Dear Paloma,
With different trials starting at different time points relative to TR onsets one usually expects different activation levels after certain moments in time. Going with a single FIR set (which rounds trial onsets to the next or nearest TR) means trying to predict different activation levels at e.g. +/- TR/2 relative to stimulus onset by a single value then. Whether this is a large confound or not is going to depend on the data. Leaving this aside, you might also introduce systematic biases if the "onset discrepancy" differs between conditions and/or subjects. But as stated, it is going to depend on the exact hypotheses whether to go with a FIR set / similar predictors or not, mixed design doesn't mean you need FIR, but it might be employed frequently due to people being interested in the exact time course of the transient response.
BTW, Dosenbach et al. don't use the canonical HRF at all, sustained activation is modeled as an unconvolved boxcar function ("with a square wave") shifted in time relative to the onsets (similar to the approach in Donaldson et al., 2001, https://dx.doi.org/10.1006/nimg.2000.0664 ). This seems to be more consistent IMO than a mixture of FIR and canonical HRF.
Best
Helmut
|