|
Abstract:
|
In longitudinal clinical trials, one often encounters informative missingness. It is well-known that informative missingness introduces fundamental identifiability issues, resulting in unidentified intention-to-treat effects; the best one can do is conduct a sensitivity analysis to assess how much of the inference is being driven by missingness, anchored at some clinically meaningful baseline assumption such as missing-at-random (MAR). We introduce a class of baseline identifying restrictions, and argue that these assumptions are more appropriate for the analysis of non-monotone missingness than MAR. Deviations from baseline assumptions are handled through either transformation approaches or exponential tilting. The class of identifying restrictions so obtained results in a model which is nonparametrically unidentified, but is sufficient to identify the effects which are of clinical interest. We implement these ideas in a Bayesian nonparametric framework, where our identifying restrictions are shown to interact conveniently with posterior consistency. Advantages of our approach include a flexible modeling framework, access to simple computational methods, and theoretical support.
|
