The AMOS software for SEM analysis offers researchers an alternative method for
estimating models in the presence of missing data that relies neither on removing
cases nor imputing data. Rather it uses a special - and effective (Byrne 2010:
359) - form of maximum likelihood (full information maximum likelihood, FIML)
estimation to compute model parameters where variables contain missing data
(Kline 2005: 56; Arbuckle 2009: 270). Shumaker and Lomax (2004: 43) note
that FIML has become the favoured approach for dealing with missing data in SEM
analyses.
The FIML method, however, is not particularly suitable for the model generating
approach to SEM utilised in this research. This is because the FIML method (in
AMOS at least) does not allow for the computation of a standardised residual
covariance matrix (SRCM) or modification indices, both of which provide
information related to unmeasured variance in the model and are of key
importance for model modification and development. A further drawback with the
FIML method (which affects some of the analyses in this research) is its inability
to calculate bootstrapped estimates that are used to provide confidence intervals
for parameter estimates where data are multivariate non-normal. Finally, the
standardised root mean square residual (SRMR) estimate - which is a useful
measure of global model fit (see Section 7.2.6) - cannot be calculated with the
FIML method.
Based on the information described above, the favoured approach for dealing with
missing data in this research, is to firstly replace the missing data using the
model-based imputation method (using the AMOS software). This is the next-
best method for dealing with missing data short of using the FIML method and it
allows for (i) the use of modification indices for model development and (ii)
bootstrapped estimates to check the robustness of parameter estimates in models
where data are multivariate non-normal. Following Schumaker and Lomax’s
‘advice for prudent researchers’, (2004: 43) the parameter estimates and
goodness of fit statistics from the models that have been developed using data
sets containing imputed data will then be compared with estimates for the same
(fully-developed) models calculated using the FIML method.
In summary, the final models developed from model-based missing value
replacement will be re-estimated using the FIML to check that the imputation
process has not adversely affected the ultimate findings from the research.