Practical Implementation of Bayesian Dose-Escalation Procedures

Abstract

This paper reviews Bayesian dose-escalation procedures for phase 1 clinical trials and describes a systematic approach to their implementation. The methodology is constructed for studies in which each subject is administered a single dose of an experimental drug and provides a single binary response, referred to here as toxicity or no toxicity. It is assumed that the probability of toxicity rises with log dose of drug according to a logistic regression model.

It is suggested that the choice of suitable prior distributions be aided via graphical representations of their properties and simulation investigations of their consequences. Possible safety constraints and stopping rules are described. Given this information, the recommended doses for the first cohort of subjects can be computed. Once their responses become available, subjective distributions can be updated, and the recommended doses for the second cohort can be determined. The procedure continues in this way until a stopping rule is reached, or until some maximum number of subjects has been observed. Clinical investigators are free to overrule the doses recommended by the procedure and to substitute those that they feel are more appropriate.