Two issues in Bayesian adaptive designs in early phase clinical trials

This dissertation includes two research topics, focusing on the use of two Bayesian adaptive methods in early phase oncology clinical trial design: (1) the use of Bayesian outcome-adaptive randomization (OAR) method in Phase II randomized controlled trials, and (2) the use of Bayesian adaptive dose-finding method in Phase I trials. ❧ For the first research topic, this dissertation investigates the influence of the use of OAR on an analysis assessing the effect of a newly identified biomarker in an ancillary study or a retrospective study that uses data from an OAR trial. The dissertation demonstrates that the use of an OAR design perturbs the patient population and treatment assignments such that analysis of a newly identified biomarker using data from an OAR trial becomes subject to significant bias. It is shown that bias would be introduced to the analysis if OAR used information on an established biomarker, and the existence of bias for a series of settings is demonstrated both analytically and with simulation studies. The extent and the magnitude of the bias are examined. This dissertation proposes a correct method to conduct analysis of biomarker effect using data from an OAR trial that can render an unbiased estimate. The dissertation also proposes a variance estimator for the unbiased estimator of biomarker effect, and assesses the properties of the proposed estimators. ❧ For the second research topic, this dissertation assesses the properties of the EWOC (Escalation with Overdose Control) dose-finding design, and demonstrates that the EWOC design has several issues. First, an EWOC trial can keep recommending the same low dose level to all future patients after observing a toxicity response in the first or the second patient enrolled on the trial and will inevitably fail. Secondly, an EWOC trial can keep recommending a low dose level to a long series of patients, and use and treat patients inefficiently. Thirdly, there are undesirable properties with the estimators of the recommended dose at the completion of the trial. As a result, the EWOC design in its original formulation is impractical to be used in clinical trials. Considering the attractive overdose control feature of the EWOC design that the other dose-finding designs do not have, the dissertation proposes modified EWOC designs that resolve the issues, and demonstrates that the modified designs have more desirable properties than existing dose-finding designs, including the original EWOC design, the 3+3 design, and the restricted CRM (Continual Reassessment Method) design. Simulation studies are conducted to compare the performance characteristics among the designs, and recommendations are made regarding their usage in Phase I clinical trials.