> It seems like it's very difficult to come up with a bad sequencing in
> the modelling.
That's probably true. Instead of hunting for an "optimal" design, I would rather view the process as a filtering process to avoid bad ones. Plus the optimality here is relative in multiple senses: (1) only among those designs randomly generated in the program; (2) depending on the modeling approach: gamma HDR, or other basis functions? (3) depending on the specific contrasts; (4) how do you balance the efficiencies when you have multiple contrasts?
> However, in real life, it seems like there are "stories" about how it is
> very easy to come up with an identity problem in your event related
> design. So, is it that this modelling isn't very reflective of real life or
> is the fear of multicollinearity a little unwarranted as long as you
> randomize and jitter?
My feeling is that the probability of encountering a multicollinearity problem is way much higher when the user mistakenly fills in two or more identical regressors in the model than when a really bad design strikes. Still design optimization is a good practice and worthwhile safety check.
Gang