The focus of this grant is on development of new statistical methods that allow inferences to be drawn from the complex data that arise in AIDS clinical trials and other studies of HIV disease. Understanding the relationship between clinical progression and longitudinal covariates such as viral load has important implications for management of HIV disease. Current methods for this problem make assumptions that that may not always be realistic. The first specific aim is to develop new, semiparametric methods for this problem that do not require such assumptions. The second specific aim focuses on adjustment for baseline covariates in analysis of treatment differences for censored, time-to-event data in AIDS clinical trials to account for chance imbalance at randomization, e.g. using the proportional hazards model. This practice raises concerns of model misspecification and the effects of post hoccovariate selection. An alternative strategy that may be less sensitive to these issues, based on modeling propensity of treatment assignment as a function of covariates, is proposed. The methods are also relevant to observational data, e.g. from large cohort studies. Although HIV patients may participate in randomized clinical trials, their actual treatment involves a series of complex decisions and deviations from protocol. Intent-to-treat analyses compare the policies of using different treatments, but another question of clinical interest is to identify the best treatment decision strategies, e.g. at what at point and based on what clinical/laboratory information should patients switch therapies? The third specific aim is to develop methods for identifying optimal time-dependent treatment strategies using complex, longitudinal, follow-up data from AIDS clinical trials. The fourth specific aim is focused on new methods for analysis of time-to-event (""""""""failure"""""""") data where the cause of failure is missing for some subjects; these methods apply in studies with complex endpoints, e.g. """"""""failure"""""""" corresponds to two consecutive viral load measurements above detectable limits, and some patients with one such observation fail to appear for the second.
Showing the most recent 10 out of 24 publications
