UC Irvine

Randomized controlled clinical trials (RCTs) serve as the gold standard to determine whether a candidate treatment has a favorable benefit-to-risk ratio for a pre-specified target patient population. Enrichment strategies are commonly employed in RCTs to identify the appropriate target population of patients who would likely benefit from a candidate treatment (predictive) and/or have the outcome of interest during the course of the trial (prognostic), including enriching based upon amyloid beta and tau protein levels in Alzheimer’s disease (U.S. Food and Drug Admministration, 2019). Currently, there is a gap in the understanding of RCTs using enrichment and adaptations to the randomized treatment assignment allocations (response-adaptive randomization), especially under model misspecification (violation of assumptions) for fixed sample and adaptive RCTs with a repeated measures (e.g., changes in activities of daily living scores) or a censored time-to-event (e.g., time to dementia) primary outcome. In this dissertation, we focus on valid estimation of estimands (a contrast of summary measures between treatment arms for an appropriate target population) from enriched and adaptively randomized clinical trials. We consider the trial-specific RCT estimand (RCT-E) and a real world estimand (RW-E) for a broader patient population for whom off-label use may be a possibility. Aim 1 quantifies the impact of enrichment in fixed sample pre-post (only two assessments; one pre- and one post-randomization) RCTs. We show that the application of standard statistical methods, such as the analysis of covariance (ANCOVA) model, yield biased estimates for the RW-E. We propose a novel bias-adjusted estimator of the RCT-E to estimate the RW-E based on an analytic derivation under model misspecification in the multivariate normal data setting. Aim 2 focuses on reliably estimating the RCT-E in fixed sample adaptively randomized time-to-event RCTs that allows for enhanced replicability in the presence of time-varying treatment effects. We propose a novel adaptive randomization censoring-robust estimator that reweights the partial likelihood score a la Boyd et al. (2012) that accounts for differential censoring patterns resulting from adaptive randomization and by incorporating the randomization scheme in the re-weighting, we gain efficiency. Importantly, our proposed estimator consistently estimates a standardized marginal hazard ratio for the RCT-E. Finally, in Aim 3 we examine how to prospectively design and plan for the monitoring of time-to-event group sequential designs that warrant using censoring-robust estimators when targeting a RCT-E. We show how the statistical information of our proposed adaptive randomization censoring-robust estimator is non-linear and has non-independent increments, thus requiring appropriate modifications to the planned timing of interim analyses and the final boundary to maintain statistical operating characteristics and scientific objectives of such trials. Overall, the statistical contributions of this research will aid in the design, conduct, and analysis of enriched and adaptively randomized clinical trials to support efforts during drug development, regulatory review, and clinical decision-making post approval.