On Fri, May 21, 2010 at 6:44 PM, Dominik Bach <[log in to unmask]> wrote:
> Can I ask you a more fundamental question?
>
> How do you statistically test source reconstructions on the group level?
>
> I know this sounds stupid because the SPM manual proposes to do t-tests on
> images at the group level. But ...
>
> I assume there is a strong bias in this approch, especially when the data is
> noisy, because you will end up with your priors, and they are of course
> consistent across subjects. So whatever I test, the whole brain is always
> significant - with different patterns, but that's not a useful statistical
> test.
>
> Do you have a suggestion for me?
>
The idea is that you do statistics to compare between conditions. It
is very likely that the differences you will find will be at one of
the locations corresponding to the group prior, but the question you
ask is at which of those locations a particular effect is expressed.
In my mind the aim of volumetric M/EEG source reconstruction is
separation and labelling of different effects that are mixed at the
sensor level rather than precise anatomical localization which is more
an fMRI thing.
If you just want to know where there are activations at a single
condition there is no good way to ask this question statistically.
Some people do activation vs. baseline test but that usually is not
very specific.
Also at the moment we are still struggling with some imperfections of
the method. While it seems that the latest improvements produce decent
localizations it is still difficult to get significant effects with
FWE correction which might have to do with the way we normalize the
data when it is converted to images. We are now looking at it and
hopefully there will be some improvements in this respect as well.
Best,
Vladimir
|