See a detailed answer provided by Guillaume below, but to sum it up it
is advised to use the modality-specific classical statistics,
especially in the latest SPM8 version as there are some modifications
specific to M/EEG. For the Bayesian statistics there is no
M/EEG-specific code any of relevant SPM authors is aware of so you can
in principle use the fMRI Bayesian option. I can send you off the list
a poster from Rik Henson and Jason Taylor where this option is
discussed. There is now also a paper under review along the same lines
which will probably be the best reference for you ones it comes out.
* overall, (second level) models are the same for all modalities but you
should run SPM in the modality from which your data come from, for the
reasons detailed below.
* a Bayesian estimation option is only available in the interface for
fMRI and does not contain any modality-specific code. If requested, this
option could be made available for M/EEG (just have callback
spm_jobman('interactive','','spm.stats.fmri_est') instead of spm_spm).
* Clasical estimation has some modality-specific parameters:
- threshold used to select voxels for non-sphericity correction, see
spm_spm.m and spm_defaults.m, defauls.stats.*.ufp.
- tweak on variance estimate to deal with region with very low
variance: currently implemented in spm_spm.m by modifying ResMS image
(in PET/VBM, M/EEG modality), and previously implemented in
spm_contrasts.m (for all modalities, last public update).
These two specificities are likely to change in SPM in the future
(selecting the x% most significant voxels, robust shrinkage) so that
they would not be modality specific any more.
If a script is used without setting the modality beforehand
(spm('defaults','eeg')), SPM behaviour is to use the hardcoded values
* In the Results section, an extra question is added for M/EEG regarding
the display in the MIP (different for source reconstruction, TF,
On Tue, Mar 8, 2011 at 7:01 PM, Linda Burns <[log in to unmask]> wrote:
> I am still naive to spm (reading manuals) but I have a question about the second level source analysis in spm5 and spm8 regarding meg data. In short what is the difference between second level analysis for meg and fmri? can both options be used for meg data? I have created volumetric images from meg data.
> In spm5 manual it states that the 'basic models' implemented for fmri can be used for meg data as well. during the estimation procedure when using the fmri option there is an prompt about classical or bayesian model estimation. In the m/eeg option such prompt does not appear but the estimation still takes place. the former option give results but the latter does not. no contrast manager is created actually. Is it safe to let classical model estimation/and or bayesian model estimation be applied to meg data by using the fmri option? what does it depend on? any theory behind that I should read? I have the same problem in spm8 as in spm5. any thoughts? my basic problem is that I would like to know what is estimated in each option behind the scenes and whether I can use the frmi option for meg data. the manual does not state the opposite. I tried to sent this msg before but it did not appear on the list. sorry for any inconvinience