In addition to the considerations described above, the selection of imputation
method is dependent on the characteristics of the missing data in terms of their
type, extent and degree of randomness (Hair et al. 2006: 54). The model-based
imputation method for replacing missing values is robust against all degrees of
randomness (Hair et al. 2006: 58), nevertheless, this research goes on to follow
Hair et al.’s (2006: 53) four-step process to evaluate the type and extent of the
missing data and to characterise its degree of randomness.
Prior to beginning the four-step process, an initial examination of the 234
returned survey forms found that of these, 8 had not been completed at all and a
further 2 were missing around half of the 53 item scale questions. These 10
survey forms were removed from the database leaving 224 survey forms to
examine.
Step 1 – Type of missing data
Step 1 of Hair et al.’s (2006) process is to determine whether or not the missing
data are ignorable.
The missing data associated with the Service Quality (SQ) latent construct are
ignorable, since these are missing by design. Specifically, the approach to
measuring SQ for this research is to record the frequency that each respondent is
able to deal with each type of service failure / negative situation, while
maintaining customer satisfaction. Because it is possible that an individual
respondent ‘never’ experiences the particular service failure situation being
described, respondents were instructed simply not to respond in these
circumstances.
Because they are ignorable, the Service Quality (SQ) data (measured using 7
item statements) are not included in the following discussion. Furthermore,
because of the inherent missing data in the measurement of the SQ latent
construct, the SQ data are analysed in a separate modelling exercise using a
truncated sample based on listwise deletion of cases.
Step 2 – Extent of missing data
The remaining missing data are non-ignorable and Step 2 is to determine the
extent of these.