Hi Jeff,
I'm a little unsure about exactly what your regressors/contrasts are,
as you seem to refer to both [1 -1] and [1 0] in the same context, so
forgive me if I am totally confused, but here goes...
At the second level, if you have two groups of first level con images
for task>rest, and you are primarily interested in the group
comparison [1 -1], then I think you can disambiguate this by first
testing the contrast [1 1] for *overall* activation, then using this
to mask your [1 -1] or [-1 1] contrast.
> A few people have just suggested to use the (1 0) contrast as a mask at the
> "use different contrast as a mask" query stage, but this appears to be
> impossible to do (at least in SPM2) as clicking on this option only opens
> the current mixed-effect matrix/contrast window. I can't seem to figure out
> how to navigate to the spm.mat housing the one-sample mixed-effect con
> images. What am I missing?
I think this is okay, as I believe you should always be able to find a
higher-level/mixed-effect contrast that is appropriate, without having
to select other (e.g. lower level) spm.mat's. E.g. as above.
(or e.g. with repeated-measures data, as I *hope* I got right in this
post: www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind0702&L=SPM&P=9410 )
> Alternatively, I thought that a solution would be to create binary masks
> from the (1 0; 0 1; etc.) con images and then apply these to the
> mixed-effect two-sample (1 -1; -1 1) group analyses to limit analyses to
> regions only observed in the individual group (1 0; 0 1; etc.) analyses.
> This appears to work well, but I'm suspecting that this process may be
> statistically/methodologically different from the "use different contrast as
> mask" option. Any problems with this approach?
I think this is probably reasonably okay (see later for a bit more
waffle), but I wait nervously to be shot down by a proper statistician
(as I always do!)
> Lastly, if I use the binary mask creation from (1 0; 0 1; etc.) con images,
> would it be more appropriate to create separate masks for each group and
> then combine the binary mask regions using Imcalc (e.g., i1+i2 / 2)
First, don't use things like (i1+i2)/2 for masks -- you'll confuse us
both! In order to avoid getting confused about things like what
boolean(0.5) means, I'd suggest treating masks with boolean/logical
operations, e.g. i1|i2 or i1&i2, depending on whether you want
one-or-more, or both. (If you're curious, I think (i1+i2)/2 would be
equivalent to i1|i2 since most SPM code looks for mask>0, though note
other software might look at mask==0 or round(mask) or something else...)
> it be more appropriate to visualize the mixed-effect two group (1 -1; -1 1)
> contrast results using the individual (1 0; etc.) masks separately?
I guess the best answer to this is probably to use contrast masking
with [1 1], since this will use both groups, and avoid the need for
worrying about how to combine the masks. I think it should be pretty
much equivalent to i1|i2, but I'm not exactly sure how the thresholds
for the two masks would compare to a single threshold for the [1 1]
contrast. Generally though, it's the threshold at the second level
that you'd be more interested in, though note that the volume of the
mask will affect FWE-correction, so check that things seem reasonable.
Phew, having said all that, you'll probably be amused to read the post
I sent a few minutes ago, basically saying "yeah, you can just look at
one of the first level contrasts". To clarify my apparent
schizophrenia, I think that if you have got higher level blobs that
you are excited about, and just want to disambiguate their cause, then
a quick check of a lower level contrast is probably adequate; on the
other hand, if you want an accurate estimate of the significance of an
unambiguous higher level contrast, then I think contrast masking is
probably the right thing to do (though I've not read enough about how
this interacts with the size of masked multiple-comparison volumes to
be 100% confident).
I hope this helps (sorry it's not the most concise answer ever...)
Ged.
|