Many Roads to Mediation: A Methodological and Empirical Comparison of Different Approaches to Statistical Mediation

Abstract

This paper provides both a theoretical foundation and a simulation analysis of different statistical approaches to mediation. Regarding theory, a brief sketch of the fundamentals of mechanism-based explanations sets the argument of adhering to a consecutive order of predictor, mediator and outcome in mediation analysis. Having summarized the statistical fundamentals of different approaches to mediation analysis including simple mediation within OLS regressions, fixed-effects (FE) regressions, generalized-method-of-moments (GMM) regressions, causal mediation analysis without (CM) and with fixed effects (CMFE), and fixed-effects cross-lagged panel models (FE-CLPMs), I provide a simulation analysis with known but variable values for the intercorrelations between predictor, mediator and outcome in presence of unobserved heterogeneity and reverse causality. The aim of the simulation study is to examine differences in the relative performance of the aforementioned statistical approaches to mediation under different scenarios of causal order. Results reveal that OLS estimates are generally upwardly biased, FE and CMFE estimates by trend downwardly biased, and the ones of CM models (without FEs) can be biased in both directions. In contrast, coefficients and confidence intervals estimated by both GMM regressions and FE-CLPMs are most accurate – particularly if the structure of lags in the empirical models met the consecutive order set up in the data-generating process. Furthermore, FE-CLPMs are least sensitive to whether the first lag of the outcome variable is included as an additional predictor. All in all, analyses imply the importance that researchers most carefully translate their theoretical assumptions into an empirical model with the appropriate causal order.