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

Area of Science:

• Biostatistics
• Spatial Epidemiology
• Disease Mapping

Background:

• Traditional Bayesian disease mapping relies on neighborhood-based Gaussian Markov random field (GMRF) models for spatial smoothing.
• These models assume spatial autocorrelation based on adjacency, which may not always be appropriate for disease determinant factors.

Purpose of the Study:

• To introduce a novel conditionally specified Gaussian random field (GRF) model utilizing a similarity-based, non-spatial weight matrix.
• To enable non-spatial smoothing in Bayesian disease mapping, particularly when disease risks and determinants lack systematic spatial variation.

Main Methods:

• Development of the similarity-based GRF model, defining similarity based on shared disease determinant factors.
• Comparative analysis using a simulation study and two case studies (alcohol abuse in Portugal, lip cancer in Scotland).
• Evaluation against the standard adjacency-based GMRF model with covariates.

Main Results:

• The similarity-based GRF demonstrated a consistent gain in efficiency compared to the adjacency-based GMRF in simulations with no positive spatial correlation.
• The model effectively handles situations where disease risk factors do not vary systematically in space.
• Case studies validated the model's performance in real-world disease mapping scenarios.

Conclusions:

• The similarity-based GRF model provides a valuable alternative for non-spatial smoothing in Bayesian disease mapping.
• This approach expands the utility of conditional autocorrelation models, offering enhanced efficiency when spatial assumptions are not met.
• It broadens the scope of statistical tools for analyzing disease distribution and risk factors.