Dear David,
A more general reply (which will not really answer your question I'm afraid), whether to take into account reaction time (and if so, in which way) is not trivial, as one would have to have an idea about the time courses of the underlying neural processes, which is unkown in many, if not most instances. Quite likely and as stated by Colin already, different regions will also behave differently. As an illustrative example, early visual regions might behave as if they were "on" for a fixed, short period, then there might be intermediate processing steps which cause the observed RT differences between trials, and then motor regions might be activated to a roughly fixed period just before the button press is recorded. In that case, the intermediate processes would best be reflected by durations that correspond to trial-specific RTs, while for the last process the time dependence would best be reflected by "on" periods of fixed durations but time-shifted onsets. Even in this simple example one can easily think of various confounds, motor regions might be pre-activated during the whole period, not just for execution, early visual cortices might be modulated by attention or working memory (possibly still "on" despite the stimulus to be "off"), the "on"-"off" pattern might just be insufficient (adaptation effects, accumulating evidence leading to building-up activations, ...). Now, depending on the processes involved (RT-dependent yes/no/combined with an RT-independent process) the "optimal" predictor might look very different, and accordingly, a certain (set of) predictor(s) might lead to a severe bias for one experiment, while it is perfect for another.
Now often, there are RT differences between conditions of different magnitude for different subjects, and there might be additional overall RT differences between subjects. Depending on predictors this will bias second level statistics in one or another direction.
Thus in general, FIR seems to best insofar as it has the least assumptions and is the most flexible. One could then look at the exact time course, one could determine the area under the curve, one could contrast the peaks, ... This still leaves unanswered how to treat trial-specific differences (as the FIR would basically average across all trials of a condition), but one could go with the beta series strategy combined with FIR, always modeling one trial separately and collapsing the rest. Now in practice, I'm not aware of any such study (and there are only very few that report FIR results on whole-brain level without considering trial-specific RT differences), likely as it would result in a massive amount of data that would have to be brought to second level. This, however, is a rather poor argument if you think about it. It would be different if another study had shown that, given a particular paradigm, regions xyz behave in this or that manner, which could be approximated by a certain (set of) predictor(s). This information is unavailable almost all the time though. Funnily, it is more common to switch to FIR when it comes to ROI analyses, often conducted after whole-brain analyses based on canonical HRF predictors. The FIR time courses often look very un-canonical, (which might, in some instances, be due to misspecifications), which challanges the whole-brain results then. Anyway. ;-)
Best regards
Helmut
|