|
Hello,
Perhaps my last email on this topic wasn't as concise or clear as it
perhaps should have been, but I remain to be convinced that it's
incorrect... though like Marko said, I too may be utterly wrong, in
which case hopefully someone will clarify, with reference to the code,
what is actually happening.
However, Mahinda, I think your observation can be explained
consistently with what Susie and I seemed to find, and what I
described previously.
>> Following from the recent SVC discussion I was under the impression
>> that the SVC when applied only includes those voxels (within whatever
>> volume you have defined) that were initially activated at the
>> previously chosen uncorrected threshold. In which case the value
>> chosen for this threshold should dictate the number of voxels within
>> the region where the SVC is applied and therefore also the resulting p
>> value for the analysis.
This is an extremely good point, which I hadn't thought about
properly. But returning to the code, I think the explanation can be
seen in the following lines from spm_VOI.m:
122 xSPM.S = length(k);
123 xSPM.R = spm_resels(FWHM,D,SPACE);
124 xSPM.Z = xSPM.Z(j);
125 xSPM.XYZ = xSPM.XYZ(:,j);
126 xSPM.XYZmm = xSPM.XYZmm(:,j);
127 xSPM.Ps = xSPM.Ps(k);
As I described in my previous email,
http://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind0708&L=SPM&P=16990
the variables k and j come respectively from the complete set of
voxels in the SVC ROI, and the subset in both the ROI and the set of
whole-brain supra-threshold voxels.
Now, in lines 122 and 123 we see that the size of the search volume,
in voxels and resels, uses the entire SVC ROI (respectively through k,
and then directly in the spm_resels function). Similarly, line 127
stores the sorted p-values used for any subsequent FDR correction
using the complete ROI set, k.
(For FDR, note that in spm_getSPM, line numbering from SPM5 latest
updates, we find
533 % P values for False Discovery FDR rate computation
% (all search voxels)
545 %-Compute mask and eliminate masked voxels
that indeed Ps is independent of the whole-brain threshold used in
lines 545 onward.)
So, the previous whole-brain threshold affects (in my/Susie's opinion)
the set of voxels which are considered (through line 124-126 above,
with the subset of voxels j), but not the characterisation of either
the search volume for FWE or the p-value distribution for FDR. So the
significance of the *reported* voxels won't depend on the previous
whole-brain threshold, but some voxels might not be reported at all,
if the whole-brain threshold was too strict.
Mahinda, if you vary the uncorrected whole-brain threshold from the
values of 0.99 and 0.01 that you mentioned in your previous email,
down to 0.001, 0.0001, 0.00001, etc. do you still find that the SVC is
independent of this prior threshold, or do the voxels disappear from
the results (even though those voxels that remain show unchanged
significance)? This would be interesting to know. If so, I'd like to
ask again the questions in my previous email:
- whether this is deliberate/correct
- if so why
- and if for good reason, does the practice of setting the whole-brain
alpha near to one violate that reason?
And if not, please do tell the mailing list, so that people can throw
rotten vegetables at me when they next see me ;-)
> I would see no good
> rationale for excluding voxels within the search volume on the basis of
> an analysis that you are not interested in anyway
Nor can I, Marko, but this is what Susie seemed to find was happening
in practice, and what matched my interpretation of the code. And this:
> but there may be factors involved that I do not see
is worrying me too... Hopefully other more knowledgeable folk can
comment on this...
All the best,
Ged.
P.S. I fear this email is also not "as concise or clear as it perhaps
should" be... Sorry about that. But perhaps I could borrow Blaise
Pascal's famous apology: "Je n'ai fait celle-ci plus longue que
parceque je n'ai pas eu le loisir de la faire plus courte." -
http://en.wikiquote.org/wiki/Blaise_Pascal
|
