Optimized variable selection via repeated data splitting.

Area of Science:

• Statistics
• Machine Learning
• Biostatistics

Background:

• Model selection in high-dimensional data is well-studied.
• Advancements in low-dimensional regression variable selection are lacking.
• Need for robust variable selection in moderate-scale datasets (n>p).

Purpose of the Study:

• Introduce a new variable selection procedure for low to moderate scale regressions.
• Provide an empirically optimized threshold for variable screening.
• Evaluate the performance of the proposed method against existing techniques.

Main Methods:

• Repeatedly split data into estimation and validation sets.
• Empirically optimize a threshold for variable inclusion.
• Compare performance with backward elimination, univariate screening, adaptive LASSO, and SCAD.

Main Results:

• The proposed method demonstrates superior performance in simulations.
• Achieves low inclusion of noisy predictors and high power for correct model detection.
• Performance is unaffected by correlations among predictors.

Conclusions:

• The new variable selection procedure is effective for low to moderate scale regressions.
• Offers a robust alternative to existing methods, especially when predictors are correlated.
• Successfully applied to a clinical dataset of hepatectomy patients.