Bayesian methods for hurdle models.

Hurdle models are often presented as an alternative to zero-inflated models for count data with excess zeros. They consist of two parts: a binary model indicating a positive response (the “hurdle”) and a zero-truncated count model. One or both parts of the model can depend on covariates, which may or may not coincide. In this dissertation, we explore the Bayesian approach to these models in detail, focusing on prior structures. Many of the Bayesian hurdle models encountered in the literature fail to incorporate expert opinion into the prior structure. We consider how prior information can be elicited from experts and incorporated into the prior structure of a hurdle model with shared covariates through the use of conditional means priors. More specifically, we propose a prior structure that assumes an inherent functional relationship between the two parts of the model. Through simulations, we explore the potential gains, as well as the shortcomings, of the approach. We also consider a simulation algorithm for Bayesian sample size determination for such models. We illustrate the use of the new methods on data from a hypothetical sleep disorder study.