Response-Adaptive Randomization for Precision Medicine Trials and the Pure-Exploration Contextual Bandit Problem

• Activity Number: 30 - Statistical Considerations in Adaptive Designs
• Type: Contributed
• Date/Time: Sunday, August 7, 2022 : 2:00 PM to 3:50 PM
• Sponsor: Biopharmaceutical Section
• Author(s): John Sperger* and Michael Kosorok
• Companies: University of North Carolina at Chapel Hill and University of North Carolina at Chapel Hill
• Keywords: response-adaptive randomization; experimental design; clinical trials; contextual bandits

Abstract

In the past decade, interest has risen substantially in the promise of data-driven precision medicine, which aims to assign treatments to patients based on their unique characteristics. Trial designs that account for substantial treatment effect heterogeneity are needed to work toward this goal. Response-adaptive randomization methods have a rich history in statistics, but to date, the literature has not focused on cases where the best treatment varies by subgroup. We propose a response-adaptive randomization technique for the case of linear decision boundaries to address this gap that periodically removes treatments at fixed intervals over the course of the study. This work was motivated by a clinical trial design with four active interventions and a strong a priori belief that the treatments will have different effects between patient subgroups. We show that by adapting the randomization based on the accumulating participant data, multi-arm clinical trials can simultaneously improve their power to identify the best treatment by patient subgroup and improve patient outcomes within the trial.