On Estimating Causal Mediation Effects from a Single Regression Model (304769)*Christina Tripp Saunders, Vanderbilt University Medical Center
Keywords: Mediation analysis, effect decomposition, portion eliminated, direct and indirect effects
We describe a classical regression framework for estimating causal mediation effects and their variance from the fit of a single regression model, rather than from a system of equations. Requiring the fit of only one model to estimate mediation effects permits the use of a rich suite of regression tools that are not easily implemented on a system of equations. Further, we show how to visualize mediation effects and we highlight situations in which using the difference and product of coefficients approaches do not yield the same estimate of the total effect. We provide extensive examples that apply our approach to complex research hypotheses, including models with multiple mediators, interactions, and nonlinearities. Finally, we show how the proposed framework extends to generalized linear models. Using an example from genetic epidemiology, we compare our formula to existing methods such as the difference of coefficients, the mediation formula, and the KHB method. For large datasets (such as GWAS data), using the single-model framework imparts substantial gains in computational efficiency.