This project will develop statistical methods for the joint analysis of time-to-event and repeated measures data collected serially in longitudinal studies. Such data are increasingly common in many settings in medicine and public health, including clinical trials designed to evaluate palliative, maintenance, or preventative therapies, in evaluation of prevention interventions, and in population studies of the effects of risk factors on the development of disease. Our methodology will be applied to clinical trial data for AIDS, schizophrenia, and contraception therapies, and to a longitudinal epidemiological study of obesity in childhood. There are two main components to the project, both of them involving the joint analysis of repeated measures and time-to-event data. The first component will focus on the case where primary interest is in describing trends in a categorical outcome over time in the presence of non-ignorable dropout. This is common in clinical trials where removal from treatment may be related to the outcome of interest and causes follow-up to cease. A new feature of this work will be the use of a mixture model to incorporate information on dropout. Mixture models offer two important advantages: they can be easily implemented and it is straightforward to characterize model assumptions. We will develop methods based on Generalized Estimating Equations (GEE) for the case where any intermediate missingness prior to dropout is Missing Completely at Random, and extend the method to del with arbitrary patterns of non-ignorable non-response. We will also develop maximum likelihood approaches using both marginal and Markov models for the case where intermediate missingness is Missing at Random. Second, we will focus on the case where the primary interest is survival and where the repeated measures are used to gain efficiency in the presence of censoring. This component will extend previous work on this problem by using mixture models for categorical repeated measures, by including covariates for the survival model with the proportional hazards assumption, and by using maximum likelihood to estimate the joint distribution of survival and repeated measures.
Showing the most recent 10 out of 76 publications
