STRUCTURAL EQUATION MODEL (SEM)

American Journal of Humanities and Social Sciences Research (AJHSSR) 2021
A J H S S R J o u r n a l P a g e | 11
American Journal of Humanities and Social Sciences Research (AJHSSR)
e-ISSN: 2378-703X
Volume-5, Issue-7, pp-11-19
www.ajhssr.com
Research Paper Open Access
STRUCTURAL EQUATION MODEL (SEM)
1
AJAYI, Lawrence Boboye Phd , 2
ADEBAYO, Adeyinka Taoheed
1
Department of Finance Ekiti State University, Ado-Ekiti, Ekiti State.
2
Department of FinanceEkiti State University, Ado -Ejiti, Ekiti State.
ABSTRACT : This paper critically examined a broad view of Structural Equation Model (SEM) with a view
of pointing out direction on how researchers can employ this model to future researches, with specific focus on
several traditional multivariate procedures like factor analysis, discriminant analysis, path analysis. This study
employed a descriptive survey and historical research design. Data was computed viaDescriptive Statistics,
Correlation Coefficient, Reliability. The study concluded that Novice researchers must take care of assumptions
and concepts of Structure Equation Modeling, while building a model to check the proposed hypothesis. SEM is
more or less an evolving technique in the research, which is expanding to new fields. Moreover, it is providing
new insights to researchers for conducting longitudinal investigations.
.
KEYWORDS: factor analysis, discriminant analysis, path analysis, descriptive statistics, and Structural
Equation Model (SEM).
I. INTRODUCTION
Structural Equation Modeling, or SEM, is a very general statistical modeling technique, which is widely used in
the behavioral sciences. The most important reason of the spread of this statistical technique is that the direct
and indirect relationships among causal variables can be measured with a single model (Meydan&Şen, 2011).
Another reason for the widespread adoption of this method is the ability of taking in to the account of the
measurement errors and the relationships between errors in the observed variables (Civelek, 2018). This goes a
long way in minimizing measurement errors.
Furthermore, it provides a very general and convenient framework for statistical analysis that includes several
traditional multivariate procedures like factor analysis, regression analysis, discriminant analysis, path analysis,
to mention but a few.The key in the regression analysis is to determine how much of the change in the
dependent variable is explained by the independent variable or variables. Differently from the regression,
structural equation modeling, allows to test research hypotheses in a single process by modeling complex
relationships among many observed and latent variables. In traditional regression analysis, only direct effects
can be detected, but in the method of structural equation modeling, direct and indirect effects are put together
(Kline, 2011).
Structural equation modeling is used to test the relationships between observed and latent variables. Observed
variables are the measured variables in the data collection process and latent variables are the variables
measured by connecting to the observed variables because they cannot be directly measured. According to
Civelek (2018), most of the statistical methods other than structural equation modeling try to discover
relationships through the data set.However, structural equation modeling confirms the correspondence of the
data of the relations in the theoretical model. For this reason, it can be said that structural equation modeling is
more suitable for testing the hypothesis than other methods (Karagöz, 2016).
II. ASSUMPTIONS OF STRUCTURAL EQUATION MODELING
Similar to regression analysis, structural equation modeling has its assumptions. But in structural equality
models, many regression equations work together, whether in the structural model part or in the measurement
model part. Therefore, the assumptions that apply to the regression models are valid for the structural equation
models. These assumptions can be summarized as follows (Bayram, 2013);
 Observed variables have multivariate normality
The multivariate normal distribution is the most important assumption of the maximum likelihood estimation
method used in structural equation modeling. This rule is often violated when ordinal and discrete scales are
used. The skewness and kurtosis values are examined to determine whether the variables in the data set are
normally distributed. These values are calculated on the basis of moments. In general, the packaged

softwarecalculates these values to be 0 as base value. In this case values between -2 and +2 are considered
normal. In addition, Kolmogorov-Smirnov and Shapiro-Wilk tests can be conducted to test whether the data set
is normally distributed (Sarstedt&Mooi, 2014).
 Latent variables have multivariate normal distribution
It refers to the endogenous latent variables have normal distribution. In practice, it is a violated assumption.
 Linearity
Linearity, which is the most important assumption of regression analysis, also applies to structural equation
modeling. In the structural equation model, it is assumed that there are linear relationships between latent
variables and also between observed and latent variables.
 Absence of outliers
The outlier affects the significance of the existence model negatively.
 Multiple measurements
In the structural equation model, three or more observed variables must be used to measure each latent variable.
 Absence ofmulti-co-linearity
It is assumed that there is no relation between the independent variables in the structural equation model.
 Sample size
In the structural equation modeling, many of the fit indices are influenced by sample size. In some sources, a
minimum sample size of 150 is recommended for structural equation models (Bentler& Chou, 1987). The
minimum sample size that should be used in the structural equation modeling method is at least 10 times the
number of parameters that can be estimated in the model. (Jayaram, Kannan,&Tan, 2004). According to some
researchers, the sample size required for structural equation modeling should be at least 200 and 200 – 500
(Çelik&Yılmaz, 2013).
 No correlation between error terms
It is assumed that there is no correlation between error terms in the structural equation modeling method.
However, if it is explicitly stated by the researcher in the conceptual model, a correlation can be made between
the error terms (Doğan, 2015).
III. VALIDITY AND RELIABILITY ANALYSIS
Reliability means that a scale always measures the same value under the same conditions consistently.Scale is
the method used to find the numerical values of the dimensions that constitute a concept. Since concepts cannot
be directly measured in social sciences, questionnaires are formed to define these concepts.For example, a
questionnaire form is reliable if the same group is given the same result when applied two different times. So if
we ask the same questions about the same people, if the conditions are not changed, they are expected to give
the same answers. Otherwise, this means that the persons in the sample either they did not understand the
questions on the questionnaire or they did not read them.
Validity is a measure of what we really want to measure. For example, if a questionnaire actually measures a
different concept than the dimension we want to measure, it is not valid. If the questions we ask about the
concept A are confused with the questions about the concept B, then it means that the concepts we consider to
measure are not perceived or perceived as different from those in the sample. In this case, the scale we use is not
a valid measurement tool for this sample. For this reason, it is necessary to test the validity and reliability of the
scale before any analysis is started. As a result of these tests, verification of unidimensionality is generally
provided. Unidimensionality means that the observed variables used to measure each dimension must measure
only one dimension (Avcılar&Varinli, 2013).
Construct validity and reliability must be determined in order to confirm unidimensionality.The construct
validity indicates that the observed variables do not measure any latent variable other than they connected in the
conceptual model. But in this case it would not be correct to say that the validity of the construct is fully realized
without confirming the reliability of the scale (Gerbing& Anderson, 1988).
1. Determination of Convergent Validity
Convergent validity indicates that the correlations between questions constituting a construct are high. In
structural equation modeling method, it is necessary to look at the results of confirmatory factor analysis to
determine the convergent validity of the scales used to measure the dimensions constituting the conceptual
model of the research. The measurement model part of structural equation models correspond to confirmatory
factor analysis (CFA). Therefore, if the measurement model fit indices are low, there is no need to test the
structural model.
The t test results of all the coefficients in the measurement model should indicate that the coefficient values are
different from zero. The standard value of each coefficient in the measurement model is the factor loadings of
the confirmatory factor analysis. Each factor load should be higher than 0.50.Otherwise, the fit indices of the

general model will beadversely affected. The fact that the factor loads are above 0.5 is evidence of convergent
validity. If the critical rate value of a question in CFA results is greater than 2 as an absolute value this means
that this item is loaded to the factor it is connected.
Before applying confirmatory factor analysis (CFA), it is first necessary to look at the results of explanatory
factor analysis (EFA) in practice. Even though scales generally accepted in the literature are used, to see if the
survey fillers correctly perceive the questions principle component analysis should be conducted in SPSS before
set up CFA model in AMOS. And how many different dimensions the questions are perceived by those who
solve the questionnaire should be clarified. At this stage, the necessary questions should be eliminated. This step
is also called the purification stage. Principle component analysis is a type of analysis that assigns the variables
in the data set into groups so that the relationship between the variables in the group is maximized. Main
purpose of this analysis is to obtain the least number of factors to represent the relationship among items at the
highest level.
2. Determination of Discriminant Validity
Discriminant validity is the measure of the level at which a structure in a measurement model differs from other
structures. It is an indicator of a low correlation between the questions that form a construct and other questions
that form other construct. To find the discriminant validity for each dimension, we first need to calculate the
Average Variance Extracted (AVE) value for each dimension. The acceptable AVE value must be greater than
0.50 or 0.50. However, this value confirms convergent validity when examined alone (Fornell&Larcker, 1981).
In order to determine discriminant validity, it is also desirable that the values of the AVE for each construct in
the data set are larger than the correlation coefficients of that construct with the other constructs. In this case, it
can be determined that the scales used have discriminant validity for each dimension. AVE value alone does not
indicate discriminant validity but the square root of the AVE value of each construct is larger than the inter-
dimensional correlation value it can be said that there is discriminate validity (Fornell&Larcker, 1981). The
AVE value is not calculated by the AMOS package program. However, it is easy to find ready-made excel files
that provide this value calculation on the internet.
3. Determination of Reliability
After determination of the validity of the scales by means of CFA reliability analysis must be conducted for
each construct. First of all, Cronbach's α value is calculated for each dimension separately. Values greater than
0.7 threshold indicate that the internal reliability of the scale used is sufficient. Cronbach’s α is a measure based
on correlations between items in a construct. It is obtained by dividing the sum of the variances of the items
constituting a scale by the general variance. It takes a value between 0 and 1. Values beyond 0.7 threshold
indicate that the scale is reliable. If it is below 0.6, the reliability of the scale is low (Karagöz, 2016).
Another value that is used to calculate the reliability of the scale for each dimension is the composite reliability
value. The composite reliability value is calculated from the factor loads found in the confirmatory factor
analysis. After CR values beyond 0.7 threshold or equals to 0.7 it can be said that there is composite reliability
(Raykov, 1997).
The Table below shows a sample table showing Cronbach's α, AVE and CR values calculated for each construct
and the correlation values between constructs. Cronbach's α value can be calculated from the scale reliability
menu in the SPSS program. The AVE and CR values are found by placing the results of the CFA factor loadings
in to the formulas. There are readymade calculation tools on the Internet.
Descriptive Statistics, Correlation Coefficient, Reliability Results and Discriminant Validity
Avr.
Std.
Dev
1 2 3 4 5 6 7 8 9
1.Constru
ct 3,25 0,81 (0,842)
2.Constru
ct 3,28 0,71 ,216* (0,711)
3.Constru
ct 3,63 0,72 ,427* ,383* (0,840)
4.Constru
ct 3,72 0,68 ,228* ,533* ,457* (0,718)
5.Constru
ct 3,62 0,70 ,449* ,192* ,378* ,298* (0,769)

6.Constru
ct 3,76 0,68 ,430* ,394* ,551* ,450* ,499* (0,734)
7.Constru
ct 3,23 0,87 ,585* ,166* ,452* ,174* ,479* (0,800) ,449*
8.Constru
ct 3,68 0,67 ,394* ,496* ,672* ,508* ,350* ,508* ,358* (0,722)
9.Constru
ct 3,02 0,77 ,340* ,374* ,353* ,335* ,209* ,302* ,219* ,410* (0,754)
Cronbach Alpha
Reliability Coefficient
0,92
7
0,86
1
0,90
1
0,85
1
0,78
1
0,77
1
0,82
8
0,80
8
0,72
1
Composite
ReliabilityCoefficient
(CR)
0,92
4
0,85
4
0,90
5
0,84
1
0,79
1
0,77
7
0,84
0
0,81
3
0,72
5
Avrerage Variance
Extracted (AVE) 0,710 0,506 0,706 0,516 0,592 0,539 0,640 0,522 0,570
* P<0,05, Note: the values written in brackets indicate the square root of the AVE values.
There are statistically significant relationships among the constructs in the sample in Table above.
Correlation is the coefficient that indicates the power of linear relationship between variables. This
coefficient must be statistically significantnordertobeabletosaythatthereisarelationship between variables.
The correlation coefficient takes a value between-1and+1(Sipahi,Yurtkoru,&Çinko,2010).
IV. TYPES OF STRUCTURAL EQUATION MODELS
There are four basic types of structural equation models. These are explained below:
 Path Analysis Models
In the method of structural equation modeling, the models established with only observed variables are called
path analysis models. The basis of the structural equation modeling depends upon path analysis. This model was
developed by biologist Sewall Wright (Taşkın&Akat, 2010) and wasfirst implemented in the 1920s.The path
analysis is similar to multiple regression as it is done with observed variables. However, it is superior to
multiple regression, because there is one dependent variable in the multiple regression. Although, there may be
more than one dependent variable in the path analysis, and a variable can be both a dependent variable and an
independent variable, more than one regression model can be analyzed at the same time, and indirect and direct
effects can be measured at the same time. Direct effect is the effect of one variable on another variable without
any mediation. However, the indirect effect arises from the intervention of a variable which is playing mediator
role between independent and dependent variables. This variable is named as the mediator variable. The sum of
the direct effect and the indirect effect of a variable on another variable is called the total effect
(Raykov&Marcoulides, 2006). Path analysis do not contain latent variables, they cannot be saved from
measurement errors (Meydan&Şen, 2011). For this reason, structural regression models generated by latent
variables give more accurate results.
Examples of Path Analysis

 Confirmatory Factor Analysis
This is divided into exploratory and confirmatory. In explanatory factor analysis, factors are revealed from
relations among variables. In explanatory factor analysis, the observed variables can be loaded on any factor or
on multiple factors. However, in the confirmatory factor analysis, the theoretically predetermined factor
structure is confirmed by the current data. In other words, in the confirmatory factor analysis, which factor will
be loaded on an observed variable is predetermined. By means of the explanatory factor analysis, the latent
variables are revealed from the observed variables. However, in the confirmatory factor analysis, previously
discovered scales are confirmed again with the collected data.
Single Factor CFA Model
Results of Confirmatory Factor Analysis
The Table below shows the way in which confirmatory factor analysis results are given. What is important here
is that the standard factor loads of the questions under each conceptual variable are over 0.50. By looking at this
table, questions with a standard factor load of less than 0.50 are discarded.
Items
Conceptual
Variable
Standardized
Factor Loads
Unstandardiz
ed Factor
Loads
Standard Error
t-Value
(Critical
Ratio)
Qestion1
X
0,818 1
Qestion2 0,906 1,104 0,049 22,523
Qestion3 0,907 1,111 0,049 22,570
Qestion4 0,825 1

Qestion5 Y 0,732 0,882 0,057 15,549
Qestion6 0,718 0,885 0,058 15,187
Qestion7
Z
0,757 1
Qestion8 0,835 1,102 0,062 17,785
Qestion9 0,939 1,255 0,062 20,176
Qestion10
W
0,676 1
Qestion11 0,799 1,131 0,083 13,555
Qestion12 0,785 1,158 0,087 13,379
Note: For all values P<0.01
 Structural Regression Models
This is formed between latent variables in structural equation models. It consists of a combination of
measurement model and structural model. Incorporating the measurement model and the structural model allows
the inclusion of measurement errors so that more accurate results can be obtained. In other words, confirmatory
factor analysis and multiple regression analysis coexist.
StructuralRegressionModelExample
These following are examples of structural regression models. Although the models in the first and second
figure are basedon the same measurement model, the path analysis created is different. In the first figure, there
are more than one exogenous variable and therefore covariance is placed between them. In the second figure
there is only one exogenous variable. In both modelsresidualtermsarelinkedtotheendogenousvariables.Careful
attention should be paid to these rules when constructing structural models. Otherwise the model will not work.
Hypothesis Test Results Table Example
The Table below shows an example of the hypothesis test results. The values in this table are in the estimates
section of the output screen of the AMOS program. The notation *** in AMOS output means that P is equal to
zero.
Relations Standard Coefficients Unstandardized

Coefficients
X → Z 0.533* 0.594*
Y → Z 0.437* 0.638*
Z → W 0.493* 0.377*
*p < 0.05
 Latent Change Models
They are also named as “latent growth curve models” or “latent curve analysis”. They are models that describe
longitudinal variation in time series (Raykov&Marcoulides, 2006). They are also used to explain the growth and
decay of an event over time, similarities or differences within and between units (Doğan, 2015). Structural
equation modeling is a very useful method for analyzing changes in time.In the figure below, two factorial
growth models are observed fortwotimepoints(T1,T2). Repeated measurements over time are needed to use the
latent change models. Such data are called longitudinal data (vertical crosssection data).These models are the
models used to explain the growth and decay of an event over time, similarities or differences within and
between units. (Doğan, 2015).
According to Baltes and Nesselroade (1979), thismodel can be used for the following purposes:
(1) Describeobservedandunobservedverticalsectiondata.
(2) Characterizethedevelopmentofindividualsandgroups.
(3) To predict individual and group differences in developmentalforms.
(4) To examine the dynamic determinants among variablesin time.
(5) To reveal the group differences of the dynamic determinants between variables intime.
Latent Change Model Example
Source:Raykov,T.,&MarcOULIDES,G.(2006).AFIRStCourSEinStructuralEquation
Modeling.Mahwah:LawrenceErlbaumASSOCIATES.

V. METHODS TO BE APPLIED IN CASE OF DATA INADEQUACY
Sometimes, there may be cases where the assumptions of the estimation methods used are not met by the
existing data set. In this case, there are methods that can be applied if it is necessary to be satisfied with the
dataset available. Leading methods among them are bootstrap partial least square.
1. Bootstrap Method
The bootstrap technique is applied when one of the assumptions of normal distribution or being continuous
variable is not met. This method was developed by B. Efron in 1979 (Efron, 1979). In many studies in the
literature the condition of normal distribution obligation is neglected. It is also seen in many studies in the
literature that X2 value is derived by maximum likelihood and generalized least squares methods. Estimation
methods which are frequently used in the structural equation model are these two methods. In particular, with
the non-normal distribution, the number of observations is also low cause X2 value to increase. At the same
time, irreversible and inadequate modifications made during the analysis of such data are not scientifically
acceptable and result in inconsistent estimations about the population. In the bootstrap method, a different data
set is obtained from the existing observations (Sacchi, 1998). This method is basically the derivation of the
sample from the sample.
There are advantages and limitations of the bootstrap process. The main advantage of the bootstrap technique is
the ability to evaluate the accuracy of the predicted parameters. The idea underlying the bootstrap technique is
to create sub-samples of the current data and look at the distribution of the parameters computed from each sub-
sample.
2. Partial Least Square Structural Equation Modeling ( PLS-SEM)
It is also called covariance-based structural equation modeling since the structural equation model that has been
examined in the previous sections is based on the covariance matrix. However partial least square structural
equation modeling is based on variance. For this reason, it is also called as the variance-based structural
equation modeling. Partial least square structural equation modeling (PLS-SEM) is an advantageous method
when the assumptions of least squares are not met. It is an alternative of covariance-based structural equation
modeling (CB-SEM). It is a second generation multivariate analysis method that enables measurement model
and structural model to be analyzed together like covariancebased structural equation modeling.
According to Civelek (2018), covariance-based structural equation modeling is a more powerful and reliable
method. For this reason, the partial least square structural equation modeling method is generally preferred in
cases where the conditions listed below are found:
1. If the sample is small.
2. If the data do not distribute normally.
3. If the number of indicators connected to the latent variable is less than three.
4. If there is a multicollinearity.
5. There is missing value.
6. If the number of observations is less than the number of explanatory variables.
If the above listed conditions are found, method PLS-SEM method is far superior to method CB-SEM, because,
in these cases, it reduces the unexplained variance to the lowest level. As the model is complex, such as in the
CB-SEM method, no larger sampling is required in PLS-SEM. However, some researchers who have done
research on the sampling sensitivity of the PLS-SEM method have raised the ten-fold rule. According to this
rule, there is a necessity to have 10 times observation of the number of indicators used to measure a construct in
the measurement model and 10 times observation the number of the path in a structural model (Barclay,
Higgins, & Thompson, 1995).
However, the PLS-SEM method is a non-parametric method because it does not have any distributional
assumption (Hair, Hult, Ringle, &Sarstedt, 2017). It is also an explanatory approach, which is why it is preferred
in exploratory research. In other words, when the theory is underdeveloped, it can be said that researchers prefer
to use partial least squares structural equation modeling. This judgment is partially correct in cases where the
structure need to be predicted and relations need to be explained (Rigdon, 2012).
When the theory needs to be tested and verified, in case of there is cycles in the structural model and if the
model needs to be verified in general with fit indices it is more accurate to use CB-SEM method, because the
PLS-SEM method cannot explain loop-related relations. In addition, it does not give general fit indices of the
model. Partial least squares method can be easily implemented by means of a packet program called SmartPLS.
SmartPLS is a packet program that allows the creation of partial least squares based structural equation models.
Structural equation modeling programs outside of SmartPLS makes the maximum likelihood estimation method
the default choice.

VI. CONCLUSION
Novice researchers must take care of assumptions and concepts of Structure Equation Modeling, while building
a model to check the proposed hypothesis. SEM is more or less an evolving technique in the research, which is
expanding to new fields. Moreover, it is providing new insights to researchers for conducting longitudinal
investigations.
REFERENCES
[1]. Avcılar, M., &Varinli, İ. (2013).PerakendeMarkaDeğerininÖlçümüveYapısalEşitlikModeliUygulaması
(1 b.). Ankara: DetayYayıncılık.
[2]. Barclay, D., Higgins, C., & Thompson, R. (1995). The partial least sequares approach to causal
modeling: Personal computer adoption and use as illustration. Technology Studies, 2, 285-309.
[3]. Bayram, N. (2013). YapısalEşitlikModellemesineGiriş. Bursa: EzgiKitapevi
[4]. Bentler, P. M., & Chou, C.-P.(1987). Practical Issues in Structural Modeling.Sociological Methods
Research, 16(1), 78-117.
[5]. Çelik, H. E.&Yılmaz, V. (2013).Lısrel 9.1 ileYapısalEşitlikModellemesi. Ankara: AnıYayıncılık.
[6]. Civelek, M. (2018). The Mediator Roles Of Attitude Toward The Web Site And User Satisfaction On
The Effect Of System Quality On Net Benefit: A Structural Equation Model On Web Site Success.
European Scientific Journal, 61-73.
[7]. Doğan, İ.
(2015).FarkliVeriYapisiveÖrneklemBüyüklüklerindeYapisalEşitlikModellerininGeçerliğiveGüvenirliği
ninDeğerlendirilmesi.Eskişehir: EskİşehİrOsmangazİÜnİversİtesİDoktoraTezi.
[8]. Efron, B. (1979). Bootstrap method: Another look at the jackknife. Annals of Statistics, 7, 1-26.
[9]. Fornell, C., &Larcker, D. (1981).Evaluating Structural Equation Models with Unobservable Variables
and Measurement Error.Journal of Marketing Research, 18(1), 39-50.
[10]. Gerbing, D., & Anderson, J. (1988).An Updated Paradigm for Scale Development Incorporating
Unidimensionality and Its.Journal of Marketing Research, 25(2), 186-192.
[11]. Hair, J., Hult, G., Ringle, C., &Sarstedt, M. (2017).A primer on partial least squares structural
equation modelingPLS-SEM. Los Angeles: SAGE.
[12]. Jayaram, J., Kannan, V., & Tan, K. (2004). Influence of initiators on supply chain value creation.
International Journal of Production Research, 42(20), 4377-4399.
[13]. Karagöz, Y. (2016). SPSS ve AMOS 23 UygulamalıİstatistikselAnalizler. Ankara: Nobel.
[14]. Kline, R. (2011). Principles and practice of structural equation modeling (3nd Ed. b.). New York:
Guilford.
[15]. Meydan, C. H. &Şen, H. (2011).YapısalEşitlikModellemesi AMOS Uygulamaları. Ankara:
DetayYayıncılık.
[16]. Raykov, T. (1997).Estimation of composite reliability for congeneric measures.Applied Psychological
Measurement, 21(2), 173-184.
[17]. Raykov, T., &Marcoulides, G. (2006).A First Course in Structural Equation Modeling. Mahwah:
Lawrence Erlbaum Associates.
[18]. Rigdon, E. (2012). Rethinking partial least squares path modelling: In praise of simple methods. Long
Range Planning, 45, 341-358.
[19]. Sacchi, M. (1998).A bootstrap procedure for high-resolution velocity analysis.Geophysics, 63(5).
[20]. Sarstedt, M., &Mooi, E. (2014).A concise guide to market research: The process, data, and methods
using IBM SPSS statistics (2nd b.). Berlin: Springer.
[21]. Taşkın, Ç.&Akat, Ö. (2010).AraştırmaYöntemlerindeYapısalEşitlikModelleme. Bursa:
EkinBasımYayın

STRUCTURAL EQUATION MODEL (SEM)

More Related Content

STRUCTURAL EQUATION MODEL (SEM)