What is Causal Mediation Analysis?
Causal Mediation Analysis studies how much of an intervention's effect operates through a specified mediator versus other pathways by defining direct and indirect counterfactual contrasts under an explicit causal model.
Quick Facts
| Specification | Official Specification |
|---|
How It Works
The mediator must be causally and temporally defined
Measure baseline confounders before treatment, the mediator after treatment, and the outcome after the mediator. A variable merely correlated with both treatment and outcome is not automatically a mechanism. Specify which intervention could change the mediator and whether mediator values or distributions are feasible, because path labels alone do not define an actionable indirect effect.
Direct and indirect effects require extra assumptions
Identification commonly needs consistency, positivity, no unmeasured treatment-outcome, treatment-mediator, or mediator-outcome confounding, plus correct temporal ordering and model specification. Exposure-induced mediator-outcome confounders make standard natural-effect formulas invalid. Nguyen and colleagues explain why the estimand and required assumptions must be chosen together.
A decomposition is not a percentage explanation by default
The identity total = direct + indirect holds for selected estimands and scales, not universally for nonlinear models, interactions, survival outcomes, or alternative effect definitions. Report both components with uncertainty, preserve the effect scale, test treatment-mediator interaction, examine overlap, and perform sensitivity analysis for mediator-outcome confounding rather than presenting a fragile proportion mediated.
Key Characteristics
- Defines pathways with ordered treatment, mediator, and outcome variables
- Distinguishes total, controlled, natural, and interventional effects
- Requires assumptions beyond those needed for a total causal effect
- Is sensitive to mediator-outcome confounding and exposure-induced confounders
- Depends on effect scale, interactions, and the chosen mediator intervention
- Requires uncertainty and sensitivity analysis for pathway claims
Common Use Cases
- Testing whether onboarding affects retention through early activation
- Separating a policy's behavioral pathway from its remaining effect
- Designing follow-up experiments that intervene on a candidate mediator
- Comparing natural and interventional effect definitions
- Auditing mechanism claims before turning them into product decisions
Example
Loading code...Frequently Asked Questions
What is the difference between a total, direct, and indirect effect?
The total effect compares outcome interventions on treatment. A direct effect holds or redistributes the mediator according to a specified rule, while an indirect effect captures the contrast attributed to changing that mediator rule. Exact definitions depend on controlled, natural, or interventional estimands.
Can regression coefficients prove mediation?
No. A smaller treatment coefficient after adding a mediator does not identify a causal indirect effect. The mediator is post-treatment, and conditioning can introduce bias. Identification requires a causal estimand, temporal design, confounder control, positivity, and suitable models.
What makes natural direct and indirect effects demanding?
They use nested counterfactuals that combine an outcome under one treatment with the mediator value under another. Identification usually invokes assumptions linking counterfactual worlds, including no unmeasured mediator-outcome confounding and no problematic exposure-induced mediator-outcome confounder.
What is exposure-induced mediator-outcome confounding?
It occurs when treatment changes a variable that then affects both mediator and outcome. Adjusting for that variable can block part of the treatment effect, while ignoring it confounds the mediator-outcome relation. Standard natural-effect formulas generally fail without specialized estimands or methods.
Should the proportion mediated always be reported?
No. It can be unstable or misleading when direct and indirect effects have opposite signs, the total effect is near zero, the scale is nonlinear, or interactions exist. Report the underlying effect estimates, scale, uncertainty, assumptions, and sensitivity analyses first.