Dear Matthew,
> Apologies that this is a well-worn topic, but I remain confused after an afternoon of reading the forum. I need some clarification regarding using randomise to look at un-paired two-group differences (Patients and Healthy Controls). Looking at group differences without covariates, I set up my model and contrasts as below:
>
> Model:
> GRP EV1 EV2
> PT 1 1 0
> HC 2 0 1
>
> Contrast:
> 1 -1
> -1 1
> 1 0
> 0 1
>
> I got some reasonable findings.
>
> Yet, when I went to add nuisance covariates (by adding EVs 3 and 4) I got the “non-separable EVs” error message. After reading through this forum, I changed my setup for the covariate analysis to look like below (change being group assignment, in bold):
>
> GRP EV1 EV2 EV3 (score1 demeaned) EV4 (score 2 demeaned)
> PT 1 1 0 4 0.5
> HC 1 0 1 -2 -0.25
>
> Contrasts:
> 1 -1 0 0
> -1 1 0 0
> 1 0 0 0 (grp mean)
> 0 1 0 0 (grp mean)
>
> I have three questions
> 1) It seems one (or both!) of these ways of setting it up is incorrect, since one declares two different groups and one does not. I am most interested, ultimately, in the more complex covariate part.
Your EV's 1 and 2 are what specifies these as two groups with potentially different means (that you then want to test using contrasts). The GRP indicator on the other hand tells FEAT that these are two groups that might potentially have different variances, and that we should therefore try to estimate separate variances for them. If you specify two (or more) groups as having potentially different variances you cannot have EV's that span both groups (that is why your first example doesn't work and your second does).
>
> 2) Is this somehow different than what I would do if I were using FEAT?
I _think_ that you cannot actually model different group variances in randomise, but I might be wrong here so maybe Matthew/Tom can pitch in if I am. In randomise there is an optional separate design.grp file which specifies the groups that it makes sense to permute the labels within, which is a different matter.
My suspicion is that in the case above where you specified a GRP EV it was simply interpreted as "yet another EV" by randomise. But I am, as I said, not certain about this.
> 3) Is there a way I can throw in a third covariate (of interest) to this model to see how strongly it correlates with activation differences or with one group's activation? (Or do I need to do this in a separate analysis, as currently planned.)
I am not sure what you want here. Do you want to look at group-by-covariate interaction, i.e. "does the correlation with my covariate differ between groups"?
If so, you can (and probably should) put it in the same analysis. You remember to split your covariate between the groups.
Good luck Jesper
|