Auto-adaptive Alpha Allocation: A Strategy to Mitigate Risk on Study Assumptions

In some clinical development programs, there are potential biomarkers with promising but uncertain predictive effect, while the probability of success in the overall population cannot be readily dismissed. It is risky to focus only on the overall population, or just the biomarker subpopulation. In 2009, Chen and Beckman proposed a Bayesian decision framework to optimize the type I error rate (alpha) allocation in a Phase III clinical study with possible predictive subset effect. The utilization of internal data in this framework is of particular interest because it provides an opportunity to mitigate the potential risk of misspecified study assumptions using an auto-adaptive strategy. In this paper, we examine this auto-adaptive strategy in detail through extensive numerical case studies and provide guidance on the appropriate use of partial current trial (internal) data in this data-driven optimization framework. We show that internal data can be used to inform the alpha allocation to hypothesis testing in the overall population and the subgroup. The resulting adaptive testing strategy is robust with respect to the uncertainty in the predictive subgroup effect and biomarker prevalence.