UC San Diego

The objective of this dissertation is to utilize statistical methods to obtain consistent estimates from longitudinal data with monotone dropouts. This dissertation is comprised of three main studies. In the first study, which aims to identify the heterogeneity of cognition profiles in probable Alzheimer’s disease (AD) and determine if cognitive profiles are systematically related to the clinical course and neuropathological features of the disease, we explored a comprehensive data set from the National Alzheimer’s Coordinating Center (NACC) and successfully classified AD patients into 80% ”typical” versus 20% ”atypical” profiles, across two independent cohorts and one subset of subjects with autopsy available. We found that the atypical cognition profile was associated with lower Braak stage at autopsy and slower cognitive decline.

Observing an increasing attrition rate after two years with apparent informative dropout in the first study, and being motivated by the informative dropout that is common in FDA regulated trials for AD, in the second study, we proposed a doubly robust imputation approach to adjust for dropout-related bias in longitudinal studies. We illustrated this approach with an application in a prodromal AD trial conducted by Alzheimer’s Disease Cooperative Study (ADCS), which is a major center for AD clinical trials. We believe the imputation approach we presented has the advantage of computational simplicity and transparency compared to existing approaches in the literature, and may be suitable for use in FDA-regulated trials and a variety of other applications.

As an extension to this topic, in the third study, we investigated the doubly robust imputation method within a pattern mixture model framework, in order to deal with sensitivity analysis for randomized trials under several missing-not-at-random scenarios. We applied the proposed approach to two ADCS trials with different stages of disease, and compared the approach with other well-known methods to evaluate the performance. This study supports that the doubly robustimputation method is a competitive method for handling longitudinal data with monotone dropout, and may be suitable for use in randomized trials to obtain valid estimates of treatment effect.