Hi Alfonso,

Thanks again for the detailed response, I promise the questions will stop soon!

Just to contextualise, we are still trying to resolve the discrepancy in results and my collaborator informs me that they didn't select the "non-parametric" option in CONN and as such I am assuming that the results they are reporting come from the univariate SPM approach. As such, I am just trying to understand this method fully so we can decide what to do.

I just want to check about the implementation of the "two-stage" approach as the guidance given in the attached paper (and by Will Penny on the SPM Wiki) do not actually give the correct F-statistics or degrees of freedom when you have a within-subject factor with > 2 levels. I have been in discussions with Guillaume Flandin at the FIL about this and I can send you scripts and examples to show this if you want.

For now I just want to clarify how the model works in CONN because in order to get an equivalent to the multivariate test across components (i.e. when the M contrast matrix is the identity) you would still need multiple components in the same model and thus would still require the inclusion of the subject factor in the design matrix in order to get the correct degrees of freedom and compute the correct error term for the F-tests. Is this being done in CONN?

Best wishes,
Martyn