Regression-based variable clustering for data reduction

Area of Science:

• Biostatistics
• Geographic Health Analysis
• Statistical Modeling

Background:

• Clustering small geographic regions is crucial for accurate disease rate estimation and concise summarization.
• Many regions have insufficient cases for reliable individual analysis, necessitating aggregation.
• Existing methods may not adequately account for subject-specific data or confounding variables.

Purpose of the Study:

• To develop and evaluate a new subject-specific clustering algorithm for geographic regions.
• To improve the accuracy and parsimony of disease rate estimates by creating larger, more informative areas.
• To incorporate confounding variables directly into the clustering process.

Main Methods:

• A novel clustering algorithm is proposed, formulated within a regression framework.
• The algorithm utilizes subject-specific data to cluster original regions into larger areas.
• Confounding variables are integrated during the clustering process, not applied post-hoc.

Main Results:

• Simulation studies indicate promising statistical properties and performance of the proposed algorithm.
• The method effectively clusters regions based on disease rates and subject-level associations.
• Controlling for confounding variables during clustering is shown to be critical for accurate results.

Conclusions:

• The developed regression-based clustering algorithm offers a robust approach for analyzing geographic health data.
• The method facilitates improved disease rate estimation and parsimonious representation of health patterns.
• Its applicability extends to various problems requiring geographic data aggregation and analysis, as demonstrated with Cardiovascular Health Study data.