Dear Yvonne,

I saw your first message and meant to reply, but Tim & Joe have done most of the work. Just to clarify/amplify:

The product of p-values approach only works (as Tim said) to test the 'global' null hypothesis (no effects real), not the conjunction null (as many as all but one effect real).
Testing the conjunction null works by creating a min z image, and making inference on it just as if it were a regular z image. This applies to uncorrected voxel-wise and corrected voxel- and cluster-wise inferences.

As a stab at this, take a look at the attached easythresh_conj. If you supply it two z images, it will do conjunction inference on them and give you the standard easythresh output.

easythresh_conj stats/zstat1 stats/zstat2 mask 2.3 0.05 example_func grot"

If you look at the code, all it does is take the min of the two z images and then do all the other easythresh stuff on the min.

One detail is that easythresh has to estimate the smoothness from the statistic image itself, which results in varying P-values for the same cluster sizes depending on the z images used. If you have a single feat directory from which the images come from, try using the -s option to specify the smoothness file (stats/smoothness), which will give more accurate (and contrast-independent) P-values. E.g.

easythresh_conj -s stats/smoothness stats/zstat1 stats/zstat2 mask 2.3 0.05 example_func grot"

Lastly, note that this is a conservative procedure... i.e. proper conjunction inference has to account for the worst case scenario, where one statistic image is wildly significant, and the other one null. Because of this, you may find it hard to attain significance with this method. Just a warning.

Let me know if you have any troubles with it.

-Tom

On Wed, Feb 13, 2008 at 9:10 PM, Yvonne Brehmer <[log in to unmask]> wrote:

Dear all,

I asked a few days ago a question regarding conjunction analysis - hopefully
it a not too trivial one.
I post the same question again in the hope that someone out in the FSL
community can help me with it.

I'm comparing two groups in a pre-post training experiment and would like to
identify brain regions that are activated in both groups for different
condition contrasts. I read in the FSL list about the simple way of
inclusive masking and that this is a more descriptive method instead of a
statistical test. However, in the current FSL version, is it possible to run
conjunction analysis between contrasts that provide me with statistical
information as copes?

Thanks for your help,
Yvonne

____________________________________________
Thomas Nichols, PhD
Director, Modelling & Genetics
GlaxoSmithKline Clinical Imaging Centre

Senior Research Fellow
Oxford University FMRIB Centre