Dear Gang,
Thank you for your response. I understand that GAM and dmBLOCK does not assume the undershoot, but around 5 s, I expected the direction of effect to match the TENT model. But anyways, it seems using TENT-basis model better catches the time course.
I married the 1D files (onset:duration) term for stim_times_IM, rather than modelling different durations of delay separately. Because I am not really interested in effects due to the delay period duration, I wanted to remove the potential differences arise from different duration.
For your second point on how to untangle the two phases as they are yoked.. So I thought about modelling them as one regressor from the onset of cue. I am still bit juggling around to decide - But I think it is important to model the delay period as I expect some amount of neural activity for WM maintenance..
Thanks for your input!