Hi Cyril,
> I'm a bit confused about FDR correction - I have data that show some
> activity let say in regions A and B using a F contrast after FDR
> 0.05 correction (regions A and B p corr = .009). Regions A and B
> show reversed effects (gp1>gp2 and gp2>gp1). If I look at the same
> data with t contrasts I have the region A for a positive t contrast
> (still FDR 0.05 and p= .007 ) but nothing for the reverse negative t
> contrast - indeed the region B is now at p corr .06. How is it
> possible to get such a difference? and in this case which contrasts
> have to be used F or T?
What you're running into here is FDR's adaptiveness. You'll get
different FDR thresholds for different distributions. In this case
you have three different distributions:
1. F dist (equiv. to two-sided t)
2. t dist, increases
3. t dist, decreases
and you will get different FDR thresholds on each.
Why? Because in each there are different distributions of signal
voxels and FDR is adapting to these, finding the lowest threshold in
each that controls the fraction of false positives detected (on
average).
Does this help?
-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
|