1 1 1 1 1
1 e23 1
e12 e11 e10 e9 e8 e7
e17 e16

Figure 10: The Second Order CFA Model for Technological Innovation

Note: As usual, one of the components namely Research & Development is selected as a
reference point.

192
 e14
e13 e15

.86

Mktg2 e4 e5 e6
.64 Mktg1 Mktg3 .11
e1 e2 e3
.80 .93
.69
.90 .33
.68 RdF1 RdF2 RdF3 .01
.01 IdeaF1 IdeaF2 IdeaF3 .34 Marketing
e22
Planning .82 .83
.11 .95 .86 .10
.58

Research
.93 e21
Idea Development
e18
Creativity

.63
.38 .79
.61
Technological
Innovation

.72 .87
.86
.51 .76

Commercial
.74 International
e19 e20
ization ization

Collaborations
.76 .45 .83 .20
.92 .56
.84 .31
.92 .84
.58 ComF1 ComF2 ComF3 .20 .69 IntnF1 IntnF2 IntnF3 .04
CollbF1 .84 CollbF2 .71

e23
e12 e11 e10 e9 e8 e7
e17 e16

Figure 11: The Factor Loading for Each Component in the Second Order CFA

The output in Figure 11 indicates that the factor loading for all components are
above 0.6. In this case, the study has proved that Technical Innovation constructs really
consists of the underlying six components.

Once proven that Technological Innovation consists of these components, the

study could proceed into further analysis by using all components in the model.

193
 In Figure 12, the study is estimating the effect of Technological Innovation
construct on each of its underlying components.

.08
.23 .56
e14
e13 e15
1
1 1
.06 .07 .29
.20 .02 .07
Mktg2 e4 e5 e6
Mktg1 Mktg3
e1 e2 e3
.92 1.00 1 1 1
1 1 1
.37 .07
RdF1 RdF2 RdF3
Marketing 1
IdeaF1 IdeaF2 IdeaF3 e22
Planning .91 1.00
.12 1.00 .14
.06
.45
.11
Research 1
2.10 e21
1 Idea Development
e18
Creativity

.10
1.00
.84
Technological
Innovation

1.32 1.56
2.88

.16 .07

1 International 1
Commercial e20
e19 ization
ization

Collaborations
.90 .49 1.00 .19
1.00 .63

1.00 .84
ComF1 ComF2 1 IntnF1 IntnF2 IntnF3
ComF3

CollbF1
.28
CollbF2
1 1 1 1 1 1
.19 .06 .31 1 1 .14 .27 .27
e23
.20 .31
e12 e11 e10 e9 e8 e7
e17 e16

Figure 12: The Regression Weights for the Second Order Model

194
 The text output is presented in Table 9.

Table 9: The Regression Weights

Beta
S.E. C.R. P
Estimate
Marketing & Technological
← 2.100 0.456 4.602 ***
Planning Innovation
Research & Technological
← 1.000 Reference point
Development Innovation
Technological
Idea & Creativity ← 0.843 0.253 3.330 ***
Innovation
Technological
Commercialization ← 1.318 0.354 3.724 ***
Innovation
Technological
Internationalization ← 1.557 0.386 4.028 ***
Innovation
Technological
Collaborations ← 2.875 0.665 4.320 ***
Innovation

The results in Table 9 indicates that Technological Innovation has significant

effects on all six components namely Marketing and Planning, Research and
Development, Idea and Cretativity, Commercialization, Internationalization, and
Collaborations.

Estimating the effect of every components on another endogenous

construct

Now, the components have become exogenous constructs in the study, and the
study could use these six exogenous constructs for further analysis. For instance, the
study is interested to estimate the effects of these exogenous constructs on the latent
endogenous construct namely NIP. The model for estimating the causal effects is
presented in Figure 13, and the result is presented in Figure 14.

195
 e2 e14 e5
e1 e3 e13 e4 e6
e15
1 1 1
1 1 1 1 1
1

IdeaF2 Mktg2 RdF2

IdeaF1 IdeaF3 RdF1 RdF3
Mktg1 Mktg3

1 1 1

Idea Marketing Research

Creativity Planning Development

1
e20 NipF1

1 1
e19 NipF2 NIP e21

1
1
e18 NipF3

Commercial International
ization ization

1 1
Collaborations

ComF1 ComF3 IntnF3

1 IntnF1
ComF2 IntnF2
1 1 1
1 1
1 CollbF1 CollbF2

e12 e10 1 1 e7
e9
e11 e8

e17 e16

Figure 13: Modeling the Effects of Every Component of TI (Technological Innovation) on

NIP (National Innovation Policy)

196
 e2 e14 e5
e1 e3 e13 e4 e6
e15

.48 .49 .43

IdeaF2 Mktg2 RdF2

.03 IdeaF1 IdeaF3 .63 1.09 RdF1 RdF3 .04
Mktg1 1.14 Mktg3 .14
.69 .70 1.04 .65
.16 1.07
.79 .19
.37

Idea Marketing Research

Creativity Planning Development

0.73
0.79
.17 0.56

e20 NipF1

.41
.58
.76 0.64
e19 NipF2 NIP e21
.54 .74

e18 NipF3

0.48
0.55 0.67

Commercial International
ization ization

.71 .38 .26 .21

1.00 Collaborations 1.72

ComF1
.51
1.00
ComF3 .14 1.13 .57 .07 IntnF1 .95 IntnF3 .04
ComF2 IntnF2

CollbF1 1.27 CollbF2 .33

e12 e10 e7
e9
e11 e8

e17 e16

Figure 14: The Regression Weight for Each Component of TI on NIP

197
 In Figure 14, the study is trying to determine the effects of six exogenous
constructs on NIP (National Innovation Policy). Note: R2 = 0.64, meaning the six
components estimate 64% of the NIP. The text output is presented in Table 10.

Table 10: The Regression Weights and Its Respective Probabilities

Estimate S.E. C.R. P

NIP ← Commercialization 0.551 0.189 2.915 ***
NIP ← Idea & Creativity 0.733 0.293 2.501 0.01
NIP ← Collaborations 1.000 Reference point
NIP ← Internationalization 0.484 0.162 2.987 ***
NIP ← Research & Development 0.785 0.336 2.336 0.01
NIP ← Marketing & Planning 0.564 0.228 2.473 0.01

The results presented in Table 10 indicate that all six exogenous constructs of
Technological Innovation have significant effects on NIP (National Innovation Policy).

From the output, the study can derive the following equation relating NIP to its
components:

NIP = 0.551* Commercialization + 0.733* Idea & Creativity + 1.000*

Collaborations +

0.484* Internationalization + 0.785* Research & Development + 0.564*

Marketing & Planning. The results also indicate that the above components
contribute significantly to NIP.

198
 CHAPTER 9

THE APPLICATION OF AMOS GRAPHIC

IN POSTGRADUATE RESEARCHES

In this chapter, the discussion is focused on analyzing the theoretical model of a research
using Structural Equation Modeling (SEM) with AMOS Graphic. As has been said
earlier, the researcher could convert the exact research model into AMOS Graphic for
further analysis.

9.1 MODELING THE LOYALTY OF BANKING CUSTOMERS

The model is shown in Figure 1. In this study the researcher was trying to analyze the
inter-relationships among four exogenous constructs, two mediating constructs, and one
endogenous construct in a model. The latent constructs involved in this study are:

1. Perceived Service Quality (PSQ). This latent construct is measured using seven
items namely ps1 to ps7. This is an exogenous construct.

2. Perceived Value (PV). This latent construct is also measured using seven items
namely pv1 to pv7. This is an exogenous construct.

3. Communication (COM). This latent construct is measured using four items

namely co1 to co4. This is an exogenous construct.

4. Corporate Image (IMG). This latent construct is measured using eight items
namely im1 to im8. This is an exogenous construct.

5. Customer Satisfaction (CS). This latent construct is measured using five items
namely cs1 to cs5. This is a mediating construct.

6. Customer Trust (TRUST). This latent construct is measured using five items
namely tr1 to tr5. This is a mediating construct.

199
7. Customer Loyalty (CL). This latent construct is measured using six items namely
cl1 to cl6. This is an endogenous construct.

The latent constructs and their respective indicators are modeled in Figure 1.

e1 e2 e3 e4 e5 e6 e7
1 1 1 1 1 1 1

ps1 ps2 ps3 ps4 ps5 ps6 ps7

1

1
e27 e28 e29 e30 e31
e22 pv1 PSQ
1 1 1 1 1
1
e21 pv2
cs1 cs2 cs3 cs4 cs5
1 1
e20 pv3
1
1 CS e44 1
e19 pv4 PV cl1 e37
1 e43
1 1
e18 pv5 cl2 e38
1
1 1
1
e17 pv6 cl3 e39
1 CL
1
e16 pv7 cl4 e40
1
e26 co1 1
cl5 e41
1
e25 co2 1
cl6 e42
1 COM 1
e24 co3 TRUST e45
1
1
1
e23 co4
tr1 tr2 tr3 tr4 tr5
1 1 1 1 1
IMG
1 e36 e35 e34 e33 e32

im1 im2 im3 im4 im5 im6 im7 im8

1 1 1 1 1 1 1 1

e15 e14 e13 e12 e11 e10 e9 e8

Figure 1: The Model of Customer Loyalty towards Service Providers in Banking Industry

200
 Reminder: The researcher should assess the measurement model for all
constructs before modeling the structural model. The assessment for Unidimensionality,
Validity and Reliability for the measurement model should be carried out using
Confirmatory Factor Analysis (CFA).

9.2 MODELING THE LOYALTY OF UNDERGRADUATES

TOWARDS THE POSTGRADUATE STUDY IN A UNIVERSITY

The schematic diagram of the model is shown in Figure 2, while the conversion into
AMOS graphic of the model is shown in Figure 4.

In this study, the researcher was trying to analyze the inter-relationships among
four constructs namely Service Quality, Corporate Image, Students Satisfaction, and
Students Loyalty. In this case, Service Quality and Corporate Image are two latent
exogenous constructs, Students Satisfaction is a mediating construct, and Students
Loyalty is the latent endogenous construct.

The hypotheses in the path of interest to be tested by the study are shown in
Figure 2 and the hypothesis statement is presented in Table 1.

201
Figure 2: Modeling the Loyalty of Outgoing Undergraduates towards University
Postgraduates

The Schematic diagram in Figure 2 presents the five hypotheses of interest in the
study.

The model has four latent constructs namely Service Quality (34 items),
Corporate Image (16 items), Student Satisfaction (9 items), and Students Loyalty (14
items). The researcher employed EFA to analyze every construct in order to assess its
dimensionality of items. The EFA found that every construct is formed by a few sub-
constructs, and every sub-construct has certain number of items. The CFA was carried
out for every construct and the redundant items in every sub-construct were deleted.
Finally, the remaining items in every sub-construct were averaged in order to simplify the
model. The sub-construct has been renamed accordingly as shown in Figure 4.

202
Table 1: The Hypothesis Statement for Every Path of Interest as shown in Figure 2

H1 Service Quality provided by the university has significant and direct effects on
Students’ Satisfaction towards their university
H2 Corporate Image of the university has significant and direct effects on Students’
Satisfaction towards their university
H3 Service Quality provided by the university has significant and direct effects on
Students’ Loyalty towards their university in the future (postgraduate study)
H4 Corporate Image of the university has significant and direct effects on Students’
Loyalty towards their university in the future (postgraduate study)
H5 Students’ Satisfaction with the university has significant and direct effects on their
Loyalty towards their university in the future (postgraduate study)

The diagram (Figure 2) is converted to AMOS graphic (Figure 4). The analysis
required to test these hypotheses are causal effects using path analysis.

Caution: The dataset that has been used during the EFA procedure to obtain the
dimensions for every construct cannot be used in the CFA. Normally in research, the
researcher runs the EFA procedure using data from pilot study or preliminary study,
while the CFA procedure is run using data from the actual study.

Question: The researcher did not carry out pilot study. He collected all data
during actual study. How he should go about doing EFA and CFA?

Supposed without doing a pilot study, a researcher collected all 400 data for the
research. In this case, he needs to split randomly these 400 data into two separate data
files – let say dataset 1 and dataset 2. Assuming dataset 1 is the data he obtained from a
pilot study (minimum 100 data), while dataset 2 was obtained from actual study. He
could run EFA procedure using dataset 1 to obtain the dimensions, and the CFA
procedure using dataset 2 to confirm the underlying dimensions.

Caution: The data that has been used to analyze EFA cannot be re-used for
analyzing CFA. In other words, data cannot be recycled.

203
 Service Corporate Students' Students'
Quality Image Satisfaction Loyalty

1 1 1 1

CI F1 CI F2 CI F3 SS F1 SS F2 SL F1 SL F2
SQ F1 SQ F2 SQ F3

1 1 1 1 1 1 1 1 1 1

e1 e2 e3 e4 e5 e6 e7 e8 e9 e10

Figure 3: The CFA Procedure to Validate All Constructs Involved in the Study

As we can observe in Figure 4, the Service Quality construct (originally 34 items)

has been factored into 3 components as a result of EFA. The researcher renamed the three
components as “Human Elements in a service”, “Academic Facilities”, and
“Infrastructure Facilities”.

The Corporate Image construct (originally 16 items) has been factored into 3
components as a result of EFA. The three components are renamed as “Image of a
University”, “Image of Academic Staffs”, and “External Recognitions”.

The Students’ Satisfaction construct (originally 14 items) has been factored into
two components as a result of EFA. The two components are renamed as “Satisfaction
with the Process” and “Satisfaction with the Outcome”.

The Students’ Loyalty construct (originally 9 items) has been factored into two
components as a result of EFA. The two components are renamed as “Market Retention”
and “Positive Recommendation”.

204
 Now, the structural model in Figure 4 has four constructs namely Service Quality,
Corporate Image, Students’ Satisfaction, and Students’ Loyalty. Every construct has
certain number of dimensions resulted from EFA procedure.

err1 err2 err3

1 1 1

Humans AcadSvc Infras

err9 err10
SERVICE
QUALITY 1 1

RetenN RecomM

err7
1
1

Process 1
STUDENTS' STUDENTS'
SATISFACTION LOYALTY
Outcome
1 1
1

Error2 Error1
err8

CORPORATE
IMAGE

UnivMge AcadMge RecogN

1 1 1

err4 err5 err6

Figure 4: The Conversion of the Schematic Diagram in Figure 2 into AMOS Graphic for
Conducting Path Analysis

205
 err1 err2 err3

.65 .91 .67

Humans AcadSvc Infras

.81 .95 .82

err9 err10
SERVICE
QUALITY
.50 .70
RetenN RecomM
.05
err7 .46
.71 .84
.62
.40 .65
Process .79
STUDENTS' .66 STUDENTS'
.78 .83 LOYALTY
.91 SATISFACTION
Outcome

.21
Error2 .17 Error1
err8

CORPORATE
IMAGE

.82 .42
.90

.68 .81 .18

UnivMge AcadMge RecogN

err4 err5 err6

Figure 5: The Standardized Coefficients for Students’ Loyalty towards University for
Postgraduates

206
 .43 .12 .52

err1 err2 err3

1 1 1

Humans AcadSvc Infras

1.00 1.22 1.16

.79 .78 .40
err9 err10
SERVICE
QUALITY 1 1

.45 RetenN RecomM

.05
err7 .44
1.00 1.10
1

Process 1.00
STUDENTS' .67 STUDENTS'
.25 LOYALTY
1.25 SATISFACTION
Outcome
1 1
1 .44 .27
.23
.51
Error2 .42 Error1
err8

.13

CORPORATE
IMAGE

2.83 1.00
2.53

UnivMge AcadMge RecogN

1 1 1
.49 .20 .60

err4 err5 err6

Figure 6: The Regression Weights for Students’ Loyalty towards University for
Postgraduates

207
 Table 2: The Path Analysis for All Constructs and Variables in the Figure 6

Beta Standard Critical

Variables PATH Variables P-value
Estimate Error Region

STUDENTS’_SATISFACTION ← SERVICE_QUALITY 0.44 0.06 7.72 0.001

STUDENTS’_SATISFACTION ← CORPORATE_IMAGE 0.51 0.15 3.51 0.001
STUDENTS’_LOYALTY ← CORPORATE_IMAGE 0.42 0.14 2.96 0.001
STUDENTS’_LOYALTY ← SERVICE_QUALITY 0.05 0.06 0.88 0.382
STUDENTS’_LOYALTY ← STUDENTS’_SATISFACTION 0.67 0.05 12.60 0.001
*** Indicate highly significant, p-value < 0.001

Table 3: The Regression Equation for the Main Constructs in the Study

Coefficient of the Error

Endogenous Construct = P-value Result
Exogenous Construct Term
Students’ Satisfaction = 0.44 * Service Quality + Error2 0.001 Significant
Students’ Satisfaction = 0.51 * Corporate Image + Error2 0.001 Significant
Students’ Loyalty = 0.67 * Students’ Satisfaction + Error1 0.001 Significant
Not
Students’ Loyalty = 0.05 * Service Quality + Error1 0.382
Significant
Students’ Loyalty = 0.42 * Corporate Image + Error1 0.001 Significant

208
9.3 MODELING CUSTOMER LOYALTY TOWARDS THEIR
SERVICE PROVIDER

In this study, the postgraduate researcher was trying to estimate the inter-relationships
among six constructs in his study. The constructs are:

1) Financial Bond (Exogenous).

2) Social Bond (Exogenous).
3) Structural Bond (Exogenous).
4) Customer Emotion (Mediating).
5) Relationship Quality (Mediating).
6) Customer Loyalty (Endogenous).

209
 e1 e2 e3 e4
1 1 1 1

Fb1 Fb2 Fb3 Fb4 e13 e14 e15

1 1 1
1
Positive Negative Neutral
Financial
Bonds 1

Customer 1
e22
Emotions
1
e8 Sc1
1
Attitudinal e19
1 1
e7 Sc2
Social Customer 1
Bonds e24
1 Loyalty
e6 Sc3 1
1
1 Behavioral e21
e5 Sc4
Relationship 1
e23
Quality
1
Structural
Bonds
1 Trust Satisfaction Commitment
1 1 1

Sb1 Sb2 Sb3 Sb4 e18 e17 e16

1 1 1 1

e12 e11 e10 e9

Figure 7: The Theoretical Model has been converted into AMOS Graphic for Analysis using
SEM

In Figure 7, the exogenous constructs are Financial Bond, Social Bond, and
Structural Bond. The Customer Emotion construct is a mediator linking the relationship
between three exogenous constructs to Relationship Quality. Customer Loyalty is the
endogenous construct in the model. Another researcher has modified this model and
presented in Figure 8.

210
 e19 e20 e21 e22 e23
1 1 1 1 1
e1 e2 e3 e4 e5
tr1 tr2 tr3 tr4 tr5
1 1 1 1 1
1
rb1 rb2 rb3 rb4 rb5 e43
1
1 e40 Trust 1 1
1 at1 e29
e16 e17 e18
Relational
1 1 1 1
Bond Attitudinal at2 e30

em1 em2 em3 1

at3 e31
1
e15 fb1 1
1 e38 st1
1
e14 fb2
1
e37 st2
1 Financial
1
e13 fb3 Emotion 1 Sat e41
Bond
e36 st3
1
e12 fb4 1 1
1
1
1 e35 st4
e11 fb5
e39 1
1 bh1 e32

1
Behavioral bh2 e33
Strustural
Bond 1
1 1 bh3 e34
1
e42 Comm
sb1 sb2 sb3 sb4 sb5 e44
1
1 1 1 1 1

e10 e9 e8 e7 e6 cm1 cm2 cm3 cm4 cm5

1 1 1 1 1

e28 e27 e26 e25 e24

Figure 8: The Theoretical Model has been converted into AMOS Graphic for Analysis using
SEM

The researchers have modified the model in Figure 7 and have resulted to model in
Figure 8.

In Figure 8, the Relationship Quality has been split into three sub-constructs; Trust,
Satisfaction, and Commitment. Also, the Customer Loyalty construct has been split into
two separate constructs; Attitudinal Loyalty and Behavioral Loyalty. The researcher is
also interested to determine the influence of Attitudinal Loyalty on Behavioral Loyalty.
The study intends to determine the mediating role of Attitudinal Loyalty in linking the
relationship between Trust and Behavioral Loyalty, between Emotion and Behavioral
Loyalty, and also between Commitment and Behavioral Loyalty.

211
9.4 MODELING INVESTOR LOYALTY TOWARDS THE
LISTED COMPANIES

e1 e2 e3
1 1 1

im1 im2 im3

1
e10 e11 e12 e13
1 1 1 1
Investor
Reputation
loy1 loy2 loy3 loy4

1 1
e9 sat1
1 Affective Bhv
e8 sat2 loyalty
1 loyalty
1 1
e7 sat3 1
e15
e14
Company
Reputation
1

im4 im5 im6 Respondents’ SES

1 1 1

(A moderator)
e6 e5 e4

Figure 9: The Theoretical Model has been converted into AMOS Graphic for Analysis using
SEM

In Figure 9, Investor Reputation and Company Reputation are exogenous

constructs, Affective Loyalty is a mediating construct, while Behavioral Loyalty is the
endogenous construct. The study is also intended to assess the effect of moderator
variable namely the investors’ Socio-Economic-Status (SES) in the relationship between
Company Reputation and Behavioral Loyalty.

212
9.5 MODELING THE EFFECTS OF ENTREPRENEURIAL
ORIENTATION ON BUSINESS PERFORMANCE
The model which consists of first order construct, second order construct, and formative
construct is presented in Figure 10.

Figure 10: The Model Consists of Reflective Construct, Formative Construct, and Second
Order Construct

213
9.6 MODELING THE EFFECTS OF UNIVERSITY
REPUTATION ON STUDENTS LOYALTY

Figure 11: Modeling the Effects of University Reputation on Students Loyalty

214
9.7 MODELING THE EFFECTS OF INDIVIDUAL FACTORS,
TRANSFER CLIMATE AND TRAINING DESIGN ON
TRAINING TRANSFER IN AN ORGANIZATION

Example: The Schematic Diagram of Training Transfer in an organization

Figure 12: The Schematic Diagram of Training Transfer Model

215
Modeling the Second Order Construct in AMOS Graphic

Sub Construct/First Order

Construct

Main Construct/Second Order

Construct

Figure 13: The Theoretical Framework of Training Transfer – Main Constructs and the
Corresponding Sub-Constructs

216
 Items of Sub-Construct

Sub- Construct

Main Construct

Figure 14: First Order Construct with Its Items, Second Order Construct with Its Sub-
Constructs and the Items for the Respective Sub-Construct

217
THE SUB-CONSTRUCT BECOMES A MEASURED VARIABLE IN THE
MODEL

Main Construct

Main Construct

Mediator

Figure 15: The Sub-Constructs have been converted into Measured Variables after CFA

218
EXECUTING THE MODEL: THE CAUSAL EFFECTS

Figure 16: The Results after the Model is Executed – The Standardized Regression Weights

Table 4: The Standardized Regression Weights

Construct Construct Estimate

Motivation_to Transfer ← Individual_Factors 0.423
Motivation_to Transfer ← Training_Design 0.237
Motivation_to Transfer ← Transfer_Climate 0.196
Training_Transfer ← Motivation_to Transfer 0.708

219
 Figure 17: The Results after the Model is Executed – The Regression Weights

Table 5: The Regression Weights and Its Significance Value

P-
Construct Construct Estimate S.E. C.R. Result
value
Motivation_to
← Individual_Factors 0.597 0.226 2.643 0.008 Significant
Transfer
Motivation_to Not
← Training_Design 0.252 0.171 1.473 0.141
Transfer Significant
Motivation_to Not
← Transfer_Climate 0.158 0.120 1.317 0.188
Transfer Significant
Training_ Motivation_to
← 0.662 0.084 7.868 0.001 Significant
Transfer Transfer

220
9.8 MODELING MOTIVATION TO TRANSFER AS A
MEDIATOR IN THE TRANSFER TRAINING MODEL

EXECUTING THE MODEL TO ANALYZE MOTIVATION TO TRANSFER AS

A MEDIATOR

Figure 18: The Results after the Model is Executed – The Standardized Regression Weights

Table 6: The Standardized Regression Weights

Construct Construct Estimate

Motivation_to Transfer ← Transfer_Climate .212

Motivation_to Transfer ← Training_Design .217

Motivation_to Transfer ← Individual_Factors .404

Training_Transfer ← Transfer_Climate .077

221
 Construct Construct Estimate

Training_Transfer ← Individual_Factors .283

Training_Transfer ← Training_Design .242

Training_Transfer ← Motivation_to Transfer .265

Figure 19: The Results after the model is executed – The Regression Weights

222
 Table 7: The Regression Weights and Its Significance Value

Construct Construct Estimate S.E. C.R. P Results

Motivation_to ← Not
Transfer_Climate 0.175 0.124 1.410 0.158
Transfer Significant
Motivation_to ← Not
Training_Design 0.238 0.174 1.366 0.172
Transfer Significant
Motivation_to ←
Individual_Factors 0.592 0.227 2.601 0.009 Significant
Transfer
Training_ ← Not
Transfer_Climate 0.058 0.117 0.494 0.622
Transfer Significant
Training_ ← Not
Individual_Factors 0.374 0.228 1.639 0.101
Transfer Significant
Training_ ← Not
Training_Design 0.240 0.164 1.460 0.144
Transfer Significant
Training_ Motivation_to
← 0.240 0.119 2.019 0.043 Significant
Transfer Transfer

223
 Bibliography

A.J. Fairchild, D.P. MacKinnon, M.P. Taborga, A.B. Taylor (2009). R2 effect size measures for
mediation analysis. Behavioral Research Methods, 41 (2009), pp. 486–498

Anderson, J.C. & Gerbing, D.W. (1988). Structural Equation Modeling in Practice: A review of
recommended two steps approach. Psychological Bulletin, 103 (3), 411-423.

Arbuckle, J.L. (2005). AMOS 6.0 User’s Guide. Spring House PA: Amos Development Corporation.

Bagozzi, R. P. & Yi, Y. (1988). On the evaluation of structural equation models. The Academy of
Marketing Science. 16(1), 74-94.

Baron, R.M. & Kenny, D.A. (1986). The moderator-mediator variable distinction in social
psychological research: Conceptual, strategic and statistical considerations. Journal of
Personality and Social Psychology, 51: 1173-1182.

Bentler, P. & Chou, C. (1987). Practical Issues in Structural Equation Modeling. Sociological Methods
and Research. 69(16): 78-117.

Bentler, P.M. & Bonett, D.G. (1980). Significance tests and goodness of fit in the analysis of
covariance structures. Psychological Bulletin. 88: 588–606.

Bentler, P.M. (1990). Comparative fit indexes in structural models. Psychological Bulletin 107: 238–
246

Bollen, K.A. (1986). Sample size and Bentler and Bonett’s non-normed fit index. Psychometrika. 51:375–
377.

Bollen, K.A. (1989b). A new incremental fit index for general structural equation models. Sociological
Methods and Research 17: 303–316.

Byrne, B.M. (2001). Structural Equation Modeling with Amos: Basic Concepts, Applications, and
Programming. Mahwah, NJ: Erlbaum.

224
Byrne, B.M. (2010). Structural Equation Modeling With AMOS: Basic Concepts, Applications, and
Programming 2nd Edition. Routledge. New York.

Cronbach. L.J. (1951). Coefficient Alpha and the Internal Structural of Tests. Pychometrica. 16(3): 297-
334

D.P. MacKinnon, C.M. Lockwood, J.M. Hoffman, S.G. West, V. Sheets (2002). A comparison of
methods to test the significance of mediation and other intervening variable. Psychological
Methods, 7 (2002), pp. 83–104

Edwards, J.R. & Lambert L. S. (2007). Methods for integrating moderation and mediation: A general
analytical framework using moderated path analysis. Psychological Methods, 12: 1-22.

Frazier, P.A., Tix, A. P. & Barron, K. E. (2004). Testing moderator and mediator effects in counseling
psychology research. Journal of Counseling Psychology, 51: 115-134.

Hair, J.F., Black, W.C., Babin, B.J., Anderson, R.E. & Tatham, R.L. (2006). Multivariate data
analysis. New Jersey: Pearson Prentice Hall.

Hair, J.F., Black, W.C., Rabin, B.J. & Anderson, R.E. (2010). Multivariate Data Analysis (7th Ed).
Englewood Cliffs, NJ: Prentice Hall.

Hair, J.F.,Anderson, R.E., Tatham, R.L. & Black, W.C. (1995). Multivariate Data Analysis with
Readings (4th Ed). Englewood Cliffs, NJ: Prentice Hall.

Hayes, A.F. & Matthes, J. (2009). Computational procedures for probing interactions in OLS and
logistic regression: SPSS and SAS implementations. Behavior Research Methods, 41: 924-936.

Holmes-Smith, P., Coote, L. & Cunningham, E. (2006). Structural Equation Modeling: From the
Fundamental to Advanced Topics. Melbourne: Streams.

J. Cohen (1988). Statistical power analysis for the behavioral sciences. Lawrence Erlbaum Associates,
New York, NY (1988)

225
Jöreskog, K.G. & Sörbom, D. (1984). LISREL-VI user’s guide (3rd ed.). Mooresville, IN: Scientific
Software.

Kenny, D.A. & Judd, C.M. (1984). Estimating the nonlinear and interactive effects of latent variables.
Psychological Bulletin, 96: 201-210.

Klein, A.G. & Moosbrugger H. (2000). Maximum likelihood estimation of latent interaction effects
with the LMS method. Psychometrika, 65: 457-474.

Kline, R.B. (2005). Principles and Practice of Structural Equation Modeling (2nd ed.).New York: The
Guilford Press.

Kraemer, H.C., Wilson G.T., Fairburn, C.G. & Agras, W.S. (2002). Mediators and moderators of
treatment effects in randomized clinical trials. Archives of General Psychiatry, 59: 877-883.

Kraemer, H.C., Stice, E., Kazdin, A., Offord, D. & Kupfer, D. (2001). How do risk factors work
together? Moderators, mediators, independent, overlapping, and proxy risk factors. American
Journal of Psychiatry, 158: 848-856.

Mahadzirah, M. & Zainudin, A. (2009). Building corporate image and securing students’ loyalty in the
Malaysian higher learning industry. Journal of International Management Studies. Vol. 4 (1): 30-
40.

MacKinnon, D.P, Lockwood, C.M, Hoffman, J.M., West, S.G., and Sheets, V. (2002). A comparison
of methods to test the significance of mediation and other intervening variable. Psychological
Methods, 7 (2002), pp. 83–104.

MacKinnon, D.P., Fairchild, A.J., Yoon, M., & Ryu, E. (2007). Evaluation of the proportion mediated
effect size measure of mediation. Unpublished manuscript.

Muller, D., Judd, C. M. & Yzerbyt, V. Y. (2005). When moderation is mediated and mediation is
moderated. Journal of Personality and Social Psychology, 89: 852-863.

226
Norsalihah, A.R., Zainudin, A. and Nor Hazreeni, H. (2010). The effects of information and
communication technology on security compliance. Journal of Statistical Modeling and
Analytics. Vol 1 (1): 28-44.

Ping, R.A. (1996). Latent variable interaction and quadratic effect estimation: A two-step technique
using structural equation analysis. Psychological Bulletin, 119, 166-175.

Salamiah A. Jamal (2011). Perceived Value, Satisfaction and Visitor Psychographics in Community-
based Homestay Tourism: Behavioural Intentions Perspective. Unpublished Thesis: Faculty of
Hotel and Tourism Management, UiTM Malaysia

Shammout, A.B. (2007). Evaluating Extended Relationship MarketingModel for Arab Guests of Five
Star Hotels. Unpublished Thesis: Faculty of Business and Law, Victoria University Melbourne.

Tanaka, J.S. & Huba, G.J. (1985). A fit index for covariance structure models under arbitrary GLS
estimation. British Journal of Mathematical and Statistical Psychology 38: 197–201.

Zainudin, A. (2010). The importance of corporate image of a university in the marketing of

postgraduate programmes. Asian Journal of University Education. Vol. 6 (1): 13-28.

Zainudin, A. & Kamaruzaman, J. (2009). The effects of corporate reputation on the competitiveness of
Telecommunication service providers. International Journal of Business and Management. Vol. 4
(5): 173-178.

Zainudin, A., Junaidah, H.A. & Nazmi, M.Z. (2010). Modeling job satisfaction and work commitment
among lecturers. Journal of Statistical Modeling and Analytics. Vol 1 (22): 45-59

Zainudin, A. (2010). Research Methodology for Business and Social Science. Shah Alam: Universiti
Teknologi MARA Publication Centre (UPENA).

Zainudin, A. (2012). Research Methodology and Data Analysis 2nd Edition. Shah Alam: Universiti
Teknologi MARA Publication Centre (UiTM Press).

Zainudin, A. (2007). Modeling the Loyalty of Outgoing Undergraduates towards Postgraduate

Education at their University. Thesis: Faculty of Management and Economics, Universiti
Malaysia Terengganu.
227
Index

a corporate reputation, 14, 146

absolute value, 81, 94 correlated errors, 20, 76

AMOS graphic, 25, 27, 32, 120, 129 correlation, 17, 31

AMOS structural model, 84 correlational effects, 84

Analysis of Moments Structures, 5, 24 correlational relationship, 50, 99

assess the moderating effects, 40 covariance estimate, 100, 103

b covariance, 100, 103

baseline comparison indices, 184, 188 critical ratio, 93, 104

c critical region, 104, 107

causal effect, 6, 10, 41, 104, 105 Cronbach’s Alpha, 63, 79

causal relationship, 31, 32, 52 d

CFA procedure, 69, 78, 107 demographic characteristics, 40, 149

Chi-Square value, 155, 161 discriminant validity, 62, 79

Comparative Fit Index, 64, 65, 70 e

complete mediation, 37, 120, 133 educational research, 12

complete moderation, 150 endogenous constructs, 23, 25, 53

concept of mediation, 124 error in the equation, 21, 23

confirmatory factor analysis, 5, 18, 46 errors in measurement, 23

estimated beta, 104

exogenous construct, 25, 30, 53 measuring items, 47, 50, 52

228
f mediating constructs, 199

factor loading, 47, 61, 62 mediating effects, 120, 124

fitness index, 47, 62, 63 mediating role, 119, 130

h mediating the relationship, 37

hypothesis of causal effects, 44 mediating variable, 25, 118

hypotheses of interest, 17, 202 mediator construct, 53, 54

hypothesis statement, 171, 201 medical and healthcare research, 9

hypothesized relationships, 25, 90 model fit, 47, 63

hypothetical concept, 18 modeling the structural model, 201

l modeling the moderating effects, 149, 154

latent constructs, 3, 4, 18, 22 moderating effects, 149, 151

latent endogenous construct, 53, 69 moderating variable, 26, 40, 41

latent exogenous construct, 53, 69, 91 moderation effects, 8, 42

level of significance, 93, 101 moderator variable, 44, 144

m modification indices, 76, 184

market research, 14 modify the measurement model, 76

measurement errors, 20, 27 multicollinearity, 20

measurement model, 33, 46, 47 multiple dependents, 17, 20

multiple indicators, 22

multiple regression, 17, 22 regression relation, 34, 148

n regression weight, 90, 91, 92

229
no mediation, 37, 39 s

Normed Fit Index, 65 schematic diagram, 17, 84

o Second Order CFA, 176, 177

Ordinary Least Square , 1, 149 Second Order Factor Analysis, 122, 137

p significance of moderating variable, 176

partial mediation, 37, 38 significant direct effect, 129

parameter estimates, 172 simple effect, 120, 147

partial moderation, 150, 153 social sciences, 144

path analysis, 34, 37 squared multiple correlations, 92

Perceived Quality, 55, 96 standard deviation, 88, 91

Perceived Value, 199 Standard Error of regression weight, 92, 93

probability value, 107, 132 Standard Error, 104, 107

q standardized coefficients, 206

questionnaire items, 29, 66 standardized estimate, 66, 68

r standardized factor loading, 183, 188

ratio scale, 148, 149 standardized regression weight, 88, 91, 95

redundant items, 7, 62 Structural Equation Modeling, 5, 17, 46, 84

Relative Fit Index, 184, 188 structural model, 20, 46

response items, 20, 25

RMSEA, 64, 65

230
t

text output, 85, 90

theoretical framework, 2, 5

theorized construct, 176

Tucker-Lewis Index, 65

underlying components, 194

unobserved variables, 4, 27

unstandardized estimate, 85, 88

Unstandardized Regression Weight, 188

validity and reliability, 61, 78

variable, 2, 8

231