Dear Andy,
I wouldn't recommend to exclude data points. However, some studies go with the canonical HRF in combination with shifted onsets to only model the period in which activation has reached a certain level, that is, instead of onsets x and durations y you would go with x + a and y - a. a might be derived from literature, preliminary findings, ...
However, if you have reasons to believe that the response does not conform to the canonical HRF an "adjusted" version might still be a crude approximation (e.g. you don't control for the "fading out" effect) and/or just arbitrary. Thus, in case you don't have good reasons (BOLD data - subjects might report that they don't smell anything any more at a certain point, but maybe there's still some activaiton going on) to justify certain shifts/adjustments I would go with a FIR model, which should be more elegant from a technical perspective and more informative as you can directly target effects that occur over time within a particular trial. In case of FIR you also wouldn't need a separate conditon "transition", instead you would adjust the length of the time period under investigation.
FIR is employed frequently when it comes to plotting responses of certain ROIs, but it is possible to go with whole-brain analyses as well - although this is less common, probably due to more complex designs as you introduce another factor "time"). See Windischberger et al. (2008, J Neuroci Methods, "Time-resolved analysis of fMRI signal changes using Brain Activation Movies") to get an impression of the approch.
FIR might not be perfect when subsequent trials are rather close together, but this should hold for a default HRF model as well then (if a "transition" regressor overlaps with the following trial).
Best,
Helmut
|