to 7.11, indicator variables will be referred to using their respective acronyms
(e.g. ML1, DSB4). This procedure is generally followed in order to achieve brevity
throughout the technical process of model modification and development. In
some cases, however, the substantive content of an indicator (i.e. the actual
statement used) will be referred to in order to clarify the theoretical implications
of a modification.
Section 7.13 is devoted to an in-depth discussion of the substantive implications
of the model modifications and explicitly discusses the wording of each relevant
indicator / item statement. Prior to Section 7.13, it is possible to identify the
content of any item statement by consulting Appendix VIII or the relevant section
of Chapter 5 where the selection of the item statements for measuring each
construct is discussed in detail.
7.2.8 Measurement model modification
Kline (2005: 186-188) describes ‘two general classes of problems that can be
considered in respecification’; these being (i) factors and (ii) indicators.
Considering the former, Kline recommends examining the strength of factor
correlations for evidence of poor discriminant validity noting (2005: 60 and 73)
that factor correlations in the region of >0.85 and >0.90 are indicative of
constructs which are not sufficiently distinct from one and other. In this model,
the largest between factor correlation is 0.442, indicating that the constructs are
distinct from each other.
The other rule of thumb with regard to discriminant validity (see Section 7.2.5
above) is that of Fornell and Larker (1981: 47) who recommend that the estimate
for the proportion of AVE (average variance extracted) for each individual latent
factor should exceed the value of each of the squared correlations between the
factors. For CFA 1:1, the lowest AVE is 0.557 (for the JP factor) and the highest
squared inter-factor correlation estimate is 0.195. We can therefore conclude
that the initially specified model performs adequately for discriminant validity.
Table 7 - 2 shows the squared correlation estimates and Table 7 - 3 integrates the
squared correlation estimates with the AVE values from Table 7 - 1 to provide an
at-a-glance comparison of the estimates for determining discriminant validity.