You are asking more from the data then the experiment
was designed for.
Using tents on a trial-by-trial basis is almost like
not modeling at all, just looking at the raw EPI data.
The model is just too flexible, and the stimulus
response periods would have to be almost completely
separate to avoid multi-collinearity. The response
periods overlap in your case as you know, certainly
evidenced by having 2151 TRs and 2400 regressors.
Even if this somehow ran, what do you plan to do with
2000 beta weights? What would the next step be?
- rick