Bayesian Hierarchical Mixture Models for High-Risk Births in California, 1968-2005

James H. Jones, Stanford University
Simon D. Jackman, Stanford University

Birthweight shows complex patterns of heterogeneity and has strong implications for infant mortality and later-life demographic outcomes. Using NCHS registration data from 1968-2005, we model the distribution of birthweight in California as a two-component Gaussian mixture. The mixture has an intuitive interpretation: the first component represents the majority of the population and the second component represents a high-risk sub-population with lower mean birthweight and higher variance. Using a Bayesian framework, we estimate the joint posterior distribution of the mixture model via MCMC simulation. The flexibility afforded by fitting the mixture model by the Gibbs sampler allows us to model the (binary) indicator for component membership as a function of covariates. Our interest focuses primarily in mother's and father's race, their interaction, and proxies of SES available from birth certificate information. Results highlight the continuing importance of race as a predictor of birth outcomes in the United States.