Variable selection in modelling clustered data via within-cluster resampling

Area of Science:

• Biostatistics
• Health Services Research
• Data Science

Background:

• Risk-adjustment models are essential in biomedical applications but face challenges with clustered, high-dimensional data.
• Existing variable selection methods are inadequate for discrete clustered data with numerous variables and large clusters.

Purpose of the Study:

• To develop and evaluate a new variable selection approach for high-dimensional clustered data.
• To address the lack of suitable methods for selecting variables in complex biomedical datasets.

Main Methods:

• A novel approach combining within-cluster resampling with penalized likelihood methods was developed.
• Theoretical properties, including an upper bound on false selections, were derived.
• Extensive simulations were used to assess finite sample performance.

Main Results:

• The proposed method demonstrates oracle properties, indicating effective variable selection.
• Simulations confirmed the method's performance in practical scenarios.
• The approach was successfully applied to a large colon surgical site infection dataset.

Conclusions:

• The new variable selection technique is effective for high-dimensional clustered data in biomedical research.
• This method enhances the development of accurate risk-adjustment models by accounting for complex data structures and interactions.