Why Kruskal-Wallis? And why 42 groups? I guess that technically it is possible, but what would you do with a "statistical significant" result (if such a result is possible at all in your study)?
With kind regards
K.
Is there a maximum number of groups in Kruskal Wallis' one-way ANOVA? I am looking at a data set with 42 separate groups and I am unsure if there would be some sort of issue in running that many groups under Kruskal Wallis.
I remember reading once that ANOVA should not be ran with more than 26 groups although I am not sure what the logic behind that number exactly is
Why Kruskal-Wallis? And why 42 groups? I guess that technically it is possible, but what would you do with a "statistical significant" result (if such a result is possible at all in your study)?
With kind regards
K.
I thought KW test would be the most appropriate given that the distributions are not normal. There are 42 groups because there are 42 very distinct classifications in a particular categorical variable. A statistical significance would mean that one or more of the groups was significantly different from rest of the groups in terms of the median or the shape of distribution was significantly different depending on whether or not the distributions differ significantly. Are there other things that I should be considering?
Are the shapes similar? That is an assumption of the KW test.
Not quite. It would mean that one or more of the groups are different from each other, not from the rest of the groups.
By comparison: The null hypothesis for a 1-way ANOVA is that all means are equal, while the null hypothesis for an ANOM is that all means are equal to the group mean. These null hypotheses are distinctly different. In ANOVA, two means significantly different from each other, regardless of the other means would provide a low p-value. In ANOM, 1 mean significantly different from the group mean would provide a low p-value. Your wording seemed to imply a KW null interpretation similar to the ANOM.
If your dependent variable is interval scaled and your sample size is large enough
(which must be the case here), then why bother about normality? You can use a
oneway anova then.
Yes. But how helful would this information be for you?A statistical significance would mean that one or more of the groups was significantly different from rest of the groups in terms of the median or the shape of distribution was significantly different depending on whether or not the distributions differ significantly.
Usually, one wants to identify which particular groups differ from
other groups, but you'd have to perform 861 pairwise
comparisons in that case.
With kind regards
K.
You're right. I thought about and decided to change the categorization schemes to create 18 groups to make the posthoc tests more manageable
my understanding is that the shapes do not have to be similar in order perform a kb test but the interpretation of the results would be different in that a significant result indicates a difference in the shape of distribution between the groups at a population level. Perhaps not a very helpful inferential insight, but something, i suppose
Tweet |