Hi Gang,
It's a within-subject design in two groups though here I'm only considering one group (group comparison soon, e.g. MannWhitney).
I ran 3dWilcoxon but I'm not sure how to interpret this test. My thoughts: 20 participants = 20 ranked pairs in a given voxel and if the pre/post beta median ranks significantly differ, then this voxel shows a group-level scaling effect. Is this correct?
My (awkward?!) use of 3dW:
1) Input are masked betas. The mask is derived from a clustered one-sample 3dMEMA which then is masked by the individual's grey matter. Scaling is carried out in each voxel of a given participant's mask which range between ~ 13000 - 22000 voxels. 3dW outputs only 906 voxels, I assume it's due to the individual masks not overlapping more.
Clustering 3dW for FWE correction again (after clustering 3dMEMA) would be strange. So better is to feed the whole brain data into 3dW and cluster this, which would be a different but related question.
Some analyses I had planned:
1) Within participant:
- Paired t-test on pre/post betas in a given cluster (Could do a Wilcoxon test on each cluster?)
- Coefficient of variation (Could do a quartile coefficient of dispersion?)
3) Between participants:
- Paired t-test on the average of pre/post betas in a given cluster
- Coefficient of variation
Except for the Wilcoxon and quartile coefficient of dispersion, these are affected by extreme post betas. So I could (at least for some analyses) use non-parametric tests or find a proper decision rule on how to remove extreme outliers should I stay with parametric tests (bc e.g. non-parametric may not be available) assuming this is defensible in some way (if so, would e.g. 3 or 4 MAD be a good choice?).
Thanks so much!
Nic​
Edited 2 time(s). Last edit at 08/12/2014 09:17PM by nic.