New estimation approaches for graphical models with elastic net penalty

TL;DR

This paper introduces three elastic net-based estimators for Gaussian graphical models, demonstrating that a two-stage procedure outperforms others in accuracy and applying it to analyze economic sector dependencies during Covid-19.

Contribution

The paper proposes three novel elastic net penalized estimators for sparse precision matrix estimation in Gaussian graphical models, including a new two-stage approach.

Findings

The two-stage estimator outperforms others in accuracy.

Simulations confirm superior performance in network structure recovery.

Application reveals Covid-19 impact on economic sector dependencies.

Abstract

In the context of undirected Gaussian graphical models, we introduce three estimators based on elastic net penalty for the underlying dependence graph. Our goal is to estimate the sparse precision matrix, from which to retrieve both the underlying conditional dependence graph and the partial correlation graph. The first estimator is derived from the direct penalization of the precision matrix in the likelihood function, while the second from using conditional penalized regressions to estimate the precision matrix. Finally, the third estimator relies on a 2-stages procedure that estimates the edge set first and then the precision matrix elements. Through simulations we investigate the performances of the proposed methods on a large set of well-known network structures. Empirical results on simulated data show that the 2-stages procedure outperforms all other estimators both w.r.t. estimating the sparsity pattern in the graph and the edges' weights. Finally, using real-world data on US economic sectors, we estimate dependencies and show the impact of Covid-19 pandemic on the network strength.