Dear Jörg,
I would recommend that you set up a single GLM in a slightly
different way from the 2x2 ANOVA example, and then test
everything directly with different contrasts. If you have
EV1 model AX, EV2 model AY, EV3 model BX, and EV4
model BY then you can easily do the tests you are interested in.
By "model" I mean that EV1 should have a one for each instance
of an AX datapoint and a zero for all others. This then associates
EV1 with the average of AX, EV2 with the average of AY, etc.
So testing AX > BX would boil down to a contrast that is:
1 0 -1 0
or the opposite sign for AX < BX.
If you want a two-sided test just put an F-test on the above contrast.
Alternatively, if you want the standard A > B then you can use
the contrast [ 1 1 -1 -1 ]. Similarly, use an F-test for a 2-sided test.
This way you benefit from estimating the variance from all the
datapoints and there is no need to do extra steps or to worry
about different skeletons.
All the best,
Mark
On 4 Jun 2012, at 18:50, Joerg Trojan wrote:
> Dear Stephen,
>
> thanks for the hint. fslsplit and fslmerge work well.
>
> I'm still struggling with my second question: I intend to use the mean skeleton based on the full sample for conducting the statistical tests in the subgroups. Any opinion on that?
>
> Thanks,
> Jörg
>
|