Written and maintained by CASRAI Editorial Board
Last updated
Levene’s test doesn’t have its own menu item in SPSS. It’s a diagnostic that rides inside three separate procedures — the independent-samples t-test, one-way ANOVA, and the general linear model’s univariate (two-way) ANOVA — and SPSS treats it differently in each one: automatic and unavoidable in one, an opt-in checkbox you have to remember to tick in the other two. Knowing which is which, and what actually changes in your reported result when it comes back significant, is the point of this guide.
What Levene’s test is checking
Levene’s test asks whether two or more groups have equal variance on your outcome variable — not whether their means differ, which is a separate question the t-test or ANOVA itself answers. It works by taking the absolute deviation of each score from its group’s center and running an ANOVA on those deviations; if the deviations themselves differ significantly by group, the groups don’t have equal variance. SPSS actually offers four versions of this, differing only in what "center" means: the group mean (the original Levene formulation), the group median, the median with an adjusted degrees of freedom, or a 5% trimmed mean. The median-based version is also known as the Brown-Forsythe test — it’s more robust when the outcome is skewed, since the mean itself is pulled around by the same skew that inflates variance estimates.
The reason it matters: an independent-samples t-test and a between-groups ANOVA both normally pool the group variances into a single estimate for the test statistic’s denominator. That’s only valid if the groups’ variances really are roughly equal. Levene’s test is the gatekeeper SPSS uses to flag when that assumption doesn’t hold.
Where Levene’s test appears in SPSS: automatic vs. you have to ask for it
This is the part that trips people up, because SPSS is not consistent about it:
| Procedure | Menu path | Levene’s test | How you get it |
|---|---|---|---|
| Independent-Samples T Test | Analyze > Compare Means > Independent-Samples T Test | Automatic | No checkbox — SPSS prints it in the Independent Samples Test table every time, with no way to turn it off |
| One-Sample T Test / Paired-Samples T Test | Analyze > Compare Means > … | Doesn’t apply | Not printed at all — there’s only one variance (or one differenced variable) in these designs, so there’s no second group to compare it against |
| One-Way ANOVA | Analyze > Compare Means > One-Way ANOVA | Opt-in | Options tab > check Homogeneity of variance test |
| Two-Way ANOVA (GLM Univariate) | Analyze > General Linear Model > Univariate | Opt-in | Options tab > check Homogeneity tests |
Independent-samples t-test: automatic, no setup required
Run Analyze > Compare Means > Independent-Samples T Test and Levene’s test is already in the output — it appears as two columns (F, Sig.) at the left of the Independent Samples Test table, ahead of the actual t-test columns. There’s nothing to request and nothing to miss; it’s printed whether you want it or not. See CASRAI’s guide to running a t-test in SPSS for the full three-menu-path walkthrough and how to read the rest of that table.
One-way ANOVA: you have to ask for it
Levene’s test is not on by default in either route to a one-way ANOVA. In the dedicated Analyze > Compare Means > One-Way ANOVA dialog, open the Options button and check Homogeneity of variance test — skip this and SPSS runs the ANOVA with no variance diagnostic at all, silently. Checking it produces a Test of Homogeneity of Variances table with four rows (Based on Mean, Based on Median, Based on Median and with adjusted df, Based on trimmed mean) rather than the single row the t-test prints. See CASRAI’s one-way ANOVA in SPSS guide for the full procedure and the Tukey-vs-Games-Howell post hoc decision this table feeds into.
Two-way ANOVA (GLM Univariate): also opt-in, and named differently
If you run your ANOVA through Analyze > General Linear Model > Univariate — the route needed for two factors, or for a single factor when you want partial eta-squared reported automatically — the checkbox lives under the Options button as Homogeneity tests. Ticking it produces a table labeled Levene’s Test of Equality of Error Variances rather than the "Test of Homogeneity of Variances" title one-way ANOVA uses, and it’s a single overall test across every cell of the design (every combination of your two factors), not a per-factor breakdown. See CASRAI’s two-way ANOVA in SPSS guide for the full Options-tab setup and how to read the interaction term.
Reading the output once you have it
Regardless of which procedure produced it, the decision rule is the same: a Sig. value at or below .05 means the variances differ significantly across groups and the equal-variances assumption is in question; above .05, the assumption holds well enough to proceed with the standard equal-variances version of the test.
Two things catch people out here:
- Which row to read in one-way ANOVA’s four-row table. The "Based on Mean" row is the traditional Levene’s test and the one most commonly reported by default; the other three (median-based, adjusted, trimmed-mean) are more robust alternatives worth checking when the outcome is visibly skewed, since a skewed distribution can itself inflate the mean-based version’s false-positive rate.
- Levene’s own F and Sig. are not the result you’re reporting. They only decide which downstream row or correction to use — the equal-variances vs. unequal-variances row in a t-test, or the standard vs. robust omnibus test in ANOVA. A frequent write-up error is quoting Levene’s F/Sig. as if it were the substantive test result.
What to do when Levene’s test comes back significant
The correction is different in each procedure — there’s no single "switch to the robust version" button that works everywhere.
Independent-samples t-test: switch rows, nothing to re-run
The Independent Samples Test table always prints two full rows — "Equal variances assumed" and "Equal variances not assumed" — regardless of what Levene’s test says. If Levene’s Sig. is ≤ .05, report the "Equal variances not assumed" row: this is the Welch-Satterthwaite correction, which adjusts the degrees of freedom (often to a non-integer value) instead of pooling the two group variances. SPSS has already calculated it in the same output; no separate procedure or menu is needed.
One-way ANOVA: two separate corrections, not one
A significant Levene’s test in one-way ANOVA affects two different parts of your results, and both need a fix:
- The omnibus F-test. Go back into Options and check Welch and/or Brown-Forsythe alongside the homogeneity test — SPSS adds a Robust Tests of Equality of Means table with variance-corrected F statistics and adjusted degrees of freedom. Report the Welch F instead of the standard ANOVA F when variances are unequal; it’s the more commonly recommended of the two robust options in the applied-statistics literature.
- The post hoc comparisons. Switch from a Tukey-family test (which assumes equal variances) to Games-Howell, which SPSS groups separately in the Post Hoc dialog precisely because it doesn’t require that assumption. Defaulting to Tukey without checking Levene’s test first is the single most common mistake in one-way ANOVA write-ups.
The omnibus F is reasonably robust to moderate variance violations when group sizes are similar-sized, which is part of why this step is easy to skip — but the post hoc test choice should follow Levene’s result regardless.
Two-way ANOVA (GLM Univariate): no built-in row-switch
This is the case with the least good news. GLM Univariate has no Welch- or Brown-Forsythe-corrected F option the way One-Way ANOVA does, so there’s no equivalent checkbox to tick. With balanced or near-balanced cell sizes (roughly equal n in every combination of your two factors), the omnibus F-tests are fairly robust to the violation and it’s defensible to proceed while noting it. With unequal cell sizes, the violation matters more: report it explicitly, treat borderline (p close to .05) effects more cautiously, and consider either a variance-stabilizing transformation of the outcome before re-running the model, or running the analysis on ranks / with a bootstrap for the standard errors (available on the Bootstrap tab of the same dialog) rather than trusting the standard F as-is.
Common mistakes
- Assuming Levene’s test runs everywhere automatically, the way it does for the t-test. It doesn’t — skip the Options checkbox in either ANOVA procedure and you get no variance diagnostic at all, with no warning.
- Reporting Levene’s F/Sig. as if it were the finding. It’s a gate, not a result in its own right — report the test it feeds into (the correct t-test row, the correct ANOVA F, the correct post hoc test), not Levene’s statistic itself, unless your reporting convention specifically calls for disclosing it too.
- Fixing the omnibus test but not the post hoc test (or vice versa) in one-way ANOVA. They’re two separate corrections that both key off the same Levene’s result — ticking Welch without also switching to Games-Howell (or the reverse) leaves half the analysis using an assumption you’ve already shown doesn’t hold.
- Looking for a two-way-ANOVA equivalent of the Welch row and not finding one. GLM Univariate genuinely doesn’t have it; the fix there is procedural (balanced design, transformation, bootstrap), not a checkbox.
Frequently asked questions
Does SPSS always run Levene’s test automatically?
No. It’s automatic only in the independent-samples t-test. Both ANOVA routes (One-Way ANOVA and GLM Univariate) require you to check a box under Options first, and use different box labels ("Homogeneity of variance test" vs. "Homogeneity tests") for what’s functionally the same diagnostic.
What’s the difference between the four Levene’s rows in one-way ANOVA’s output?
They differ only in what "center" the deviations are measured from: group mean (the classic version), group median, median with adjusted degrees of freedom, or a 5% trimmed mean. The mean-based row is the one conventionally reported; the median-based rows (sometimes called the Brown-Forsythe test) are more robust when your outcome is skewed.
What do I report if Levene’s test is significant in a two-way ANOVA?
There’s no separate corrected F to switch to inside GLM Univariate. Report the violation alongside your results, note whether your cell sizes were balanced (which limits the practical impact), and interpret borderline effects more cautiously — or transform the outcome and re-run the model if the violation is severe.
Is Levene’s test the same as the Brown-Forsythe test?
Closely related but not identical. Levene’s original formulation centers on the group mean; the Brown-Forsythe test is a modification that centers on the group median instead, making it more robust to skewed data. SPSS reports both as rows in the same homogeneity-of-variance table in One-Way ANOVA.
Can I run Levene’s test on its own, without also running a t-test or ANOVA?
Not as a standalone menu command — SPSS only produces it as an embedded diagnostic inside the t-test and ANOVA procedures above. If you need a homogeneity-of-variance check unattached to a mean-comparison test, run one-way ANOVA with your grouping variable and the Homogeneity of variance test box checked, and read only that table.
