ABSTRACT
Non-ignorable missing data are the common type of missing data in HRQoL studies where there is dropout as a result of toxicity, disease progression, or even therapeutic effectiveness. Studies with this type of missing data are also the most difficult to analyze. The primary reasons are (1) there are a large number of possible models and (2) it is impossible to verify statistically the correctness of any model because the data required to distinguish between models are missing. In this chapter, we describe one group of models, pattern mixture models, that has been proposed for this problem. In pattern mixture models, the portion of the model specifying the missing data mechanism (f [R|X,Ψ]) does not depend on the missing values. Thus, for pattern mixture models, we only need to know the proportion of subjects with each pattern of missing data and we do not need to specify how missingness depends on Y misi . This advantage is balanced by (1) the large number of potential patterns of missing data and (2) the difficulties of estimating all parameters in each pattern.
