Thank you for your comments. I'm not sure I understand the statistics
entirely but at least what I've found is not impossible. Is the
'variance component estimation' pooled over the whole brain? It seems
very counterintuitive to me that changing a mask mainly over the
cerebellum should have effects on stats in the parietal lobe - what is
the rationale for this?
I am attaching a small file illustrating the scale of the problem - my
F values have gone down by 5% with the change of mask, which seems
like quite a lot to me (the area I'm interested in is highlighted).
More pragmatically, can anyone explain how the standard mask would be
calculated if I had done my smoothing at the first level, rather than
after estimation. Then maybe I can use that method to get a better
mask. At the moment I'm just using masks generated by a different
study (also spm2, whole brain, same acquisition parameters etc),
because I was assuming it didn't matter.
Thanks,
Antonia.
On 5/18/05, Jesper Andersson <[log in to unmask]> wrote:
> Hi Antonia,
>
>
> > I am confused by an analysis where applying an explicit mask seems to
> > change my uncorrected statistics. To explain fully: I did my first
> > level analysis with unsmoothed data, and then smoothed the resulting
> > betas. My smoothed betas for the second level were completely
> > unmasked - non-zero values filling the whole bounding box, so when I
> > did the second level analysis I applied an explict mask shaped like
> > the whole brain. And I got some reasonable activations in the
> > expected areas. Then I decided my first choice of explicit mask was
> > not great and applied a slightly different one - about 2 voxels
> > smaller all over, with quite a lot less cerebellum. I expected my
> > uncorrected stats for cortical areas to be identical under both masks,
> > because the mask is identical at this point. But it seems this isn't
> > the case - all my F values have got a little smaller with the bigger
> > mask and areas which used to be significant (even ones deep in a
> > sulcus and nowhere near the changes in the mask) are now subthreshold.
>
> That sounds interesting, and a little worrying. SPM used to be (almost still is) "massively univariate", meaning that when estimating a statistic (t or F) in any one voxel it does so without in any way considering neighbouring voxels. I.e. that statistic would have been the same had it been the only voxel in the brain.
>
> Now, as of "variance component estimation" in SPM2 this is no longer 100% true. The parameters (e.g. the different variances of two different groups) that are used to build the variance-covariance matrix to be used in the weighted least-squares (WLS) are based on a set of voxels, and hence the parameters can be said to be pooled over those voxels.
>
> My suspicion (unless you have unearthed a proper bug) is that this is what you see. When including more voxels by increasing the mask the variance components are estimated from a different set of voxels and the estimates change. And those estimates go straight into the estimation of the parameters and statistics for the individual voxels, and hence those would be expectecd to change.
>
> >
> > Looking at my spm outputs, there are small differences in smoothness
> > and larger differences in search vol and resels between the analyses:
> >
> > mask 1: FWHM 12.3, 12.6, 11.7 mm, search vol 226850 voxels,
> > 932.3 resels
> > mask 2: FWHM 12.3, 12.6, 11.8 mm, search vol 195379 voxels,
> > 790.3 resels
> >
>
> This bit isn't really surprising. Of course search volume changes, and the smoothness that is reported is pooled across all voxels in the search volume so that will also change.
>
> > And how should I make a principled decision about what mask to use?
>
> Oh dear, I am afraid I dont have any great ideas there. Someone else?
>
> Good luck Jesper
>
>
|