Heterogeneity in Response-Adaptive Randomization

Population heterogeneity is widely observed in various clinical trials. However, most current response-adaptive designs in clinical trials are based on the assumption of homogeneous patient responses, which could potentially undermine the solidness of the results of the trial. This dissertation aims to investigate the performance of response-adaptive randomization, especially doubly-adaptive biased coin design (DBCD) and covariate-adjusted response-adaptive (CARA) randomization under population heterogeneity.;Two properties of response-adaptive designs of major interest are the strong consistency to the desired target allocation and asymptotical normality. We first study these properties of DBCD under patient heterogeneity, and we find when certain conditions are imposed on patient responses the asymptotical properties of DBCD will be maintained. We also propose a modified DBCD which incorporates population heterogeneity into the allocation rule, and can potentially extend the application of DBCD to various nonhomogeneous clinical trials. In addition, the impact of heterogeneity on CARA randomization is examined, and the robustness of CARA randomization is proved under some widely satisfied conditions on population heterogeneity.;We also study some related issues in the inference such as the detection of time trends in patient responses and possible improvement of the estimation of unknown parameters. The weighted likelihood method is proposed to reduce estimation bias. Some remaining issues of future interest are discussed in the end.