Doubly regularized generalized linear models for spatial observations with high-dimensional covariates.

Area of Science:

• Statistics
• Spatial Analysis
• Machine Learning

Background:

• Spatial data often exhibits correlations across locations.
• High-dimensional datasets present challenges for traditional statistical models.
• Network structures can capture spatial relationships and feature similarities.

Purpose of the Study:

• To develop a novel doubly regularized regression framework for analyzing doubly-structured high-dimensional data.
• To incorporate both spatial network and feature network structures into statistical models.
• To improve predictive power and feature identification in spatial analysis.

Main Methods:

• Developed a doubly regularized regression framework.
• Utilized convex optimization algorithms for implementation.
• Proposed a procedure for asymptotically valid confidence intervals and hypothesis tests.

Main Results:

• The proposed framework demonstrated improved predictive accuracy and inferential power over existing methods.
• The method showed advantages even with partially misspecified or uninformative network structures.
• Empirical results confirmed the framework's effectiveness in high-dimensional spatial analysis.

Conclusions:

• The doubly regularized regression framework effectively integrates network structures for enhanced spatial data analysis.
• The method offers superior predictive and inferential capabilities compared to current high-dimensional spatial techniques.
• The framework shows promise for applications like disease mapping, as evidenced by COVID-19 mortality data analysis.