Missing data are categorised by Hair et al. (2006: 56) according to three different
types:
(i) non-random missing values – also referred to as Not Missing at Random
(NMAR, Byrne 2010: 354) and non-ignorable missing data (Schumaker and
Lomax 2004: 43; Garson 2011a);
(ii) Missing at Random (MAR); and
(iii) Missing Completely at Random (MCAR).
Non-random missingness in variable X can be determined where there are
statistically significant differences in the values of an independent variable Y when
the observations on Y are disaggregated using X = missing and X = not missing.
For example, where a categorical variable (Y) is measuring gender and X is a
attitudinal scale variable, a new, dichotomous missing/not missing variable (X 1 )
can be recoded from X. A chi square (χ^2 ) test can then reveal if there is a
statistically significant measure of association between X 1 and Y. A positive result
indicates that gender may be having an effect on whether or not variable X is
missing. Simply removing cases where X is missing might then bias the survey
findings with regard to gender dimensions. This is one reason (aside from
attrition of sample size) why it is often preferable to replace missing values rather
than simply remove them.
Using the same example, variables X (an attitudinal scale variable) and Y (a
categorical gender variable) as above, where data are Missing at Random
(MAR), we see that while there is not a random distribution of missing data
between females and males, within each sub-group (female and male) the
missing data are distributed at random.
Where data are Missing Completely at Random (MCAR), there is no correlation
between missingness of data and any other variable.
Following the guidelines in Hair et al. (2006: 57) each of the 29 item statements
with missing data was cross-tabulated with the demographic variables in the
survey to assess whether or not these demographic variables influence the
presence of the missing data. The demographic variables are:
gender;
age;
full-time / part-time;