Dear Anette,
Just to add to this, the reason that the newest update says that it
"permits F-contrasts", is that the release notes for the previous
update said in the section on "Group inversion":
"Note that F-contrasts should not be used"
http://www.fil.ion.ucl.ac.uk/spm/software/spm8/SPM8_Release_Notes_r4010.pdf
The reason for that, is that the residual variance for source
reconstructed data can be extremely low away from regions with
substantial reconstructed signal, and these low variance (and low
signal) regions can have artefactually high F-statistics. In some
cases, it can be very hard to judge which regions are wrongly
significant in this way, as it might well include even the most
significant voxel in the image. This is why we would recommend that
you rerun the statistics in the newest version.
Having said that, the F-statistics under r4010 are *valid where the
signal is large*, so you might be able to avoid rerunning (if you are
in a hurry to submit!) by using "Check Reg" and simultaneously viewing
your spmF and its underlying ESS image, to check that you are only
trying to interpret regions that are significant due to large ESS
rather than small ESS and even smaller ResMS (you might want to
simultaneously view ResMS too to be sure). If you have multiple
contrasts to look at and/or multiple blobs to interpret within each,
you might find that it is quicker to redo things with the update than
to do lots of manual work in check reg, but I thought I should mention
this for completeness.
Best wishes,
Ged
On 12 April 2011 13:41, Vladimir Litvak <[log in to unmask]> wrote:
> Dear Anette,
> Yes F-tests were not valid in 4010. I'm surprised that you haven't seen it
> yourself because the results usually looked quite weird. This only pertains
> to SPM imaging source reconstructions results. Sensor level or beamformer
> images should be OK.
> Vladimir
> On 12 Apr 2011, at 12:49, Anette Giani <[log in to unmask]>
> wrote:
>
> Dear Vladimir,
>
> In the release note of the latest spm8-update it has been written:
>
> (ii) It is now enacted by modifying the residual mean squares image (ResMS)
> saved on disk. This permits F-contrasts for M/EEG data (and reduces leakage
> outside the brain in VBM F-contrasts with generous analysis masks). It also
> has the advantage that the tstatistics can be easily reproduced using the
> saved con_xxxx and ResMS images (though they still of course differ slightly
> to statistics recomputed from the original input images outside SPM).
>
>
>
> However, I do not exactly get what this means. Especially the sentences:
> “This permits F-contrasts for M/EEG data” confuses me. Does this imply that
> F-test on M/EEG data have been invalid for earlier spm8 releases? I am
> asking, since we are about to submit a paper and I am wondering if I have to
> redo statistics using the latest update? So far we calculated F-test on
> M/EEG data using SPM-r4010.
>
> Thanks in advance!
>
> LG,
>
> Anette
>
>
|