Dear Martin,
On 06/01/17 13:05, Martin Dyrba wrote:
> Dear Guillaume,
>
> Thank you for your reply. Actually, I talked to a colleague who gave me some advice how to properly enter the contrasts:
> Main effect of time: 1 -1 1 -1
> Interaction of group by time: 1 -1 -1 1
> ...
Yes, these contrasts are fine, when you look at differential effects of
time, but it gets more complicated for the main effect of group (ie when
you average over time) as 1 1 -1 -1 is not an estimable contrast neither
the one you suggested with non zero weights on the subject effects:
1/2 1/2 -1/2 -1/2 0 0 0 1/N1 ... 1/N1 -1/N2 ... -1/N2
That one would be estimable if you removed the subject-specific
covariates. And, again, in this model, these extra regressors are anyway
redundant and don't increase the explanatory power of the model (compare
rank(SPM.xX.X) with and without them).
> When I saw it right, Donald MacLaren stated various times that assessing the group effect using such a design does not give valid estimates... However, we are more interested in the interaction of group by time etc. so that this should not be a problem.
If, at some point you want to test for the main effect of group, best is
to form the partitioned error model, being here a two sample t-test (and
where the covariates will not be redundant any more).
> Finally, I wondered if the results change if I add a main effect for the group factor to the design?
It shouldn't change the overall model and will only increase redundancy
in the design (again, compute rank(SPM.xX.X) with and without it). It
will also unnecessarily make your life more difficult to define some of
the contrasts.
Best regards,
Guillaume.
> Best regards,
> Martin
>
> PS: Using your MATLAB code gives the following visualization:
>>> load SPM
>>> X=SPM.xX.X;X(X&X~=1)=2;c=' |~';
>>> c(X+1)
> | ~|~|
> | ~|~|
> | ~|~ |
> | ~|~ |
> | ~|~ |
> | ~|~ |
> | ~|~ |
> | ~|~ |
> | ~|~ |
> | ~|~ |
> | ~|~ |
> | ~|~ |
> | ~|~ |
> | ~|~ |
> | ~|~ |
> | ~|~ |
> | ~ ~ |
> | ~ ~ |
> | ~ ~ |
> | ~ ~ |
> | ~ ~ |
> | ~ ~ |
> | ~ ~ |
> | ~ ~ |
> | ~ ~ |
> | ~ ~ |
> | ~ ~ |
> |~ ~ |
> | ~ ~ |
> |~ ~ |
> with columns: g1t1, g1t2, g2t1, g2t2, TIV, Site, Age, Subject1, Subject2 ... Subject 15
>
--
Guillaume Flandin, PhD
Wellcome Trust Centre for Neuroimaging
University College London
12 Queen Square
London WC1N 3BG
|