Fitting models to incomplete categorical data requires more care than fitti
ng models to the complete data counterparts, not only in the setting of mis
sing data that are non-randomly missing, but even in the familiar missing a
t random setting. Various aspects of this point of view have been considere
d in the literature. We review it using data from a multi-centre trial on t
he relief of psychiatric symptoms. First, it is shown how the usual expecte
d information matrix (referred to as naive information) is biased even unde
r a missing at random mechanism. Second, issues that arise under non-random
missingness assumptions are illustrated. It is argued that at least some o
f these problems can be avoided using contextual information. Copyright (C)
1999 John Wiley & Sons, Ltd.