Would agree with Rik. Cluster level inference assumes adjacent voxels are more or less independent. In MEG all of the voxel values come from linear combinations of N (eg 275) channels. So there will always be large regions where voxel values are similar. This will generally not be due to a consistent electrical change across a volume, but simply because they are formed by the same channel weightings.
So I would say cluster level inference not appropriate for MEG, unless you can establish the independence of the elements within the cluster. This is what the non-stationary correction is trying to do, but in my experience (and this depends on the kind of inversion you use), to get true estimate of MEG fwhm you need to sample at an very fine level in some areas (fwhm goes from order of 1mm to a few cm), which takes ages.
Most practical solutions for MEG stats I have come across are randomisation testing (on peak values within the image). And have also heard of, but not tried, simply bonferroni correcting the parametric p value by the total number of channels.
A couple of related papers.
A comparison of random field theory and permutation methods for the statistical analysis of MEG data.
Neuroimage. 2005 Apr 1;25(2):383-94.
Statistical flattening of MEG beamformer images.
Hum Brain Mapp. 2003 Jan;18(1):1-12.
Best wishes
Gareth
-----Original Message-----
From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]] On Behalf Of Rik Henson
Sent: 11 November 2008 08:42
To: [log in to unmask]
Subject: Re: [SPM] MEG analysis in SPM8b
>> 2) Is reporting meg data (voxel-level uncorrected p 0.05, cluster level
>> corrected p 0.05)
>> acceptable you think?
>
> Unlike the fMRI community we do not have well established standards
> for publishing findings of 3D source reconstructions. From what I'm
> hearing from fMRI people I understand that whole brain FWE correction
> is very conservative and is not usually used when people have an idea
> what they are looking for. They usually do a small volume correction
> around the area where they expect the response (of course this should
> be decided in advance and not post-hoc). I'd be glad if our colleagues
> contribute their thoughts on this, especially those who have
> experience with publishing results of distributed source
> reconstructions.
>
>
My (probably out-of-date) understanding is that one should be wary of
cluster-level correction when
the initial height threshold is too liberal (ie p<.05 uncorrected in
your example; p<.001 uncorrected
probably safer). But an RFT expert might correct me. You might also want
to consider using the
"nonstationary" extension of Hayasaka:
http://www.fil.ion.ucl.ac.uk/spm/ext/#NS
for cluster-level inference, particularly if you are using F-statistics.
Rik
--
-------------------------------------------------------
DR RICHARD HENSON
MRC Cognition & Brain Sciences Unit
15 Chaucer Road
Cambridge, CB2 7EF
England
EMAIL: [log in to unmask]
URL: http://www.mrc-cbu.cam.ac.uk/people/rik.henson/personal
TEL +44 (0)1223 355 294 x522
FAX +44 (0)1223 359 062
MOB +44 (0)794 1377 345
-------------------------------------------------------
|