The “true” indirect effect won't (always) stand up: When and why reverse mediation testing fails

Abstract

Many social psychological studies aim to test whether an independent variable (X) affects a dependent variable (Y) via one (or more) intervening variable(s) or "mediator(s)" (M). One way to test such a mediation model (X➔M➔Y) is to manipulate X, measure both M and Y, and test statistically whether the indirect effect of X on Y via M is significantly different from zero. However, since the causal order between M and Y is unclear, alternative models (such as X➔Y➔M) are also compatible with the data. Scholars have argued that comparing such models statistically against each other can help decide which model is “correct.” In the present article, we scrutinize the tenability of this “reverse mediation testing” approach via Monte Carlo simulations. Our findings show that reverse mediation testing often fails—especially when the mediator is measured less reliably than the dependent variable.