Testing Heteroskedasticity in High-Dimensional Linear Regression

Abstract

A new procedure that is based on the residuals of the Lasso is proposed for testing heteroskedasticity in high-dimensional linear regression, where the number of covariates can be larger than the sample size. The theoretical analysis demonstrates that the test statistic exhibits asymptotic normality under the null hypothesis of homoskedasticity, and the simulation results reveal that the proposed testing procedure obtains accurate empirical sizes and powers. Finally, the procedure is applied to real economic data.