Fast Bayesian Estimation of Spatial Count Data Models

摘要: Spatial count data models are used to explain and predict the frequency of phenomena such as traffic accidents in geographically distinct entities census tracts or road segments. These typically estimated using Bayesian Markov chain Monte Carlo (MCMC) simulation methods, which, however, computationally expensive do not scale well large datasets. Variational Bayes (VB), a method from machine learning, addresses shortcomings MCMC by casting estimation an optimisation problem instead problem. In this paper, we derive VB for posterior inference negative binomial with unobserved parameter heterogeneity spatial dependence. The proposed uses Polya-Gamma augmentation deal non-conjugacy likelihood integrated non-factorised specification variational distribution capture dependencies. We demonstrate benefits approach simulated real on youth pedestrian injury counts New York City boroughs Bronx Manhattan. empirical analysis suggests that is between 7 13 times faster than regular eight-core processor, while offering similar predictive accuracy. Conditional availability computational resources, embarrassingly parallel architecture can be exploited further accelerate up 100 times.