Ben Rogers: Bayesian Covariance Modeling for Longitudinal Zero-Inflated Count Data
Abstract
We develop three classes of Bayesian two-part regression models for longitudinal zero-inflated data. To model the number of days of heavy drinking out of the previous 90 days we develop a longitudinal binomial hurdle model. We also develop a longitudinal Poisson hurdle model and a multivariate longitudinal zero-inflated Poisson model for modeling counts of medical visits with unbalanced observation times. Time varying random effects with a parameterized covariance matrix are used to model within-individual correlation over time and are shown to improve fit over standard models. Unbalanced observation times are accounted for using partially observed latent time effects and regression equation offsets. We model data coming from two randomized controlled trials, one measuring the effect of an intervention to reduce levels of alcohol abuse and the other measuring the effect of an intervention to improve linkage and engagement in medical care among previously incarcerated people with HIV.