Marginalized LASSO in the low-dimensional difference-based partially linear model for variable selection.

Area of Science:

• Statistics
• Econometrics

Background:

• Partially linear models are useful for data with both linear and nonlinear predictors.
• Variable selection in difference-based methods is challenging due to low-variance predictors.

Purpose of the Study:

• To develop a novel methodology for variable selection in difference-based partially linear models.
• To address challenges posed by low-variance predictors in low-dimensional settings.

Main Methods:

• Proposed a marginalized LASSO estimator with a modified penalty term.
• Conducted comprehensive simulation experiments for small sample performance analysis.
• Utilized a bootstrapped method for prediction evaluation on the King House dataset.

Main Results:

• The proposed method demonstrates superior performance in estimation and prediction compared to standard LASSO.
• Effectively handles mixed linear and nonlinear relationships in data.
• Shows robustness in low-dimensional setups with low-variance predictors.

Conclusions:

• The novel marginalized LASSO approach offers an effective solution for variable selection in difference-based partially linear models.
• Provides a reliable tool for analyzing complex datasets with mixed predictor types.
• Validated through simulations and a real-world housing dataset analysis.