Content area
Some New Advancements in Adaptive Randomization for Clinical Trials Clinical trial methodology continues to evolve as researchers strive to enhance efficiency, ethics, flexibility, and accuracy in response to the growing complexity and stringent regulatory requirements of modern clinical research. Despite these important methodological advances, significant practical challenges persist. Among the most pressing concerns are the development of trial designs that can adapt effectively to real-world conditions, inference methods capable of accommodating diverse data types, and statistical approaches that align closely with evolving regulatory standards. These ongoing issues highlight a gap between theoretical developments and practical application in clinical research. This dissertation addresses these critical gaps by identifying key limitations in current clinical trial methods and proposing practical, theoretically grounded solutions. The contributions presented here are intended not as incremental improvements but as meaningful advances capable of transforming clinical trial practice to better serve both researchers and patients.
In Chapter 3, we focus on Response-adaptive Randomization (RAR). Traditional RAR requires updates after every individual patient response, which is often impractical in real-world clinical trials. To address this limitation, we introduce new group-based RAR procedures. These methods update randomization periodically, either in groups or at fixed intervals like weekly or biweekly. Our theoretical analyses show these group-based methods maintain the strengths of traditional doubly adaptive biased coin designs (Hu and Zhang (2004)) and effectively handle delayed or missing responses. We illustrate the practicality and effectiveness of this approach through a real clinical trial example.
Chapter 4 explores statistical inference challenges in covariate-adaptive randomization (CAR) for multi-arm trials with general outcomes. Most existing inference approaches address continuous outcomes, leaving a notable gap for common trial endpoints such as binary or count data. We assess the performance of the standard Wald test under model misspecification, highlighting problems like inflated or overly conservative Type I error rates due to omitted covariates and correlation among tests from shared control groups. We propose adjusted test statistics that effectively control Type I error and may improve statistical power. This research represents the first comprehensive study on inference for multi-arm CAR trials with GLM endpoints.
In Chapter 5, we address seamless phase II/III clinical trial designs, highly relevant in oncology due to their potential to accelerate drug development. Current seamless methods predominantly target continuous endpoints, which do not align with oncology trials often relying on time-to-event primary outcomes and binary surrogate endpoints during interim stages. We propose a covariate-adaptive seamless phase II/III trial design specifically for trials using binary surrogate endpoints for interim analyses and time-to-event outcomes for final assessments. The accompanying inference method we introduce enhances statistical power, remains robust against model misspecification, and reliably controls Type I error. Simulation studies validate the practical effectiveness and regulatory suitability of our approach, emphasizing its usefulness for oncology trials.
Overall, these methodological improvements contribute significantly to clinical trial practice, addressing key limitations in current approaches. The proposed designs and inference methods enhance the practicality, adaptability, and accuracy of clinical trials, paving the way for more efficient drug development, improved regulatory alignment, and better patient outcomes across various therapeutic fields.
