Hi Will
I don't know if I can address everything you ask, but I'll see if I
can help a bit. The short answer is probably that there is no single
correct way to do things; the important thing is to provide enough
information so readers know exactly what threshold you used.
> The first set of questions involve VBM.
> When carrying out voxel based morphometry in SPM5 after pressing for results
> out of the options FWE, FDR and none, I select none. The results I have do
> not survive this correction for multiple comparisons.
> I then choose e.g. p<0.01 and an extent threshold of zero.
> In the results table that I get, I can report the z-values (or T-values) which
> correspond to certain coordinates, but if I wanted to report p-values as well
> which column should I report in my manuscript? As far as I know there is an
> issue with using cluster p-values when using VBM. Should I therefore report
> the voxel P FWE-corr, the voxel P FDR-corr, or the P uncorrected voxel value?
First, you are correct that the random field theory (RFT)-based
cluster level p values should generally not be used for VBM because
the data aren't uniformly smooth (i.e. there is a nonstationarity
problem). There is, however, a nonstationarity toolbox that can
correct for this, allowing you to use cluster-level correction:
http://fmri.wfubmc.edu/cms/NS-General
Regarding what to report, generally you will want to report the
voxel-level statistic that was used for all of your reported voxels,
and then in addition, some additional information about a
cluster-level p value (if you use one). In your case as you describe,
you would just describe the results as using a voxelwise threshold of
p < .01, with no correction for multiple comparisons. If you use a
cluster-level threshold, include that as well. In other words, the
same standard information you would find in most fMRI papers.
If you had done an analysis using FWE or FDR to correct for multiple
comparisons, this would apply to all voxels in your results, and you
would report it accordingly.
In the case in which an uncorrected value is used (like you have), the
FWE or FDR values for the peaks are sometimes (but I think not often)
used as a sort of additional criteria. For example, sometimes people
will use a voxelwise threshold of p < .01, but only report clusters in
which the peak voxel is FWE < .05. (I always find this a little bit
odd as typically conclusions are drawn about the whole extent of the
cluster, but only the peak voxel passes a more stringent threshold.)
Generally I think it is more straightforward to have a voxelwise p
value, and optionally, a cluster-level p value, and leave it at that.
> Also, if some are significant at the voxel P FWE-corr level and others are only
> significant at voxel P uncorrected level what should I do here? E.g. report P
> FWE-corr only and disregard the others? Or report a combination of p-values
> (showing the distinction between thesein the paper in some way)?
Generally you would want to pick a single criterion and stick with
that for the whole analysis. For example, if you decide p < .01
(uncorrected) is appropriate, you would just use this, and not bother
noting which voxels surpass this. However, if you are (probably
correctly) worried that reviewers will think this threshold is too
lenient, you might want to use a voxelwise threshold of FWE < .05, in
which case you would disregard voxels that don't reach this level.
I will leave the SVC questions to someone with more experience in
those matters. Good luck! Hope this helps.
Jonathan
|