Mediation and Indirect Effects Analysis

Core Idea

Mediation analysis decomposes a causal effect into direct effects (X→Y) and indirect effects operating through a mediator (X→M→Y). Understanding mechanisms requires identifying the causal pathway through which an independent variable influences an outcome. Modern mediation analysis uses causal inference frameworks: the natural indirect effect (NIE) and direct effect (NDE) are defined under counterfactual logic, accounting for treatment-mediator interactions and sequential ignorability assumptions.

Explainer

From your linear regression background, you know that regression estimates the average relationship between a predictor and an outcome while holding other variables constant. Mediation analysis takes the next step: instead of just asking *whether* X affects Y, it asks *how* — through what pathway does the effect travel? This distinction between "does it work?" and "how does it work?" is the difference between establishing an effect and understanding a mechanism.

The basic setup has three variables. You have an independent variable X (a treatment, policy, or cause), an outcome Y, and a mediator M — an intermediate variable that lies on the causal path from X to Y. For example: does attending college increase lifetime earnings (X→Y)? Part of that effect might operate directly (employers value degrees per se), and part might operate through the skills and networks college develops (X→M→Y). Mediation analysis partitions the total effect into these pieces. The direct effect is the effect of X on Y that does not go through M. The indirect effect is the portion that travels through M. The two sum to the total effect.

The classical approach (Baron and Kenny's "causal steps" procedure) estimated these pieces using a series of regression equations: regress M on X, regress Y on X and M, and interpret coefficients. The indirect effect equals the product of two coefficients — the effect of X on M and the effect of M on Y controlling for X. This product-of-coefficients approach is still the core computational intuition. But modern mediation analysis, built on the counterfactual framework you may recognize from causal inference, is considerably more demanding. It requires sequential ignorability: X must be effectively randomized (no unmeasured confounders of X→Y), and M must also be effectively randomized conditional on X (no unmeasured confounders of M→Y). In observational research, neither assumption is easily satisfied, which is why mediation claims from purely observational data are often overstated.

The modern definitions of the natural direct effect (NDE) and natural indirect effect (NIE) handle the case where X modifies the effect of M on Y — that is, when the pathway through M works differently depending on the value of X. In this interaction case, the simple product-of-coefficients formula gives misleading results; counterfactual definitions correctly partition the total effect. In practice, this means testing for X×M interactions and using bootstrapping to construct confidence intervals for indirect effects, since the product of two regression coefficients doesn't follow a simple known distribution. The upshot for applied research: mediation analysis is a powerful tool for investigating mechanisms, but its causal interpretation requires strong assumptions that should be stated explicitly and probed with sensitivity analyses rather than assumed away.