Robust Bayesian approaches for clinical trial monitoring.

The interim monitoring and final analysis of data arising from a clinical trial require an inferential method capable of convincing a broad group of potential consumers: doctors; patients; politicians; members of the media, and so on. While Bayesian methods offer a powerful and flexible analytic framework in this setting, this need to convince a diverse community necessitates a practical approach for studying and communicating the robustness of conclusions to the prior specification. In this paper we attempt to characterize the class of priors leading to a given decision (such as stopping the trial and rejecting the null hypothesis) conditional on the observed data. We evaluate the practicality and effectiveness of this procedure over a range of smoothness conditions on the prior class. First, we consider a non-parametric class of priors restricted only in that its elements must have certain prespecified quantiles. We then obtain more precise results by further restricting the prior class, first to a non-parametric class whose members are quasi-unimodal, then to a semi-parametric normal mixture class, and finally to the fully parametric normal family. We illustrate all of our comparisons with a dataset from an AIDS clinical trial that compared the effectiveness of the drug pyrimethamine and a placebo in preventing toxoplasmic encephalitis.