Vladimir,
> I am searching for a tool that will allow me to make statistical
> comparisons for analysis of EEG and EEG derived measures. We do several
> kinds of studies but in most of them the design is such that we have
> multiple subjects and for each subjects there are two conditions which
> we would like to compare. For each conditions there are multiple
> trials (usually hundreds) but they are averaged to get a single
> dataset.
[...]
> From what I've read at SnPM pages I see that the method has few
> assumptions and I know it has already been used in LORETA package for
> similar purposes but I would like to try to adapt the MATLAB code
> you released.
As you've seen, some people have used SnPM for LORETA data. The
best way to summarize SnPM use is as follows:
o If, for a one-sample t data (one group study), you can reduce each
subject's data to one *difference* image, SnPM ('MultiSub: 1
condition, 1 scan per subject') will work.
o If, for two-sample t data (two-group study), you can reduce each
subject's data to one image, SnPM ('MultiGroup: 2 groups, 1 scan
per subject') will work.
You can also do correlation for difference data (counter-intuitively,
run it as 'SingleSub: Single covariate of interest'). Note that,
above, 'image' can mean reconstructed source density map, or elements
in the original measurement domain).
> Therefore before I make further efforts I would like to ask you whether
> you are familiar with any special issues I should be aware about when
> I use the kinds of data mentioned below and not fMRI/PET.
This is what this means for your data...
> 1) ERP traces (values in uV)
-> This data cannot be currently used with SnPM; it involves
multiple measures per subject. We've done recent work
[1] which will allow this type of data, but isn't currently
implemented in SnPM.
-> In the hopefully-near future, you will be able to use such
data *if* it consists of a sets of *difference* measures
for each subject (e.g. *difference* of ERP traces between
tasks).
> 2) Event related power (values can be in % relative to baseline).
-> If summarized in one measure per subject per spatial location,
this can be analyzed, as per the conditions listed at the
top of the message.
> 3) Event related coherence (values between 0 and 1).
-> See comment for (2).
> 4) Source localization images (these are more similar to PET or fMRI
> but there can be great degree of dependency between the neighboring
> and even not neighboring voxels).
-> Ah! I thought you were talking about these all along.
Thanks to the nonparametric approach, it makes no difference
if there is subtle correlation or heavy correlation.
-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
[1] D Pantazis, TE Nichols, S Baillet and RM Leahy.
"A comparison of random field theory and permutation methods for
the statistical analysis of MEG data."
NeuroImage, In Press, Corrected Proof available.
|