Abstract:
In dose-finding studies with toxicity-efficacy responses, utility functions and Bayesian procedures are used to find a single optimal dose with ethical toxicity-efficacy trade-offs. We demonstrate that the design can have convergence issues when the prior information is misspecified. We propose to incorporate an information penalty to obtain multiple-dose allocation with efficient ethical measures. A coefficient is introduced to control the trade-off between information gain and ethical gain. We conduct simulations using Markov chain Monte Carlo (MCMC) algorithms to examine the convergence of Bayesian dose finding designs and investigate their operating characteristics. Different stopping rules are considered and their benefits are demonstrated by simulation. Guidance on how to select design parameters are given in two hypothetical trial redesigns.
