Causal Mediation Analysis: Warning! Assumptions Ahead

Causal Question / Estimand

Not just whether a treatment works but why — decomposing a total effect into direct and indirect effects through a mediator, and what it takes to identify them.

Identification Strategy

A cautionary review of causal mediation in the Potential-Outcomes framework. The core message: estimating mediation requires assumptions beyond those needed for the total effect. Even when the treatment is randomized, identifying the indirect effect requires sequential ignorability — no unmeasured confounding of the mediator–outcome relationship (and no treatment-induced confounding) — which randomization of the treatment does not provide. The traditional Baron–Kenny regression approach hides these assumptions; the paper urges explicit assumptions and sensitivity analysis.

Key Assumptions

Mediation (direct/indirect effects), sequential ignorability (Ignorability of the mediator given treatment and covariates), Potential-Outcomes, SUTVA.

Threats to Validity

Unmeasured mediator–outcome confounding (the dominant threat); treatment-induced confounders of the mediator; over-reliance on Baron–Kenny regressions that obscure the identifying assumptions. Recommends sensitivity analysis as standard.

Setting / Data

n/a — methodological essay; illustrated with a classroom intervention to raise test scores.

Key Claims

  • Mediation analysis rests on strong, often implausible, untestable assumptions that randomizing the treatment does not satisfy.
  • The popular Baron–Kenny approach buries these assumptions; modern causal mediation makes them explicit.
  • Report sensitivity analyses for the mediator–outcome confounding assumption.

Connections

Citation

Keele, L. (2015). Causal Mediation Analysis: Warning! Assumptions Ahead. American Journal of Evaluation, 36(4), 500–513.