Even for a "fantastic voxel" the seed is not a linear
combination of the task regressors, it contains noise.
Regardless, the goal is to find a fit to the interaction
term, perhaps while ruling out a fit to just the seed.
Even if the seed happened to be a combination of task
regressors, we would not care (multicolliearity aside),
because those terms are all considered "noise" in the
model. Only the interaction result matters.
But the aspect of the seed that we really care about is
above and beyond the task regressors, so task regressors
should not really matter in any case.
Along these lines, it is generally a nice idea to start
with the residual time series when extracting the seed
to begin with. There is no reason to leave the seed
contaminated with any signals that one might consider as
noise from the view of the new model.
- rick