Statistics Glossary → Indirect effect (mediation)
Indirect effect (mediation)concept
The part of the effect of a variable (X) on an outcome (Y) that passes through one or more mediators (M); in the simple model it is measured as the product of the a and b paths (a·b).
In the simple mediation model, a is the effect of X on M, and b is the effect of M on Y holding X constant. The indirect effect is a·b; the direct effect c' is the effect of X on Y holding M constant; the total effect is c = c' + a·b (in linear models). With several mediators, the parallel model gives each mediator's contribution a_i·b_i and their sum, while the serial model gives the products along chained paths such as X → M1 → M2 → Y.
The sampling distribution of a·b is not normal (the product of two normal variables is skewed), so the Sobel test, which relies on a normal approximation, can have low power and mislead. The current recommendation is a bootstrap confidence interval for the indirect effect; if the interval excludes zero, the indirect effect is considered significant. The 'total effect must be significant first' condition of the older Baron-Kenny approach is no longer considered necessary. Mediation analysis is a causal claim: it rests on strong assumptions such as the time ordering of X, M and Y and the absence of unmeasured confounding, especially of the M-Y relationship; with cross-sectional data these assumptions are hard to defend.
An everyday example
A company observes that flexible working hours (X) improve job performance (Y) and wonders whether this works through better sleep (M). Flexible hours increase sleep by 0.5 hours on average (a = 0.5), and each extra hour of sleep raises the performance score by 4 (b = 4). The indirect effect is 0.5 × 4 = 2 points; since the total effect is 5 points, the remaining 3 points come through paths other than sleep (the direct effect).
In MerQur
Available in mediation analysis through the 'X (independent), M (mediator), Y (dependent)' selections and the 'Additional mediators and Multiple-mediator model (Parallel / Serial)' option; indirect effects are given with bootstrap confidence intervals. Structural equation modeling, path analysis and PLS-SEM also report indirect and total effects.
Analyses where this concept appears
Structural Equation Modeling (SEM)Mediation AnalysisPath AnalysisPLS-SEM (Partial Least Squares Path Modeling)