I am comparing duration of a behaviour between difference age classes (4 different classes) in squirrels for part of my thesis. I am restricted to using non-parametric tests because no transformation seems to normalize my data.
When I run the Kuskal-Wallis test in SPSS I get significant differences between age classes (0.007).
My question: is there a way to determine which age classes are significantly different from one-another without running a bunch of pairwise Mann-Whitney tests?
As far as I know there is no other way in SPSS other than running a 'bunch of pairwise Mann-Whitney tests' (corrected for multiple comparisons).
In GraphPad, the Dunn's test is available for post-hoc comparisons after a Kruskall-Wallis. It might not be too difficult to implement though I would not know how. Alternatively, you can get GraphPad, it is a nice little package which draws much prettier graphs than SPSS!
I hope it helps,
Molly, if you're using a recent version of SPSS (18>) then post-hoc tests are automatically performed for the KW test when the omnibus statistic is significant. Unfortunately, they didn't make this obvious. You need to double click the test summary in your output and in the model viewer that opens click on the drop down menu on the bottom right and select the pairwise comparisons view and scroll down to see the results of the post-hoc tests.
Thank you for the helpful suggestions. I unfortunately only have 11.5 on the lab computers. However, I was able to download a free Graphpad trial for 30 days. I am using that for now and it has options to perform the Dunn's post test which is perfect for what I am needing.
I just found about that way of post-hoc-Kruskal-Wallis-tests and am wondering now which post-hoc-test SPSS performs there and what's the difference between significance and adjusted significance. Since I ran Mann-Whitney-U-tests and also did Bonferroni-corrections, it doesn's seem to correlate exactly with one of them. I'm really helpless at the moment and I'm looking forward to a smart answer from your side ;-)
but since I'm not a Statistics-Crack and I'm just using it for the analysis of my experiments, I'm wondering which value I should take now. Since I did the Bonferroni correction for my data, which should be more conservative than the post-hoc-tests of the Kruskal-Wallis, and it gives me more significant values than the Kruskal-Wallis-post-hocs, I'm really unsure which is the best way to interpret my data. Of course I want to have significant results, but they shouldn't be false positive. Thanks in advance for every useful hint!
I must admit that now I'm stuck. If you make all pairwise comparisons
using two-tailed tests then indeed I would expect that Bonferroni's
p/15 would not give more statistically significant results than Dunn's
procedure. Since Bonferroni correction is still the standard for K-W post
hoc (at least as far as I can see) and since it is conservative, I guess
there's enough reason to stick to Bonferroni in the present case.