Statistics Glossary → Dunnett's test (comparison with a control)
Dunnett's test (comparison with a control)concept
A multiple-comparison test that compares several treatment groups with a single control group only, while controlling the familywise error rate across those comparisons.
With k groups, all-pairs tests such as Tukey make k(k − 1)/2 comparisons, whereas Dunnett makes only k − 1 (each treatment versus the control). Because the comparisons share the same control mean, they are correlated, and Dunnett accounts for this correlation through the multivariate t distribution. At the same familywise error rate it is therefore more powerful than Bonferroni or Tukey.
The test fits when the research question really is 'which treatments differ from the control?'; if the treatments are also to be compared with each other, Tukey or another method is needed. It can be one- or two-sided. The classical version assumes equal variances. The control group should be decided before the analysis.
An everyday example
In a field trial, untreated control plots are compared with four different fertilisers. The researcher is not interested in how the fertilisers differ from one another, only in whether each raises yield compared with no fertiliser. Dunnett's test makes these four comparisons and keeps the overall error rate at 5% for all of them together.
In MerQur
In one-way ANOVA, the 'Dunnett control group' option sets the control as the first or last level in natural order (numeric labels sorted numerically, others alphabetically).