It is not unusual for error variances to be linked (creating error covariances) in
model generating SEM, and this is often done during the model modification stage
(Kline 2005: 64-65) in order to improve model fit. Indeed, Bentler and Chou
(1987: 108), 1987, p. 108) have argued that forcing errors to be uncorrelated is
highly restrictive and “rarely appropriate for real data”. When introducing error
covariances in this way, however, researchers should always give consideration to
the theoretical implications it has for the model - not least because including
previously unspecified error covariances introduces an unknown and unmeasured
common factor into the model which can only serve to weaken the a priori theory
(see e.g. Kline 2005: 318).
The specification for measurement model CFA 1:1 is illustrated in Figure 7 - 2.
Figure 7-2 Specification for the measurement model CFA 1:1
Two constructs, Job Performance (JP) and Discretionary Service Behaviour(DSB),
are specified with 4 indicators while the Motivational Leadership construct (ML)
has 5 indicator variables. This means that the model exceeds the requirements
for model identification (see Section 7.2.2 below). Each indicator corresponds
with an individual item statement in the questionnaire (see Appendix VIII) and
the inclusion of each set of item statements (i.e. the measurement scale for each
construct) is based on the discussions in Chapter 5.
7.2.2 Measurement model identification
Model identification is essentially concerned with satisfying the requirements that:
(a) it should be possible to solve the equations (identify a unique solution) for the
relationships that are described by the theoretical model; and (b) every latent