A Flexible Zero-Inflated Conway--Maxwell--Poisson Regression Model for Spatiotemporal Data of US Vaccine Refusal

Vaccination is widely acknowledged as one of the most effective tools for preventing disease. However, there has been a rise in parental refusal and delay of childhood vaccination in recent years in the United States. This trend undermines the maintenance of herd immunity and elevates the likelihood of outbreaks of vaccine-preventable diseases. Our aim is to identify demographic or socioeconomic characteristics associated with vaccine refusal, which could help public health professionals and medical providers develop interventions targeted to concerned parents. We examine US county-level vaccine refusal data for patients under five years of age collected on a monthly basis during the period 2012--2015. These data exhibit challenging features: zero inflation, spatial dependence, seasonal variation, spatially-varying dispersion, and a large sample size (approximately 3,000 counties per month). We propose a flexible zero-inflated Conway--Maxwell--Poisson (ZICOMP) regression model that addresses these challenges. Because ZICOMP models have an intractable normalizing function, it is challenging to do Bayesian inference for these models. We propose a new hybrid Monte Carlo algorithm that permits efficient sampling and provides asymptotically exact estimates of model parameters.