Hello everyone,
I would like to perform a post-hoc Dunn's multiple comparison test
following a Kruskall-Wallis test. I've already used multcomp package and dunn.test package but I obtain different results. Even using the "mixtox" library with the same data, the result is different. Can you please suggest a way to use Dunn's test after K-W?
Thank you in advance!
How did you use multcomp for Dunn test? I think you applied Dunnett's test in multcomp instead of Dunn's test. Even it seems that Dunnett's test is implemented in mixtox package. So dunn.test package is the correct package to use.
An appropriate test is, like the Kruskall-Wallis test, non parametric. The PMCMR package contains a suitable (assume that you called your data spray then use something like:
> posthoc.kruskal.nemenyi.test(x=count, g=spray, dist="Tukey")
You will probably have to install the PMCMR package from CRAN.
The article at,
https://cran.r-project.org/web/packages/PMCMR/vignettes/PMCMR.pdf
describes the test, provides a bibleography, and might give you additional ideas.
https://cran.r-project.org/web/packages/PMCMR/vignettes/PMCMR.pdf
Thank you everyone!
When I use dunnTest (FSA) I obtain different results from dunn.test even though they should be the same since I used "two.sided=FALSE". I will try again! And I will use nemenyi.test as well.
Thank you again for all of your answers.
Zar, in Biostatistics (2010), mentions these tests. He notes that the Nemenyi is applicable only if there are equal sample sizes.
I am now using posthoc.kruskal.nemenyi.test. After running it, I got a warning message:"Ties are present, p-values are not corrected". I think they can be present because I have several observations that share the same value.
Do you think I can correct it?
Zar (5th ed. section 11.5) doesn't have a correction for ties for the Nemenyi test.
In my experience, the Dunn test is the standard post-hoc test, probably because it can handle both ties and unequal sample sizes. You may want to check out that section in Zar.
When in doubt, you can always go back to the original literature and compare the procedure to that in the functions. Functions for nonparametric tests are usually relatively easy to follow.
If you have questions about the implementation of Dunn test in FSA, I might also suggest emailing the author. He's pretty good about being responsive (if he isn't swamped with work, etc.)
Maybe you can reduce the ties by eliminating rounding in your raw data input. Although it might be inappropriate for parametric tests in general, it is considered permissible in nonparametric testing.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics