Past Event: Oden Institute Seminar

Bandit Problems, Clinical Trials, and Computational Ethics

Bill Press, Professor, ICES ICES and the Institute for Cellular and Molecular Biology (ICMB)

Abstract

As electronic medical records enable increasingly ambitious studies of treatment outcomes, ethical issues previously important only to limited clinical trials become relevant to unlimited whole populations. For randomized clinical trials, adaptive assignment strategies are known that expose substantially fewer patients to avoidable treatment failures than strategies with fixed assignments (e.g., equal sample sizes). An idealized adaptive case, the two-armed Bernoulli bandit problem, can be exactly optimized for a variety of ethically motivated cost functions that embody principles of duty-to-patient; but the solutions have been thought computationally infeasible when the numbers of patients in the study (the "horizon'') is large. We derive from numerical experiment a heuristic approximation that applies even for very large horizons, and propose a near-optimal strategy that remains valid even when the horizon is unknown or unbounded, thus applicable to comparative effectiveness studies on large populations or to standard-of-care recommendations.