Year: 2006
Type: report
Abstract: We study the variable selection problem in generalized linear models, and propose a penalized likelihood approach for joint variable selection and parameter estimation.Unlike many shrinkage methods such as the LASSO, the new method advocates different penalties for different coefficients: unimportant covariates receive larger penalties than important ones.Consequently, the coefficients of important variables can be efficiently estimated while those of unimportant ones are more shrunk to zeros.We show that the proposed penalized likelihood estimator has desired theoretical properties including root-n consistency and oracle properties.Another advantage of the new procedure lies in its easy implementation; one efficient algorithm is provided in the paper.We illustrate performance of the proposed procedure through numerous simulations, and then apply it to the Wisconsin Epidemiological Study of Diabetic Retinopathy.
Authors: Hao Helen Zhang, Wenbin Lu
Language: English
Subfield: Statistics and Probability
Field: Mathematics
Domain: Physical Sciences
Open Access status: green
Funder: National Science Foundation
OpenAlex ID: W2528081019