There have recently been substantial developments in the analysis of incomp
lete data. Modeling tools are now available for nonrandom missingness and t
hese methods are finding their way into the broad statistical community. Th
e computational and interpretational issues that surround such models are l
ess well known. This article provides an exposition of several of these iss
ues in a categorical data setting. It is argued that the use of contextual
information can aid the modeler in discriminating among models that are ind
istinguishable purely on statistical grounds.