Applied Structural Equation Modelling for Researchers

and Practitioners
Definition of SEM
Article information:
To cite this document: "Definition of SEM" In Applied Structural Equation Modelling
for Researchers and Practitioners. Published online: 16 Jan 2017; 1-12.
Permanent link to this document:
https://doi.org/10.1108/978-1-78635-883-720161001
Downloaded by Goethe-Universität Frankfurt At 09:23 30 March 2018 (PT)

Downloaded on: 30 March 2018, At: 09:23 (PT)

References: this document contains references to 0 other documents.
To copy this document: [email protected]
The fulltext of this document has been downloaded 13 times since 2017*
Access to this document was granted through an Emerald subscription provided by
Goethe-Universität Frankfurt
For Authors
If you would like to write for this, or any other Emerald publication, then please
use our Emerald for Authors service information about how to choose which
publication to write for and submission guidelines are available for all. Please visit
www.emeraldinsight.com/authors for more information.
About Emerald www.emeraldinsight.com
Emerald is a global publisher linking research and practice to the benefit of society.
The company manages a portfolio of more than 290 journals and over 2,350 books
and book series volumes, as well as providing an extensive range of online products
and additional customer resources and services.
Emerald is both COUNTER 4 and TRANSFER compliant. The organization is a partner
of the Committee on Publication Ethics (COPE) and also works with Portico and the
LOCKSS initiative for digital archive preservation.

*Related content and download information correct at time of download.

Downloaded by Goethe-Universität Frankfurt At 09:23 30 March 2018 (PT)

Definition of SEM

1.1. Introduction

▾
Known as causal models with a conspicuous presence in the
1
field of consumer psychology, structural equation model (SEM)
allows complex modelling of correlated multivariate data in
order to sieve out their interrelationships among observed and
latent variables. SEM constitutes a flexible and comprehensive
methodology for representing, estimating and testing a theoreti-
cal model with the objective of explaining as much of their var-
iance as possible. In simple terms, SEM is nothing more than an
analysis of the covariance structure. SEM incorporates various
statistical models such as regression analysis, factor analysis and
variance/covariance analysis. Under SEM, a clear demarcation
line is established between observed and latent variables. SEM
can handle complex relationships as it can simultaneously factor
in a measurement equation and a structural equation. Moreover,
SEM represents a large sample technique, widely known under
its rule of thumb, that is to have at least 10 observations per
variable. SEM represents a vital multivariate data analysis techni-
que, widely employed to answer distinct types of research ques-
tions in statistical analysis. Other names are associated with SEM
such as simultaneous equation modelling, path analysis, latent
variable analysis and confirmatory factor analysis. Technically
speaking, SEM can be defined as a combination of two types of
statistical technique, namely, factor analysis and simultaneous

1
 equation models. SEM is so much coveted by researchers that
there is even a journal in this area, namely, Journal of Structural
Equation Modelling. In a nutshell, SEM can best be described as a
powerful multivariate tool to study interrelationships among
both observed and latent variables.

SEM = 2 types of statistical techniques:

Confirmatory Factor Analysis + Simultaneous Equation Models
Downloaded by Goethe-Universität Frankfurt At 09:23 30 March 2018 (PT)

While the objective of the measurement model is to relate the

latent variables to the observed variables, the aim under structural
equation focuses on the relationship between dependent and inde-
pendent latent variables, the effects of explanatory (independent)
latent variables on dependent (outcome) latent variables.
Alternatively stated, the need to sieve observed variables to cap-
ture latent variables is effected in the measurement equation. In
that respect, the measurement equation in SEM constitutes an
exploratory tool, more specifically, a confirmatory factor analysis
tool. Under SEM, there is the need to compare several structural
equations via model comparison statistics to sieve out the most
appropriate model. Latent variables are also known as unobserved
variables, intangible variables, ‘directly unmeasured’ variables, unknown
variables or simply constructs, whereas manifest variables are called as
observed variables, tangible variables, indicator variables or known
variables.
Latent variables are inherently linked to observed variables as
they can only be captured by observed variables or indicators.
Latent variables are inferred from observed variables. This can be
best explained when it comes to LISREL software application
whereby observed variables are input in such a way as to respect
the order of the data input, while for latent variables, they can be
defined in any order. Moreover, ellipses or circles are associated
with latent variables while rectangles are inherently associated
with the observed variables. Latent variables can be dependent or

2 Applied Structural Equation Modelling for Researchers and Practitioners

 independent variables. As a matter of fact, SEM comprises
observed and latent variables, whether dependent or independent.
Overall, SEM includes observed variables, latent variables and
measurement error terms. Pure latent variables are those which
are uncontaminated by measurement error. Second-order latent vari-
ables are functions of other latent variables while first-order latent vari-
ables do not depend on any other latent variables.
Examples of latent variables are intelligence, market psychol-
ogy, achievement level and economic confidence. Examples of
observed variables are economic performance, scores obtained
and number of items sold.
Downloaded by Goethe-Universität Frankfurt At 09:23 30 March 2018 (PT)

Various versions of SEM prevail with the most basic version

being the linear SEM. Other types of SEM consist of Bayesian
SEM, non-linear SEM and hierarchical SEM. The main ingredient
used in SEM application pertains to the covariance or the correla-
tion matrices. It is of paramount significance to gain a proper
insight into their inherent difference. In essence, covariance consti-
tutes an unstandardised1 form of correlation. SEM constitutes an
exploratory factor analysis tool for gauging on the factors that
underlie a set of variables. It assesses which items should be
grouped together to form a scale. While exploratory factor analysis
allows all the loadings to freely vary, confirmatory factor analysis
constrains certain loadings to be zero. SEM is widely preferred to
regression analysis by virtue of its powerful distinctive features
such as the ability to incorporate multiple independent and depen-
dent variables, inclusion of latent constructs and measurement
errors being duly recognised. SEM is widely applied in non-
experimental data (Figure 1.1).

1. Under STATA SEM estimation, unstandardized estimates pertain to covariance and stan-
dardized estimates for correlation coefficient.

Definition of SE M 3
 Two main variable types
in SEM

Observed (indicator) Latent (construct)

variables variables

“directly unmeasured” or
Directly observable or measured using several
Downloaded by Goethe-Universität Frankfurt At 09:23 30 March 2018 (PT)

measured observed (indicator)

variables

Figure 1.1: Variables Type under SEM.

Intelligence→ Achievement1→ Achievement2

A mediating latent variable

1.2. Regression and SEM

Compared to regression analysis, SEM is imbued with various dis-
tinctive features that render it particularly appealing. SEM incor-
porates multiple independent and dependent variables while a
multivariate regression analysis can merely have one dependent
variable. SEM factors in hypothetical latent constructs so that the
proxies employed in the studies are equal to the latent variables
with measurement errors. Proxies measure the impacts of many
different attributes. While regression analysis is directed towards
generating a causal relation, SEM pertains to modelling a causal
relation. SEM is flexible enough to account for measurement
errors, whereas regression assumes perfect measurement. SEM is
more powerful than regression analysis as it can simultaneously
handle indirect, multiple and reverse relationships. Under SEM,
researchers are interested in a large p-value and not a small
p-value as it is the case for regression analysis. Causal assumptions

4 Applied Structural Equation Modelling for Researchers and Practitioners

 are made implicit in SEM and not explicit as in the case of regres-
sions. SEM tools are not only more rigorous but also more flexible
than conventional regression analysis.

1.3. Data Pre-Processing in SEM

Under SEM, the models are conceptually derived first. To test the
specified theory/model, data is required, in particular, the correla-
tion/covariance matrices. In fact, the heart of SEM is related to the
covariance/correlation matrix. In that respect, there is the need to
use data to generate these matrices. Consequently, prior to deriv-
Downloaded by Goethe-Universität Frankfurt At 09:23 30 March 2018 (PT)

ing these covariance/correlation matrices, it is of paramount sig-

nificance to have recourse towards data screening to remove
outliers, ensure that normality conditions are being fulfilled and
that there is no missing data. Data imputation techniques can be
used to deal with missing data.

1.4. Confirmatory Factor Analysis versus

Exploratory Factor Analysis
Confirmatory factor analysis, developed by Karl Je0 0308; oreskog
in the 1960s, is extremely important in SEM applications as it is
used to test the measurement model. Under the measurement
model, the objective is geared towards testing the relationship
between a latent variable and its corresponding observed vari-
ables. Only after the measurement model has been derived that the
focus is then shifted towards the structural model estimation, also
known as the path analysis. Confirmatory factor analysis is not
same as exploratory factor analysis. The underlying rationale is
based on the fact that, in the case of exploratory factor analysis,
there is no stipulated a priori hypothesis being laid out when it
comes to establishing relationships among the variables so that
principle components are applied. Explanatory factor analysis is
deficient as it tends to be guided by ad hoc rules and intuition,
whereas confirmatory factor analysis is imbued with robust dual
tests for both parameters (factor loadings) and the quality of the
factor.

Definition of SE M 5
 Confirmatory Factor Analysis ≠ Exploratory Factor Analysis CFA: Based
on model specification (Theory based approach) Exploratory Factor
Analysis is based on principle component analysis

EFA: Data driven approach

Two main model types in SEM: Measurement model and structural

model
Downloaded by Goethe-Universität Frankfurt At 09:23 30 March 2018 (PT)

Anderson and Gerbing (1988) pointed out a dual approach in

SEM modelling as encompassing the measurement model
(under confirmatory factor analysis) and the structural model.
Under the confirmatory factor analysis, there is the need to
develop an acceptable measurement model that adequately
describes the relationship between a number of latent variables
and their corresponding observed variables. Under the measure-
ment model, no causal relationships are established among the
latent variables of interest. Under the structural model (also
known as the causal model), the relationships among the latent
variables are shown, usually between the dependent and inde-
pendent variables.

1.5. Measurement and Structural Models

The two components of SEM are:

(a) Structural model capturing causal relationship between the

endogenous and the exogenous variables
(b) Measurement model showing relationship between latent
factor variables and the observed variables

6 Applied Structural Equation Modelling for Researchers and Practitioners

 Examples of measurement and structural models are depicted
below with their respective components.

(i) Measurement Model

Vi ¼ λi Fi þ ei ð1:1Þ

Measurement model consists of latent variables.

Vi: vector of observed variables
Fi: vector of latent constructs
Downloaded by Goethe-Universität Frankfurt At 09:23 30 March 2018 (PT)

λi: vector of parameters

ei: vector of measurement errors

(ii) Structural or causal model

Fi ¼ βi Mi þ τi Fi þ di ð1:2Þ

βi and Гi are parameter vectors

Fi : endogenous variables
Mi : mediating variable
Fi: exogenous variables
di: residual terms

Latent-independent variables are measured by observed indepen-

dent variables through a confirmatory factor analysis measure-
ment model (Figure 1.2).

Definition of SE M 7
 Latent-independent variable Latent-dependent variable

Measured by Measured by
observed observed
independent dependent
variables variables

via via
Downloaded by Goethe-Universität Frankfurt At 09:23 30 March 2018 (PT)

Confirmatory Factor Analysis

Figure 1.2: Confirmatory Factor Analysis under SEM.

1.6. Software for SEM

The following constitutes a list of software widely used in SEM.

• Open-source software. R has several contributed packages

dealing with SEM (e.g. ‘lavaan’). Lavaan was developed by
Rosseel (2012).
• STATA 13 (powerful and flexible enough to do sound SEM
estimation)
• LISREL (linear structural relations model) first developed in
1973 by Jöreskog and Sörbom.
• AMOS: Arbuckle (2011)
• SAS PROC CALIS
• OpenMx: Neale, Boker, Xie, and Maes (2003)
• EQS: Bentler (2004)
• Mplus: Muthén and Muthén (2010)

8 Applied Structural Equation Modelling for Researchers and Practitioners

 • SEPATH
• RAMONA

1.7. Non-Experimental Data (Absence of

Treatments and Controls)
SEM is mainly used in the case of non-experimental (observa-
tional) data. Experimental data pertains to data generated by a for-
mal measurement approach where controls and treatments are
introduced. Non-experimental data are widely used in social
Downloaded by Goethe-Universität Frankfurt At 09:23 30 March 2018 (PT)

science research by virtue of the fact that experimental data analy-

sis can be very costly. Because the researcher cannot control
assignment of subjects to the treatment and control groups, non-
experimental data are more difficult than experimental data to
analyse. Instances of non-experimental data include survey data.

1.8. Theoretical Construct

SEM represents a statistical technique geared towards gauging on
the extent to which a theoretical model is endorsed by data via
correlations. Indeed, in the case that the data endorses the theoreti-
cal model, this signifies possibility to test more complicated theo-
retical models. In the case that data does not corroborate the
theoretical model, there is the need to either change the model or
develop other theoretical models. After all, any theory holds when
the relationships among variables (captured through correlations/
covariances) are compatible with the propositions laid down by
the theory.

1.9. Conclusion
SEM constitutes a correlation research or statistical method which
has been widely used in consumer psychology, social and beha-
vioural sciences, but is now gaining ground in other fields such as

Definition of SE M 9
 finance (Chang, Lee, & Lee, 2009) as it is endowed with the ability
to frame and answer complex questions relating to the data. Other
names are used to capture SEM such as simultaneous equation
modelling, path analysis, latent variable analysis, confirmatory fac-
tor analysis. A latent-dependent variable is determined from other
independent latent variables. Alternatively stated, a latent-
dependent variable is a latent variable that is predicted by other
independent latent variables in SEM. Any latent variable (whether
dependent/independent) is always measured by observed variables
through confirmatory factor analysis. As a matter of fact, under
SEM, measurement models are developed to define latent variables.
Downloaded by Goethe-Universität Frankfurt At 09:23 30 March 2018 (PT)

1.10. Key Elements to Retain Whenever

Working with SEM
Variable definitions follow from LISREL convention.

ξ: (Ksis): latent-independent variables

Xs: observed-independent variables
ns(Etas): latent-dependent variables
ys: indicators for the dependent variables
SEM requires that sample size >200
General rule for SEM: 510 observations needed for each model
parameter estimated
Sample covariance matrix
Implied covariance matrix as generated by the theoretical
model
Correlation matrix is equal to standardized scaling among the
observed variables
S: Sample/observed covariance matrix

10 Applied Structural Equation Modelling for Researchers and Practitioners

 P
: Implied covariance matrix based on estimated parameters
based on theoretical foundation
P
S  = residual matrix; also called fitted residuals in LISREL
Examine residual values to ensure that they are small

When working with SEM, there is the need to report both

model fit and parameter estimates. The problem is that though a
model may generate sound model fit, yet, its parameter estimates
may be problematic.
Even if the SEM model fits to the data, this does not guarantee the exis-
Downloaded by Goethe-Universität Frankfurt At 09:23 30 March 2018 (PT)

tence of causal inference. Consequently, it can be argued that, under

SEM modelling, the statistical significance of the parameters predomi-
nate over the model fit. The underlying rationale is that, despite the fact
that a model fits the data well, yet, if some of the parameters are not sig-
nificant, this would blow up or discontinue the theory.
A correlation matrix simplifies the interpretation of results by rescaling
all variables to have unit variance.
A structural model, though, links only the latent variables, can also con-
tain independent observed variables. This can also be the case for a mea-
surement model. By default, the measurement model links the latent
variables to the dependent observed variables. However, additional
observed variables can still be incorporated as independent variables in
the case of a measurement model which purports to measure a dependent
latent variable.
SEM is not always scale free so that a model that fits the correla-
tion matrix may not fit the covariance matrix. In this case, it is
recommended to have recourse towards the covariance matrix.
Two major types of statistical tests are used in SEM:

(a) Assess the overidentifying restrictions placed on the model→χ2

likelihood test for goodness of fit.
(b) t-test for the parameters.

Def initio n of SEM 11

 The real interpretation of correlation analysis under SEM
Under SEM, the correlation coefficient analysis has three distinct
meanings as shown below.
Corr (x, y)→ three meanings

(i) X→Y
(ii) Y→X
→X
(iii) Z→Y (Mediating effects usually overlooked)
Downloaded by Goethe-Universität Frankfurt At 09:23 30 March 2018 (PT)

Some of the benefits of SEM over regression analysis:

1. Regression analysis does not control for measurement errors

and can merely deal with one dependent variable at a time.
2. SEM controls measurement errors.
3. SEM can handle several dependent variables.
4. SEM allows several independent variables without causing
multicollinearity problems.
5. Regression analysis assumes that the observable proxies are
exact measures of the theoretical constructs, which may not
be true due to measurement errors. Alternatively stated, only
SEM deal with observable variables but not latent constructs.
6. Assumptions of regression analysis can easily be violated.
However, under SEM, the normality distribution required by
maximum likelihood can be achieved by a normal score
transformation.

A separate chapter of the book probes further into the benefits

attached to SEM.

12 Applied Structural Equation Modelling for Researchers and Practitioners