Interactive module guide

Welch ANOVA Calculator — Unequal Variances

This Welch ANOVA calculator compares group means without assuming equal variances and follows up with Games-Howell pairwise comparisons.

Open calculator

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

Welch’s ANOVA compares two or more group means without assuming equal variances.

Each group is weighted by its precision wᵢ = nᵢ/sᵢ², the weighted means produce Welch’s F* statistic, and the reference F distribution gets a fractional denominator degrees of freedom from the Welch (1951) approximation.

When the omnibus test rejects, Games-Howell pairwise comparisons—Welch t statistics with Satterthwaite degrees of freedom referred to the studentized range—identify which groups differ, with multiplicity control built in.

This calculator supports two to twelve groups and reports an ω²-style effect size alongside the omnibus result.

Use the Welch ANOVA calculator whenever group spreads differ or sizes are unbalanced; read the Games-Howell comparisons only after a significant omnibus result.

When to use this Welch ANOVA calculator

Use it when

Use it as the default replacement for classical one-way ANOVA: it costs almost nothing when variances happen to be equal and remains valid when they are not, which removes the need for a variance pre-test entirely.
Use it when group spreads visibly differ, particularly when group sizes are also unbalanced—the combination in which classical ANOVA’s pooled error term is most vulnerable.
Use it with Games-Howell follow-up when the research question is which groups differ, since those comparisons inherit the same no-equal-variance philosophy pair by pair.

Choose another method when

Avoid it for repeated measures or blocked designs—independence between groups is still assumed, and within-subject structures need their own machinery.
Avoid it when group distributions are heavily skewed with small samples; Welch fixes the variance assumption, not the normality one, and the Kruskal-Wallis rank test may represent such data better.
Avoid reading the omnibus rejection as “all groups differ”; it licenses the pairwise Games-Howell reading, which frequently flags only some pairs.

Interactive tool

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

Example preview

Means under unequal variances

Inputs: Three groups with visibly different spreads; Welch weighting.
Representative result: Welch F≈22.98 with df=(2, 9.9) and p≈0.00019; Games-Howell flags every pairwise gap.

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

Watch the explanation

3:52 min

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

Read the complete transcript

Welch ANOVA. Welch ANOVA compares two or more independent group means without assuming equal variances. It targets location while letting every group keep its own uncertainty. This page accepts two to twelve groups, at least two finite observations per group, positive sample variance everywhere, and a pre-specified alpha. For group i, compute its size, mean, and sample variance. The precision weight is sample size divided by variance, so stable groups count more. Combine group means using those weights. The resulting weighted grand mean moves toward precise groups rather than treating every group as equally informative. Welch F compares weighted mean separation with a heterogeneity correction. The numerator degrees of freedom are k minus one; the denominator degrees of freedom are usually fractional. Only the upper F tail is used. Rejection says at least one group mean differs, not that every mean or every pair differs. If Games Howell is selected and the omnibus test rejects, each pair gets its own Welch t, Satterthwaite degrees of freedom, and multiplicity-adjusted p-value.

Use Welch as a planned mean comparison when spreads may differ, especially with unequal group sizes. A preliminary variance test is not required. Unequal-variance robustness is not outlier robustness. Extremes still change means and variances, while dependence, clusters, heavy skew, and selection problems remain unresolved. Now compare three machines with sizes seven, six, and eight. Their means are thirteen point two seven, eighteen point three eight, and fifteen point three four. Their sample variances are one point four two two, one point nine eight two, and point five five seven. Machine Three receives almost sixty-four percent of the precision weight. The weighted grand mean is fifteen point two nine five. The resulting Welch F is twenty-two point nine eight four with degrees of freedom two and nine point nine two. The critical value is four point one one four and p is point zero zero zero one eight nine. Reject equal means; the approximate omega-squared style effect is point six seven seven. Games Howell marks all three pairs significant here, but with different pairwise degrees of freedom and adjusted p-values. The omnibus result alone did not provide that detail.

Add one hundred to every observation and Welch F is unchanged. Multiply every observation by two and the common scale cancels from all weights and differences. Center every group at the same mean while preserving its variance. F becomes essentially zero and p equals one, proving unequal variances alone are not the target. Append forty to Group One and the test still rejects, but for changed means and variances. Welch does not identify whether an extreme is scientifically valid. Fail validation for undersized groups, nonfinite values, invalid alpha, or any zero sample variance. Never invent infinite precision for a constant group. Report the independent design, group sizes, means, variances, Welch F, both degrees of freedom, upper-tail p, alpha, effect estimate, and selected follow-up. Interpret practical differences from the group means and context, not p alone. The omega-squared style value is approximate here and has no confidence interval. Use repeated-measures ANOVA for paired conditions, Kruskal Wallis for rank-based independent groups, and graphics for outliers. Run Welch ANOVA free at Distri Scope dot com.

How to read the result

The fractional df₂ is not an error: Welch’s approximation produces non-integer denominator degrees of freedom, which the F distribution accepts directly; the overall reference remains an approximation.

Precision weighting means small, noisy groups influence the weighted grand mean less—groups are not treated as equally informative, which is the method’s core honesty.

In the Games-Howell table, each pair carries its own degrees of freedom; a pair of tight groups can reach significance while a pair involving a noisy group does not, at identical mean differences.

How to use the Welch ANOVA calculator

Enter each group in its own field (2 to 12 groups); every group needs at least two observations and positive variance, and sizes may differ freely.

The post-hoc selector controls Games-Howell; comparisons run only after a significant omnibus result, never before.

The group summary reports each group’s mean and variance—the raw material of the precision weights.

Formula, hypotheses, and assumptions

F^* = \frac{\sum w_i(\bar{x}_i-\bar{x}_w)^2/(k-1)}{1+\frac{2(k-2)}{k^2-1}\Lambda}, \quad w_i = n_i/s_i^2

Conditions to review

Independent observations within and between groups
Approximate normality within each group
No equal-variance assumption — that is the point
Every group needs positive sample variance

Calculator parameters

Significance Level (α): default 0.05.
Post-hoc Test: default Games-Howell.

What the method is doing

The omnibus statistic follows Welch (1951): F* = [Σwᵢ(x̄ᵢ−x̄w)²/(k−1)] / [1 + 2(k−2)Λ/(k²−1)] with df₂ = (k²−1)/(3Λ), where Λ sums (1−wᵢ/W)²/(nᵢ−1); the implementation is pinned against an independent scipy computation.

Games-Howell compares each pair with t = (x̄ᵢ−x̄ⱼ)/√(sᵢ²/nᵢ+sⱼ²/nⱼ), Satterthwaite degrees of freedom, and p-values from the studentized range at q = |t|·√2—verified against scipy’s studentized_range distribution.

The reported effect size is the standard ω²-style estimate df₁(F−1)/(df₁(F−1)+N), floored at zero.

Worked example: three groups with unequal spreads

Three independent groups with visibly different variability are compared. Group 1: 12.1, 14.3, 11.8, 13.5, 12.9, 15.2, 13.1. Group 2: 18.4, 16.2, 19.1, 17.5, 20.3, 18.8. Group 3: 14.9, 15.5, 16.1, 14.2, 15.8, 16.4, 15.1, 14.7. Run Welch’s ANOVA at α = 0.05 with Games-Howell follow-up.

1Compute group means (≈13.27, 18.38, 15.34) and variances (≈1.40, 2.07, 0.60); the largest variance is more than triple the smallest, exactly the anti-pooled-ANOVA situation.
2Form precision weights and the Welch statistic: F* ≈ 22.98 with df = (2, 9.9).
3The p-value is ≈ 0.00019, so the hypothesis of equal means is rejected decisively; the ω²-style effect is large.
4Games-Howell flags all three pairs: group 1 vs 2 (t ≈ −6.999, df ≈ 9.9, p ≈ 0.0001), group 2 vs 3 (p ≈ 0.005), and group 1 vs 3 also significant—each with its own Satterthwaite df.

Interpretation

The three production settings differ in mean output, and the differences survive a method that never assumed their spreads were comparable. Because every pairwise comparison carries its own degrees of freedom, the evidence strength varies by pair even though all three reject—the kind of nuance a pooled Tukey analysis would have flattened, or gotten wrong, under these unequal variances.

Common mistakes

Do not run classical ANOVA first and switch to Welch only on a failed variance test; the two-stage ritual distorts error rates, and Welch-by-default avoids it.
Do not compare Games-Howell p-values across pairs as if on one scale of “effect strength”; each pair has its own df and its own power.
Do not treat the method as outlier-proof; single extreme values still distort group means and variances, and robustness to outliers is a different property than robustness to unequal variances.

Limits and independent validation

Normality within groups is still assumed for the F approximation; heavy skew at small n degrades the p-value.

The ω²-style effect size is an approximation carried over from the classical framework; no confidence interval for it is reported.

Trimmed-mean and bootstrap variants of heteroscedastic ANOVA are not implemented.

Before using the result

Verify the omnibus against an independent implementation (R’s oneway.test with var.equal = FALSE, or the Welch formulas in scipy) — F, df₂, and p should agree to several decimals.
Check one Games-Howell pair by hand: the Welch t and Satterthwaite df use only the two groups’ means, variances, and sizes.
Compare with classical ANOVA on the same data; a large discrepancy is itself evidence that the equal-variance assumption was doing real damage.

one way ANOVA →Use the classical pooled version when variances are genuinely comparable.two sample t test →See the k = 2 special case, where Welch’s t is already the default.

← Brown-Forsythe Test Repeated Measures 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.