The finite sample performance of estimators for mediation analysis under sequential conditional independence

Using a comprehensive simulation study based on empirical data, this paper investigates the finite sample properties of different classes of parametric and semi-parametric estimators of (natural or pure) direct and indirect causal effects used in mediation analysis under sequential conditional independence assumptions. The estimators are based on regression, inverse probability weighting, and combinations thereof. Our simulation design uses a large population of Swiss jobseekers and considers variations of several features of the data generating process and the implementation of the estimators that are of practical relevance. We find that no estimator performs uniformly best (in terms of root mean squared error) in all simulations. Overall, so-called 'g-computation' dominates. However, differences between estimators are often (but not always) minor in the various setups and the relative performance of the methods often (but not always) varies with the features of the data generating process.