Gang Chen wrote:

> My suggestion is that you consider running 3dDeconvolve with
> this stimulus as usual for this subject's data instead one of
> all zeros for two reasons: (1) If you feed in all zeros for
> this stimulus, 3dDeconvolve would assume that those moments as
> resting state and would use the data points for baseline
> estimation, which is not what you want.

Gang,

This confuses me. The problem is that the vectors are created from the behavior and sometimes a particular behavior (say, a kind of error -- call it Event Type X) never occurs. What would be good would be to be able to have this vector coded as all zeros (at no time point did we have this stimulus type) that is entirely ignored by 3dDeconvolve. That way, we can keep our big, nasty 3dDeconvolve calls and .mat files the same for each participant and not need to customize them.

Now, your point (1) I don't understand. At each point in time (on each TR, say), some event did happen. It's just that Event Type X never happens. So, summing across the rows, there is a "1" somewhere, for a different event type did happen (e.g., they didn't make an error). So, I don't see how anything else is going into the baseline here. Really, it's like you take an existing, perfectly good 3dDeconvolve call and say, "I want to have a vector that will code for TRs in which a meteor happens to strike planet Earth at that point in time". It never happens and your call and results should be the same -- any beta times a vector of all zeros will be zero and add nothing into the model.

> (2) With a usual
> stimulus file without zeroing anything, you could keep
> everything consistent, and still run all the glt analyses as
> usual. Since you are doing deconvolution, the worst part is
> that you would not want to look at the results of the impulse
> response function of this stimulus and any glt tests that
> involve this stimulus.

Again, if the event type never happened, we could assume a zero beta vector for this stimulus. Granted, doing cross-participant tests with these artificial zeros would be incorrect, but our goal is really not to ever analyze these vectors in a cross-participant analysis. The goal is simply to account for variance within the EPI data as best we can. Some times, an odd error pops up, and we want 3dDeconvolve to have a shot at pulling that variance out of the baseline in an intelligent way that's also convenient (one could lump a bunch of error types together, but if the truth is that their activities differ, you'd end up having more unaccounted for variance go into the baseline).

Craig