I liked Dan's post very much. The magnitude of the forfeiture is important in considering the balance. If bending timing drops predictive validity from .5 down to .42, but improves representativeness, I'm okay with that. If we're talking a drop from .6 to .2, that's a whole other story.