Uplift Modeling & Heterogeneous Effects

The ATE — average treatment effect — is a scalar summary of a heterogeneous reality. 'The promo adds 2% conversion on average' can mean 'everyone +2%' or 'some +15%, some −8%, and the rest nothing'. Those worlds need opposite policies: blanket vs targeted.

CATE — the conditional average treatment effect, E[Y(1)−Y(0) | X] — is the effect for this user. It's the quantity targeting decisions actually consume, and it's unobservable per-unit (you never see both potential outcomes — the 'fundamental problem of causal inference'). Everything in uplift modeling is machinery for estimating it anyway, from randomized or credibly-adjusted data.