ISSN (Print) : 2231-1734

Journal of Business Thought, Vol 10, DOI: 10.18311/jbt/2019/23452, April 2019 – March 2020 ISSN (Online) : 2581-8104

Structural Equation Modelling: A Powerful Antibiotic

H. K. Dangi1*, Ashmeet Kaur2 and Juhi Jham2
1
Associate Professor, Department of Commerce, Delhi School of Economics, University of Delhi, Delhi – 110021,
India; [email protected]
2
M.Phil Research Scholar, Department of Commerce, Delhi School of Economics, University of Delhi,
Delhi – 110021, India; [email protected], [email protected]

Abstract
This article is an attempt to scrutinize the applicability of the widely used statistical technique of Structural Equation Modelling
(SEM). SEM is a comprehensive technique to test the model adequacy. SEM is considered as an important advancement in
social science research as it combines measurement with substantive theories. It has been observed that many studies pay
attention to statistical mechanisation of SEM rather than the assumptions on which it is based. The history of SEM can be
traced to Regression Analysis, Path Analysis and Confirmatory Factor Analysis. SEM is popularly applied because of its use in
estimating multiple dependence relationships. It is able to measure the unobserved variables, define the model representing
the set of relationships and also corrects the measurement error. The technique is commonly applied in disciplines including
sociology, psychology and other fields of behavioural science. The availability of various user-friendly software programmes
like LISREL, AMOS, EQS, Mx, Mplus and PISTE is an added advantage. However, one should be careful while using SEM for
causal inferences. In comparison to other common standard statistical techniques, SEM is based on several assumptions. The
technique requires a priori knowledge of all the parameters to be estimated and a substantial amount of data pertaining to
covariances, variances and path coefficients. It also requires relationships to be specified in the model. The model inherently
assumes temporal precedence and is heavily dependent on researcher’s judgements about exogeneity and directionality.
Normality is yet another important assumption of SEM. The mismatch between data characteristics and assumptions imperils
inference and accuracy. Like antibiotics are a boon to mankind yet one needs to judiciously use them. Similarly, SEM is a
powerful technique however, researchers are suggested to apply cautiously.

Keywords: Confirmatory Factor Analysis, Latent Variables, Path Analysis, Regression Analysis, Structural Equation
Modelling
JEL classification: C36, C38

1. Introduction to support the hypotheses. SEM offers a major advantage

of the latent variables being free of random error, leaving
Structural Equation Modelling (SEM) is a statistical tech- only a common variance. The technique involves building
nique that subsumes and extends correlation, regression, a model and starts with specification of a model. Model
factor analysis and path analysis. The technique focuses specification is probably the most critical as well as chal-
on the fit of the data to the theory. SEM is analysis of lenging part because it requires adequate knowledge and
variance/covariance matrices of observed variables4. It understanding of the theoretical models.
yields an implied variance/covariance matrix which can
SEM is gaining popularity amongst researchers since
be compared to observed matrix. It tests how well the
it is a comprehensive technique that tests model ade-
constructed model fits the data as the model may be
quateness. SEM has become an essential statistical tool
theoretically identified but empirically unidentified. It
and technique for academicians and practitioners. It is
comprises testing two models namely measurement and
gaining popularity in its applications because of various
structural model. Measurement model gives empirical
reasons.
evidences, while the structural model provides framework

*Author for correspondence

 Structural Equation Modelling: A Powerful Antibiotic

Firstly, multiple observed variables can be studied 2. Historical Perspective of SEM

simultaneously. It makes SEM efficient to deal with the
modelling and testing of complex theories and phe- It is important to trace the history of SEM. Basically the
nomena. It also provides an edge to SEM over other premise of SEM rests on four kinds of related models. They
basic statistical techniques using only limited number are; Regression Analysis, Path Analysis and Confirmatory
of variables. Secondly, greater relevance is given to the Factor Analysis (CFA) and Structural Equation Models.
measurement error and the validity and reliability of the Path Analysis and CFA are actually two special types of
observed scores. This adds to the popularity of SEM tech- SEM. Regression model was developed by Francis Galton.
niques since the measurement error is taken explicitly It uses the criterion of Ordinary Least Squares and the
into account while statistically analysing the data. Thirdly, coefficient of correlation to depict a relationship between
several indices are reported for goodness and badness of two variables. Multivariate Regression models are used to
model fit. depict the scores of the dependent observed variable (Y)
Over the past 30 years, SEM has matured in its abil- through a set of independent observed variables (Xs) in
ity of analysing and testing advanced theoretical models. a way that the sum of the squared residual error values
In recent times, SEM has witnessed some major advance- is minimized. Structural Equation Models are different
ments such as the addition of new features of multilevel from regression models as they are relational and additive
structural equation modelling, growth curve change in nature.
modelling, generalized linear and mixed modelling, Path Analysis is an extension of multiple regression9.
meta-analysis, and partial least squares. In those theoreti- It was first developed in the 1930s by Sewall Wright to be
cal models wherein group differences are to be assessed, used in phylogenetic studies. It uses linear equation sys-
multiple-group SEM models can be used11. The devel- tem to examine causal relationships between two or more
opment of multilevel SEM, which allows analysing data variables. In the 1960s, Path Analysis was adopted in social
collected at more than one level (e.g. educational data) is sciences and since 1970s it is also being used in ecological
another reason for its popularity. sciences to a larger extent. In Path Analysis, we can have
All these advanced SEM models and techniques are more than one dependent variable at a time. The variables
can be both dependent and independent at the same time.
an improvement over the basic statistical methods. A
There can be a chain of association wherein one variable
variety of SEM methods have been used in various fields
can influence another variable, which, in turn, can influ-
such as business and sciences. But now, SEM techniques
ence the third variable. However, we substitute the terms
are widely used in the disciplines of biology, operation independent and dependent variables with exogenous
research, social, health and behavioural science. SEM is variables (which are not influenced by any other vari-
also evolving in the field of longitudinal investigations. ables) and endogenous variables (which are influenced by
All these improvements have enhanced the applications other variables) respectively. A major limitation of Path
of SEM and are referred to as Second Generation SEM. Analysis is its inability to measure and analyse unobserv-
The various software programmes available for SEM able variables. Path Analysis is a subset of SEM. In SEM,
are LISREL, AMOS, EQS, Mx, Mplus, PISTE. LISREL has we deal with unobservable variables, known as latent
been the most widely used programme since the 1970s. variables. These include the concepts which we encounter
All these software programmes are user-friendly and in everyday life such as anxiety, depression, quality of life
allow the researchers to get the results conveniently. and happiness. SEM extends Path Analysis and allows us
The rest of the paper is organised as follows: Section to examine relations among the variables, both latent and
2 discusses the history behind the evolution of SEM from measured.
various statistical techniques. Section 3 reviews SEM as an The applications of factor analysis were developed
important advancement in social science research and its by D. N. Lawley in the year 19408. Over the years, factor
applicability in various disciplines. Section 4 examines the analysis has been used in various disciplines to develop
rigid assumptions associated with the technique and how measurement instruments. It is being used to test theo-
they limit its applicability. Concluding remarks on SEM retical models and constructs and assess whether the
being a powerful antibiotic are presented in Section 5. specified model fits the data or not. It is used when a

20 Vol 10 | April 2019-March 2020 | www.informaticsjournals.com/index.php/jbt/index Journal of Business Thought

 H. K. Dangi, Ashmeet Kaur and Juhi Jham

priori hypothesis is framed in regard to how the vari- istics of using covariance as the basic statistic. Covariance
ables will cluster together on a factor1. It can be used to statistic conveys more information than regression as in
determine if the scale performs in the same manner while the latter, the differences between observed and expected
working with different population groups. The psycho- individual cases are minimized. While, in SEM, the dif-
metric properties of varied versions of the scale can be ferences between observed and expected covariance
compared using CFA. CFA is different from Exploratory matrices are minimized14. The analysis focuses on the fit
Factor Analysis (EFA). In EFA, the software performs its of the data to the theoretical model. SEM allows us to dis-
statistical operations and produces the best combination tinguish between direct and indirect relationships among
of variables clustering together to form a factor, even if it the variables by examining mediation and moderation.
is different from the combinations of variables that have SEM also indicates the group differences. Hence, it can
been hypothesised. In those cases where the model does be used to compare the results of separate models devel-
not fit the data, certain ideas and clues are available to oped for different groups through multiple-group SEM
guide the shuffling of variables so that the model fits the models. The longitudinal data for measuring the change
data in a better way. SEM is a combination of CFA and in the growth of variables over a period of time can be
multiple in a broad sense. collected. It augments in providing new ideas to research-
ers in the field of SEM. The availability of user-friendly
software programmes is an added advantage. Many of the
3. Applications of SEM software programmes are Windows-based and generate
SEM is a multivariate procedure that, as defined by the programme syntax internally, and thus, are easier to
Ullman12, “Allows examination of a set of relationships apply.
between one or more independent variables, either con- SEM is popular in its applications because of the req-
tinuous or discrete, and one or more dependent variables, uisite use of multiple observed variables by researchers
either continuous or discrete.” In SEM, the oval shape for better understanding. It is capable of dealing with
depicts latent variables while squares represent measured the sophisticated theories that are statistically modelled
variables10. Each latent variable, also called construct or and tested. All SEM have three main features3. First is the
unobserved variable, is in fact a small CFA. Lines are used characteristic of estimating multiple dependence rela-
to indicate relationships between variables. Lines have tionships. Second is the representation of unobserved
either one arrow for depicting a hypothesised direct rela- concepts in the relationships and correction of the mea-
tionship between two variables or two arrows indicating a surement errors. And third is the ability to define the
covariance between the two variables13. model representing the set of relationships. Another con-
SEM is considered as an important advancement in tribution to its popularity is the measurement error taken
social science research as it combines measurement with explicitly into account while statistically analysing the
substantive theories. SEM analysis deals with testing of a data. This ensures a greater importance to the reliability
model, testing a hypothesis about a model or modifica- and validity of the observed scores from the measurement
tion of an existing model. The analysis makes it possible instruments.
to simultaneously test all the relationships in case of com-
plex and multidimensional constructs. SEM is commonly
4. Limitations of SEM
applied in disciplines including sociology, psychology
and other behavioural science because of its capability to SEM is a powerful technique for testing models but the
test relationships between latent and measured variables. modelling process, at times, is complicated. The tech-
Further, availability of many user-friendly software have nique requires a priori knowledge of all the parameters
increased the popularity of SEM amongst researchers. to be estimated. Significant amount of data pertaining to
To check model adequacy, goodness of fit test is used5. covariances, variances, path coefficients and the relation-
SEM is useful in understanding the relational data in mul- ships is needed to be specified in the model. With the
tivariate systems and in examining the variances in the availability of user-friendly statistical softwares such as
variables. It can be distinguished from other conventional LISREL, AMOS and EQS, SEM is being used widely It is
methods of statistical analysis due to its distinct character- being excessively reported in social work journals without

Vol 10 | April 2019-March 2020 | www.informaticsjournals.com/index.php/jbt/index Journal of Business Thought 21

 Structural Equation Modelling: A Powerful Antibiotic

adequately validating the assumptions on which the tech- research are mostly composed of less than a few hundred
nique is fundamentally based. SEM inherently assumes cases, thus, violating the basic assumption.
temporal precedence i.e., presumed cause occurs before the It has been observed that in many research studies,
presumed effect6. Temporal precedence must be checked efforts made to model modifications for better goodness
through random assignment of cases to conditions in of fit statistic, are unnecessary. Significant modifications
experimental studies and by measuring cause and effect are made to improve a model fit. These include dropping
relationships at different points over time in non-experi- indicators, allowing cross-loadings and including numer-
mental studies. It is not possible to demonstrate temporal ous correlated error terms, which not only challenge the
precedence if all variables are measured simultaneously, reliability and validity of some studies, but also make it
which is true in case of most of the studies, thus, rendering impossible to replicate the findings with new data.
little justification for the inferences made. Studies that make excessive modifications to a model
It further assumes strong association or co-variation also hamper meaningful interpretations as interpreting
backed by both theories and results of empirical studies. a model with numerous correlated error terms or cross-
This assumption significantly affects the inferences made loadings or both, is inconclusive2. These modifications
as data may indicate spurious association. Association have improved the empirical fit of the model, however,
between two variables could be strong even if there is theoretical consistency is compromised. Testing direc-
no causal relation as both variables might have a com- tionality in relationships is yet another challenge in the
mon third factor causing them. Prior knowledge of causal model. It is the researcher’s hypotheses of causality that
relations is assumed in interpreting path coefficients. In
form the model. Recreation of the variance patterns,
various fields of research like behavioural science, one
observed in nature, is not possible by using SEM since the
barely knows the causal model, rather one hypothesizes
working of the model is limited by the researcher’s choice
the model. If the model fits the data one may conclude
of variables and paths. This makes several models to fit
that model is consistent with the data but one cannot
the data equally well. Eliciting desired results is relatively
claim about the applicability of the model as it is not
easy indicating why the model is overly applied.
proven to be true.
Structural models are heavily dependent on the
researcher’s judgments about exogeneity and directional- 5. Conclusion
ity. Exogeneity implies that the variable presumed to be
exogenous must not affect the endogenous variable in any One should rather be careful while using SEM for causal
other way than prescribed, directly or indirectly. Such a inferences. Assumptions are critical in specification,
variable must be uncorrelated with any other unmeasured analysis and interpretation. All statistical tests make cer-
cause of the endogenous variable. It is difficult to validate tain assumptions about the data or model. It has been
assumptions in the absence of robust empirical evidence. observed that inadequate attention is paid to assumptions.
Multivariate normality of the observed variable is yet Mismatch between data characteristics and assumptions
another important assumption. Before building a model of a particular method used, imperils inference and accu-
it is important to ensure that all the observations must racy of results. Conclusions that are extrapolated from
be drawn from a continuous and multivariate normal a model based on a small sample size are unreliable.
population. The Maximum Likelihood (ML) technique Omission of crucial variables is another major cause of
of approximation, which assumes normality, is used to poorly fit models.
estimate the parameters7. This makes SEM a large sample SEM is being overly applied and hence, grossly misused.
technique. Thus, drawing conclusions from a small sam- Unwarranted usage of SEM makes the validity of inferences
ple size makes them unreliable. questionable. It is being applied without an exhaustive
Use of non-continuous data in the model also leads knowledge of the variables and the model constructed, and
to profusely biased results. Many studies have used this without checking whether, or not, it is justified to use SEM.
technique on dichotomous or ordinal data, which is an Further, the technique is deployed most of the times with-
incorrect estimation method, as it may give inconsistent out validating the compliance of all the assumptions.
results. Estimation via ML technique further requires One may consider applying another suitable statistical
large sample size while sample sizes in social sciences test depending on the data properties. For non-normal

22 Vol 10 | April 2019-March 2020 | www.informaticsjournals.com/index.php/jbt/index Journal of Business Thought

 H. K. Dangi, Ashmeet Kaur and Juhi Jham

data, alternative methods of estimation like Ordinary 5. Kenny DA, McCoach DB. Effect of the number of vari-
Least Squares can be used with large sample size require- ableson measures of fit in structural equation modeling.
ments. There should be no missing data in any variable. Structural Equation Modeling. 2003; 10(3):333–51. https://
There are various methods dealing with such issues like doi.org/10.1207/S15328007SEM1003_1
Missing Completely at Random approach, Missing at 6. Kline RB. Assumptions in structural equation modeling. R.
H. Hoyle, Ed. Handbook of structural equation modeling.
Random approach and Imputation approach. One may
The Guilford Press; 2012. p. 111–25.
use Partial Least Squares in place of AMOS for forma-
7. Kumar S. Structural equation modeling basic assump-
tive constructs. Thus, if assumptions are not met, suitable tions and concepts: A novices guide. Asian Journal of
alternatives should be explored and employed. Management Sciences. 2015; 03(07):25–8.
Therefore, just like antibiotics, which are a big gift to 8. Lawley D. VI - The estimation of factor loadings by the
mankind, but have been extensively misused resulting in method of maximum likelihood. Proceedings of the Royal
drug resistance; SEM is also widely misused. Society of Edinburgh. 1940; 60(1):64–82. https://doi.
However, if used judiciously, it is a very powerful tech- org/10.1017/S037016460002006X
nique. Thus, there is a need for ensuring a more prudent 9. Streiner DL. Building a better model: An introduc-
use of SEM amongst researchers, warranting it is applied tion to structural equation modelling. Can J Psychiatry.
where apt and for the purpose it is meant for. 2006; 51(5):317–24. PMid: 16986821. https://doi.
org/10.1177/070674370605100507
10. Stoelting R. Structural Equation modeling/Path Analysis.
6. References 2002 Sept. http://userwww.sfsu.edu/efc/classes/biol710/
path/SEMwebpage.htm
1. Anderson JC, Gerbing DW. Structural equation modeling
11. Tomarken AJ, Waller NG. Structural equation modeling:
in practice: A review and recommended two-step approach.
Strengths, limitations and misconceptions. 2005; 1:31–65.
APA PsycNET Direct. 1988; 103(3):411–23. https://doi.
https://doi.org/10.1146/annurev.clinpsy.1.102803.144239
org/10.1037//0033-2909.103.3.411
12. Ullman JB. Structural equation modeling. New York, NY:
2. Guo B, Perron BE, Gillespie DF. A systematic review of
Harper Collins College Publishers; 1996. p. 709–819.
structural equation modelling in social work research. Br.
13. Ullman JB, Bentler PM. Structural equation modelling.
J. Soc. Work 2009; 39(8):1556–74. https://doi.org/10.1093/
Research Methods in Psychology. 2012; 2:663–83. https://
bjsw/bcn101
doi.org/10.1002/9781118133880.hop202023
3. Hair JF, Sarstedt M, Ringle CM, Mena JA. An assessment
14. Valluzzi JL, Larson SL, Miller GE. Indications and limita-
of the use of partial least squares structural equation
tions of structural equation modeling in complex surveys:
modeling in marketing research. Journal of the Academy
Application in the Medical Expenditure Panel Survey
of Marketing Science. 2012; 40:414–33. https://doi.
(MEPS). Agency for Healthcare Research and Quality,
org/10.1007/s11747-011-0261-6
Center for Financing, Access, and Cost Trends. 2003.
4. Hox JJ, Bechger TM. An introduction to structural equa-
tion modeling. Family Science Review. 1998; 11:354–73.

Vol 10 | April 2019-March 2020 | www.informaticsjournals.com/index.php/jbt/index Journal of Business Thought 23