After conducting a single subject analysis with surface data using 3dDeconvolve I found that the p and q values were very similar. I then used 3dcalc to create new contrasts based on the 3dDeconvolve results and recalculated the FDR curves with 3drefit. To my surprise I found that the corrected q values were now smaller than p. I am aware that this is theoretically possible with FDR but I’m not sure how verify if it is correct or if I have made a mistake somewhere.
Even if it is theoretically possible, it means that every
voxel (in your mask) is below some limit that you are
If you look at a small ROI, and/or with a relatively
large amount of blurring, perhaps that might happen.
How many voxels are in your mask? And consider
Thank you for the response.
As pertinent to the research question I had to remove voxels the showed deactivations in any of the the conditions. That leaves me a total of 31653 nodes (from 3dBrickStats) which significantly reduced from the original 3dDeconvolve results (8176252). Based on what you have said, it seems the voxel count is the reason for the FDR results.
It isn’t exactly the voxel count that is the reason, it is the
very heavy bias in selected voxels. It sounds like you are
running FDR restricted to voxels that have small p-values,
which will almost force such strange results (q ~= p) until
p goes well below the original threshold.
In any case, it does not seem appropriate to run FDR after
thresholding. The FDR mask should be independent of the
tests in question.
I realize that you are exactly right and my approach was wrong. By calculating the FDR curves after thresholding I was biasing the results and obtaining odd q values. Thank you for the timely responses.