Bootstrap methods for adaptive designs

Adaptive designs generate dependent sequences of random variables that are not exchangeable. Therefore, it is not obvious how to employ a resampling s cheme for confidence interval estimation. We propose a simple procedure whe re observed response rates from an adaptive experiment are input to a simul ation program. The program then generates sequences from the adaptive sampl ing scheme. We compare, via simulation, three bootstrap confidence interval s with the asymptotic confidence interval for two adaptive designs useful f or clinical trials. A simple ranking of simulated response rates yields a c onfidence interval approximation with coverage close to 1-alpha in most cas es. The method allows us to incorporate such complexities as staggered entr y and delayed response. We give an example of its utility on a clinical tri al of fluoxetine in depression. Copyright (C) 1999 John Wiley & Sons, Ltd.