An empirical investigation of tests for mediation with respect to four statistical properties.

摘要

Mediation analysis is increasingly common in the social and behavioral sciences. In these contexts, merely establishing that a relationship between variables exists is often insufficient for the research questions being asked. Importantly, mediation analysis can be used to identify mechanisms underlying observed relationships by attempting to explain how or why given variables are related. Nonetheless, research on the statistical properties of tests for mediation has been limited, leaving educational researchers unclear about relevant strengths and weaknesses of various methods. In part, previous studies have failed to establish a consensus on which procedures are ideal because of inconsistencies in the statistical properties examined (e.g., Type I error rate, bias), differing perspectives on the value of estimating the indirect effect itself, and a focus on normally distributed data. The current study evaluates methods for mediation analysis including the Causal Steps method, the Test of Joint Significance, and bootstrap procedures by (1) simultaneously examining Type I error rate, power, confidence interval coverage rate, and bias, (2) considering normal, symmetric nonnormal, and asymmetric nonnormal distributions, and (3) introducing new extensions for the Test of Joint Significance: the Serlin-Harwell Aligned Ranks Procedure (SHARP) and the Product of Interval Endpoints (PIE). Results indicate that the Test of Joint Significance avoids the statistical pitfalls of other prominent methods, along with increased simplicity in calculation and interpretation.