Treed Variance

Abstract

This article proposes a data-driven tree method, called “treed variance” (TV), to model heteroscedasticity in linear regression. Specifically, we use a score test statistic to recursively bisect data into heterogenous groups, and then adopt the pruning methodology of CART to determine the best tree size. The proposed method provides not only a piecewise constant modeling of the error variance, but also facilitates a natural check of homoscedasticity. We assess the performance of the TV method via simulation studies and illustrate its use with an empirical example.