196
allows more flexible assumptions, has the ability of testing
models overall rather than
coefficients individually, is able to model mediating variables, and can handle difficult data
(Anderson and Gerbing 1988, Hair
et al.
2006). Schumacker and Lomax (2004) describe the
following five basic-building blocks of all SEM analyses:
model specification, model
identification, model estimation, model fit testing and model modification. These basic-
building blocks are absolutely essential to both the measurement and structural models.
SEM is particularly valuable in inferential data analysis and hypothesis testing where the
pattern of relationships among the study constructs is specified
a priori and grounded in
established theory. It allows the researchers to test prior theoretical assumptions against
empirical data statistically (MacCallum and Austin 2000, Schumacker and Lomax 2004,
Arbuckle 2011). For these reasons, the present study employed SEM (using the AMOS 21.0
software program) to test the proposed hypothesized models. However, there are a number of
issues that must be addressed when using the SEM technique.
4.9.1 Analysis Approach
The first issue in the application of the SEM technique is the sequence
in which structural and
measurement analysis should occur. Although SEM is capable of testing the measurement
model and structural model simultaneously, Anderson and Gerbing (1988) recommend a two-
stage model-building approach and that the measurement model should be tested separately,
using confirmatory factor analysis (CFA) in order to detect any
inadequacy in fit, prior to
testing the full structural model. They suggest that the measurement model provides an
assessment of convergent and discriminant validity, while structural model provides an
assessment of predictive validity. By using a sequential approach in analysis,
the researcher is
able to pinpoint where a model is misspecified (Anderson and Gerbing 1988). The present
197
study adopts this approach by measuring the fit of a model with the current data, before
testing the structural relationships among the constructs in the model.
In addition, Mulaik and Millsap (2000) and Byrne (2009) suggest that it is best to have a few
good indicators for each of the latent variables in order to check
more easily how well each
observed variable measures a latent variable. Therefore, rather than using individual items as
indicator variables, the present study also uses EFA for each of the proposed sixteen latent
variables and so reduce a large number of related items to a
manageable number prior to
using them in the measurement and structural analyses (See the next chapter).
Do'stlaringiz bilan baham: