Dear Anette,
This is what could be expected. Look at the recent 'SVC vs. Explicit
Mask' thread on the list.
On Tue, Feb 1, 2011 at 8:54 AM, Anette Giani
<[log in to unmask]> wrote:
> I have a question concerning the statistical analysis in SPM.
>
> I have scalp-frequency MEG data which I smooth [20 20 0] (i.e. in space but
> not in frequency) trying two options: (1) using an implicit mask and (2)
> using no implicit mask for smoothing. I then do a simple t-contrast
> comparing A-B. Not surprisingly, t-values, z-scores and peak activation
> coordinates do not change according to the type of smoothing I am using.
> However, p-values (FWE-corrected) do change.
>
> For whole brain corrected data:
>
> - p-values (FWE-corrected) are smaller for implicitly masked data (1) than
> for non-masked data (2)
>
This is because your search volume is smaller. In the non-masked data
you also have the corners which are meaningless but SPM doesn't know
that.
>
> For small volume corrected data (corrected for all voxels at frequencies of
> interest):
>
> - p-values (FWE-corrected) are smaller for non-masked data (2) than for
> implicitly masked data (1)
>
This because of the subtle difference between SVC and masking. When
you use SVC the smoothness and and error variance are computed over
the whole image and therefore the estimates are more reliable which
results in lower p-values.
>
> I would like to understand how this difference in p-values comes up.
> Moreover, I would like to know when applying an implicit masking is useful.
>
I think in this case you should apply implicit mask for the scalp
because the corners do not provide independent data. If you want to
also limit the frequency, it might be better to use SVC.
Best,
Vladimir
>
>
> Thanks a lot in advance,
>
> Anette
>
>
|