Stephan,
I think that conjunction could be used here, but I actually prefer the
approach you give for its simplicity. But how you proceed depends on
the contrasts.
> Assume that I have studied 12 subjects, which provided me with 12
> contrast A and 12 contrast B images respectively. I would like to
> test for an effect of contrast A in only those voxels that also show
> an effect in contrast B.
> For each subject, the following criteria should apply at each voxel:
> Contrast A: effect, Contrast B: effect - include
> Contrast A: effect, Contrast B: no effect - exclude
>
> Is it possible to limit the one-sample t-test of the 12 contrast A images
> to only those voxels that, at the individual subject level, also show an
> effect in contrast B?
If contrast estimates A and B are independent (i.e., the contrast vectors
are orthogonal, e.g. A: [1 0] B: [0 1], or A: [-1 1] B: [1 1]), then
you can use a mask based on contrast B to restrict inference on
contrast A (this is OK, since B tells you nothing about A). Easiest
way to do this would be with SVC using a image mask you've created by
thresholding an spm_T with ImCalc.
If the contrasts are not independent then you can't do this, and
you're limited to just doing 'visualization masking' (as Dave
McGonigle put it) in the results contrast specification stage.
-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
PS: The whole conrast independence issue gets sticky under a random
effects model, where even orthogonal within subject contrast estimates
may be dependent due to dependence of the signal (e.g. contrast A:
effect of verbs, contrast B: effect of nouns). As with the
conjunction...
http://www.jiscmail.ac.uk/cgi-bin/wa.exe?A2=ind0212&L=spm&P=17956
...under the Null the test will be valid. Under positive dependence
between the random effects the significance will be inflated, under
negative dependence the signifance will be underestimated. (For
comparison, visualization masking can only underestimate your
significance.)
|