Is this what you want to do:
Run separate regressions with various models, then pick out the model that "wins" in each voxel or region.
If so, one way to proceed would be to create the separate time series models with waver (e.g., with the -GAM option and various parameters to stretch it out, or with the -EXPR option), then do some separate 3dDeconvolve runs, then do the voting based on the overall F statistics.
If the above
is what you want to do, then we can try to fill out the details.
bob cox