Error Variance Estimation in Ultrahigh-Dimensional Additive Models.

Area of Science:

• Statistics
• Machine Learning
• Econometrics

Background:

• Error variance estimation is vital for statistical inference in high-dimensional regression.
• Traditional methods like naive mean squared error can underestimate variance in sparse additive models due to spurious correlations.
• Nonparametric models exhibit higher susceptibility to spurious correlations than linear models.

Purpose of the Study:

• To address the underestimation of error variance in high-dimensional sparse additive models.
• To propose a novel, accurate error variance estimation method for ultrahigh-dimensional settings.
• To establish the theoretical properties and practical performance of the new estimation technique.

Main Methods:

• Investigated the asymptotic behavior of traditional mean squared errors.
• Developed a new error variance estimator by integrating sure independence screening and refitted cross-validation.
• Established the root-n consistency and asymptotic normality of the proposed estimator.

Main Results:

• Demonstrated that naive estimates significantly underestimate error variance in high-dimensional sparse additive models.
• The proposed method effectively corrects for spurious correlations, providing accurate variance estimation.
• Theoretical analysis confirmed the consistency and asymptotic normality of the new estimator.

Conclusions:

• The proposed error variance estimation method is accurate and theoretically sound for ultrahigh-dimensional sparse additive models.
• The integration of sure independence screening and refitted cross-validation offers a robust solution.
• Simulation studies and a real data example validate the practical utility of the methodology.