Statistical Estimation and Innovation in Oncology Trials

This thesis investigates estimation in innovative oncology trial designs, with emphasis on basket and master protocols. It links design choices to estimators and quantifies trade-offs between bias, variance, and interpretability. Chapter 2 develops an unbiased UMVUE and exact confidence intervals for response rates in stopped Simon's two-stage designs, showing that early stopping can save resources while modestly reducing precision. Chapter 3 studies the Restricted Mean Duration of Response (RMDoR) under right/interval censoring, clarifying the role of the truncation time τ, proposing practical estimators, and demonstrating that RMDoR ratios provide robust, comparable efficacy summaries. Chapter 4 benchmarks Bayesian information-borrowing estimators for cohort-specific response in single-stage basket trials, revealing context-dependent performance and sensitivity to prior specification. Chapter 5 extends the evaluation to two-stage basket designs (motivated by DRUP), introducing simulation-based tuning to select models and hyperparameters under mean and worst-case utilities; 4 methods was used on this research, EXNEX is generally stable, while Berry, Psioda, and Fujikawa can excel when carefully tuned. Overall, the work argues that estimation is integral to design, advocates prespecifying borrowing strategies and parameter grids, and highlights the need for clearer guidance on transparent prior/tuning choices in precision-oncology trials.