STOCHASTIC OPTIMIZATION OF ADAPTIVE ENRICHMENT DESIGNS FOR TWO SUBPOPULATIONS

Aaron Joel Fisher, Johns Hopkins University, Bloomberg School of Public Health
Michael Rosenblum, Johns Hopkins University Bloomberg School of Public Health

Abstract

An adaptive enrichment design is a randomized trial that allows enrollment criteria to be modified at interim analyses, based on preset decision rules. When there is prior uncertainty regarding treatment effect heterogeneity, these trials can provide improved power for detecting treatment effects in subpopulations. An obstacle to using these designs is that there is no general approach to determine what decision rules and other design parameters will lead to good performance for a given research problem. To address this, we present a simulated annealing approach for optimizing the parameters of an adaptive enrichment design for a given scientific application. Optimization is done with respect to either expected sample size or expected trial duration, and subject to constraints on power and Type I error rate. We use this optimization framework to compare the performance of two types of multiple testing procedures. We also compare against conventional choices for design parameters that approximate O'Brien-Fleming boundaries and Pocock boundaries. We find that optimized designs can be substantially more efficient than simpler designs using Pocock or O'Brien-Fleming boundaries. Much of this added benefit comes from optimizing the decision rules concerning when to stop a subpopulation's enrollment, or the entire trial, due to futility.