Brief comments on computational issues with multiple imputation

MI is usually performed under the assumption that the mechanism causing the missing data is ‘Missing At Random’. We discuss the practical implications of this elsewhere (Carpenter and Kenward, 2008). Here we note that although this assumption may be plausible, it cannot be verified from the data at hand, and therefore the analysis of a partially observed data set under the MAR assumption can never have the same status as the analysis of the fully observed dataset would have had. Thus, it is helpful to present any analysis carried out on the partially observed data (e.g. using MI) alongside the analysis based on only those units/individuals with no missing data (so called ‘Complete Cases’ (CC)), to see how the conclusions differ. If there should be differences, it is then important, if possible, to provide an explanation for these, as this increases confidence in any conclusions drawn.