Methods for the Design and Analysis of Clinical Trials: Uncertainty Directed Randomization and Data Synthesis

Abstract

Randomized trials are one of society's most powerful tools for testing novel therapeutics. There are both ethical imperatives and economic incentives to ensure that their design and analysis is as efficient as possible. This dissertation focuses on trials that test multiple treatments and those that use external or shared data during analysis. In Chapter 1, the focus is on factorial trials that test treatment combinations. We present a novel Bayes-adaptive randomization design for factorial trials with binary outcomes. Our approach enables the investigator to specify a utility function representative of the aims of the trial and the Bayesian response-adaptive randomization algorithm aims to maximize this utility function. We explore several utility functions and corresponding factorial designs, and conduct a simulation study to illustrate relevant differences in key operating characteristics between the resulting designs. We also investigate the asymptotic behavior of the proposed designs. We then use data summaries from three recent factorial trials in perioperative care, smoking cessation and infectious disease prevention to define realistic simulation scenarios and illustrate advantages of the introduced trial designs compared to designs from the previous literature. In Chapter 2, we switch our attention to trials that use external patient-level data to augment the final analysis dataset. We describe and formalize the key sources of distortion that arise in this setting. We then review key methods that have been proposed to analyze randomized trials that use external data. Using simulations we conduct a comparative analysis of these methods, in the presence of different distortions in the data. We then use a glioblastoma data collection to generate a realistic trial and external datasets. We conduct our comparative analysis of methods in this realistic setting, in order to provide recommendations for future glioblastoma trials. Chapter 3 considers why randomized trials do not currently use external datasets from both a statistical and logistic perspective. We suggest a trial strategy that overcomes many of the key difficulties associated with data sharing. We use simulations and a data collection from glioblastoma to quantify the benefits of our strategy compared to conventional approaches, illustrating the benefit in terms of patients and time saved.