Adaptive designs for phase II clinical trials with binary endpoints.

Abstract

Because the sample size is varying while the estimate of sample size is changing, the quality of an approximation of the binomial by the Gaussian is variable and thus not desirable. Also, adaptive designs do not follow the tradition of evaluating an entire sample of patients before analyzing the data. For an adaptive design with early stopping for either futility or efficacy, experimental designs are provided for some of the more common error probabilities. Following the footsteps of Simon (1989) two different levels of clinical significance are assumed for each experimental design. There are two problems with original design that are addressed. The first problem is that no explicit stopping boundary for efficacy was provided in the early stages. The other problem lay in assuming Bernoulli observations with identical probabilities of success. This is addressed by assuming the probability is a random variable and that each outcome is Bernoulli conditioned on the probability. When evaluating drugs, it is important to address both the efficacy and toxicity jointly. A joint model for both efficacy and toxicity is proposed and evaluated. Also a method for dichotomizing efficacy and toxicity events in a way that incorporates their severity, duration, and type.