Structural Equation Modelling (SEM) in Research: Narrative Literature Review

The structural equation modelling (SEM) method has stronger predicting power than path analysis and multiple regression because SEM is able to analyze at the deepest level the variables or constructs studied. This literature review aimed to describe the use of structural equation modelling in research. In general, SEM can be used to analyze research models that have several independent (exogenous) and dependent (endogenous) variables, as well as moderating or intervening variables. SEM provides several benefits and advantages for researchers, including building research models with many variables, examining variables or constructs that cannot be observed or cannot be measured directly (unobserved), testing measurement errors (measurement errors) for observed variables or constructs ( observed) and confirmatory factor analysis. Broadly speaking, SEM methods can be classified into two types, namely covariance-based structural equation modelling (CB-SEM) and variance or component-based SEM (VB-SEM), which includes partial least squares (PLS) and generalized structured component analysis (GSCA). This literature review aimed to describe the use of structural equation modelling in research.


Introduction
Human nature wants to continue to progress and develop in order to achieve a better quality of life. This also happens in the world of research. Experts in the social or behavioral sciences, including management, consistently develop research methods that can be used to obtain better, perfect, fast, accurate, effective, and efficient quality research results (Burhan, 2011).
Experts in the field of social or behavioral sciences, including management, have developed a research method called structural equation modelling (SEM) (Byrne, 2013). At first, the SEM method was only good at the conception level. At that time, the SEM method could not be operationalized due to technological limitations. With the rapid development of computer technology, the SEM method is now becoming increasingly recognized and widely used in behavioral and management research (Capmourteres, 2016). The SEM method is a development of path analysis and multiple regression, which are both forms of multivariate analysis models. In an associative, multivariate-correlational, or causal-effect analysis, the SEM method seems to break the domination of the use of path analysis and multiple regression, which have been used for decades. Compared to path analysis and multiple regression, the SEM method is superior because it can analyze data more comprehensively (Chang, 1981). Data analysis in path analysis and multiple regression was only carried out on the total variable score data, which is the sum of https://doi.org/10.37275/oaijss.v5i6.141

A B S T R A C T
The structural equation modelling (SEM) method has stronger predicting power than path analysis and multiple regression because SEM is able to analyze at the deepest level the variables or constructs studied. This literature review aimed to describe the use of structural equation modelling in research. In general, SEM can be used to analyze research models that have several independent (exogenous) and dependent (endogenous) variables, as well as moderating or intervening variables. SEM provides several benefits and advantages for researchers, including building research models with many variables, examining variables or constructs that cannot be observed or cannot be measured directly (unobserved), testing measurement errors (measurement errors) for observed variables or constructs ( observed) and confirmatory factor analysis. Broadly speaking, SEM methods can be classified into two types, namely covariance-based structural equation modelling (CB-SEM) and variance or component-based SEM (VB-SEM), which includes partial least squares (PLS) and generalized structured component analysis (GSCA  (Chen, 2010).
The SEM method has stronger predicting power than path analysis and multiple regression because SEM is able to analyze at the deepest level the variables or constructs studied (Cohen, 2013). The  (Cudeck, 1994). In comparison, the SEM method can be likened to being able to reach as well as parse and analyze the deepest entrails of a research model. The SEM method is expected to be able to answer the weaknesses and impasses faced by the previous generation of multivariate methods, namely path analysis and multiple regression (Curran, 2003). The development of SEM methods is becoming increasingly significant in the practice of social, behavioral, and management research, along with advances in information technology (Duncan et al., 2013). Many multivariate statistical methods which were difficult to operate manually in the 1950s, such as factor analysis, multiple regression with more than three independent variables, path analysis, and discriminant analysis, gradually became necessary because of the invention of computer programs such as SPSS (Statistical Package for Social Science), Minitab, Prostat, QSB, SAZAM, etc. The SEM method is currently estimated to be the most dominant multivariate method.
Computer programs that can currently be used to process data in SEM research methods include AMOS, LISREL, PLS, GSCA, and TETRAD. This literature review aims to describe the use of structural equation modeling in research (Eisenhauer et al., 2015).

The benefits of SEM in research
In general, SEM can be used to analyze research models that have several independents (exogenous) and dependent (endogenous) variables, as well as moderating or intervening variables (Fan et al., 1999).
SEM provides several benefits and advantages for researchers, including building research models with many variables, examining variables or constructs that cannot be observed or cannot be measured directly (unobserved), testing measurement errors (measurement errors) for observed variables or constructs (observed), confirming the theory in accordance with research data (confirmatory factor analysis), being able to answer various research problems in a more systematic and comprehensive analysis set; more illustrative, robust and reliable than the regression model when modeling interaction, nonlinearity, measurement error, correlation of error terms, and correlation between multiple independent latent variables; used as an alternative to path analysis and covariate-based time series data analysis; factor, path and regression analysis; explain the complex interrelationships of variables and direct or indirect effects of one or several variables on other variables; and has higher flexibility for researchers to relate the theory with data (Fritz et al., 2007;Grace, 2006).

Types of SEM
As stated above, in general, the SEM method can be classified into two types, namely covariance-based structural equation modelling (CB-SEM) and variance or component-based SEM (VB-SEM), which includes partial least squares (PLS) and generalized structured component analysis (GSCA) (Grace, 2008;Grace, 2010). A variant is the deviation of the data from the mean (average) value of the sample data. Variance measures the deviation of data from the mean value of a sample, so it is a measure of metric variables.
Mathematically, the variance is the average of the squared differences between each observation and the mean, so the variance is the average squared value of the standard deviation (Haavelmo, 1943). A variable must have a variance that is always positive. If it is zero, then it is not a variable but a constant.
Meanwhile, covariance shows a linear relationship that occurs between two variables, namely X and Y. If a variable has a positive linear relationship, then the covariance is positive. If the relationship between X and Y is opposite, then the covariance is negative. If there is no relationship between the two variables, X and Y, then the covariance is zero.

Covariance-based structural equation modelling (CB-SEM)
Covariance-based SEM (CB-SEM) was first developed by Joreskog (1973), Keesling (1972), andWiley (1973). CB-SEM became popular after the availability of the LISREL III program developed by Joreskog and Sorbom in the mid-1970s. By using the maximum likelihood (ML) function, CB-SEM tries to minimize the difference between the sample covariance matrix and the covariance matrix predicted by the theoretical model so that the estimation process produces a residual covariance matrix with a small value close to zero. Some things that need to be considered in CB-SEM analysis include the following: a) The assumption of using CB-SEM is like the parametric analysis. The assumptions that must be met are that the observed variables must have a multivariate normal distribution, and the observations must be independent of one another. If the sample is small and not asymptotic, it will give poor parameter estimates and statistical models or even produce a negative variance, which is called the Heywood Case.
b) A small sample size will potentially result in a Type II error, i.e., a bad model will still result in a fit model.

PLS-SEM
PLS-SEM aims to test predictive relationships between constructs by seeing whether there is a relationship or influence between these constructs (Hair et al., 2013). The logical consequence of using PLS-SEM is that testing can be carried out without a strong theoretical basis, ignoring some assumptions (non-parametric) and the parameter accuracy of the prediction model seen from the value of the coefficient of determination (R 2 ). PLS-SEM is very appropriate for use in research that aims to develop theory. PLS-SEM was developed to overcome tests that cannot be done with CB-SEM. (Harrington, (2009) . This provides an ideal picture scientifically in data analysis. However, the data to be analyzed does not always meet the ideal criteria, so it cannot be analyzed by hard modeling (Hu, 1999). As a solution, soft modeling tries to analyze data that is not ideal. Literally, soft actually means soft or soft, but in the research context, soft is defined as not based on assumptions on the scale of measurement, data distribution, and sample size (Iacobucci, 2010). The main purpose of analysis with hard modeling is to test the causal relationship between those that have been built based on the theory and whether the model can be confirmed with empirical data. In comparison, the main objective of soft modeling analysis aims to find predictive linear relationships between latent constructs. It should be understood that a causality or estimation relationship is not the same as a predictive relationship (Jackson et al., 2009) (Joreskog, 1993). So that the existing events can not be fully controlled, if the data to be analyzed meets all the assumptions required by CB-SEM, then the researcher should analyze the data by hard modelling using appropriate software, such as AMOS and LISREL (Kim, 2005).
If the data does not meet all the required assumptions, but the researcher still uses hard modelling or CB-SEM analysis, then several problems may be encountered, an improper solution or an imperfect solution because of the Heywood Case, which is a symptom of a negative variance value; the model becomes unidentified due to indeterminacy; and non-convergence algorithms. If that conditions occur and we still want to analyze the data, then our goal is not to change causality between variables but to find optimal predictive linear relationships using component or variance-based SEM (Lamb et al., 2014).  (Mulaik et al., 1989;Murtaugh, 2009

Conclusion
SEM can be used to analyze research models that have several independent and dependent variables as well as moderating or intervening variables.

References
Burnham 2008. An empirical evaluation of the use of fixed