Cyril,
> I read your article in neuroimage and looked at the multFDR.m and
> FDRill.m matlab scripts, and I have several questions.
>
> A) In the article, you wrote that as we know that voxels out of the brain
> or in ventricles are not active and that FDR procedures operate on all
> voxels included in the analysis, we need to remove this voxels. In SPM,
> except for the ventricles, the analysis is performed only on the brain
> using the mask .. and we don't need to remove 0 signal voxels out of the
> brain, do we?
That's right, SPM does that for you automatically as part of implicit
masking.
> B) The FDR corresponds to the rate of false positives among
> 'activiated' voxels (false + true positives) ...
Don't forget that it's the long-run average proportion of false
positives. It doesn't guarantee that the false positive fraction will
be controlled for a particular experiment (just as with FWER control
you never know if a particular statistic image has a FWE in it).
> In SPM, I expected
> to have to set the threshold of false positives q for the implicit
> T-map threshold (default is .001 in spm) but it is a p value that we
> have to chose. What that p value represent? Is it the value for
> non-corrected maps on which an FDR correction is done (but in this
> case how can we choose the q value) or is it the minimum p value
> after correction (and again how can we choose the q value)?
In SPM if you select 'FDR' when asked for a corrected threshold, then
you are entering the q value. This is then the threshold for what SPM
calls FDR-corrected P-values. (Note, in SPM2 make sure you have all
the updates to ensure you are asked this question.)
> C) Imagine I've understand precedent questions ... Can I use the p-value
> calculated from the distribution of T scores in each voxel for one
> experimental condition (FDRill.m)
Yes, this is exactly what the FDR-corrected threshold does for you.
> or the mean p-value for multiple experimental conditions (multFDR.m)
> to correct my images? (with q=.05 in your scripts or something else
> if I want to change)
multFDR isn't the mean p-value, rather, it computes a FDR threshold
for a union of statistic images. Use multFDR cautiously: It says
that you're willing to control false positives as a fraction of all
detections in *both* images. This is potentially bad, as one image
might have really strong, extensive signals, and the other image no
signals, and then you'll have more false positives in one than the
other.
> Thank you in advance for your reply, I'm really sorry to bother you
> with that but the fact that we only have the possibility to chose p
> values (and not q) really disquiet me.
I'm not sure I understand... even in univariate statistics we have a
choice of P-value threshold (alpha); some disciplines use 0.001, some
0.01, many 0.05, and in the social sciences, a 0.1 P-value might get
you published.
The correct FDR threshold also may vary; in particular, if you have
huge activations a smaller FDR q maybe desirable.
You simply have to interpret FDR results as per its definition... the
long run average of a false positive proportion.
Hope this helps!
-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
