Simple Missing Data Methods

As suggested in Section 15.2, missing data nearly always entail problems for the practicing statistician. First, inference will often be invalidated when the observed measurements do not constitute a simple random subset of the complete set of measurements. Second, even when correct inference follows, it is not always an easy task to trick standard software into operation on a ragged data structure.

Even in the simple case of a one-way ANOVA design (Neter, Wasserman, and Kutner 1990) and under an MCAR mechanism operating, problems occur since missingness destroys the balance between the sizes of the sub-samples. This implies that a slightly more complicated least squares analysis has to be invoked. Of course, a regression module for the latter analysis is included in most statistical software packages. The trouble is that the researcher has to know which tool to choose for particular classes of incomplete data.