Factor correlation Correlation estimate Squared correlation
estimate
ML JP 0.442^ 0.195^
ML DSB 0.375^ 0.141^
JP DSB 0.331^ 0.110^
specified measurement model (CFA 1:1) Table 7-2 Factor correlations and squared correlation estimates for the initially
(CFA 1:1) Table 7-4 Modifications made to the initially specified measurement model
In Table 7 - 3 AVE estimates are on the diagonal and squared correlations are
displayed beneath the diagonal.
ML JP DSB
ML (^) 0.560 - -
JP 0.195 0.557 -
DSB 0.141 0.110 0.585
AVE values are on the diagonal and squared correlation estimates below the diagonal
Table 7-3 Discriminant validity estimates for CFA 1:1
Turning to an examination of the indicator variables. Hair et al. (2006: 796)
recommend that low loading items (<0.7) are “candidates for deletion” and that
the decision to remove such indicators from the model should be made with
reference to other model diagnostics including the relevant standardised residual
covariance values. Of course, as noted above, theoretical considerations must
also guide the model modification process. Schumaker and Lomax (2004: 71)
write that substantive interest should be the ‘guiding force’ in model modification
and go as far as to say that, even where a particular parameter exhibits a
problem, if it is of ‘sufficient substantive interest then it should probably remain in
the model (Schumaker and Lomax 2004: 71, emphasis added).
Before considering the substantive implications of removing the low-loading
indicators, an examination of the standardised residual covariances is undertaken
to identify which indicators are exhibiting large volumes of unmeasured variance.
There is a range of approaches to interpreting the standardised residual
covariance matrix: Schumaker and Lomax (2004: 71) write that values >1.96 or
2.58 “indicate that a particular covariance is not well explained by the model”;
Byrne (2010: 86) notes that values >2.58 “are considered to be large”; and Hair
et al. (2006: 797) say that a value >4 “flags a problem” while values between 2.5
and 4 deserve attention unless there are no other concerns specific to that
variable’s diagnostics. Considering these guidelines, for this research, SRC values
greater than 2.58 will be addressed by seeking to remove an associated indicator