Predictive Partly Conditional Models for Longitudinal Ordinal Outcomes with Applications to Alzheimer's Disease

Abstract

There are a variety of methods for modeling disease progression. The choice of which model to use should be motivated by the question of interest and the disease pathology. In Alzheimer’s disease (AD), which is progressive, much of the focus has been on short-term transitions or relationships between baseline values of predictors and the final disease stage of the patient. In this dissertation, we examine the predictive partly conditional model (PPCM), a flexible model that can be used to answer a variety of questions, while allowing for time-dependent covariates and time-varying effects. We first examine the statistical properties of the method, including the conditions necessary for consistent unbiased estimates. These conditions include the use of an independence covariance matrix and a fixed time window between the measurement of predictors and outcomes. We then extend the method to incorporate potentially informative death and dropout, often present in aging studies, using inverse probability weighting. We show the additional conditions necessary for unbiased and consistent estimation, namely the assumption of independence between the probabilities of death and study dropout. Finally, we conduct a case study examining three questions of interest related to the relationship between everyday cognition and disease progression using data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI). Overall, we found patient-assessed cognitive scores become less reliable as a predictor of progression once a patient progresses along the AD pathway, while the study-partner assessment of cognition retains a strong relationship with disease progression regardless of the disease stage of the patient.