Revisiting the Methodological Schools of Response-Adaptive Randomization

This session offers a unique and timely comparative overview of the major methodological paradigms for the design of Response-Adaptive Randomization (RAR), which aims to optimize patient benefit and statistical efficiency in clinical trials. The field's adoption has been hindered by the fragmentation of methodological approaches across different literatures (clinical trials, Bayesian methods, and Machine Learning). This session is designed to be unique in the breadth of literature put together, with the singular goal of identifying common foundational ideas for the design of experiments. Leading experts will present recent work in three dominant areas-Frequentist optimal allocation, Bayesian adaptive designs in theory and in practice, and Multi-Armed Bandit (MAB) methods adapted for clinical trials-allowing for a deep discussion of their statistical underpinnings, critical assumptions, and practical differences. By revisiting these schools, the session will focus on reconciling them through common language and objectives. The goal is to build robust trial designs that satisfy both ethical mandates (patient-centricity) and regulatory requirements (e.g., strong Type I error control), charting a path toward a more harmonized methodology for achieving superior patient outcomes. The session will feature four presentations, followed by a discussion. Lorenzo Trippa will begin by presenting Uncertainty-Directed Designs, a modern Bayesian approach that uses uncertainty about treatment effects to govern randomization, focusing on maximizing both information gain and patient benefit. Following this, Feifang Hu will address frequentist optimal adaptive strategies for navigating A/B testing settings. Emilie Kaufmann will then introduce some ideas from the growing literature on pure exploration in multi-armed bandit models and their potential benefit for limiting the number of patients needed to attain some type I and power control (instead of maximizing patient benefit as Thompson Sampling does). The final presentation will come from Sofía S. Villar, who will present hybrid approaches to Bayesian design and non-parametric analysis based on allocation probabilities tests. The session will conclude with a synthesizing discussion led by Professor William Rosenberger (The George Washington University), who will summarize the methodological overlap and divergence among the three schools (Bayes, Frequentist, MAB) and offer views on the practical pathways for harmonization and regulatory acceptance of these advanced tools in modern clinical trials.