After great help from SPM experts in email and above, I've been able to see that what I thought was strange behavior in my original post is exactly what you want SPM to do!
Thank you to the message board! - Guillaume Flandin and Rik Henson emailed with clarification and Donald McLaren replied above.
On the face of it, the adjustments to regressors due to T0 can be a bit counter-intuitive, and that (plus a reinforcing comment from someone in my dept.) unfortunately kept me from seeing what SPM should be doing (which in retrospect is embarrassingly obvious).
Here's part of a note from Rik that may help future users:
"As you know, SPM simulates evoked responses with (SPM.xBF.)T timebins every TR. The value of (SPM.xBF.)T0 is then the specific bin taken to represent the predicted response for a given scan in a regressor (see spm_fMRI_design.m). Thus if you increase T0=1 to T0=T, then you are sampling that simulated response *later*, so that the downsampled regressor will appear to be shifted *earlier* in time (as you found)."
For a specific example, imagine there are just two slices, acquired 2 s apart (without slice-time correction). An event occurred and you're looking at the resulting signal. If the first slice captures the HRF at the peak, the second slice will capture the HRF 2 s after the peak. Regressors that aren't adjusted will capture the first slice well, but in the second slice, the regressor will peak 2 s after the actual HRF peak. To capture the HRF peak in second slice, and generally to align to the expected signal in that slice overall, you will need to shift the regressor 2 s *earlier* in time.
Best regards,
Elliott
|