Yes, it seems like making a separate regressor might be the way to go.
Yes, that would mean putting those events in a separate timing file.
The other difficulty here would be in addressing whether some subjects have no such events.
In that case, you might still prefer to have that (possibly empty) timing file for all subjects, just to be consistent with the regression output. But that would also mean adding -GOFORIT options, to tell 3dDeconvolve/3dREMLfit to proceed, even with bad-looking input.
Without all subjects having the non-AM event
This gets slightly messy...
- rick