Diagnosing multicollinearity in stochastic actor oriented models
Scott Duxbury · OSF Preprints (OSF Preprints) · 2018
Stochastic actor-oriented models (SAOM) have become a staple approach for longitudinal network analysis. However, SAOM’s tolerance to correlations is poorly understood and, consequently, multicollinearity often goes undiagnosed in empirical research. Multicollinearity may prevent models from converging and problematize the estimation procedure, inhibiting inference from parameters. This study proposes a measure to detect multicollinearity in SAOM and evaluates SAOM’s tolerance to correlations. It uses a Monte Carlo experiment to evaluate the association between the correlation in the SAOM parameter set and model instability. The sample space of the experiment is also examined to identify model properties that may increase or reduce the risk of multicollinearity in SAOM. Guidelines for interpreting the measure and diagnosing multicollinearity in SAOM are provided. Results are discussed for choosing between statistical network methods and for reducing multicollinearity in afflicted models.