Dear Etienne,
> Thank you for your answer. Determining extend threshold (p-value) in
> conjunction analyses (in minimum fields) appear to be a quite tough
> issue. My basic problem is may be simpler: I have two groups of
> subjects (actually they are the same subjects, but this is two distinct
> fmri sessions) and I wanted to know which regions are involved in the
> same contrast in both groups. A conjunction analysis seems appropriate
> (the two contrasts are orthogonal). My problem is that as a result I
> got clusters including e.g. 1, 2, 10 or 50 voxels. Which extend
> threshold to use to select regions with significant joint activations?
> (in the example, 1 and 2 voxels appear too small, 50 voxels
> significant, and 10 voxels in between). My first answer would be to use
> the same threshold as in single contrast (e.g. 15 voxels, i.e. p<.05),
> because the conjunction has already taken into account the joint test
> (at least on height threshold). However, it seems to be inapproriate in
> some instance: Consider for example that each single contrast activates
> a cluster including 20 voxels and the resulting conjunction analysis
> provides a cluster of 12 voxels within each single-contrast clusters.
> Using the same theshold (15 voxels) would reject this conjoint
> activation, but 12 voxels over 20 voxels which co-jointly activated
> seems to be quite significant. I was thinking that in this instance, we
> indeed use the implicit assumption that we are testing for joint
> activations within given clusters and a standard masking analysis might
> be more appropriate. But, the same spatial extend threshold problem
> appears to occur again, if I am right. What do you think?
The corrected p values are sensitive to the spatial extent threshold
for single contrasts. In other words the p<0.05 corrected height
threshold is lower if you specifiy an extent threhold of 8 voxels than
if you use 0 voxels. For conjunctions you are forced to use 0 voxels
because, as Darren points out, the theory does not exist for > 0.
Therefore even a cluster with 1 voxel in a conjunction SPM is
significant. Remember the conjunction SPM finds the overlap among a
series of component regions. This overlap can be small but very
significant.
I hope this helps - Karl
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|