[FieldTrip] one-sample cluster test
Ulrich.Pomper at charite.de
Fri Mar 30 10:37:06 CEST 2012
Thanks Mark + Jörn for the replies!
I'll have a look at the NAI.
On 27.03.2012 15:48, "Jörn M. Horschig" wrote:
First of all, you might want to use the Neural Activity Index (NAI), as
also described in the tutorial:
Then, to answer your question, I am tempted to say that this would be an
invalid approach when dealing with power, so in frequency-space. Power
obviously has a lower bound, so that the average of any random
collection of power values will never be 0.
That sounds plausible for raw power data, but actually, I want to compare whether my power change to baseline activity is different from zero (maybe that wasn't evident in my first post). I think that's a correct approach.
Though, what you are
suggesting is the same as a regular t-test, isn't it?
In case you are dealing e.g. with am LCMV beamformer, I am not quite
sure, but the null-hypothesis you suggested (amplitude==0) sounds fair.
In any case, building a surrogate distribution same mean and variance
than your data might be a better way to deal with your problem. That
should be easily doable in Matlab without any need to implement this in
FieldTrip. And note that the MC problem is always present, but a
cluster-based correction (similar to Bonferroni, but correcting by the
number of voxels in the cluster not by the number of all voxels) sounds
legitimate to me. Afair, that's how it's done in FT.
Are you sure? I am not aware of any correction to the alpha level that is done by the cluster-based correction in FT. Afaik all it does is look for a certain number of significant voxels that are close in time, freq and space.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the fieldtrip