The remainder of this section (7.2) describes the issues associated with construct
validity and model fit and goes on to report the diagnostic information for the
initial specification of Model 1. The following sections (7.2.8 to 7.2.10) then
describe the process of measurement model modification. Having established a
measurement model which exhibits good construct reliability and model fit, the
next stage is to undertake work on the structural equation model. This process is
reported in Section 7.3.
7.2.5 Assessment of construct validity
The rationale for conducting CFA (confirmatory factor analysis) in SEM is to
represent (using a range of observable indicator variables) latent constructs which
are not directly observable. Therefore, it is important to assess construct validity
- the extent to which the latent constructs actually represent the concepts they
are designed to measure. Hair et al. ( 2006 : 776) describe this as assessing “the
extent to which a set of measured items actually reflects the theoretical latent
construct those items are designed to measure”. There is not a single and
definitive test for construct validity (Kline 2005: 60) and Hair et al. (2006: 776-
779 ) provide a detailed discussion of four measures of construct validity:
(i) convergent validity – the extent to which items (indicators) of the latent
construct converge, that is, “share a high proportion of variance in common” (Hair
et al. 2006: 771). Three ways to estimate convergent validity are detailed:
factor loadings – the amount of variance in each indicator explained by the
latent factor
average variance extracted – the average variance extracted (AVE) is the
variance accounted for by the latent factor averaged across the indicators
construct reliability (CR) – high measures (>0.7) for CR are indicative that
all of the indicators consistently represent the same latent factor.
(ii) discriminant validity – the extent to which the latent variable is distinct from
other latent variables in the model.
(iii) nomological validity – is concerned with whether or not the correlations
among latent factors in the CFA make sense (with respect to the hypothesised
relationships or prior research).