Dear Marko,
as the paired t-test is probably based on the beta images you will indeed run into some scaling issues, at least if "event-related" means a duration of 0 s. See the picture in https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=spm;71399e6e.1402 , all other durations are scaled in a certain, compatible way, but not 0 s (as pointed out by Donald).
But leaving this aside, I'm not sure what the message of this test is. Are you interested in a sort of "goodness of fit"? This would probably require some other form of statistics (Bayesian model comparison?).
Or do you want to test for different types of activation, like "transient activation" that is directly linked to a certain well-defined event (button press) vs. more diffuse "sustained activation" like motor preparation (which might be better captured by a block regressor)? In that case I'd suggest to integrate both the block regressors and the event-/epoch regressors into a single model. There should be quite some papers in different contexts, but searching for sustained/transient should be a good start. One paper I explicitely remember is Chawla et al. (1999, Nat Neurosci, "The physiological basis of attentional modulation in extrastriate visual areas"), which is based on block regressors for the sustained activation ("set-related baseline activity") and event-related regressors for transient activation ("stimulus-evoked activity"). Their methods might be a little confusing, as they don't go with event regressors A and B, but rather an "average regressor" modeling both A and B plus a "differential regressor" A - B, same for the block part (actually it remains unclear whether they did include an "average" block regressor as well or just a "differential" block regressor, but this is another issue). You might also want to look at Mechelli et al. (2003, Neuroimage, "Comparing event-related and epoch analysis in blocked design fMRI") and Petersen & Dubis (2012, Neuroimage, "The mixed block/event-related design"), the latter especially for caveats/limitations.
Finally, make sure the regressors don't correlate too much, if events are close enough the predicted time course is very similar to a block regressor. Depending on purpose it might be worth to think about orthogonalizing one regressor to the other, or instead of block regressor B and event-related regressor E go with B and a regressor that reflects the difference B - E (which should avoid issues with multicollinearity).
Best,
Helmut
|