Mediation analysis

Description

Decomposes the effect of a cause on an outcome into a direct path and an indirect path through a mechanism, quantifying how a design change works.

Mediation analysis estimates whether X affects Y through an intermediate variable M, for example a redesign increases adoption because it raises perceived trust. The indirect effect (the product of the X-to-M and M-to-Y paths) is tested with bootstrap confidence intervals rather than the outdated Baron and Kenny causal-steps logic. Modern causal mediation makes the assumptions explicit: no unmeasured confounding of X-M, M-Y and X-Y, and no M-Y confounder affected by X. Rigour requires temporal ordering (M measured after X and before Y), ideally a randomised X, and sensitivity analysis for unmeasured mediator-outcome confounding. It is the quantitative bridge between a qualitative mechanism story and a measurable causal claim.

Attributes

Loss type
Latent
Effort
Weeks

Steps

  1. Specify the mechanism from qualitative evidence and draw it as a causal diagram including likely confounders of M and Y.
  2. Design measurement so X precedes M and M precedes Y; randomise X where possible.
  3. Fit the mediator model (M on X plus covariates) and the outcome model (Y on X and M plus covariates).
  4. Estimate the indirect effect with 5,000 or more bootstrap resamples and report the 95% percentile or bias-corrected interval.
  5. Report the proportion mediated with caution and run a sensitivity analysis (e.g. the rho parameter of Imai et al.) for unmeasured M-Y confounding.
  6. Interpret the mechanism in design terms and state the assumptions that would overturn it.

Sources

  • Applied Predictive Modeling (Max Kuhn & Kjell Johnson)
  • R for Marketing Research and Analytics (Chris Chapman & Elea McDonnell Feit)
  • Introduction to Mediation, Moderation, and Conditional Process Analysis (Andrew Hayes)
  • Explanation in Causal Inference (Tyler VanderWeele)
  • Applied Predictive Modeling (Kuhn & Johnson)
  • R for Marketing Research (Chapman & Feit)
  • The Inquiry Series, Ch 43 — Regression, Moderation & Mediation for UX Researchers

Also known as

  • Indirect effect analysis
  • Causal mediation analysis
  • Process analysis

Categories

  • Analytics & Behaviour
  • Statistics & Modelling