Dear Anette,
This is a very interesting question of common interest so I'm copying
my answer to the list.
On Mon, Jul 5, 2010 at 4:46 PM, Anette Giani
<[log in to unmask]> wrote:
> I have a question concerning MEEG source reconstruction (using: MSP/ GS,
> group inversion.
>
> In the manual you mention that, for statistical reasons, all experimental
> conditions should be inverted simultaneously. Hence, I started inverting all
> conditions at once. Notably however, activation seemed to be localized in
> similar cortical regions throughout all conditions (even though I would
> expect them to be different for different conditions). This phenomenon was
> especially strong for control conditions (i.e. conditions in which subjects
> where just fixating, i.e. no region in particular (except maybe visual
> cortex) would be assumed to be activated). For example, inverting auditory
> and control conditions would 'bias?' control conditions to become localized
> in auditory cortex, while inverting control conditions together with visual
> conditions, biased control conditions to become localized in visual cortex.
>
> I therefore started inverting all conditions separately. Interestingly,
> activation patterns became more as expected (and extremely different from
> previous analysis!). In control conditions, seemingly random activations
> patterns appeared and different areas became active for different
> experimental conditions.
>
> I was wondering how these huge differences can come up? Moreover, I would
> like to know the exact reason for inverting all conditions simultaneously.
> Is it statistically invalid to invert conditions separately?
>
I think what you observe is not very susprising. The input to the
inversion algorithm is the channel covariance matrix computed from all
the data. The sources should model the interesting patterns in this
matrix. So if in one condition you have interesting responses and the
other is basically noise (or at least it doesn't average to anything
interesting) the sources that you will get will be specific to the
interesting condition. What should then happen is that if you do
statistics you should still get significant differences between the
activation and the baseline condition, although because of the way the
images are normalized this might actually not happen if one of the
conditions is just noise. The idea behind what's written in the manual
is that you have some conditions that basically activate the same
areas but to a different extent and you are interested to find which
of those areas are modulated in a specific way. Then you should invert
them together to avoid the situation when because of some random
localization errors the activations don't overlap between conditions
and you can't do meaningful statistics.
We once discussed similar issues at our group meeting and what Karl
said is that it's not very meaningful to do a comparison with
'nothing' because the null hypothesis is clearly false. If you just
want to see what's significant in a single condition you should look
at PPMs (this is what's displayed after the inversion). The problem
with that is that those PPMs are difficult to summarize accross
subjects.
What I'd say if you ask me is that none of these things is 'clearly
wrong' or 'statistically invalid'. You should find a protocol that
works for your purposes and clearly document it in your paper.
Best,
Vladimir
|