Neighborhood dependence in Bayesian spatial models.

Area of Science:

• Spatial statistics
• Bayesian modeling
• Geostatistics

Background:

• Conditional autoregressive (CAR) and intrinsic autoregressive (IAR) models are common priors for spatial random effects in Bayesian analyses.
• Previous research has highlighted impractical or counterintuitive findings regarding the covariance matrices of these spatial effects.

Purpose of the Study:

• To elucidate the reasons behind seemingly anomalous behaviors in CAR and IAR models.
• To demonstrate the critical role of graph structure in understanding these models' properties.

Main Methods:

• Analysis of the eigenvalues and eigenvectors of matrices derived from the adjacency matrix of the neighborhood graph.
• Theoretical clarification of prior and posterior covariance matrix properties.

Main Results:

• The structure of the neighborhood graph, specifically its spectral properties (eigenvalues and eigenvectors), is shown to be the primary driver of unusual results.
• Demonstration that these graph-based properties govern the behavior of spatial random effects.

Conclusions:

• The apparent anomalies in CAR and IAR models are not inherent flaws but are systematically explained by the underlying graph topology.
• Understanding the spectral properties of the neighborhood graph is key to correctly interpreting and applying these spatial Bayesian models.