[R] post-hoc test with kruskal.test()

[Thomas Lumley]

On Wed, 14 Oct 2009, Meyners, Michael, LAUSANNE, AppliedMathematics wrote:

> Robert,
>
> What do you mean by "not symmetric"? If you mean unbalanced in terms of sample size, that's not a
> problem if you choose the right specifications for wilcox.test. The Kruskal-Wallis-Test is a
> generalization of the unpaired Wilcoxon test for more than two groups. Not sure whether
> kruskal.test works with just two groups, but if so, it should give the same results as wilcox.test if
> you set the arguments accordingly.
>
> Having said that, I should mention that unlike some normality-based post-hoc tests, the proposed
> approch is not based on a common error term. The paired comparisons will ignore the fact that you
> had a third group, and this will in particular result in (possibly quite) different power of the three
> comparisons, depending on the sample sizes and the noise given in just these two groups. I
> wouldn't know what to do about that, though.


It's worse than that: you don't necessarily even get the test in the same *direction* when you ignore 
the third group, though it takes some effort to produce a good example.  There's a nice paper by 
Brown & Hettmansperger in ANZ J Stat a few years ago where they look at the decomposition of the 
KW test into paired tests and 'non-transitivity' components.

        -thomas

>
> Michael
>
>> -----Original Message-----
>> From: Robert Kalicki
>> Sent: Mittwoch, 14. Oktober 2009 14:11
>> To: Meyners,Michael,LAUSANNE,AppliedMathematics
>> Subject: RE: [R] post-hoc test with kruskal.test()
>>
>> Hi Michael,
>> Thank you very much for your clear and prompt answer.
>> Is it still valid if I use an unpaired comparison with
>> wilcox.test() since my groups are not symmetric.
>> Many thanks
>>
>> Robert
>>
>> -----Message d'origine-----
>> De : Meyners,Michael,LAUSANNE,AppliedMathematics
>> Envoyé : mercredi 14 octobre 2009 10:30
>> À : Robert Kalicki; r-help at r-project.org Objet : RE: [R]
>> post-hoc test with kruskal.test()
>>
>> Robert,
>>
>> you can do the corresponding paired comparisons using
>> wilcox.test. As far as I know, there is no such general
>> correction as Tukey's HSD for the Kruskal-Wallis-Test.
>> However, if you have indeed only 3 groups (resulting in
>> 3 paired comparisons), the intersection-union principle and
>> the theory of closed test procedures should allow you to do
>> these test without further correction, given the global test
>> was statistically significant.
>>
>> HTH, Michael
>>
>>
>>
>>> -----Original Message-----
>>> From: r-help-bounces at r-project.org
>>> [mailto:r-help-bounces at r-project.org] On Behalf Of Robert Kalicki
>>> Sent: Mittwoch, 14. Oktober 2009 09:17
>>> To: r-help at r-project.org
>>> Subject: [R] post-hoc test with kruskal.test()
>>>
>>> Dear R users,
>>>
>>> I would like to know if there is a way in R to execute a
>> post-hoc test
>>> (factor levels comparison, like Tukey for
>>> ANOVA) of a non-parametric analysis of variance with
>>> kruskal.test() function. I am comparing three different groups. The
>>> preliminary analysis using the kruskal-wallis-test show
>> significance,
>>> but I still don't know the relationship and the significance level
>>> between each group?
>>>
>>>
>>>
>>> Do you have any suggestion?
>>>
>>>
>>>
>>> Many thanks in advance!
>>>
>>>
>>>
>>> Robert
>>>
>>>
>>>
>>>
>>>
>>> ___________________________________________
>>> Robert M. Kalicki, MD
>>>
>>> Postdoctoral Fellow
>>>
>>> Department of Nephrology and Hypertension
>>>
>>> Inselspital
>>>
>>> University of Bern
>>>
>>> Switzerland
>>>
>>>
>>>
>>> Address:
>>>
>>> Klinik und Poliklinik für Nephrologie und Hypertonie
>>>
>>> KiKl G6
>>>
>>> Freiburgstrasse 15
>>>
>>> CH-3010 Inselspital Bern
>>>
>>>
>>>
>>> Tel     +41(0)31 632 96 63
>>>
>>> Fax    +41(0)31 632 14 58
>>>
>>>
>>>
>>>
>>> 	[[alternative HTML version deleted]]
>>>
>>>
>>
>>
>
> ______________________________________________
> R-help at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
>

Thomas Lumley			Assoc. Professor, Biostatistics
tlumley at u.washington.edu	University of Washington, Seattle