Slow drifts due to motion are probably modeled well enough.
Using a slow event related design does not really change
things. If the 2 peak TRs in the middle of a reponse curve
happen to get censored, that does not really matter either
(beyond losing a couple of high variance TRs for the model
estimates).
Also, I do not see any gain in censoring more TRs from that
particular reponse. The other response TRs still help to
estimate the beta of interest, which is to say, we do not
expect the BOLD response from your stimulus of interest to
be affected by the motion (unless the response is in an
area of the brain where there is an actual BOLD response
due to the motion, but that is typically not a concern).
- rick