Multiple outcomes and multivariate process data are collected routinely in HIV and AIDS research. Examples include multiple markers of disease state, such as CD4 cell counts and HIV-RNA Levels (viral load); multiple endpoints for disease progression, such as change in CD4 and time to AIDS progression; bivariate treatment-response processes, such as timevarying antiretroviral therapy and viral load response; or several measures of an underlying construct, such as multiple scales for assessing neurocognitive functioning. Moreover, in long-term studies such as natural history studies, dropout (attrition) can be considered as a process unto itself, and adjusting analyses of the primary endpoint for potential selection biases frequently requires framing the estimation procedure in terms of the joint distribution of the endpoint and the dropout mechanism. There exists a wide range of statistical tools for analyzing one outcome or process at a time, such as models for event histories (e.g. Cox proportional hazards model) and repeated measures (e.g. random effects models or generalized estimating equations), but far fewer that allow an integrated analysis of several outcomes simultaneously. The primary goal of this project is to develop and disseminate new biostatistical methods that will enable researchers in HIV and AIDS research to address, in meaningful and interpretable ways, centrally important questions from studies that generate complex arrays of outcomes. This objective will be met in three ways. First, statistical models for classifying HIV disease stage based on the joint evolution of CD4 cell count and plasma viral load will be developed: this includes deriving a univariate measure of progression risk, and empirical classifications of HIV stage. The latter will be developed in the context of latent class models. The classification methods will be applied to other settings where multiple indicators are encountered, such as longitudinal patterns of depression and multiple indicators of neurocognive ability. Second, state-of-the-art methods for causal inference will be compared and applied for studying longitudinal effects of highly-active antiretroviral therapy (HAART) regimens on several aspects of HIV natural history, including variations in CD4 and viral load, health services utilization, and distribution of body mass. Third, a highly flexible class of mixture models for estimating covariate effects from longitudinal data with outcome-related dropout will be developed. These models are designed to have transparent assumptions about dropout, and allow sensitivity analyses for inspecting the possible range of selection bias. The research on new statistical methodology has been motivated by analytic issues that arise in longitudinal cohort studies in HIV and AIDS. As such, we will use our methods to address key questions from three studies of HIV natural history: HERS, a cohort of 1300 women followed for seven years; ALIVE, a cohort study of 3000 intravenous drug users; and the Nutrition for Healthy Living Study, a cohort study of about 700 in New England.
Showing the most recent 10 out of 13 publications
