Dear Takanori,
The issue of scaling is something we've been dealing with here a lot
recently and indeed there have been some changes to that that can lead
to different results in different SPM releases. I also did some
testing on a group dataset and clearly scaling for each condition
separately (which is what was done in the latest release) leads to
better results than other ways of scaling. I'm glad to hear that you
reached the same conclusion from a different dataset.
Not scaling at all would be difficult with the present version of
spm_eeg_invert because it needs to handle in a consistent way the
multimodal and the unimodal cases (see R.N.A. Henson, E.
Mouchlianitis, and K.J. Friston. MEG and EEG data fusion: Simultaneous
localisation of face-evoked responses. NeuroImage, 2009). Also there
are some steps of conditioning the units of data and leadfields that
would make the units of the output meaningless in any case. As far as
I know Karl also tested various ways of handling the output and
reached the conclusion that mean scaling is the best way to get robust
results.
But in general the details of the inversion routines are still
undergoing changes and they will be slightly different again in the
next release so we'd be glad to get feedback from users who have a
good idea of what they are expecting to see in their analysis.
Best,
Vladimir
On Thu, Mar 4, 2010 at 2:54 PM, Takanori Kochiyama <[log in to unmask]> wrote:
> Dear Vladimir and SPMers,
>
> Thank you for your reponse to my data converting problem.
> I have an another question about SPM MEEG,
> scaling of the evoked/induced power image in spm_eeg_inv_Mesh2Voxels.
>
> Why does SPM conduct the global scaling of the power image?
>
> For exmaple, in PET-CBF measurements,
> scaling is important to evaluate the local neuronal activities
> because the local signal changes covaried with the global signal changes.
> For fMRI with global signal fluctuations, we often choose scaling option
> to increase sensitivity for signal detection.
>
> Are these true of MEEG power data?
> Or is there another rationale for scaling?
>
> I have learned the importance of scaling over some test analyses,
> however, I would like to confirm a theoretical rationale.
>
> The following are my test analyses:
> I performed the source localization on multi-subject MEG data
> with 2 task and 2 control conditions.
> Smoothed power images were written out for each condition and
> they were entered into the second level repeated measures ANOVA.
> Contrast of task vs. control was evaluated.
> (The issue of non-normality put aside so far.)
>
> Results were varied among 3 different scaling methods,
> scaling none, by whole conditions used in r3408, and by each condition used in r3684.
> Localization was best for r3684 (scaling by each condition),
> the rest have large clusters but too may activation foci to localize them.
> I also tried another approach: GLM scaling option,
> grand mean scaling by subject and global scaling of each image.
> Localization was best for global scaling as was expected.
>
> Thanks in advance for your help.
>
> Takanori Kochiyama.
>
> -------------------------------------------------------------
> Takanori Kochiyama, Ph.D.
> Advanced Telecommunications Research Institute International
> Brain Activity Imaging Center
> E-mail: [log in to unmask]
> -------------------------------------------------------------
>
|