Kruskal Wallis Test - Minimum Group Size?

#1
I have 23 subjects, and I want to make comparisons of the distribution of a measurement (with non-normal distribution) across several categories. In one category there are 3 groups (n1= 3, n2= 10, n3= 10), and another category 4 groups (n1= 1, n2=3, n3=4, n4=15). I'm using SPSS 20 to undertake Kruskal Wallis testing to make comparisons for each category.

I've carried out the tests, but I'm concerned of their appropriateness. Does SPSS 20 allow for such small minimum group sizes for this? :confused:

If it does, would I be right in thinking though that such small, unequal group sizes within categories means the study is likely to be underpowered to determine a true difference between the distributions of group measurements?

Any help much appreciated!!
 
Last edited:

Dragan

Super Moderator
#2
I have 23 subjects, and I want to make comparisons of the distribution of a measurement (with non-normal distribution

Does SPSS 20 allow for such small minimum group sizes for this? :confused:


Any help much appreciated!!
Well, no, the results that SPSS provides for such small sample sizes are not very good. The reason is that the results that SPSS provides are based on asymptotic chi-square statistics i.e. for large sample sizes and thus more conservative p-values.

In short, you should be using exact critical values for the K-W test. See here:

http://faculty.virginia.edu/kruskal-wallis/table/KW-expanded-tables-3groups.pdf
 
#3
Thanks for that. If I had group sizes of 5 or more could I use SPSS? Have a couple more categories with larger groups I also want to look at. Worked out how to get the H statistic from SPSS to compare to the critical value in those tables for the rest, obviously I can't report an exact p value so I presume I would document p<0.05 or whatever the lowest significance I find in the table?

Would I still be correct in my discussion saying "small, unequal group sizes within categories means the study is likely to be underpowered to determine a true difference between the distributions of group measurements"?

Thanks in advance!