Statistics Glossary → Welch correction
Welch correctionconcept
A method that corrects the standard error and degrees of freedom of the t-test and ANOVA when the group variances are not equal.
The classical (Student) t-test pools the two groups' variances into one, assuming both groups are equally spread. If one group is far more spread out and the group sizes differ, pooling is misleading: the p value may come out too small or too large.
Welch's approach uses each group's variance separately and recomputes the degrees of freedom with the Welch-Satterthwaite formula, which is why they are usually not a whole number (e.g. 169.71). Because it changes the result very little when the variances are equal, many statisticians recommend Welch as the default.
An everyday example
A factory compares the length of screws made by two machines. The new machine produces screws of almost exactly the same length, while the old machine's screws are sometimes short and sometimes long. Because the two machines differ so much in spread, the Welch correction is used to compare their means.
In MerQur
In the independent-samples t-test, the 'Variance' option: Automatic (decided by Levene's test), Equal (classical t) or Welch. In one-way ANOVA it is offered as Welch ANOVA.
Analyses where this concept appears
Independent-Samples t-TestOne-Way ANOVA