> it is recommended for FIR modeling to start the events on the TR because of the way the models are constructed
Well, that's the critical aspect. We can indeed couple trial onsets and FIR time bin onsets, or we can use a restricted number of "gaps" modelled separately, and be happy ;-) My question is aiming at random presentation, in which many separate (sets of) regressors would be required, making this approach rather ineffective, leaving aside that it would result in many amplitude estimates for different time points, each of them based on just a small number of trials, which might not necessarily be what one wants to obtain.
Thus, if trial onset is random (within the limitations due to frame rates), shifting time bin onsets to the nearest full TR is not ideal, as e.g. onsets at 1 s, 2 s, 2.999 s would result in the same series of predictors based on a TR and FIR time bin resolution of 2 s, microtime resolution 1000. Assuming very brief stimuli resulting in a BOLD response similar to a canonical HRF then the average of the three time courses is actually quite similar to that of the 2 s onset event (the deviation of the 1 s and 2.999 s functions from the 2 s function is quite symmetric), thus the FIR approach might also be sufficient to model an average response. However, there's some remaining error, and it's exactly this +/- 1 s which we sometimes try to cover with the temporal derivative in standard GLMs.
Now, within the GLM it would also be possible to adjust the predictors. This is also suboptimal, but maybe to a different extent, as it would allow to take into account the size of the deviation (is the gap just 10 ms or almost 1 s).
Best
Helmut
|