Primary Submission Category: Mediation
Regression-based mediation sample size and power determinations
Authors: Yingjin Zhang, Chung-Chou Ho Chang,
Presenting Author: Yingjin Zhang*
Mediation analysis has been widely applied in many disciplines to better understand the underlying mechanism. Sufficient sample size and statistical power are essential in the study design stage in order to ensure reliable results in research. Methods to calculate sample-size and power for mediation analysis, including the Sobel test for statistical significance of the indirect effect, Monte Carlo simulations for power, and bootstrap assessment of confidence interval, have been hampered by the lack of closed forms and thus require substantial amounts of computational simulations. Therefore, these existing methods have rarely been adopted by researchers due to computational complexity, absence of software options, and limited settings on the prespecified causal pathways. In this study, we propose and derive regression-based analytic formulas for sample size and power estimations associated to the inferences on the direct effect, indirect effect, and mediation proportion under the counterfactual mediation setting with 15 different combinations of types of the main exposure, mediator, and outcome variables. Our methods focus on the cross-sectional settings with one mediator and possible multiple measured exposure–outcome and mediator–outcome confounders. Our methods rely on the estimations of covariate effects and their variance-covariance matrices in the involved regression models, which can be either provided or calculated from pilot data sets.