Comment by mjburgess

If the confounders are fat-tail distributed etc. then arbitrarily large samples can still be inadequate.

The idea that even thousands a data points in subgroups are going to be 'well mixed' relies on extremely strong assumptions about the distribution of those traits.