Handling Problematic Between-person Estimates in Panel Network Models: A Comparative Simulation Study

A recurring challenge in the estimation of the graphical Vector Autoregressive (GVAR) model from panel data lies in the between-person network structure that sometimes contains unrealistic estimates (i.e., saturated edge weights of -1 or 1) or results in entirely empty networks. Different solutions have been proposed to overcome this challenge through 1) the use of a saturated Gaussian Graphical Model (GGMs) that can be estimated with Cholesky decomposition for modeling between-person covariances, 2) the standard use of thresholded GGMs for between-person covariances, or 3) using empty GGMs that only model between-person variances in a diagonal matrix approach. We use a simulation approach to evaluate the performance of these three approaches across different conditions, including varying sample sizes and varying number of assessment waves. Our simulation study reproduced the problematic between-person estimates and showed that saturated GGM estimation (with Cholesky decomposition) yields credible contemporaneous and temporal estimates, corroborating its use in prior empirical research. Not modeling between-person covariances entirely resulted in low specificity. This study highlights the importance of data volume, showing that larger sample sizes or increased assessment waves significantly improve model performance. These insights may inform study design and data collection strategies and provide guidelines for selecting estimation methods.