Dear Gang Chen,

Thanks for your interest in my question.
My five DTI datasets with each 8 subjects originate from five different software pipelines with different processing steps turned on and off (say for instance the b-table rotation based on subject motion was done in one pipeline and omitted in the other).
I had 8 participants, so I processed their data in the same way for one pipeline giving me a whole dataset of 8 subjects for each pipeline.

All data was registered to the FMRIB FA 1mm template, so it is ensured that I can perform comparisons and no spatial problems will arise. I went further with the processing later on to generate skeletonized FA and MD reproducibility error maps.
We expect there will be differences in the subject group's test-retest errors skeleton maps according to which pipeline was used. I already know the data is non-normally distributed with unequal variance, so Kruskal-Wallis is the test I should perform.

The problem with the zeros is, that every estimated subject's skeleton is slightly different from the next, so I would like to avoid comparing a higher reproducibility error with a reproducibility error of zero on the group level. I was wondering if then my method of avoiding this was correct and valid for 3dKruskalWallis.

Thanks a lot, all the best,
Barbara