Abhinay Joshi wrote:
> 1) Could some one please provide some doc. for FWD and FDR terms and how
> can it affect the threshold.
Tom Nichol's website is a great source of info here,
http://www.sph.umich.edu/~nichols/FDR/
look particularly at the paper linked in the second paragraph.
> 2) RPV.img and RPV.hdr are the resel per voxel values,what is "resel"?
Resolution element. Basically the idea is that in smooth data
neighbouring voxels will be correlated, so we consider a larger
collection of voxels which is more likely to be independent from
neighbouring collections. Resels per voxel is the inverse of the
number of voxels considered to be in a resel.
There is a better description of this in here:
http://imaging.mrc-cbu.cam.ac.uk/imaging/PrinciplesMultipleComparisons
> And by referring the intensity values
> in the RPV.img do we decide the P-values to be used (i am asking this
> question because---in the RPV image in my case,
> I got the intensity values in the range of 0.01 and 0.08 and so I
> decided the p-value as 0.01 and not 0.001, is that right?
No, you can't derive significance from looking at the RPV values, they
tell you how "rough" your residuals are (i.e. smaller RPV corresponds
to smoother data).
FWE correction could use the information in the RPV image, but in SPM,
it doesn't seem to -- see this recent thread:
http://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind0612&L=spm&P=8750
> 3) Could some one please tell me something more about F-test?
Well, in brief (and in my personal, possibly dodgy, interpretation)
t-tests compare a contrast of parameter estimates to the contrast's
standard deviation, where the contrast must be a vector (so that
c'beta is a scalar). F-tests compare the variance of a reduced model
(or more specifically, the increase in variance due to the
reduction/hypothesis) to that of the full model, where the reduced
model can be specified by a matrix contrast.
For example, if you have two samples, you can model separate means for
them, and then the standard two-sample t-test is simply a t-contrast
of one mean minus the other. With three samples, you might be
interested in the null hypothesis that all three means are equal, in
which case you can model the three means and then test the reduced
model given by mean1-mean2=0 and mean2-mean3 = 0, or the matrix
F-contrast of [1 -1 0; 0 1 -1].
Another point to note is that F-contrasts are always two-tailed, and
in SPM t-contrasts are assumed to be right-tailed, so you might use a
vector F-contrast in SPM simply to get a two-tailed equivalent of a
t-test.
If you need more detail than this -- or more guarantee of it being
correct ;-) -- then there is probably no substitute for reading a
statistics book!
> 4) Last but not the least, the threshold intensity is the intensity
> taken from the global mean, so if I say 0.8 then it is 80% of global
> mean but does this also mean that 80% of the brain region is "Gray"
No. If you are analysing GM images then the 80% threshold is based on
the GM intensities, so hopefully the results will be almost
exclusively GM. Incidentally, if you're doing VBM then I would
recommend using an explicit mask rather than thresholding (and prefer
absolute thresholding to relative anyway).
> so if I say 0.8 then it will consider all the gray regions in
> brain and it won't consider anything for the white matter and so if I
> need to consider the "white matter" then I should select the intensity
> as 0.2?
If you are interested in the white matter, then you should probably
test it separately, with a separate mask (produced from 0.8 relative,
or e.g. 0.05 absolute, or better still an explicit mask that looks
reasonable). I think it would be possible to add grey and white
segmentations, and then perform VBM on this combined "brain" image, in
which case any absolute or relative thresholding will be in terms of
the brain probabilities, and hence will include both GM and WM.
However, it is probably undesirable to smooth GM and WM together, so
analysing them completely separately is likely to be best.
Hope that helps,
Ged.
|