> If we know the reaction times (RTs) for trials in an
> event-related paradigm, what is the best way to include RT
> in 3dDeconvolve as a nuisance variable?
I am not sure about your case with reaction times, but a regressor in a regression model with 3dDeconvolve has to have the same length as your stimulus timing files. So if you want to co-vary out some effect of no interest, you would have to code such an effect in terms of each TR. I wonder if you really do it this way you may need to code the reaction time for the whole duration of each event/block response.
> Or, should we create a separate vector file for fast and
> slow responses and then produce a GLT for each
> separately (and then compare them)?
If you can somehow categorize your tasks/events based on the length of reaction times, that may not be a bad idea. The disadvantage of doing it this way: (1) you double the number of your regressors, losing some degrees of freedom; (2) such a categorization is approximate and could be inaccuate.
Gang