How to run a Welch ANOVA
- Copy a group name column and a measurement column together with the header.
- Select the response column and the factor (group) column (3 or more levels). The Games-Howell post hoc test is included by default.
- Press Analyze and check the p-value in the Welch ANOVA summary and the Games-Howell table.
Worked example with sample data
The sample data are strength measurements (MPa) of material from 3 suppliers (10 from Supplier A, 12 from B, 8 from C); the sample sizes and spreads differ by supplier.
The means are 409.4 (A), 404.5 (B) and 393.2 MPa (C), and the standard deviations 3.469, 5.32 and 15.76, so Supplier C spreads a lot. Welch ANOVA gives F = 6.36 with (2, 13.99) degrees of freedom and p-value 0.01085, so the supplier means differ (η² = 0.3566).
In the Games-Howell post hoc test only Supplier A - Supplier B (mean difference 4.897, p=0.0448) is significant. A - C has the largest mean difference, 16.14, but Supplier C's large spread makes it not significant (p=0.05375).
A one-way ANOVA that assumes equal variances gives a much smaller p=0.002599 for the same data. When the spreads differ this much, as the standard deviation comparison chart shows, that p-value is overstated, so use the Welch ANOVA result.
How it is calculated
Each group gets a weight n/s² in the between-group variation, and the statistic is compared with an F distribution whose denominator degrees of freedom (a decimal) are adjusted in the Welch-Satterthwaite way. With unequal group sizes and variances it controls the type I error better than the usual ANOVA.
Games-Howell post hoc test
Each pair uses a Welch t statistic and Welch degrees of freedom, corrected for multiple comparisons with the studentized range distribution. It does not assume equal variances, so it pairs naturally with Welch ANOVA.
What the result shows
Descriptive statistics by group, the Welch ANOVA summary, the Games-Howell post hoc test, a box plot, an individual value plot, means with confidence intervals and a standard deviation comparison by group.
How to read the Welch ANOVA result
- If the p-value is below the significance level, at least one group mean differs; check the 'Significant' rows of the Games-Howell table to see which pairs.
- If the standard deviation comparison shows large differences in spread, this result is the right one to use instead of the usual ANOVA.
- If DF2 (adjusted) is much smaller than the group sizes, the group with the large spread is driving the result.
Frequently asked questions
Which do I trust if it disagrees with the usual ANOVA?
Trust Welch ANOVA when the group variances differ. When the variances are similar, the two results are almost the same.
What if there are 2 groups?
Welch ANOVA needs 3 or more groups. Two groups are the same problem as the 2-sample t test (Welch).
Why are the degrees of freedom a decimal?
They are the Welch-adjusted degrees of freedom, so that is expected.