The use of mixture models for identifying high risks in disease mapping

Area of Science:

• Biostatistics
• Epidemiology
• Spatial Analysis

Background:

• Traditional disease mapping and mortality studies often use Poisson inference.
• Overdispersion, or extra variability, is common and typically addressed with random effects.
• Accurate identification of high-risk regions is crucial for public health interventions and epidemiological research.

Purpose of the Study:

• To compare the effectiveness of two computationally simple random effects models: a discrete mixture model and a Poisson-normal mixture model.
• To evaluate how well each model identifies regions with extreme mortality risks.
• To assess the impact of spatial autocorrelation on model performance.

Main Methods:

• Comparison of a non-parametric discrete mixture model with a Poisson-normal mixture model.
• The Poisson-normal mixture model was analyzed from both frequentist (empirical Bayes) and fully Bayesian perspectives.
• Model performance was evaluated using infant mortality data from British Columbia, breast cancer data from Sardinia, and lip cancer data from Scotland, along with a simulation study.

Main Results:

• Both discrete and normal mixture models effectively identify regions with high risks.
• Normal mixture models demonstrated superior performance, particularly when spatial autocorrelation was present in the data.
• The study confirmed the utility of these models in detecting extreme mortality risks across different datasets.

Conclusions:

• Discrete mixture models are capable of locating high-risk areas in disease mapping.
• Poisson-normal mixture models offer robust performance and are especially advantageous when spatial dependencies exist.
• These findings support the use of mixture models for precise risk estimation and targeted public health strategies.