Title:Two-Stage Penalized Least Squares Method for Constructing Large Systems of Structural Equations

Abstract:Linear systems of structural equations have been recently investigated to reveal the structures of genome-wide gene interactions in biological systems. However, building such a system usually involves a huge number of endogenous variables and even more exogenous variables, and hence demands a powerful statistical method which limits memory consumption and avoids intensive computation. We propose a two-stage penalized least squares method to build large systems of structural equations. Fitting one linear model for each endogenous variable at each stage, the method employs the $L_2$ penalty at the first stage to obtain consistent estimation of a set of well-defined surrogate variables, and the $L_1$ penalty at the second stage to consistently select regulatory endogenous variables among massive candidates. Without fitting a full information model, the method is computationally fast and allows for parallel implementation. The resultant estimates of the regulatory effects enjoy the oracle properties, that is, they perform as well as if the true regulatory endogenous variables were known. We also demonstrated the effectiveness of the method by conducting simulation studies, showing its improvements over other methods. Our method was applied to construct a yeast gene regulatory network with a genetical genomics data.