I can see how this would work for larger minorities, but it doesn't seem to scale (down): the smaller a minority, the higher variance you will see in your data.

I can see how this would work for larger minorities, but it doesn't seem to scale (down): the smaller a minority, the higher variance you will see in your data. I'm not sure it helps at all when you get down to the point where you have only a single member of a group in your population.