Comparison of different software implementations for spatial disease mapping

Disease mapping is a scientific field that aims to understand and predict disease risk based on counts of observed cases within small regions of a study area of interest. Hierarchical model-based approaches that borrow information from neighbouring areas via conditional autoregressive (CAR) random effects on the local disease rates have gained a lot of popularity, thanks to the readily implemented Markov chain Monte Carlo methods. Nowadays, many software implementations to model risk distributions exist. Many of these applications differ, to varying degrees, in the underlying methodology. This paper provides an in-depth comparison between analysis results, coming from R-packages CARBayes, R2OpenBUGS, NIMBLE, R2BayesX, R-INLA, and RStan. We investigate CAR models typically used in disease mapping for spatially discrete count data. Data about diabetics in children and young adults in Belgium are used in a case study, while simulation studies are undertaken to assess software performance in different settings.