MacDATA Seminar: Discovering Structure in Multiple Outcomes Models for Environmental Exposure Effects by Dr. Tanzy Love

April 23rd 2019
L.R. Wilson Hall 1003
3:00pm to 4:30pm

Bayesian model-based clustering provides a powerful and flexible tool that can be incorporated into regression models to explore several different questions related to the grouping of observations. In our application, we explore the effects of prenatal methylmercury exposure on childhood neurodevelopment. Rather than cluster individual subjects, we cluster test outcomes within a multiple outcomes model to improve estimation of the exposure effect and the model fit diagnostics. By using information on both exposures in the data to nest the outcomes into groups called domains, the model more accurately reflects the shared characteristics of neurodevelopmental domains. The paradigm allows for sampling from the posterior distribution of the grouping parameters; thus, inference can be made on group membership and their defining characteristics. We avoid the often difficult requirement of a priori identification of the total number of groups by incorporating a Dirichlet process prior. In doing so, we estimate exposure effects on neurodevelopment by shrinking effects within and between the domains selected by the data.