Hello,
I am working on applying multiple-comparison corrections to resting-state data that has been subjected to ANCOVA. I have used AlphaSim to determine the cluster-extent thresholds for various p-values (e.g., p=.005 yields a cluster size threshold of 21 voxels), and I have not had any difficulty in finding significant clusters when viewing the F-statistic maps on the Rest Slice Viewer and using the AlphaSim criteria.
I would also like to view the results that emerge under FDR-correction, however, but when I use the FDR correction (from the drop down menu in the slice viewer), I am consistently told that no significant voxels exist after FDR. This is surprising to me, as I often see results using the AlphaSim criteria. I was wondering if anyone has any insight as to why this may be happening, or if perhaps I have been trying to implement FDR incorrectly?
Any help with this would be much appreciated!
In short, why might FDR and AlphaSim results differ so, and what is the optimal method for implementing FDR within REST?
Thank you,
- Ben Mooneyham
Graduate Student
Department of Pscyhological & Brain Sciences
University of California, Santa Barbara
Submitted by Qing Zhao on Sun, 10/27/2013 - 20:31 Permalink
Re: FDR vs AlphaSim
these two methods depend on quit different assumptions, so they should and indeed give different results.
It quite difficult to say which is better.If a cluster is quite big, but all voxels within it have a p value just slightly smaller than threshold (like 0.05), FDR may kill the whole cluster, but Alphasim seems to believe it is a "significant cluster". However, if you make a whole brain search comparison, and a small region (like amygdala) has a very small (seems very significant) p value, Alphasin may kill it because it is not big enough cluster, but FDR may think it is significant enough, especially at least those peak voxels.