It sounds like you have sufficient correlation between the BOLD pre-post differences and
the test score to mean that the GLM can no longer uniquely determine the source of
the change. In this situation you will lose significance, as the GLM takes the conservative
approach that if it does not know how to assign the observed data (as a pre-post difference
or as something driven by the test scores) then it will assign it to whatever is not involved in
the contrast being calculated.
I would recommend you try an F-test over contrasts 1 and 3 (as the signs do not matter
for an F-test) in order to see what is jointly explained by either pre-post differences *or*
the test score. This should show you your original activations (obtained before adding the
test scores) and if so it indicates (in conjunction with your t-contrast results) that these
activations cannot be uniquely assigned as resulting from changes induced by the test
score or by the pre-post difference. The interpretation of what this is likely to mean is
then up to you and evidence from other sources and theories. In this situation you cannot
determine from the data alone which is the "true" source.
I hope this helps.
All the best,
On 7 Jun 2012, at 07:06, Vishwadeep Ahluwalia wrote:
> I'm trying to understand the meaning of contrasts used in a paired design with behavioral measure.
> Patients were scanned at pre medication and post medication timepoints. Using a paired ttest i saw differences in activation between the two timepoints. Then i included a psychometric test scores done outside the scanner at both timepoints, demeaned it and added it to the model. contrasts specified were
> 1. Pre>Post 1 0 0 0 0 ...
> 2. Post>Pre -1 0 0 0 0...
> 3. Test+ve 0 0 0 0 1
> 4. Test-ve 0 0 0 0 -1
> The pre>post differences disappear with the test scores in the model and contrast for test+ve and test-ve contrasts show active voxels( not in the same areas ) I'm not completely sure of what conclusion i can draw from these results. Do these results suggest that test score differences are correlated with BOLD differences between the two timepoints? I also did a group mean with and without this test measure and found no significant correlation between test scores and BOLD change. Please enlighten.