SEM is a statistical technique that allows the simultaneous analysis of a series of structural equations and is particularly useful when a dependent variable in one equation becomes an independent variable in another equation (D. Smith &
Langfield-Smith, 2004). There are two-stages in SEM process, i.e., the analysis of the measurement models and analysis of the structural model (Schumacker &
Lomax, 1996; D. Smith & Langfield-Smith, 2004). The measurement model specifies relations between manifest (observed) variables and latent variables. The structural model is a model of relations between latent variables, incorporating specified measurement error. According to Hair, Black, Babin, Anderson and Tatham (2006) SEM involves a six-stage decision process as outlined in Figure 3.2.
Stages one to three of the process are discussed throughout this chapter, while stages four to six are discussed with the data analysis in Chapter Six.
No Yes
Figure 3.2
Six-Stages Decision Process in SEM (Source: Hair et al., 2006)
Stage 1
Defining the Individual Constructs
‐ What items are to be used as measured variables
Stage 2
Develop and Specify the Measurement Model
‐ Make measured variables with constructs
‐ Draw a path diagram for the measurement model
Stage 3
Designing a Study to Produce Empirical Results
‐ Assess the adequacy of sample size
‐ Select the estimation method and missing data approach
Stage 6
Assess Structural Model Validity
‐ Assess the GOF and significance, direction, and size of structural parameter estimates
Stage 4
Assessing the Measurement Model Validity
‐ Assess line GOF and construct validity of measurement model
Stage 5
Specify Structural Model
‐ Convert measurement model to structural model
Structural Model Valid?
Refine model and test with new data
Draw substantive conclusions and recommendations
SEM is considered the most appropriate method when the research stream has a relatively sound theory. There is a reasonable strong body of knowledge in modelling relations between environment, strategy and organizational structure, and a considerable body of accounting literature that has explored relations between strategy and non-financial measures (D. Smith & Langfield-Smith, 2004). Moreover, SEM can be used to specify causal direction in specific situations. However, it should be noted that although SEM is often referred to as “causal modelling”, it can only provide evidence of causality, not establish causality (Hult et al., 2006).
SEM emphasizes the analysis of sample variances and covariances rather than individual cases. Instead of minimizing the sum of squared differences between the predicted and observed scores for each case, the SEM technique involves minimizing the difference between the matrix of sample variances and covariances and the matrix of predicted variances and covariances generated from using a set of parameters that describe the causal model underlying the relationship amongst the variables. Thus, SEM develops a comprehensive model to test hypotheses in this study.
Compared to other traditional analyses, for example multiple regressions, results of SEM are more informative for management accounting theoreticians. SEM allows a range of relations between variables to be recognized in the analysis. Thus, SEM provides the researcher with an opportunity to adopt a more holistic approach to model building. Other than that, a major difference between SEM and other traditional analyses is the ability to account for the effects of estimated measurement error of latent variables. This is particularly relevant to management accounting research when composite measures are often used to measure the construct. The use of interaction terms in multiple regressions may encompass significant measurement error, particularly when used with composite variables. These problems have led prominent management accounting researchers to suggest that multiple regression techniques are inappropriate in many situations (D. Smith & Langfield-Smith, 2004).
There are two main types of SEM:
1. Covariance-based structural equation modelling (CBSEM), such as Linear Structural Relations (LISREL).
2. Variance-based (or component-based) approach, for instance Partial Least Square (PLS).
This study uses a CBSEM approach and employs LISREL for Windows version 8.80 to analyse the data. The CBSEM approach enables researchers to construct unobservable latent variables, model errors in measurement, and statistically test a priori theoretical and measurement assumptions against empirical data. As compared to PLS which is a softer modelling approach used to determine values of the latent variables for predictive purposes (Chin, 1998), CBSEM involves analysis using a set of parameters that describe the causal model underlying the relationship amongst the variables. Under this condition, indicators are viewed as being influenced by the underlying latent construct (reflective mode).
This study aims to examine the effect of changes in MAP as well as organizational structure and strategy on performance, which caused by the changes in competitive environment and AMT. Hence, CBSEM is the best method for analysing the hypotheses developed from the conceptual framework in this study.