Smoking and obesity are two of the leading precursors to adverse health outcomes such as cancer, heart disease, and type-II diabetes, and therefore are a focal point of clinical and public health research on behavior change. Many longitudinal clinical trials in behavioral medicine are designed to effect positive and healthy changes in related behaviors, for example by promoting smoking cessation, exercise, and sensible diets. The analysis and reporting of data from these trials can be complicated by missing data, subject noncompliance with therapy, and loss to follow up. Furthermore, researchers are frequently interested in outcomes beyond the usual simple summaries such as 'quit smoking' (yes/no) or 'average weight change.' Because trials are longitudinal, the data can give insight into the process of behavior change, and can be used to classify different patterns of behavior change. Moreover, the availability of longitudinal data provides the opportunity to develop predictions of eventual outcomes that can be used dynamically in the treatment of certain behaviors. This project will develop innovative approaches to the analysis of longitudinal data specifically for behavioral medicine trials. The first two components of the project will develop latent class and latent variable models for classifying patterns of behavior change in smoking cessation and weight change. For smoking cessation trials in particular, the usual approaches to latent class and random effects modeling are inadequate because the standard assumption that random effects are normal rarely is adequate, and the usual exchangeability assumptions are not typically met in practice.
The third aim of the study addresses the issue of informative missing data, presenting a unifying and coherent framework for pattern-mixture modeling and associated sensitivity analyses. This part of the proposed work will tie together several recent proposals for PM modeling of longitudinal data. Finally, our fourth aim will address the important issue of noncompliance by developing both instrumental variable (IV) and G-computation approaches to causal inference. All of our work will be carried out in a Bayesian framework, so a major part of the proposal will be concerned with developing and implementing appropriate algorithms for simulating posterior distributions of fairly complex models. All of the methods will be developed on and applied to data from recently completed or ongoing trials in behavioral medicine, particularly in smoking cessation and weight change, run by recognized leaders in the field. ? ?
Showing the most recent 10 out of 15 publications
