MerQuR+Statistics Glossary

Statistics Glossary → Sobel test

Sobel testconcept

A method in mediation analysis that tests whether the indirect effect (a × b) differs from zero with a z test based on the approximate standard error of the product.

In a simple mediation model, a is the effect of X on the mediator M and b the effect of M on Y controlling for X; the indirect effect is a·b. The Sobel test approximates the standard error of this product with the delta method: SE = √(b²s_a² + a²s_b²). Then z = a·b / SE is compared with the normal distribution. The Aroian version adds the term s_a²s_b² inside the root and the Goodman version subtracts it.

The test assumes that the sampling distribution of a·b is normal, whereas the product of two normal variables is often skewed. The Sobel test therefore has low power, especially in small and medium samples, and its confidence intervals can be misplaced. Today bootstrap confidence intervals (percentile or bias-corrected) or Monte Carlo intervals are preferred for the indirect effect, and the Sobel test is reported mainly for comparison.

An everyday example

A study expects an exercise programme (X) to lower depression scores (Y) through sleep quality (M). With a = 0.50 (SE 0.10) and b = −0.40 (SE 0.12), the indirect effect is −0.20, the Sobel SE is √(0.16 · 0.01 + 0.25 · 0.0144) ≈ 0.072 and z ≈ −2.8. If the bootstrap interval also excludes zero, the mediation finding is better supported.

In MerQur

'Sobel test (Extra output)' in Mediation Analysis gives the Sobel, Aroian and Goodman normal-theory tests, using the OLS standard errors of the M ~ X model for a and the Y ~ M | X model for b.

Analyses where this concept appears

Mediation Analysis

Related concepts

Indirect effect (mediation)