unwilling to answer sensitive questions relating to, for example, age or income.
More significantly in the context of this research, is the issue of individual item
statements (or even whole sections of the questionnaire containing item
statements) not being completed.
Specific concern over non-response to individual item statements relates to the
limitations that this can place on the range of methods available for developing
the structural equation models. These concerns are described in greater detail
below.
Methods to deal with missing item statement data include (i) removing all cases
containing missing data (listwise deletion), (ii) excluding cases from a specific
analysis where a variable in that analysis is affected by missing data (pairwise
deletion) and (iii) replacing the missing data using estimates of what the values
might have been had they been entered by respondents.
Pairwise and (especially) listwise deletions both result in sample size attrition
(Schumaker and Lomax 2004: 25-26) and pairwise deletions can cause other
problems with SEM covariance matrices (see e.g. Kline 2005: 53-54). Because of
the sample size issue, Schumaker and Lomax (2004: 26) note that it is generally
preferable to replace missing values rather than simply remove cases with
missing values.
Missing value replacement (imputation) techniques include mean imputation
where missing values are replaced by the mean value (representing the most
likely value for that observation) for that variable, calculated from the values of
the completed responses. This method has negative consequences for SEM since
increasing the number of observations with mean values reduces the amount of
variance. Alternatively, regression imputation uses the actual observed values in
that variable as predictors (in regression equations) of the replacement values.
Once again, however, for SEM analyses, issues can arise with the undesirable
effects that this technique can have on variances (and covariances) (Byrne
2010:357). Finally, in recent years, model-based imputation methods have
emerged that improve upon both mean and regression methods. The model-
based methods replace missing data with values that are calculated from model-
specific statistics (see e.g. Kline 2005: 55) and are generally regarded as superior
to means substitution and regression imputation, particularly when sample sizes
are below 250 (Hair et al. 2006: 739-740).