Antonia,
> I am trying to look for a conjunction between a regression and a main
> effect at the second level. When I do the regression alone, I model the
> factor I am interested in (called bias) and a constant term using the
> simple regression option in SPM99 (design A attached) and I find several
> activations. Then I put exactly the same images into a bigger design
> matrix (using multiple regression w/o constant), in which the upper left
> corner is identical to my regression design, the lower right corner is a
> constant term for my additional effect and the remainder is filled in with
> zeros (design B attached). Using Imcalc, I have confirmed that the two
> designs give the same beta images and con images for the first column of
> each design. But when I use a 1 0 or 1 0 0 contrast to look at the voxels
> with activity related to the first column, they are completely different.
As you have figured out (rest of message omitted), the difference is
in the variance estimate. In model A you just have one group
of subjects, in model B your estimate of residual variance is pooled
over both the regression subjects and the 2nd group of subjects.
The dramatic difference tells you that you have a dramatically different
pattern of intersubject lack-of-fit between the two groups. I would
analyze each group of subjects separately and look at the two residual
standard deviation images (use ImCalc & 'sqrt(i1)' to create ResStd imgs).
I'm sure you'll find marked differences in the variance in the areas
where you found activation in the regression.
This is an example of where one might want to do a manual conjunction.
I'll put that in a separate message in response to your following
message.
-Tom
-- Thomas Nichols -------------------- Department of Biostatistics
http://www.sph.umich.edu/~nichols University of Michigan
[log in to unmask] 1420 Washington Heights
-------------------------------------- Ann Arbor, MI 48109-2029
|