Bivariate Spatial Analysis of Birth Weight and Gestational Age
*Brian Neelon, Duke University 

Keywords: Bivariate Conditionally Autoregressive (CAR) Prior, Finite Mixture Model, Bivariate Probit Model, Spatial Analysis, Birth Outcomes

We propose a set of bivariate spatial models for the joint analysis of birth weight and gestational age – two highly dependent pregnancy outcomes. For continuous measures of birth weight and gestational age, we develop a bivariate mixture model that addresses the potential non-normality of the outcomes. The model is expressed as a novel finite mixture of Guassian Markov random fields. For categorical measures, such as low birth weight and preterm birth, we propose an analogous spatial bivariate probit model. The models have a hierarchical structure that incorporates individual- and area-level predictors as well as random effects for each spatial unit. The random effects are assigned conditional autoregressive (CAR) priors that allow the shape of the bivariate distributions to change in flexible ways across areal units. For posterior computation, we develop an efficient Markov chain Monte Carlo (MCMC) algorithm that relies on full-conditional Gibbs and Metropolis steps. We apply the models to data from the 2008 North Carolina Detailed Birth Record.