Direction of effects in mediation analysis

Abstract

Data collected in the social sciences are rarely normally distributed. The linear regression methods that are usually employed to test mediation hypotheses consider moments no higher than second order. Recently discussed methods of direction dependence do consider higher moments. After a review of commonly used methods for mediation analysis, the present article demonstrates that these methods do not allow one to make decisions about competing mediation models, that is, models in which the reverse flow of causality is considered. Then, direction of dependence methodology is introduced which allows one to evaluate hypotheses of direction of effects, and extend its application to mediation analysis. Significance tests for statistical inference on direction of effects are proposed and discussed. Results of a Monte-Carlo simulation of the performance of the tests under various data scenarios are presented. An empirical example from research on intimate partner violence is given. Finally, possible limitations of these methods are addressed, issues of implicit assumptions concerning the origin of observed skewness are discussed, and the new methodology is embedded into the larger framework of causal inference.