ScholarWorks@성균관대학교

초록

The penalized least squares and maximum likelihood methods have been successfully employed for simultaneous parameter estimation and variable selection. However, outlying observations can severely affect the quality of the estimator and selection performance. Although some robust methods for variable selection have been proposed in the literature, they often lose substantial efficiency. This is primarily attributed to the excessive dependence on choosing additional tuning parameters or modifying the original objective functions as tools to enhance robustness. In response to these challenges, we use a nonparametric Gaussian scale mixture distribution for the regression error distribution. This approach allows the error distributions in the model to achieve great flexibility and provides data-adaptive robustness. Our proposed estimator exhibits desirable theoretical properties, including sparsity and oracle properties. In the estimation process, we employ a combination of expectation-maximization and gradient-based algorithms for the parametric and nonparametric components, respectively. Through comprehensive numerical studies, encompassing simulation studies and real data analysis, we substantiate the robust performance of the proposed method. © 2025 Elsevier B.V.

키워드