Neighborhood dependence in Bayesian spatial models

The conditional autoregressive model and the intrinsic autoregressive model are widely used as prior distribution for random spatial effects in Bayesian models. Several authors have pointed out impractical or counterintuitive consequences on the prior covariance matrix or the posterior covariance matrix of the spatial random effects. This article clarifies many of these puzzling results. We show that the neighborhood graph structure, synthesized in eigenvalues and eigenvectors structure of a matrix associated with the adjacency matrix, determines most of the apparently anomalous behavior. We illustrate our conclusions with regular and irregular lattices including lines, grids, and lattices based on real maps.