Bayesian mixture designs for phase II clinical trials

In a typical phase II efficacy trial of a new test drug, it is often desirable to assess patient response sequentially. Simon developed frequentist two-stage designs that have the desirable property of minimizing patient exposure to an ineffective test drug. However, his designs require fixed interim and final-stage sample sizes, which are often difficult to achieve. Bayesian designs allow continuous monitoring of data. In existing Bayesian designs, a single parametric model is often chosen to describe the prior belief(s) on the efficacy of a test drug. A single parametric model might not capture fully the variability associated with multiple prior beliefs of different clinical experts or multiple historic data sources. We propose using a class of mixture distributions to model multiple prior beliefs. Our goal is to develop practical and robust Bayesian sequential stopping regions that are comparable to optimal frequentist designs. We propose systematic methods to construct mixture models and compare early stopping rules of the proposed designs with existing frequentist and Bayesian designs.