GRASS-NB: Group-structured variable selection for spatial negative binomial data with applications to cancer registry...

Area of Science:

• Biostatistics
• Spatial Epidemiology
• Computational Biology

Background:

• Analyzing spatially structured, overdispersed count data with many predictors is common in epidemiology and spatial omics.
• Effective feature selection is crucial for identifying risk factors or biomarkers in such datasets.
• Existing Bayesian negative binomial regression models with variable selection priors often lack spatial considerations and group structure.

Purpose of the Study:

• To propose a flexible Bayesian negative binomial regression model for spatially autocorrelated count data.
• To introduce a novel group-structured prior by combining spike-and-slab and continuous horseshoe methods for enhanced feature selection.
• To evaluate the model's performance, including specificity, precision, and computational efficiency, particularly in high-dimensional settings.

Main Methods:

• Developed a Bayesian negative binomial regression framework incorporating spatial autocorrelation.
• Introduced a hybrid group-structured prior combining spike-and-slab and continuous horseshoe shrinkage.
• Evaluated model performance using simulations, including 'large p, small n' scenarios, and applied it to real-world datasets.

Main Results:

• The proposed model effectively handles spatial autocorrelation and high-dimensional predictors in count data.
• The novel group-structured prior demonstrated robust performance in feature selection.
• The model was successfully applied to identify cancer risk factors and predict gene expression in spatial omics data.

Conclusions:

• The developed Bayesian negative binomial regression model with a novel group-structured prior offers a flexible and powerful tool for analyzing complex spatial count data.
• This approach enhances the identification of influential predictors in fields ranging from population epidemiology to spatial omics.
• An R package is available to facilitate the application of this methodology.