Comparison of mediation analysis methods

In most studies researchers aimed to examine the relationships between two or more variables. If a significant relationship between the variables is detected, researchers apply various statistical methods to determine the structure and direction of the relationship. By adding a third variable to analyze the relationship between the independent and dependent variables, more information is obtained regarding the direction, structure, and status of the relationship. In mediation analysis, the third variable that explains how or why an independent variable affects the dependent variable is called the mediator variable. In mediation analysis, the total effect of an independent variable on a dependent variable distinguish into direct and indirect components and tries to understand how much of the effect of an independent variable on the dependent variable is transmitted through mediator pathways. In this thesis, the performances of mediation analysis methods in case of full mediation and partial mediation, at different skewness and kurtosis coefficient levels, for different sample sizes and direct and indirect effect sizes were examined in terms of Type-I error rates and statistical power. In conclusion, when researchers are only interested in Type-I error rate, in cases where the indirect effect coefficients α=β=0, the Essential Mediation Components (EMC) and Bias-corrected bootstrap methods are recommended. When the indirect effect coefficients are α≠0, β=0, EMC, Monte Carlo (MC) product, percentile bootstrap, and bias-corrected bootstrap methods are recommended. In cases where the indirect effect coefficients are α=0 and β≠0, the EMC, percentile bootstrap, and bias-corrected bootstrap methods are recommended. When researchers are only interested in statistical power, considering both full and partial mediation cases, Freedman & Schatzkin, Clogg, EMC, percentage bootstrap, and bias-corrected bootstrap methods are recommended. When researchers are interested in both Type-I error rate close to the nominal value and statistical power, the EMC and Bias-corrected bootstrap methods are recommended.