Quantile regression shrinkage and selection via the Lqsso.

Area of Science:

• Statistics
• Machine Learning
• Econometrics

Background:

• Quantile regression models are increasingly sophisticated, with ongoing research into novel penalty and loss functions.
• Existing lasso quantile regression models may overlook randomness in penalty components.
• Adaptive lasso quantile regression models perform simultaneous estimation and variable selection.

Purpose of the Study:

• To introduce a novel penalty for quantile regression models that accounts for randomness.
• To develop the least quantile shrinkage and selection operator quantile regression (lqsso-QR) model.
• To establish conditions for consistent variable selection in lasso quantile regression and demonstrate oracle properties of lqsso-QR.

Main Methods:

• Proposing a new penalty for quantile regression, termed lqsso-QR.
• Deriving sufficient and necessary conditions for consistent variable selection in lasso quantile regression.
• Applying an efficient algorithm for computation, adapted from lasso quantile regression.
• Conducting simulation studies to compare lqsso-QR with other lasso-type penalties.

Main Results:

• The proposed lqsso-QR model demonstrates oracle properties under mild conditions.
• Simulation studies indicate the superiority of lqsso-QR, particularly in reducing relative prediction error.
• The method was successfully applied to the real-life rat eye dataset.

Conclusions:

• The lqsso-QR model offers an effective approach for simultaneous estimation and variable selection in quantile regression.
• The proposed method outperforms existing lasso-type penalties in predictive accuracy.
• lqsso-QR provides a valuable advancement in statistical modeling for complex datasets.