A General Measure of Effect Size for Mediation Analysis

Abstract

Mediation analysis is an increasingly popular statistical method in the social sciences. However, currently available effect size measures for mediation have substantial limitations. In particular, complex mediation models (e.g, multiple mediators, moderators, and latent variables) have no established effect size measures. In this dissertation I will develop an effect size measure that will address these limitations. This work will have several components: 1) derive a generalizable expression of the effect size using a structural equation modeling framework, 2) evaluate the measure in terms of established criteria for good effect sizes, 3) derive a general expression for the sample bias of the effect size estimator, and propose a bias-adjusted estimator, 4) evaluate the sampling properties of the effect size estimators using Monte Carlo simulation studies, and 5) demonstrate the use and interpretation of the effect size in empirical examples. The work proposed in this dissertation will fill a substantial gap in the methodological literature, and provide researchers with a flexible method for communicating the practical importance of mediation results.