Parsa Jamshidian, Zhiwei Zhang
Seamless phase II/III clinical trials offer an efficient way to select an experimental treatment and perform a confirmatory analysis within a single trial. Despite their operational advantages, such designs pose serious statistical challenges including control of the family-wise type I error rate and adjusting for selection-induced bias in treatment-effect estimation. We propose a nonparametric bootstrap method for estimating the phase III treatment effect that leverages data from both phases while avoiding reliance on potentially unrealistic parametric assumptions. The proposed estimator is simple and general, and explicitly adjusts for the bias induced by data-driven treatment selection in phase II. Its wide applicability is illustrated through examples involving different types of outcomes. The operating characteristics of the estimator are further evaluated through simulation studies, demonstrating substantially reduced bias compared with the naïve estimator that simply pools data from phases II and III.