Glossary
Mediation analysis
What is mediation analysis?
Mediation analysis estimates a proposed indirect pathway between a treatment or exposure and an outcome through an intermediate variable. It can distinguish direct and indirect effects under a specified model and set of assumptions.
Interpretation
Imai, Keele, and Tingley explain why statistical estimation and causal identification are separate problems. Unmeasured causes of both the mediator and outcome can bias a causal interpretation even when treatment was randomized. Resampling or structural equation modeling does not remove this problem automatically.
Testing the proposed pathway
For example, a study might ask whether training improves performance partly by increasing confidence. An estimated indirect effect can support that hypothesis, subject to the model’s assumptions. Its interpretation also requires a defensible temporal order, measurement, and assumptions about confounding. Sensitivity analysis can assess how conclusions change when an identifying assumption is relaxed.
Total, direct, and indirect paths
In the hypothetical training example,
- the total effect is the overall difference training makes to later performance.
- An indirect effect through confidence is the part assigned to training changing confidence and that change affecting performance.
- A direct effect relative to this mediator represents pathways not operating through the confidence variable in the model—for example, training might teach task knowledge.
“Direct” does not mean that no other intermediate process exists.
Estimation tools
Bootstrapping repeatedly resamples the observed data and re-estimates the pathway to assess uncertainty, such as a confidence interval for an indirect effect. Structural equation modeling specifies and estimates a system of relationships among variables, which can include proposed mediation paths. These tools concern estimation and model structure. They cannot establish the direction of causation or remove unmeasured mediator–outcome confounding by themselves.