Subsampling based variable selection for generalized linear models.

Area of Science:

• Statistics
• Machine Learning
• Computational Statistics

Background:

• Variable selection is crucial for building accurate generalized linear models.
• Existing methods may lack robustness or consistent performance across diverse datasets.
• Low-dimensional generalized linear models require specialized selection techniques.

Purpose of the Study:

• Introduce a novel, robust variable selection method for low-dimensional generalized linear models.
• Evaluate the performance of the new method against existing approaches.
• Provide theoretical guarantees for the proposed method.

Main Methods:

• Developed AIC OPTimization via STABility Selection (OPT-STABS).
• Employed data subsampling and Akaike's Information Criterion (AIC) minimization.
• Introduced methods for optimal variable selection cutoff determination.

Main Results:

• OPT-STABS demonstrated consistently strong performance across various simulation settings.
• The method proved competitive, outperforming many existing variable selection techniques.
• Asymptotic properties, including root-n consistency and asymptotic normality, were derived and proved.

Conclusions:

• OPT-STABS offers a reliable and effective variable selection strategy for generalized linear models.
• The method exhibits superior robustness and consistent performance compared to alternatives.
• The theoretical underpinnings support the practical application of OPT-STABS in statistical modeling.