Assessing Mediational Models: Testing and Interval Estimation for Indirect Effects.

Area of Science:

• Social sciences
• Statistical modeling
• Psychometrics

Background:

• Indirect or mediated effects are crucial in social science theoretical models.
• Traditional mediation assessment (Baron & Kenny, Sobel) is now supplemented by computationally intensive methods.
• Performance of these methods, especially with non-normal or incomplete data, is not well understood.

Purpose of the Study:

• To conduct an extensive Monte Carlo simulation comparing various mediation assessment approaches.
• To evaluate method performance regarding Type I error rates, statistical power, and coverage.
• To assess methods under normal/nonnormal and complete/incomplete data conditions.

Main Methods:

• Monte Carlo simulation comparing bootstrapping, distribution of the product, and hierarchical Bayesian Markov chain Monte Carlo (MCMC) methods.
• Evaluation of Type I error rates, power, and coverage.
• Adaptation and testing of a novel inferential method (partial posterior p-value) not relying on confidence intervals.

Main Results:

• The partial posterior p-value method demonstrated superior Type I error control and maximized power, particularly with incomplete data.
• Bias-corrected accelerated (BCa) bootstrapping showed inflated Type I errors and inconsistent coverage, making it not recommended.
• Bootstrapped percentile confidence intervals and hierarchical Bayesian MCMC methods performed best overall, offering good error control, power, and coverage.

Conclusions:

• The partial posterior p-value is a promising new method for mediation analysis, especially in challenging data situations.
• Bootstrapped percentile confidence intervals and hierarchical Bayesian MCMC methods are recommended for robust mediation analysis.
• Researchers should carefully consider method choice based on data characteristics (completeness, normality) to ensure valid indirect effect assessment.