A Gaussian random field model for similarity-based smoothing in Bayesian disease mapping

Conditionally specified Gaussian Markov random field (GMRF) models with adjacency- or distance-based neighbourhood weight matrix, commonly known as neighbourhood- based GMRF models, have been the mainstream approach to spatial smoothing in Bayesian Disease mapping (DM). In the present work, we propose a conditionally specified Gaussian random field (GRF) model with a similarity-based non-spatial weight matrix to facilitate non-spatial smoothing in Bayesian DM. The model, named similarity-based GRF, is motivated for modeling DM data in situations where the underlying small area relative risks and the associated determinant factors do not varying systematically in space, and the similarity is defined by “similarity” with respect to the associated disease determinant factors. The neighbourhood-based GMRF and the similarity-based GRF are compared and assessed via a simulation study and by two case studies, using new data on alcohol abuse in Portugal collected by the World Mental Health Survey Initiative (WMHSI) and the well-known lip cancer data in Scotland. In the presence of disease data with no evidence of positive spatial correlation, the simulation study showed a consistent gain in efficiency from the similarity-based GRF, compared with the adjacency-based GMRF with the determinant risk factors as covariate. This new approach broadens the scope of the existing Conditional autocorrelation (CAR) models.