> As such, I have first-level canonical HRF > zero baseline .con images (1 0)
> for Remembered and Forgotten activity events and "deactivation" events (-1
> 0), ...
At the first level, estimating the model involves generating betas for all of
the contrasts that are in the model. As an additional step, to combine the
betas together, you specify contrast vectors which result in the generation of
con images (like, con_0002.img). These con images are just weighted
combinations of betas-- a con vector of (1 0) will be the same as B1, and (-1
0) will be -1*B1. So, even at the first level, the "deactivation" contrasts
are the negative of the "activation" contrasts.
To test this theory, try taking an "activation" and a "deactivation" T image
from the first level, and opening them in an editor that lets you see the
neagtive part of the t-image, which is usually colored blue instead of orange.
MRIcro is a particularly good tool for this. The blue parts in the
"activation"
image should be the orange parts in the "deactivation" image.
> When I run a random-effects two-sample t-test between Group A and B for the
> Remembered activation > zero baseline .con images, the 1 -1 contrast
> (assuming this means Group A > Group B for Remembered activity > zero
> baseline) ends up being the same result as a two-sample t-test between Group
> A and B for Remembered activation < zero baseline ("deactivation") where the
> contrast is -1 1. I'm not certain how to contrast "deactivation" or why
> these results mirror each other.
These results follow from the first part. Does that make sense?
> Also, my head is swimming. What exactly is a second-level, random-effects
> of individual, fixed effect 1 -1 (Remembered vs. Forgotten) contrasts
> telling you? Am I correct in assuming that the two-sample t-test contrast
> (1 -1) reflects activity that is greater for Group A than B for
> event-related activity that is greater for remembered vs. forgotten?
> Conversely, does the two-sample t-test contrast (-1 1) yield activity that
> is greater for Group B than A for event-related activity that is greater for
> remembered vs. forgotten?
In either case, your random effects analysis asks whether the effect of a
difference between remembered and forgotten is significantly different between
groups. The sign will flip according to the way you set it up. And if you
flip both contrasts, you get the same thing back.
In other words, an effect of (R > F) on (A > B) is exactly equivalent
to (R < F)
for (A < B).
Ken Roberts
|