Hi Dan,
This is akin to resting state analysis, except that you are correct, censoring is not an option here (unless it were combined across the 2 runs, would be more difficult to script). Censoring is okay for seed-based correlations, but that is because all voxels would have the same censoring. Here, you might want to just skip it.
Using 'basic' and 'demean' should lead to the exact same results, except for the polort 0 terms (which you will probably never see). The 'deriv' case also regresses the 6 first differences, so those results will differ. In this case, akin to resting state, it would probably be better to include the deriv terms, too, especially since you will probably not be censoring.
- rick