Assessing the impact of neighborhood structures in Bayesian disease mapping

Area of Science:

• Spatial statistics
• Epidemiology
• Computational statistics

Background:

• Defining neighborhood structure is critical for Bayesian disease mapping using conditional auto-regressive (CAR) models.
• Little research exists on how neighborhood structures impact model performance with fine-scale data.

Purpose of the Study:

• To assess the effect of different neighborhood structures on CAR model performance for fine-scale disease mapping.
• To compare neighborhood structures using COVID-19 mortality data in Limburg, Belgium.

Main Methods:

• Modeled 2020 COVID-19 mortality vs. pre-pandemic rates in small areas using BYM and BYM2 models.
• Implemented three queen-neighborhood structures (up to fifth-order) and two weight schemes.
• Conducted a simulation study to evaluate spatial correlation reproduction and compared models using WAIC, goodness-of-fit, parameter estimates, and computation time.

Main Results:

• Order-based weight matrices outperformed binary matrices.
• First-order neighborhood structures showed comparable performance to higher-order structures but with significantly less computation time.
• The BYM model was more sensitive to neighborhood structure choice than the BYM2 model.

Conclusions:

• Higher-order neighborhood matrices offer minimal advantages in Bayesian disease mapping with fine-scale data.
• A simple first-order neighborhood structure is a pragmatic and suitable choice for CAR models in this context.