In Computing Science and Statistics 23, (1991), pp. 421-424.

Bandit Strategies for Ethical Sequential Allocation

Janis Hardwick
Statistics Department, University of Michigan

Quentin F. Stout
EECS Department, University of Michigan

Abstract: The problem of allocating patients in a two treatment clinical trial with dichotomous response is considered. The trial goal is to determine the better treatment while incurring as few patient losses as possible. Several allocation rules are compared and it is found that bandit strategies perform well on both criteria in that they achieve nearly optimal power while keeping expected trial failures nearly minimal. The rules are also evaluated according to their computational complexity.

Keywords: sequential allocation, dynamic programming, ethics, Gittens index, power, probability of correct selection, indifference region

Full paper (in compressed Postscript)

Copyright © 1995-1997 Last modified: 5 Mar 1997