Hi Don -
Jostling my elbow, eh? ;)
Let's see if I'm understanding you correctly - you are saying I should not be modeling individual subjects in EVs 8-14? The reason it is necessary to do so is because this is a bit more than your standard fixed effects analysis of between-subject variables (as in the fsl wiki example http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/GLM#ANOVA:_3-factors_2-levels_.283-way_between-subjects_ANOVA.29). Instead, this design is meant as a mixed ANOVA that incorporates fixed and random effects from between-subject (group and gender) and within-subject (time) factors. Building in EVs to model individual subjects is needed to account for sameness within subjects across sessions (as in the other fsl wiki example http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/GLM#Single-Group_Paired_Difference_.28Paired_T-Test.29). The question is how to do this in a single design matrix. At the moment it doesn't seem to be working (maybe because of your N-1 issue, but I don't see what that wouldn't also be an issue for the linked paired t-test example).
If I've misinterpreted your message and you are instead referring to a cell in the sense of a single combination of each factor level (e.g A1B1C1...) then I actually have two subjects per cell here (2 females in group A1 with both time points, 2 males in group A1 with both time points, 2 females in group A2 with both time points, and 2 males in group A2 with both time points) so the N-1 issue shouldn't be a problem in this simplistic 8 person toy model.
Also, I'm not sure how FSL estimates variance components in higher level analyses. I have a feeling my solution might lie in using a more sophisticated technique than the default ordinary least squares method of variance estimation (as in tps://www.fmrib.ox.ac.uk/analysis/techrep/tr01cb1/tr01cb1/node9.html)... but I have no idea how to force FSL to so do.
Thoughts?
Jessica
|