Hi Philippe, (hope you don't mind, I posted this to the message board as well)
(previous posts start with: [
afni.nimh.nih.gov])
I don't think you want to multiply two thresholded (p<0.01) t-maps, because who knows what the distribution of t*t looks like. The threshold is not really 0.0001. That's a result of your method. Since your group data is based on separate baseline and activation conditions for each subject group you could do a somewhat more statistical "conjunction" analysis like I outlined in the previous post (http://afni.nimh.nih.gov/afni/phorum/read.php?f=1&i=1791&t=1744).
A better method for two data sets that have already been thresholded would be the following. Let's call the data sets
non-depressed+orig
depressed+orig
Take a thresholded t-map at 0.001 (the p you choose should be "significant" in each map separately). First, use 3dmerge to make maps that contain non-zero values only where p<0.001.
3dmerge -prefix non-depressed_p001 -1thresh 0.001 non-depressed+orig
3dmerge -prefix depressed_p001 -1thresh 0.001 depressed+orig
The output data sets should contain t-scores for voxels where p < 0.001 and zero elsewhere. You should check the results to make sure that any t-score with a p>0.001 is not included in the resulting mask. Then, to create a map of overlap and unique activation for the two conditions, use 3dcalc.
3dcalc -a non-depressed_p001+orig -b depressed_p001+orig -expr "(1*(a/a))+(2*(b/b))"
This will yield a map that has value 1 where only the non-depressed map is significant, 2 where only the depressed map is significant, 3 where both are significant, 0 where neither is significant. Again, this approach is NOT statistical, so I suggest this method to avoid a statistic in the "conjunction" map altogether. The method in the post from earlier is a way to make it statistical, I think, based on the Price and Friston paper.
good luck,
-jim
At 11:43 AM 10/1/2003, you wrote:
Jim:
Thank you very much for the explication.
I will try to implement your suggestions.
With SPM i did the following. I had two groups of subjects: depressed and non-depressed. i created a one-sample t-map for each group of the contrast of say BOLD response to Fear vs Neutral faces thresholded at p<.001. Then i re-did this with a more liberal threshold p<.01 and used the equivalent of 3dcalc to multiply the two one-sample t-maps so that only the overlap of voxels in activated clusters remained. Thus if there was a 0 value in a voxel in either group map then the conjunction map has a 0 value, only voxels with non-0 values in both group maps survives in the conjunction map. This is, as you pointed out, a qualitative result and not a statistical result in that the conjunction map only indicates voxels that overlap in both group maps. I was worried that multiplying a p<.01 x p<.01 results in a map that is essentially thresholded at p<.0001, and thus overly conservative. But i am not sure if that is correct? what do you think? If the resultant conjunction map is p<.0001, then i am definitely not visualizing areas that may be truly active, thus false negatives. I believe what i describe above is orthogonal because the each group has its own separate baseline or comparison task. If i was using a single group of subjects and generating a conjunction map for task a vs b and task c vs b, then the result is definitely not an orthogonal conjunction because of the identical baseline/comparison task. Do you agree with this logic?
I will try the 3dRegAna approach.
Let me know what you think.
philippe
At 10:26 AM 10/1/2003 -0400, you wrote:
Hi Philippe,
I tried to respond to your query on the afni message board. Wasn't sure if you'd see it, since it's been a week and your post migrated to the older messages.
-jim
Philippe Goldin
Department of Psychology
Stanford University
Jordan Hall, Bldg. 420
Stanford, CA 93405
Tel: 650/ 723-5977
Fax: 650-725-5699
E-mail:
pgoldin@stanford.edu
