Mediation Analysis: A Practitioner’s Guide

Causal Question / Estimand

The decomposition of a treatment’s total effect into [[Mediation|natural direct and indirect effects]] (and controlled direct effects) operating through a mediator — a practical guide to estimating “how much” of an effect runs through a given pathway.

Identification Strategy

Sets out the modern causal-inference approach to Mediation in Potential-Outcomes terms, contrasting it with traditional regression (Baron–Kenny) methods and showing how the two coincide only under restrictive conditions (no interaction, linearity). Identification of natural direct/indirect effects requires four no-unmeasured-confounding assumptions — for the treatment–outcome, mediator– outcome, treatment–mediator relationships, and no treatment-induced mediator–outcome confounder. Provides regression-based and weighting estimators, handles exposure– mediator interaction, and supplies sensitivity analysis formulas.

Key Assumptions

Mediation; four confounding-control (Ignorability-type) assumptions for the treatment, mediator, and their relationships; Potential-Outcomes, SUTVA.

Threats to Validity

Unmeasured mediator–outcome confounding; treatment-induced confounders of the mediator; ignoring exposure–mediator interaction. Recommends the proposed sensitivity-analysis bounds for unmeasured confounding.

Setting / Data

n/a — methodological practitioner guide; epidemiology/public-health framing with illustrative examples.

Key Claims

  • Natural direct and indirect effects give a rigorous, assumption-explicit decomposition that subsumes the traditional approach as a special case.
  • Valid mediation requires multiple confounding-control assumptions beyond randomizing treatment; these should be stated and probed.
  • Sensitivity analysis for unmeasured mediator–outcome confounding should accompany every mediation estimate.

Connections

Citation

VanderWeele, T. J. (2016). Mediation Analysis: A Practitioner’s Guide. Annual Review of Public Health, 37, 17–32.