Hi Ged, yes you are right. The way the model is parameterized we could
use ANOVA with constant and obtain the appropriate mask contrasts.
Thanks for clarifying this--it seems a better way to go.
though for B>A (-1 1 0) wouldnt you want to use a similar approach and
use a 0 1 1 as the mask ? this should give voxels where B is active and
also greater than A. using 1 1 2 as a mask seems more restrictive
because differences between B and A would only be tested where A+B was
significant.
cheers,
scj
>>> Ged Ridgway <[log in to unmask]> 2/20/2007 9:03 AM >>>
Sterling Johnson wrote:
> regarding the 'mask with other contrasts' question where you want to
> mask with 1 0:
> to use 1 0 as a mask you need to use anova without constant (even
> though you really have a two sample t, the t-test design matrix puts
the
> constant term in, so you cant specify contrasts that sum to anything
but
> zero).
Hi Sterling,
Completely agree with you, but I think it should also be possible to
do everything with an ANOVA-with-constant model still, though I might
have totally lost the plot...
As I see it, for a two-group design, with a constant, you might have
something like this:
test = [kron(eye(2), ones(3,1)) ones(6,1)]
where SPM automatically right-pads a contrast of [-1 1] with a zero to
make it the correct length. This [-1 1 0] contrast is then B>A.
The first beta on its own is inestimable in such a model, but I think
the effect of A, equivalent to a [1 0] contrast in a model without the
constant, would be [1 0 1], and that the A+B effect that I think it
makes sense to mask the B>A contrast of interest with, would then be
[1 1 2], which is estimable, despite not summing to zero:
spm_SpUtil('isCon', test, [1 1 2]')
Any help?
Ged.
|