For example Henson et al. (2001, http://www.fil.ion.ucl.ac.uk/spm/doc/papers/rnah_choice.pdf ) use short stimuli in combination with a window length of 24 s, corresponding to 12 time bins. The exact number of time bins = "order" won't make much of a difference probably. I think there's also no particular rule. Concerning window length / order, I would stay with one time bin a TR though, otherwise you probably overfit the model.
For evaluation purpose, it is certainly useful if the overall length of the FIR model *somewhat* corresponds to the length of the expected BOLD response based on the canonical HRF model, even if you are unable to detect anything with the last few time bins (as the activation is already close to 0 again). You might also add some more if you expect some special, more prolonged responses. But this depends on the research question and your reasons why you've decided to run a FIR model.
However, I think in your case it should actually be a window length of about 50 s. Are you sure it's really 25 s and not 25 * TR with a TR of 2 s = 50 s? The unit in that SPM plot is volumes = TR units. At least 50 s would correspond nicely to a block length of 18 s block plus these mentioned 30 s afterwards.
Best,
Helmut
|