In conducting efficacy analyses with missing data in clinical trials, the NAS Report suggests that a missing data assumption such as MAR be chosen for the primary analysis and also as an anchor for sensitivity analyses. In the meantime, the reasons for missing data should be collected and considered in the formulation of sensitivity analyses. However, the suggested statistical methods seem to ignore the important information of patient dropout reasons that have already been collected, e.g., lack of efficacy, adverse effect and others. How to make use of such information may have great impact on the analysis of treatment efficacy.

In schizophrenia trials, for example, often accompanied with rampant patient dropouts, patient populations often consist of several different homogeneous subpopulations. Each dropout reason may be associated with a different subpopulation. Such imbalance may cause informative missing data; therefore, the assumption of MAR may be inappropriate. Most existing models for MNAR data do not address such issues. In this talk, we will discuss how to detect such informative missingness and how to make use of dropout reasons in efficacy analysis accordingly.