Interpreting posterior relative risk estimates in disease-mapping studies

Area of Science:

• Spatial epidemiology
• Biostatistics
• Geographic Information Systems (GIS) in health

Background:

• Small-area health outcome analysis requires sophisticated Bayesian models to smooth sparse data.
• Concerns exist regarding model performance, prior influence, and smoothing levels in risk estimation.

Purpose of the Study:

• To evaluate the performance of Bayesian disease-mapping models in recovering true health risk surfaces.
• To assess the impact of prior structures and smoothing on risk estimates.
• To compare different Bayesian models and propose improved detection methods.

Main Methods:

• Comprehensive simulation study of Bayesian disease-mapping models.
• Comparison of conditional autoregressive (CAR) and semiparametric spatial mixture models.
• Evaluation of model specificity and sensitivity with varying data sparsity and risk levels.
• Development and testing of a decision rule based on posterior probability of relative risk.

Main Results:

• Bayesian models are conservative with high specificity but low sensitivity for moderate excess risks (<2-fold) or low expected counts (<50).
• Semiparametric models smooth less than CAR models when data provide sufficient information (moderate counts or high risk).
• A decision rule using posterior probability (70-80% cutoff) offers reasonable sensitivity for moderate excess risks (1.5-2 fold) and moderate counts (~20).

Conclusions:

• Bayesian disease-mapping models require careful consideration of prior structure and smoothing.
• Exploiting the full posterior distribution enhances sensitivity for detecting true raised-risk areas.
• The proposed probability-based decision rule improves the detection of moderate excess risks in spatial health analyses.