Comparison of methods for statistical mediation analysis within epidemiological research

For many years, epidemiologists were mainly focused on the estimation of exposure-outcome effects. Nowadays, epidemiologists are also interested in assessing the causal mechanisms underlying exposure-outcome effects. Mediation analysis can be used to gain insight in these causal mechanisms, as it decomposes the total exposure-outcome effect into a direct effect and an indirect effect through a mediator. Traditionally, mediation analysis is performed based on a sequence of three linear regression equations. Subsequently, the indirect effect is estimated based on the product-of-coefficients estimator or the difference-in-coefficients estimator. In recent years, causal mediation analysis gained in popularity. Causal mediation analysis is based on the potential outcomes framework and counterfactual framework and defines the direct, indirect, and total effect as the difference between two potential outcomes. There are some important differences between traditional and causal mediation analysis. Traditional mediation analysis does not incorporate exposure-mediator interaction in its effect definitions while causal mediation analysis does incorporate exposure-mediator interaction in its effect definitions. Another important difference between traditional and causal mediation analysis is that the causal effect definitions are not dependent on a specific estimation method, while the traditional effects are defined and estimated based on (linear) regression coefficients. The causal effect definitions can therefore be applied to any mediation model, and do not necessarily depend on parametric assumptions. Based on these differences between traditional and causal mediation analysis, the question arises when and why the effect estimates from these two methods are the same or different. The aim of this thesis is to assess the similarities and differences in the effect estimates yielded by traditional mediation analysis and causal mediation analysis for mediation models frequently encountered in epidemiological research. Simulation studies and empirical data examples were used to compare the traditional and causal effect estimates for mediation models with 1) a continuous mediator and a continuous outcome, 2) a continuous mediator and a binary outcome, and 3) a binary mediator and a binary outcome. This thesis shows that the traditional and causal effect estimates based on linear regression analysis are the same, while this does not necessarily hold when logistic regression is used. When logistic regression analysis is used, the traditional product-of-coefficients estimator and difference-in-coefficients estimator yield different indirect effect estimates. Standardization of the underlying regression coefficients minimizes the difference between these two indirect effect estimates, but does not fully dissolve the difference. The difference between the effect estimates based on these two traditional estimators is caused by non-collapsibility. Furthermore, in the presence of an exposure-mediator interaction, the traditional and causal estimates of the direct and total effect differ. In this situation, the traditional direct effect estimates conditional on the average mediator value under the two exposure levels of interest are similar to the estimate of the controlled direct effect rather than the natural direct effects from causal mediation analysis. Causal mediation analysis is the preferred method for mediation analysis, as this method provides general definitions of causal effects that can be applied to any mediation model. In some situations, traditional mediation analysis can be used to estimate causal effects, for example when the mediator and outcome are both continuous, while in other situations the causal estimators are needed to ensure a causal interpretation of the effect estimates. Epidemiologists use mediation analysis to unravel the causal mechanisms underlying exposure-outcome effects. Causal mediation analysis provides general definitions of the causal effect that constitute these causal mechanisms. These causal effect definitions can be applied to derive the causal effect estimators for any type of mediation model. Therefore, causal mediation analysis should become the default method for mediation analysis in epidemiological research