This adds a lot of complications, and is not a great idea. I think cohorts is a better version of split_out.

It would be a pretty decent simplification of the code.