Estimation of graphical models for skew continuous data

Abstract

We consider a new approach for estimating non-Gaussian undirected graphical models. Specifically, we model continuous data from a class of multivariate skewed distributions, whose conditional dependence structure depends on both a precision matrix and a shape vector. To estimate the graph, we propose a novel estimation method based on nodewise regression: we first fit a linear model, and then fit a one component projection pursuit regression model to the residuals obtained from the linear model, and finally threshold appropriate quantities. Theoretically, we establish error bounds for each nodewise regression and prove the consistency of the estimated graph when the number of variables diverges with the sample size. Simulation results demonstrate the strong finite sample performance of our new method over existing methods for estimating Gaussian and non-Gaussian graphical models. Finally, we demonstrate an application of the proposed method on observations of physicochemical properties of wine.