Estimating cross-validatory predictive p-values with integrated importance sampling for disease mapping models.

Area of Science:

• Biostatistics
• Epidemiology
• Spatial Analysis

Background:

• Disease mapping requires identifying regions with unusually high or low disease risk.
• Leave-one-out cross-validatory (LOOCV) model assessment is the standard for estimating predictive p-values to flag these divergent regions.
• Traditional LOOCV is computationally intensive due to repeated Markov chain Monte Carlo (MCMC) analyses.

Purpose of the Study:

• To introduce a novel, computationally efficient method for estimating LOOCV predictive p-values.
• To provide an alternative to traditional, time-consuming LOOCV in disease mapping.
• To accurately flag divergent disease risk regions.

Main Methods:

• Developed integrated importance sampling (iIS) to estimate LOOCV predictive p-values using existing MCMC samples.
• Integrated away latent variables associated with test observations without referencing the actual observation.
• Compared iIS against posterior predictive checking, ordinary importance sampling, and the ghosting method using two disease mapping datasets.

Main Results:

• The predictive p-values estimated using iIS were nearly identical to those from actual LOOCV.
• iIS demonstrated superior performance compared to three existing methods in disease mapping analyses.
• The proposed iIS method offers a computationally efficient and accurate approach for disease risk assessment.

Conclusions:

• Integrated importance sampling (iIS) provides an efficient and accurate method for estimating LOOCV predictive p-values in disease mapping.
• iIS significantly reduces the computational burden associated with traditional LOOCV.
• This method enhances the ability to identify and flag divergent disease risk regions effectively.