Modified adaptive group lasso for high-dimensional varying coefficient models

Abstract

This article focuses on variable selection for varying coefficient models in the case of the number of covariates being larger than the sample size. Combining B-spline basis function approximations with the modified adaptive group lasso, we establish selection consistency, convergence rate and asymptotic normality. Our contribution is that the marginal nonparametric estimates are used as weights of the adaptive group lasso. Simulation studies and two real data applications show that our method performs better than the method of Wei, Huang, and Li (2011).