Novel Methods for the Design, Monitoring, and Analysis of Clinical Trials

Abstract

Clinical trials can be divided into three phases conducted before, during, and after the trial: 1) trial design; 2) data monitoring; and 3) statistical analysis, respectfully. The scientific accuracy of any clinical trial is inherently tied to the validity of the statistical assumptions made in each phase. I present methods that seek to both measure and account for violations of such statistical assumptions. For 1), a novel ordinal endpoint was recently proposed for evaluating new treatments for patients hospitalized by influenza. I investigate the power of the ordinal endpoint under model misspecification and compare it to the power of other endpoints derived from the same information. For 2), in some trials participants are not obligated to comply with the treatment protocol, and compliance may be unobserved. Biomarker measurements, however, can provide objective information on compliance. I propose a novel method that models a longitudinal biomarker as a longitudinal mixture density to estimate various probabilities of compliance. For 3), in completed trials participant noncompliance may create a difference between the intention-to-treat estimate and the treatment effect if all participants had complied (i.e., the causal effect), as noncompliant participants may systematically differ from compliant participants on some confounding variable. The G-computation algorithm can estimate the causal effect for longitudinal trials, but typically assumes that both compliance is observed and that the confounder models can be correctly specified. I propose a modified G-computation algorithm estimator in the scenario where both assumptions are violated.