Hi Marko,
> Number one, if I try to analyze simple spm-first-level con-images,
> during calculating the correct number of permutations, I get "No voxels
> in brain" from snpm_cp. I use the multi-subject - single image standard
> model, and altering the options does not change anything. I am quite
> suprised since, on the first level, results are credible with 16 scans,
> and all con-images look ok. Any hints?
This must be a masking/analysis-threshold problem. SnPM is not even
permuting anything because it finds no voxels in the brain (as defined
by the thresholding options). Are you using all of the same options as
you use in SPM? If it is a 2nd level analysis you should have 'Threshold
Masking' none, and no 'grand mean scaling'.
(And by "calculating the correct number of permutations", I think you mean
"Working on correct permutation", which means it is simply working on
the correctly labeled data (as opposed to permuted or sign-swapped data).
> Number two, I started to analyze cross-correlation images generated
> with another software in snpm2. In some cases (n=16) I get a hugely
> significant cluster (t=41.72) covering the whole brain. Again,
> all images look technically ok. I thought that I do not have to do
> a conversion of the r-images since snpm is non-parametric anyway,
> but perhaps I am wrong? What else do I not see?
SnPM, whatever statistic it uses, is non-parametric, and so always valid
and controls the false positive rate (assuming exchangeability under
the null hypothesis). However, different statistics will be optimal for
different types of data. Standard statistical theory tells us that the
t-statistic is optimal for normally-distributed data. But if your data
are highly non-normal (e.g. consists of data bounded between 0 and 1 and
using that entire range), then the t-statistic may be quite sub-optimal
with respect to power.
So, if you are using data which you know to be quite non-normal, I
recommend either a rough transformation to make it more normal (for
correlations, ye-ole arcsine trasformation should do it), or use a
different statistic. Currently there are no other statistics
implimented in SnPM, but last May I posted a function to convert
image data to ranks:
http://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind04&L=spm&P=R167575&I=-1
Rank-transforming your correlation images may improve the sensitivity.
(If you try both rank transforming and the arc-sine transformation,
please let me know what works better).
-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
|