These findings show that there have been no substantial or adverse consequences
on the model parameter estimates as a result of employing the model-based data
imputation method in preference to the generally favoured FIML estimation
method.
A general rule of thumb for missing data approaches provided by Hair et al.
(2006: 55-56) is that, before undertaking any data imputation, it should be
possible to undertake the planned analysis effectively using only the cases with no
missing values at all. There are 177 cases with no missing data in the current
data set - according to the guidelines for SEM sample sizes described in Section
6.5.2, such a sample should be adequate for analysis to proceed. To provide an
empirical check on the adequacy of the 177 cases for SEM analysis, Model 5b was
re-estimated using only the data from these 177 cases.
The model fit statistics for this model are described in Table 7 - 32 (alongside the
results of the FIML estimates) and show that, once again, there is no appreciable
difference in the overall fit. All factor loadings and structural coefficients remain
statistically significant. All of the factor loadings and structural coefficients vary
to some extent from the values in the original Model 5b, and this reflects the fact
that the data set has changed significantly. The variations are greater than in the
previous model comparison, in this case ranging between 0.001 and 0.6.
However, the only substantive difference between the model estimated with 177
cases and the original Model 5b is that the squared multiple correlation value for
the Job Performance (JP) construct has dropped to 0.18 (from 0.22 in Model 5b
with 213 cases).
The one model fit measure that has changed somewhat is Hoelter’s Critical N
which assesses adequacy of sampling size. Reflecting the smaller sample size
(177 cases versus 213 cases in the full data set), this estimate has dropped
slightly below the preferred level of 200 to 194. Nevertheless, this estimate is
very close to 200 and is considerably above the minimum of 75 suggested by
Garson (2011b). Based on these findings, it is concluded that the sample size of
177 cases with no missing values was adequate for undertaking SEM analysis. As
a consequence, the general rule of thumb for missing data approaches provided
by Hair et al. (2006: 55-56) is satisfied.