Response to Thoemmes and Lemmer (2019) regarding Mediation Analysis Revisited

Kim et al. (2018) raised four issues along with possible remedies that could affect statistical conclusions from mediation analysis. The issues addressed were: omitted mediators, reverse mediation, multi-item measures, and experimental methods. This prompted an insightful response from Thoemmes and Lemmer (2019), who largely agreed upon and further developed three of the issues. But there was disagreement concerning the appropriateness of conducting reverse mediation, their argument being the likelihood of returning equivalent models. Here, it is acknowledged that mediation analysis should be theory-driven and not conducted just to see which model is better. Further, in some contexts such as mediation analysis based on survey data the benefits of conducting reverse mediation are likely to be minimal. However, it is not uncommon to measure the mediator after having measured dependent variables, particularly in experimental settings. Hence, it is argued here to check for artificial mediation by reversing the arrows. A simulation is conducted to support this argument.