Dear Jakub,
the best way to model such a "non-event" is indeed to place a single onset somewhere outside the time range of your data. I would go for specifying an onset at the end, since this does contribute nothing to your timeseries model.
Excluding scans would destroy the temporal order of your timeseries not only for invalid trials, but also for neighbouring valid trials (an hemodynamic response has a response function that lasts for ~20 seconds or more before returning to baseline). Also, estimation of temporal autocorrelation would be impossible if scans are missing.
Best,
Volkmar
|