Monitoring and Analysis of Cluster-Randomized Trials With Interval-Censored Endpoints

Abstract

Cluster-randomized trials (CRTs) have seen widespread use in such fields as health policy, hospital administration and management, and infectious disease epidemiology, where ethical, pragmatic, and scientific considerations make desirable the randomization of entire groups of observations. These randomization groups are not formed by investigators, and so are comprised of individuals who might reasonably share geographic or social bonds. This results in correlation within the final study sample. If the primary outcome is the time to some asymptomatic (or otherwise not directly observable) event, then these observations are also interval-censored: the exact timing of the event is known only up to the interval between study monitoring visits. Both clustering and interval censoring are associated with a loss of statistical information and study power. Thus the task of designing, monitoring, and analyzing CRTs with these features requires efficiently leveraging all available information while making as few assumptions as possible about the outcome process and underlying dependence structure. This dissertation addresses these challenges in two particular facets of CRT conduct: interim monitoring for study futility (Chapter 1), and final analysis via semiparametric regression methods (Chapters 2 and 3). In Chapter 1, we propose a flexible framework for conditional power estimation when outcomes are clustered and interval-censored; this represents the first interim monitoring method to directly account for both of these data structures. Chapters 2 and 3 then adopt techniques from the missing data literature in order to facilitate semiparametric estimation and inference for CRTs under cluster-conditional and marginal proportional hazards models, respectively.