[VoxBo] Question about multiple contrasts in a glm

[Daniel Y Kimberg]

David January wrote:
> I do want a positive slope to the line plotting activation as a Y
> against condition as an X.  The reason I specified B > Rest in my
> initial email is because I wanted to rule out getting a good fit
> with a contrast that had A < Rest, B = Rest, and C > Rest.  The 1*A,
> 1*B, and 1*C contrast precludes that, right?  Or would you recommend
> that I find a region where (1*A, 1*B, 1*C) > rest and then look for
> the (-1*A, 0*B, 1*C) contrast is significant?

Including the 0/1/1/1 covariate in your model means that none of the
other covariates' weights will be affected by task-rest differences in
signal.  If you test a contrast weighting just that covariate, you can
test the null hypothesis that there is no difference between rest and
the average of A, B, and C.  When it's significant, it doesn't
guarantee that any specific one of the three is individually different
from rest.

If you test your 0/-1/0/1 covariate, your t test won't be sensitive to
where A, B, and C are relative to rest, but the test will tell you if
your slope is significantly positive.  As I mentioned before, a
positive slope is not the same thing as A < B < C, it could easily be
consistent with a true state of affairs where B < A < C, for example.

I can't think of a meaningful interpretation of a single contrast that
includes both of these covariates.

> My problem with that is I wanted to do a whole brain analysis to
> find regions that showed the parametric effect--I just don't want to
> find regions where A elicits less activation than Rest.

None of the covariates in your model directly compares A and rest.
How about a model with just the following covariate: 1/-1/0/0.  The
right decision rule depends on what you're trying to accomplish.

dan