We're running an fMRI experiment with a mixed block-event related design, and trying to analyze it with a model in SPM which includes both events and block regressors, to examine transient and sustained activity. Visscher (2003) and some other authors recommend modelling blocks with canonical HRF but events with FIR. However we are unsure about some points regarding the most appropriate way to analyze these data:
(1) A previous discussion (https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=SPM;3e5782f7.0806) proposed a method to include in SPM regressors modeled differently (FIR+ canonical HRF) within the same GLM. However, we're not sure whether extracting the HRF modeled regressors in the model specification phase, and then including them as regressors (not conditions) in a second model in which we specify FIR regressors is the proper way to implement this. We followed these instructions, but we found bizarre results. Is this right procedure or is there a better alternative?
(2) Each trial is composed by events with different durations (and one of them is jittered across trials). Should we use different durations of the intervals modeled by the FIR functions, depending on the events’ duration? Or is it ok if we just make a total window length of 20 s (HDR duration) for all types of events?
(3) Finally, a couple of doubts come to us referring the statistical analysis with FIR. We've thought of using 9 time bins to model the events, having 9 regressors for each of them. If we'd like to analyse differences in activity associated with two different event classes (which we would normally do with a T contrast 1, -1 if we only had 1 regressor), how should we proceed?. I've seen many papers where they use ANOVA, but, should we use an ANOVA for both the first and second levels? Or is it better to use T tests for the first level and then the ANOVA for the second one?