Online Program

Hierarchical Models and Computing for Joint Longitudinal-Survival and other Multiple Component or Endpoint Data

*Bradley P. Carlin, University of Minnesota 
*Laura Hatfield, Harvard 

Keywords: Bayesian methods, joint modeling, latent variables, MCMC, CER

Hierarchical modeling is well-known for its ability to properly account for all sources of uncertainty and correlation in complex, high-dimensional data sets. The BUGS language is especially adept at implementing such an approach, since model components may be developed independently and then assembled into arbitrarily complex models. Perhaps the most useful example of this approach is in joint modeling of longitudinal and survival data, where relatively simple survival and longitudinal model components may be connected using latent variables that induce appropriate within-subject correlation. In this workshop, we describe accessible methods and software for handling this problem, as well as a variety of other settings where we also seek to link multiple components in a single larger model, and thus capture complex relationships among variables of different types. For example, we will describe the joint modeling of an exposure and corresponding multiple outcomes, necessary when the exposure is not directly measured and must instead be modeled, is measured with error, or when the outcomes themselves have important relationships (say, when one informatively censors another). We also describe methods for multiple treatment comparisons (meta-analysis) and its application in comparative effectiveness research (CER) when different treatments emerge as better for different outcomes (say, safety versus efficacy). Our presentation will include methods appropriate for carefully designed clinical studies, as well as approaches suited for observational data, including post-marketing surveillance studies of drugs and medical devices.