Bayesian Mediation Analysis for Crossed Random Effects Designs

• In Event: Factor Analysis and Structural Equation Modeling in Educational Research

Abstract

In educational and psychological research, it is common that participants respond to the same stimuli, which result in observations nested in two random factors. Crossed-random-effects models (CREM) can account for variations in participants and stimuli. However, CREMs often encounter convergence failures with maximum-likelihood estimation (MLE). The usual response is to simplify the model specification for model convergence, which can hinder the research. The current study used simulations to examine whether Bayesian estimation would be a viable alternative to MLE in analyzing multilevel mediation models with crossed random factors and a zero-inflated Poisson mediator. Bayesian estimation achieved near perfect convergence rates, and the estimates were accurate and efficient. Bayesian estimation is feasible and especially favorable over MLE for complex models.

Authors

• Pei-Yu Marian Pan, University of Illinois at Urbana-Champaign

• Carolyn J. Anderson, University of Illinois at Urbana-Champaign