Dear Valeria,
>I should see a time by group interaction where only
>the scan 2 of task 1 for the case will be different from scan 1 of case
>and scan 1 and scan 2 of controls. I am not particularly interested in the
>main time effect.
>That is why I thought of doing this type of coding:
>
>
> day 1: scan 1 1
>case
> day 6: scan 2 -3
>
> day 1:scan 1 1
>controls
> day 6:scan 2 1
>
>
>Is it more clear now?
Thank you; it is much clearer now and it sounds like a fascinating
study. As you say, you are really interested in a time by group
interaction. The contrast that you have applied is not quite right
for the effect that interests you. In particular, it will show
voxels in which there is a main effect of time (which you do not
want). If, for example, a given voxel gave values of 1 -1 1 -1 in
the four conditions (i.e. it just showed a main effect of time), then
the value for day 6 scan 2 would be -1, whereas the average of the
other three values would be +1/3, so this voxel would show up in your
contrast (according to the criterion given in my previous message).
You really want differences between day 6 and day 1 for the case
group which do not occur in the control group. Since the main effect
of time (or session) is the same in both groups, this can be removed
by comparing the effect of time between the two groups.
Your contrast should therefore be -1 1 1 -1, using the order of
covariates as listed above. Except that obviously this contrast has
to be applied in parallel to every subject. Thus, if you had four
subjects in each group, and you listed all of your case subjects
first, and then your control subjects, the contrast would in fact
look like this: -1 1 -1 1 -1 1 -1 1 1 -1 1 -1 1 -1 1 -1. Obviously
you should have a look at the inverse of this as well (1 -1 1 -1 ...
etc.).
This doesn't explain why SPM99 listed the contrast which you tried as
invalid. If you were using fMRI, then I would wonder if the problem
was the 'session effects' which are added by SMP99 (if these were the
same as your effects of interest then the parameter estimates would
not be estimable). But I don't imagine that PET models work in this
way, although this probably just shows my ignorance of PET, a method
which I have never used. Even though the contrast which you tried
was not really the right one anyway, I think that you should still
try to get to the bottom of why the other contrast was listed as
invalid, as you might learn something important about your design,
Best wishes,
Richard.
--
from: Dr Richard Perry,
Clinical Lecturer, Wellcome Department of Cognitive Neurology,
Institute of Neurology, Darwin Building, University College London,
Gower Street, London WC1E 6BT.
Tel: 0207 679 2187; e mail: [log in to unmask]
|