AI Glossary / Statistics and probability / Advanced
Sobel test
Definition
The Sobel test is a normal-approximation test of an indirect effect expressed as the product of two estimated regression paths. In a simple mediation model, path a links an explanatory variable to a mediator, and path b links the mediator to the outcome while adjusting for the explanatory variable. The test compares the estimated product ab with an approximate standard error. It is a statistical test within a specified model, not evidence by itself that a causal mechanism has been established.
What it means
A mediation question asks whether part of a relationship might operate through an intermediate variable. For example, training might influence task performance partly through greater familiarity with a tool. A fitted indirect effect summarizes the two linked associations. The Sobel approach assesses that product using a first-order approximation to its uncertainty. Its normal approximation can be poor, particularly in smaller samples or when both component paths are close to zero. The underlying study design and measurement quality remain essential.
How it works
Estimate the two paths using compatible models and obtain their standard errors, sa and sb. The familiar Sobel standard error is the square root of b squared times sa squared plus a squared times sb squared. Divide ab by this quantity to obtain a z statistic and compare it with a standard normal reference distribution. This usual expression omits a covariance term; an analysis in which the path estimates have relevant covariance needs an appropriate variance calculation. Check model assumptions, sampling dependence and the suitability of the approximation before interpreting the result.
Three practical examples
Example 1: A researcher examines whether an association between training and performance is partly represented by measured confidence. Example 2: A product analyst studies a proposed pathway from onboarding support through feature familiarity to continued usage. Example 3: An AI evaluation team examines whether an interface change is associated with better task completion through reduced navigation time. Each example requires a defensible temporal and substantive model. Observing these variables together does not establish that changing the proposed mediator would change the outcome.
Think of it this way
Think of an indirect route with two connected road segments. The estimated strength of the route depends on both segments, so uncertainty about either contributes to uncertainty about the whole route. Multiplication makes that uncertainty different from uncertainty about an ordinary single coefficient. The analogy does not imply that the route is causally valid: there may be unmeasured roads connecting the same destinations.
Worked example
Suppose compatible fitted models give a = 0.6 with sa = 0.2 and b = 0.5 with sb = 0.1. The estimated indirect effect is 0.6 times 0.5 = 0.3. Under the usual Sobel variance expression, the variance approximation is 0.5 squared times 0.2 squared plus 0.6 squared times 0.1 squared = 0.01 + 0.0036 = 0.0136. Its square root is approximately 0.11662, so z is approximately 2.5725. A two-sided standard normal calculation gives a p-value between 0.010 and 0.011. This is an approximate model-based result, not an exact probability that mediation is absent. It does not establish causal identification, nor does it show that the effect is practically large. The coefficient units and the model specification are needed to interpret the value 0.3.
Illustrative case study
In this fictional case study, Harbor Learning evaluates whether a training program improves task performance through greater software familiarity. An analyst reports a Sobel calculation but notices that familiarity and performance were measured at the same time. The report therefore describes an estimated indirect association and explains the uncertainty about temporal order and unmeasured confounding. The team plans a stronger follow-up design and compares an appropriate bootstrap analysis with the normal approximation. Managers receive a qualified finding rather than a claim that the mechanism has been proven.
Common misconceptions and mistakes
Common mistakes include treating a significant Sobel result as proof of causation, assuming the product of two estimates must have a normal distribution, and confusing the Sobel formula with Aroian or Goodman variants. Another mistake is requiring a statistically significant total effect before any indirect effect can be studied; different pathways can offset one another. Conversely, a nonsignificant result does not demonstrate that an indirect pathway is absent. Small samples, measurement error, omitted variables and incorrectly handled clustering can undermine interpretation.
Business use
Businesses can use mediation analysis to organize hypotheses about how an intervention might work, but should distinguish exploratory pathway analysis from reliable evidence for action. Report the indirect estimate, its uncertainty, the assumed causal structure and the analysis method. Resampling methods may better represent a skewed product distribution in suitable settings, yet they do not remove confounding or repair an unsuitable design. Decisions about changing a mediator need stronger support than the significance of a coefficient product alone.
People also ask
Is the Sobel test exact?
No; the usual version relies on a large-sample normal approximation.
Does significance prove mediation is causal?
No; causal interpretation requires additional design and identification assumptions.
Is a bootstrap automatically assumption-free?
No; resampling must reflect the sampling design and cannot correct every modeling problem.
Are Sobel, Aroian and Goodman tests identical?
No; their standard-error expressions differ, so the method should be named explicitly.
Check your understanding
What is the estimated indirect effect in the example?
Answer: 0.3.
What is the approximate variance used in the displayed Sobel calculation?
Answer: 0.0136.
Why is the p-value described as approximate?
Answer: It comes from a normal reference approximation for a product-based statistic.
Would a small p-value fix uncertainty about the temporal ordering of mediator and outcome?
Answer: No; that is a design and interpretation problem.
Related terms
- Path analysis (statistics)
- Bootstrap confidence interval
- Confounding
- Related terms include mediation
- indirect effect
- delta method
- path analysis
- total effect and direct effect
Learn next
Learn next how the delta method approximates the variance of a function of estimates and why a product distribution can be asymmetric. Then examine causal mediation assumptions, temporal measurement and resampling methods that respect clustered or longitudinal designs. Compare effect estimates and uncertainty rather than relying on a significance label alone.
More in Statistics and probability
- Sliced inverse regression - Sliced inverse regression, or SIR, is a supervised dimension-reduction method that seeks informative linear combinations of predictors for explaining a...
- Smearing retransformation - Smearing retransformation is a method for estimating an expected response on its original scale after fitting a regression on a transformed scale.
- Smoothing spline - A smoothing spline estimates a smooth relationship by balancing fit to observed data against a penalty for excessive curvature.
- Solomon four-group design - The Solomon four-group design is a randomized experimental design that crosses an intervention with the presence or absence of a pretest.
- Sparse binary polynomial hashing - Sparse binary polynomial hashing, or SBPH, is a text-feature generation technique associated with the CRM114 classifier.
- Sparse model - A sparse model is a model whose representation contains relatively few active or nonzero components compared with the available possibilities.
Sources
Sources: Preacher and Leonardelli, Calculation for the Sobel test, https://quantpsy.org/sobel/sobel; Preacher and Hayes, SPSS and SAS procedures for estimating indirect effects in simple mediation models, https://www.quantpsy.org/pubs/preacher_hayes_2004.pdf. These primary methodological resources explain the product calculation and alternatives. The numerical example is original and independently calculated; the organisational case is fictional.
Previous: Sobel edge detector Next: Social cloud computing