A simulation study of disaggregation regression for spatial disease mapping.

Area of Science:

• Spatial epidemiology
• Statistical modeling
• Geographic information systems

Background:

• Disaggregation regression is crucial for spatial disease mapping, predicting risk from aggregated data at fine scales.
• High-resolution covariate data and fine-scale process modeling aim to capture covariate-response relationships.
• Validating fine-scale predictions is challenging due to a lack of observed data at that resolution.

Purpose of the Study:

• To evaluate disaggregation regression performance on simulated data across various settings.
• To assess the impact of data volume, aggregation area size, and model misspecification on prediction accuracy.
• To investigate the utility of aggregate-level cross-validation for predicting fine-scale performance.

Main Methods:

• Disaggregation regression applied to simulated data under different scenarios.
• Comparison of fine-scale predictions against simulated ground truth.
• Analysis of performance variations based on data points, aggregated area sizes, and model misspecification levels.

Main Results:

• Predictive performance increased with more observations and smaller aggregated areas.
• Well-specified models yielded accurate fine-scale predictions even with limited data and large aggregation areas.
• Model misspecification degraded performance significantly for large areas but less so for smaller ones.
• Aggregate-level cross-validation showed moderate correlation with fine-scale predictive performance.

Conclusions:

• Disaggregation regression is effective for fine-scale disease risk prediction, with performance sensitive to data aggregation levels and model specification.
• Increasing data points and decreasing aggregated area size enhance prediction accuracy.
• Aggregate-level cross-validation offers a moderately reliable assessment of fine-scale predictive capabilities.