Referring to Table 7 - 18 , all factor loadings and factor covariances are statistically
significant although the model fit diagnostics for the initially-specified
measurement model indicate that the model is not a close fitting one. There are,
however, several low factor loadings that explain this poor fit. More importantly,
the initial model demonstrates discriminant validity with the highest squared
correlation estimate at 0.416 (ML EA) and the lowest AVE is 0.500.
The low loading items are recognisable from previous model iterations (ML4, ML5,
ME5, ME7, JS2, JS3, JS4 and JP3). These were all associated with high SRCs
(standardised residual covariances) and were removed and the measurement
model re-estimated. The results of this, a several subsequent model
modifications are summarised as follows.
Model CFA 3b:2
DSB4 shares a high (3.505) SRC with JP4; DSB4 removed (JP4 retained on
substantive grounds, as previously)
JS5 is flagged for observation owing to a high SRC (2.560) with ML3
Model CFA 3b:3
DSB3 removed owing to a moderate loading (0.611) and also a high SRC
(2.701) with JP4
JS5 remains on observation owing to 2.565 SRC with ML3
Prior to moving on to the modifications made to iteration CFA 3b:4, it is useful to
say a few words about the use of modification indices (MIs) for SEM model
development. For each parameter that is fixed to zero in a CFA or SEM model
(i.e. a potential path between two variables that is not included) a modification
index value is calculated by AMOS (and most other SEM software programmes).
For each pair of variables that could be connected by a path, the modification
index value provides an estimate of the improvement in model chi square (χ^2 M)
that will be observed if this path is freed. In practical terms, a high MI value
indicates that these two variables are linked by a common, but unmeasured,
influence (i.e. a common factor that is not included in the model) (Byrne 2010:
110).