Thanks for the advice, Rick. I'm working on the project with Diana, and the above pictured IRFs are averaged across a functionally-determined ROI of >20 voxels. We used a TR of 3s, presuming that a lag of 18 seconds would be sufficient for estimating the response. We did not use the TENT function, and had randomized the stimuli with two conditions (there are four IRFs pictured because we are comparing two different groups of subjects), and we had a variable ISI of 6-18 seconds, randomly determined. We did not run into any multi-collinearity problems, and tested the design in 3dDeconvolve with a dummy dataset prior to data acquisition.
We can try longer lags to see if it affects the results.
cheers,
George