Discover Contemporaneous Connections in Dynamic Network: Regularized Unified Structural Equation Modeling as a Hybrid Vector Autoregressive Model

• In Event: Innovative Methods in Structural Equation Modeling

Abstract

The availability of intensive longitudinal data draws increasing interest in subject-specific dynamic networks using vector autoregressive (VAR) modeling. However, due to identification and computation challenge, current VAR representations are restricted in how contemporaneous connections are incorporated (bidirectional-only or unidirectional-only), preventing the discovery of accurate dynamic relations. We demonstrated that traditional representations are special cases of a more general form of VAR, where hybrid contemporaneous connections are possible and recovered via a data-driven approach. Our simulation study compared VARs using various estimation regimes and revealed that the hybrid-VAR with LASSO-regularization performed optimally under varying conditions in recovering connections with precise directionality and removing false ones of any kind. This is essential to infer reliable Granger casual interactions in dynamic networks.

Authors

• Awella Ye, University of North Carolina - Chapel Hill

• Kathleen M. Gates, University of North Carolina - Chapel Hill

• Teague Henry, University of North Carolina - Chapel Hill