Written and maintained by CASRAI Editorial Board
Last updated
The Sobel test is the classic significance test for a mediated (indirect) effect — the product term ab in an X→M→Y mediation model, where a is the X-to-M path and b is the M-to-Y path controlling for X. For roughly two decades after its introduction it was the default way to ask whether that indirect effect was distinguishable from zero. It has since been displaced by bootstrapped confidence intervals as the field’s default test. This page is scoped narrowly to that one transition: the formula the Sobel test uses, the assumption baked into that formula, why the assumption fails often enough to matter, and what a seeded simulation shows about the power cost. It is not a general introduction to mediation — for the mechanics of setting up and reporting a mediation model, see Mediation Analysis: Methods and Reporting; for the causal-inference (potential-outcomes) treatment of the same problem, see Causal Mediation Analysis. This page assumes you already know what a and b are and picks up at the question of how ab gets tested.
What the Sobel test actually tests
The null hypothesis is H0: ab = 0 — no indirect effect. The Sobel test answers this with a single z-statistic built from the point estimates and standard errors of the two component paths, rather than by resampling the data or by requiring each path individually to clear its own significance threshold first. That second contrast is the historical reason it mattered: it replaced the older Baron and Kenny (1986) causal-steps procedure, which tested significance path-by-path and required X to predict Y before mediation could even be evaluated — a requirement later shown to have low power and no solid theoretical basis, since a genuine indirect effect can exist even when the total effect is null (competing mediators can cancel, or direct and indirect effects can carry opposite signs). Sobel’s method was, in that sense, a real improvement on what came before it: it tests the actual quantity of interest, ab, directly, in one test, rather than inferring it from two other tests.
The formula: a delta-method standard error for ab
Sobel’s method (Sobel, M. E., 1982, “Asymptotic Confidence Intervals for Indirect Effects in Structural Equation Models,” Sociological Methodology 13, pp. 290–312) treats ab as a nonlinear function of two estimated regression coefficients and applies the multivariate delta method — a first-order Taylor-series approximation — to derive an approximate standard error for that product:
SESobel = √(b²·SEa² + a²·SEb²)
where a and b are the fitted path coefficients and SEa, SEb are their ordinary regression standard errors. The test statistic is z = ab / SESobel, compared against the standard normal distribution — counted significant past the familiar ±1.96 threshold at α = .05.
Applied to real numbers, from one replication of the seeded simulation described below (n = 60): the fitted paths were a = 0.437 (SE = 0.126) and b = 0.159 (SE = 0.141), giving ab = 0.070. Plugging in: SESobel = √(0.159²×0.126² + 0.437²×0.141²) = 0.0648, so z = 0.070 / 0.0648 = 1.08 — not significant, despite a real underlying indirect effect built into the simulation (true ab = 0.09). That’s not an isolated unlucky draw; it’s the pattern the power comparison further down quantifies across 1,000 such samples.
Why it held up for two decades
Two things made the Sobel test the field’s default for a long stretch of the 1980s through the 2000s. First, it directly tested ab rather than requiring a chain of path-by-path significance tests, which was already an improvement on causal steps. Second, it’s a closed-form calculation — two coefficients, two standard errors, one square root — producible instantly by hand or by any of the small online Sobel-test calculators built specifically around it, at a time when resampling thousands of bootstrap replicates per analysis was a meaningfully heavier computational ask than it is today. Structural equation modeling software built it in as standard output. It became the default not because the field had settled the underlying assumption question, but because nothing more practical was on offer yet.
The assumption that breaks: a product of two normal variables isn’t normal
The delta-method SE above is only as good as the assumption it rests on: that ab, standardized by its SE, follows a normal distribution — the same assumption every ordinary z-test makes. But while a and b are each approximately normally distributed on their own, the distribution of the product of two normal random variables is not itself normal — it’s asymmetric and heavier-tailed than a normal curve, most noticeably when the true indirect effect is small to moderate relative to its standard error, which describes most published mediation effects. David Kenny’s own summary of the method states plainly that the Sobel test “is very conservative…because it falsely assumes that the indirect effect has a normal distribution, when in fact it is highly skewed,” attributing the finding to MacKinnon and colleagues’ 1995 comparison of mediation tests, and concludes: “It should no longer be used.”
The seeded simulation below makes the skew concrete rather than asserted: across 1,000 bootstrap resamples of a single dataset, the resampling distribution of ab had a skewness coefficient of 0.52 at n = 60 and 0.65 at n = 250 (a perfectly symmetric normal distribution has a skewness of 0). That asymmetry is exactly what a z-test built on a symmetric normal reference distribution mishandles: the true sampling distribution has a longer tail on one side, so a symmetric ±1.96 cutoff is either too strict or too lenient depending on which side of zero the true effect sits — and empirically, for the small-to-moderate effects most mediation studies involve, it comes out too strict, i.e. conservative, more often than not.
The power problem, demonstrated
To quantify the power cost rather than just describe it, this page runs its own seeded Monte Carlo simulation (Node.js, a mulberry32 pseudo-random generator seeded for reproducibility, Box-Muller normal transform, no external statistics packages). The data-generating model: X→M with true a = 0.30, M→Y with true b = 0.30 controlling for a direct effect c′ = 0.20, so the true indirect effect is ab = 0.09 — a small-to-moderate effect, the range Fritz and MacKinnon (2007, Psychological Science 18(3), 233–239) identify as where most published mediation effects actually fall. At each of two sample sizes, 1,000 independent samples were drawn; each sample was tested both by the Sobel z-test and by a 1,000-resample percentile bootstrap of ab, and power was the proportion of the 1,000 samples in which each test correctly rejected the false null.
| Sample size | Sobel test power | Percentile bootstrap power |
|---|---|---|
| n = 60 | 18.2% | 35.0% |
| n = 250 | 97.7% | 98.5% |
At n = 60, the bootstrap detects the real, built-in effect nearly twice as often as the Sobel test does, from identical data. At n = 250 the gap almost closes — both methods have enough data that the asymptotic normality the Sobel test assumes is a closer approximation. That pattern lines up with Fritz and MacKinnon’s own published finding that the bias-corrected bootstrap reaches 80% power around N = 71 for medium-sized path components: the Sobel test’s disadvantage is concentrated in the small-to-moderate sample sizes most single-mediator studies actually run at, not a fixed, sample-size-independent gap.
The named variants: Aroian and Goodman
Two adjustments to the Sobel formula circulated alongside it, each changing only the variance term under the square root:
- Aroian test: SEAroian = √(b²SEa² + a²SEb² + SEa²SEb²) — adds a second-order cross term that Sobel’s first-order approximation drops, giving the exact variance of a product of two independent normal variables rather than an approximation of it.
- Goodman test: SEGoodman = √(b²SEa² + a²SEb² − SEa²SEb²) — subtracts the same term, from an unbiased-estimator derivation instead. Because it subtracts rather than adds, it can produce a negative value under the square root with small, noisy path estimates — a real practical failure mode, not a rounding edge case.
Neither fixes the core problem. Both are still z-tests against a symmetric normal reference distribution for a statistic whose true sampling distribution is skewed — they refine the variance estimate, not the shape assumption that estimate feeds into. In practice the three formulas tend to produce very similar p-values on the same data, and none of the three is preferred over resampling methods in current guidance.
What replaced it
Two families of methods now stand in for the Sobel test. Neither is re-taught in depth here, since both already have dedicated coverage on this site:
- Bootstrapping — resample the observed data with replacement, recompute ab on each resample, and build a confidence interval directly from the resulting distribution (percentile or bias-corrected) instead of assuming it’s normal. This is the current default; see Bootstrapping in Statistics for the general resampling mechanics and Mediation Analysis: Methods and Reporting for how it’s applied specifically to ab, including Hayes’ PROCESS macro implementation and sample-size planning.
- Distribution-of-the-product methods (implemented as PRODCLIN, or in the R package RMediation) — instead of resampling or assuming normality, these compute the actual analytic sampling distribution of a product of two normal variables and build a confidence interval directly from that distribution’s known, asymmetric shape. They perform comparably to the bootstrap and, unlike the bootstrap, don’t require the raw data — only a, b, and their standard errors, which is useful for re-analyzing older published results that only ever reported a Sobel statistic.
Both approaches share the same underlying fix: they stop assuming ab is normally distributed and instead work with, or closely approximate, its actual skewed sampling distribution.
When you might still legitimately see a Sobel test today
Not every Sobel-test citation in current literature is an error. It still shows up, defensibly, in a few specific situations: replicating or extending an older study that used it, for direct comparability with the original result; meta-analyses harmonizing effect sizes across decades of published mediation studies that only ever reported a Sobel z (its closed-form nature means it can be reconstructed from a published regression table — a, b, and their standard errors — without access to the original raw data, which resampling methods cannot do); and as a quick supplementary sanity check reported alongside, never instead of, a bootstrap or distribution-of-product confidence interval. What it should no longer be is the sole or primary test in a newly planned analysis; a reviewer at most quantitative journals will now expect a bootstrapped or distribution-of-product interval, not a bare Sobel z reported on its own.
Frequently asked questions
Is the Sobel test still valid to report?
The formula is computed correctly and the math is not “wrong” — it’s a legitimate delta-method approximation. What’s outdated is treating it as the primary or sole test: its normality assumption is a known, demonstrated poor fit for the actual sampling distribution of a product term, and it costs real statistical power relative to bootstrapped or distribution-of-product alternatives, especially at the sample sizes most single-mediator studies run at. Report it as a supplementary figure alongside a bootstrap CI if you want the historical comparability; don’t rely on it alone.
What’s the practical difference between the Sobel, Aroian, and Goodman tests?
All three use the same z = ab / SE structure and differ only in how they compute the variance under the square root — Sobel uses a first-order delta-method approximation, Aroian adds a second-order cross term for the exact product variance, and Goodman subtracts that same term from an unbiased-estimator derivation (which can occasionally produce a negative value under the square root). In practice the three give very similar results on the same data, and the choice between them doesn’t address the shared underlying problem: all three assume a normal sampling distribution for a statistic that isn’t normally distributed.
Why does the Sobel test need more data than the bootstrap for the same effect?
Because its normality assumption is a worse approximation exactly when samples are small — that’s when the skew in the true sampling distribution of ab is most pronounced. A seeded simulation on this page found the Sobel test detecting a true small-to-moderate indirect effect in 18.2% of samples at n = 60 versus 35.0% for a percentile bootstrap on the same data; by n = 250 the gap narrows to 97.7% versus 98.5%, since at larger samples the asymptotic normality the Sobel test relies on becomes a much closer approximation.
Can I compute a Sobel test from a published table without the raw data?
Yes — that’s its one durable practical advantage over resampling methods. Because it’s a closed-form function of a, b, and their standard errors, all four of which are routinely reported in a regression or path-analysis table, you can compute (or reconstruct) a Sobel statistic for a published result you don’t have raw data access to. This is why it still appears in meta-analyses and re-analyses of older literature even though it’s not recommended for a newly planned study.
