Interactive module guide

Brown-Forsythe Test Calculator — Robust

This Brown Forsythe test calculator checks variance equality through median-centered deviations, staying reliable for skewed and outlier-prone data.

Open calculator

Free · No sign-up · Calculations stay in your browser

The Brown-Forsythe test is the more robust member of the Levene family: it checks whether groups share a common spread by running a one-way ANOVA F on absolute deviations from each group’s median rather than its mean.

Median centering usually reduces sensitivity to skew and isolated outliers, although it does not make the F reference distribution-free or guarantee its nominal error rate under every shape.

This calculator supports two to twelve groups and matches scipy’s levene with center="median" to numerical precision.

Use the Brown Forsythe test calculator as the default variance-equality check; its median centering keeps the stated error rate under skew and outliers.

When to use this Brown Forsythe test calculator

Use it when

Use it as a robust spread-equality diagnostic when distributional shape is uncertain, especially when skew or isolated outliers make mean centering questionable.
Use it for spread questions on skewed measurements: reaction times, incomes, biomarker concentrations, session durations, where a mean-centered check would confuse skew with variance inequality.
Use it when spread itself matters or when documenting assumptions, while recognizing that shape and tail differences can still affect absolute deviations.

Choose another method when

Avoid reading a non-significant result as license to pool blindly; small samples give little power, and the group SD column deserves at least as much attention as the p-value.
Avoid it when the real question is about means—this family tests spread; Welch ANOVA compares means without needing this gate at all.
Avoid pairwise conclusions from the omnibus result; a significant W says spreads differ somewhere, not which groups differ.

Interactive tool

The calculator loads as you approach this section so the guide remains fast on mobile connections.

Example preview

Robust spread check under skew

Inputs: The same three groups re-centered at their medians.
Representative result: W≈0.670 with df=(2, 18) and p≈0.524, agreeing with the mean-centered version here.

Illustrative only. Load the interactive tool to enter your own values and review assumptions.

Watch the explanation

5:03 min

The player loads only after you press play. You can also watch on YouTube.

Read the complete transcript

The Brown-Forsythe Test. The Brown Forsythe test asks whether independent groups show a common population spread. It is the median-centered, more robust member of the Levene family. Keep observations in their original groups. This page accepts two to twelve groups, at least two finite values per group, and a pre-specified alpha. Find each group median. Replace every observation by its absolute distance from that median, creating one nonnegative deviation strip for every group. Median centering is the robustness mechanism. One extreme value can grow far away while the median, and the other deviations, usually move very little. Run ordinary one-way ANOVA on the deviation groups. The resulting F statistic is called W, with degrees of freedom k minus one and N minus k. The p-value comes from the upper F tail. The omnibus alternative says at least one group differs in spread, not that every group differs. A rejection does not locate a particular pair. This calculator provides no pairwise spread follow-up, variance-ratio interval, or automatic explanation for the difference.

Use this median-centered test when skew or isolated outliers make arithmetic means fragile. The separate Levene page keeps the original mean-centered construction. More robust does not mean invulnerable or distribution-free. Different tail weights, strong skew, mixtures, discretization, and small samples can still affect the deviations. Strictly, the null compares expected absolute deviations from group medians. Interpreting that result as equal numerical variances is cleanest when population shapes are comparable. A nonsignificant p-value does not prove equal spread. Inspect group standard deviations, raw shapes, and sample sizes because weak power can hide important differences. The transform cannot repair dependent observations, clusters, repeated measures, incorrect group labels, selection bias, or measurements recorded on incomparable scales. Do not let this preliminary p-value automatically choose pooled or Welch mean analysis. Welch can be pre-specified without conditioning the main test on a noisy gate. Compare Brown Forsythe with mean-centered Levene as a sensitivity analysis. Agreement is reassuring; disagreement flags center sensitivity that deserves inspection, not instant diagnosis.

Now compare three machines with seven, six, and eight measurements, twenty-one observations total, using alpha point zero five. The group medians are thirteen point one, eighteen point six, and fifteen point three. Each machine folds around its own resistant center. Their sample standard deviations are about one point one nine three, one point four zero eight, and point seven four six. Machine Three looks tightest. After median centering, the transformed group means are point eight eight six, one point zero one seven, and point six one three. ANOVA on those deviations gives W equal to point six seven zero four, with numerator degrees of freedom two and denominator degrees of freedom eighteen. The point zero five upper-tail critical value is three point five five five. The observed W remains well below that rejection boundary. The p-value is point five two four, above alpha. Fail to reject a common spread, while keeping the standard deviation differences and limited sample size visible.

Add one hundred to every Machine One value and nothing changes. The median shifts by one hundred, so every absolute deviation stays fixed. Multiply Machine Two by two and its deviations double. Here W becomes three point three three three and p is point zero five eight seven. Replace one Machine One value by forty. Brown Forsythe stays nonsignificant at p point three four three, while mean-centered Levene rejects at point zero two nine nine. Give three groups an identical shape but shift their locations far apart. W becomes essentially zero and p equals one because location is not the target. With heavily discrete groups, every median-centered deviation can become identical. Then within-group transformed variability is zero, so the statistic is undefined and validation stops. Report the independent design, median center, group sizes and standard deviations, W, both degrees of freedom, upper-tail p, alpha, and shape or power cautions. Use mean-centered Levene for symmetric data, Welch for unequal-variance mean comparisons, and direct graphics for shape. Run the Brown Forsythe Test free at Distri Scope dot com.

How to read the result

Compare this page’s result with the mean-centered Levene on the same data: agreement means shape effects are minor; a Levene rejection that Brown-Forsythe does not reproduce points at skew or outliers rather than genuine spread differences.

The SD ratio across groups is the practical effect size; significance at a huge N with a 1.05 SD ratio is a sample-size story, not a variance story.

A rejection here matters most as a warning for downstream pooling—its constructive reading is “use Welch-type procedures”, not merely “assumption violated”.

How to use the Brown Forsythe test calculator

Enter each group in its own field (2 to 12 groups, at least two observations each, unequal sizes welcome).

The group summary reports each group’s standard deviation for context; the test itself works on median-centered absolute deviations.

Only α is configurable; the alternative is inherently “at least one spread differs”.

Formula, hypotheses, and assumptions

W = F\text{-statistic on } Z_{ij}=|x_{ij}-\tilde{x}_i|

Conditions to review

Independent observations within and between groups
Groups measured on the same interval scale
Median centering is more robust to skew and outliers, but the F reference remains approximate
Two or more groups with at least two observations each

Calculator parameters

Significance Level (α): default 0.05.

What the method is doing

The calculator computes Zᵢⱼ = |xᵢⱼ − medianᵢ| and refers the one-way ANOVA F on Z to F(k−1, N−k), reproducing scipy’s levene(center="median") to numerical precision.

Median centering is the entire robustness mechanism: a gross outlier moves its group mean (and every mean-centered deviation) but barely moves the median, so contamination stays localized.

As in the Levene page, the statistic is written W by convention while being an F statistic in construction.

Strictly, the null concerns equality of the groups’ expected absolute deviations from their medians. Reading that result specifically as equality of variances needs similarly shaped populations; different tail weights or skew can change absolute deviations even when numerical variances coincide.

Worked example: the same three machines, robustly

The same three machine samples from the Levene page (n = 7, 6, 8) are re-analyzed with median centering to check that the earlier conclusion was not an artifact of mean sensitivity. Test at α = 0.05.

1Center each group at its median and take absolute deviations; medians and means nearly coincide here because the groups are roughly symmetric.
2Run the ANOVA F on the deviations: W ≈ 0.670 with df = (2, 18).
3The p-value is P(F₂,₁₈ > 0.670) ≈ 0.524—again far above α, so equal spreads are not rejected.
4Compare with the mean-centered result (W ≈ 0.799, p ≈ 0.465): both versions agree, so the earlier conclusion did not depend on the centering choice.

Interpretation

Both members of the Levene family tell one story: no detectable spread differences among the machines at these sample sizes. The agreement itself is the robustness check—had the mean-centered version rejected while this one did not, the correct suspicion would have been an outlier or skew masquerading as variance inequality, not a real precision difference.

Common mistakes

Do not claim that median centering supplies a universal error-rate guarantee; it improves robustness but remains an approximate procedure.
Do not chain accept-then-pool decisions mechanically; conditioning the main analysis on an assumption test distorts both procedures’ error rates.
Do not interpret rejection as a difference in variances specifically; any spread-of-deviations difference, including tail-weight differences, can drive it.

Limits and independent validation

Power at small group sizes is limited—the familiar assumption-checking irony applies here too.

No pairwise spread comparisons or intervals for SD ratios are produced.

With heavily discretized data, median centering can produce many zero deviations and reduce sensitivity.

Before using the result

Cross-check with scipy: levene(..., center="median") should reproduce W and p to ordinary floating-point tolerance.
Recompute one group’s median and a few absolute deviations by hand to confirm data entry.
Run both centering versions and report them together; their agreement or disagreement is itself diagnostic information.

Kruskal-Wallis test →Compare group locations through ranks when shapes are questionable.

← Levene Test Welch ANOVA →

See every option in the Statistical Hypothesis Test Calculator or review the DistriScope methodology. Educational information; last reviewed 2026-08-09. Verify consequential calculations independently.