I am receiving a result that seems very counter-intuitive using the Dunn's test. My data is illustrated in the plot below. I have 5 groups (labelled muscle1a, muscle1b, muscle2, muscle3 and muscle4).
The groups are not normally distributed, so I used a Dunn's test with a Bonferroni correction to compare the values between them. (The statistics were done in python using the scikit_posthocs package). The output p values are here:
I received the result that there was no difference between 1a and 1b. I found this very surprising since there is no overlap of 1a with 1b. Since as I understand it the Dunn's test is based on rankings, how can this not lead to a different result? It seems especially surprising when considering that the comparison between Muscle1b and Muscle2 has a much lower p value (0.058) than the comparison between 1a and 1b (0.796), despite the fact that there are many more overlapping ranks between 1b and 2 than between 1a and 1b.
Could someone help with what I am missing here? Am I doing something wrong, or is my understanding of how the test works wrong? And if so, intuitively looking at this data the 1a group seems different to me than the rest of the groups, so what statistical test would be more appropriate to capture that difference?
