0

As far as I understand, the classic Dunnett test is based on the t-test, applied multiple times to all comparisons vs. control and then corrected. So the assumptions of the Dunnett test must agree with those for the t-test. Normal distribution, homogeneity of variances, equal sample sizes (in the idealized case).

So how it is possible, that various statistical packages, like R or SAS allow to use Dunnett test on completely different models, like GLM, GLS, mixed models, GEE? These models may have completely different assumptions, for example GLM or GEE, which allow for non-normal distribution of the reponse variable or heteroscedascity and does not care about equal sample sizes!

I saw, just to focus the attention, that in R statistical package, there is a function multcomp::glht, for package emmeans, which can perform Dunnett analysis on more than 30 models! How does it relate to the classic Dunnett test? Is this all the same Dunnett procedure? Exactly the same would refer to Tukey procedure. How is that possible to have Tukey (or Dunnett) classic test with its strict assumptions and Tukey (or Dunnett) procedure in those advanced multiple-comparison engines working on very liberal models?

kjetil b halvorsen
  • 63,378
  • 26
  • 142
  • 467
Kazakh
  • 1
  • One thing to be careful of... In `glht`, "Tukey" refers to looking at contrasts that compare each level (e.g. mean) to each other level. It's not about a Tukey test or a Tukey adjustment. And the default p-value adjustment when using "Tukey" there is not a Tukey adjustment. ... On the other hand, `emmeans` can use a Tukey adjustment for the comparison of E.M. means. – Sal Mangiafico Feb 13 '20 at 16:27
  • The glht gives mvt, as far, as I recall (please correct me if I'm wrong), which gives the close-to-exact value based on the multivariate t distribution, which is exactly how it should be addressed. – GibbsSampler10 Dec 09 '20 at 04:22

0 Answers0