Jung,
> I use SPM and traditional statistic interface to make T
> image. Than I use FDRill and it give me the t = 2.38 and p=0.012. I
> think if I use SPM statistic interface and set FDR p = 0.01, I will
> get the result that activation FDR is below 1%. Is it a correct
> step?
I'm not sure I understand your question exactly, but: Yes, it can
happen that if you ask for a FDR threshold of 0.01 you might not get
any voxel with exactly FDR-p of 0.01, but only smaller. This is due
to the discreteness of the FDR P-values, which are determined from the
emperical distribution of P-values in your data.
Also, the function FDRill uses a FDR p-value threshold 0.05 by
default. To change that you have to specify all of the options;
however, I've just changed it to make it easier. Download the revised
version
http://www.sph.umich.edu/~nichols/FDR/FDRill.m
and then do
FDRill('',[],0.01)
to see the 0.01 threshold.
> Previously I use traditional statistic interface and I set voxel
> level uncorrect p < 0.01 and cluster level correct p < 0.01 to
> choice activation cluster. If I use FDR and set p < 0.01, can I use
> the same criteria in spm table list ? or I can even say my result is
> 1)cluster level correct p < 0.01 and false discover rate also < 0.01 ?
The FDR in SPM is purely a voxel-wise inference; there is no good way
to mix it with cluster-size inference. I would go with one or the
other. (However, you could also use an arbitrary cluster-size
threshold to eliminate tiny clusters you don't belive in; there's no
statistical characterization of such a 'filtering' but it has face
validity).
-Tom
-- Thomas Nichols -------------------- Department of Biostatistics
http://www.sph.umich.edu/~nichols University of Michigan
[log in to unmask] 1420 Washington Heights
-------------------------------------- Ann Arbor, MI 48109-2029
|