I'm running Kruskal Wallis (KW) tests for my dataset, and I'm trying to do post-hoc analysis of my results. Bonferroni, and Dunn's test appears to be the most cited post-hoc test for KW. However, despite searching online, I cannot find relevant post-hoc R scripts for KW.
I also have access to Minitab 17, but unless I'm mistaken, that program doesn't run the desired post-hoc tests for KW.
Lastly, I am wondering if I could use the Mood's Median test (which has less power than KW) as a way to find significance between groups?
I'm attaching some of my data in case someone wants to try running it through R.
Hi Gary,
according to the vignette, the PMCMR package implements the post-hoc tests you mention (plus more) for Kruskal Wallis test: https://cran.r-project.org/web/packages/PMCMR/vignettes/PMCMR.pdf
Mood's median test itself does not provide post-hoc tests. The advantage of Mood's median test is that it actually tests the medians among groups. Kruskal-Wallis tests the medians only if the distributions of each group are similar in shape and spread. Otherwise, it is really testing the distribution of values for each group.
See the Kruskal–Wallis Test chapter in http://rcompanion.org/documents/RHandbookProgramEvaluation.pdf for more explanation.
The first example on the following page gives an example of three groups with the same medians (and means!) that result in a significant result for Kruskal-Wallis because the distributions are different.
http://rcompanion.org/rcompanion/d_06.html
An alternative to the PMCMR package is the dunnTest function in the FSA package. It has a different format to the output, and allows formulae input.
Hi Salvatore,
Thank you for the additional supplementary material. I've only relatively recently completed semester on R, and I'm looking for additional info on running statistics on this program.
Besides this topic, I'm trying to run discriminant analyses on my data. I've found conflicting opinions of using discriminant analysis non-normal distributions. What are the assumptions of discriminant analyses?
Thanks,
Gary
Gary,
I'm not super familiar with DFA, so I'll pass on a definitive answer, but I imagine that it assumes multivariate normality and a few other assumptions. If you can find the multivariate statistics book by Tabachnick and Fidell, that's a good one. Unfortunately I don't have a copy anymore.
Good luck on studying R. I learned on my own over several years, after being reasonably educated on SAS, and found it very frustrating. That's why I wrote the book linked above. I was rather horrified at how few sources even attempted to mention post-hoc tests or, for example, that the anova function by default uses type I sum of squares, which are not desirable in a whole bunch of cases. Hopefully the examples in my book are clear, and the test assumptions, post-hoc analyses, etc. are covered at least well enough.
Hello!
In R, You must download subprogram dunn.test who do Kruskal-Wallis test too:):
dunn.test(variable, variable with group, method = "bonferroni")
If you need to perform more significance tests, I can recommend the DescTools package. It has the dunn test and many others.
Also, it provides the formula interface
DunnTest(x, group)
DunnTest(x ~ group)
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Graphs