Effects of regularization on the stability and behavioral relevance of VAR-based functional connectivity

Vector autoregressive modeling (VAR) is a powerful method for estimating dynamic functional connectivity (dFC) from resting-state functional magnetic resonance imaging (fMRI) data. However, VAR-based dFC can become unstable when only limited data are available. We compared multiple regularization approaches (PLS, PCA, Ridge, and Lasso) to examine whether they can stabilize VAR-based dFC without substantially compromising behaviorally relevant information. All regularization methods improved stability. Behavioral measure prediction improved and reached a level comparable to Pearson correlation-based FC, except Lasso. These findings suggest that regularized VAR-based dFC can provide behaviorally relevant information about temporal dynamics not represented in Pearson correlation-based FC.