Chapter 39 of 57

ATE vs CATE: When the Average Effect Hides Everyone

Created Jul 5, 2026 Updated Oct 25, 2026

"The email campaign lifted conversion by 3 percentage points." That 3 pp is the Average Treatment Effect (ATE) — the effect averaged over the whole population. It's the right headline number for should we do this at all. It's the wrong number for almost every decision after that, because an average is silent about spread.

The Conditional Average Treatment Effect (CATE) is the effect as a function of who you are:

$\tau(x) = E[\,Y(1) - Y(0) \mid X = x\,]$

— the average effect among units whose covariates are x. Note the word average: it is still a group quantity, not a promise about any one person, and for a binary outcome it is a change in the probability of converting. And here is the uncomfortable part — two treatments with the identical ATE can have completely different τ(X):

Treatment A: τ(X) ≈ +3 pp for every group. Raises everyone's conversion probability, mildly.
Treatment B: τ(X) = +30 pp for half the population, −24 pp for the other half. The same 3 pp average — but for half the population it lowers the probability of converting.

The ATE cannot distinguish these. Roll out B to everyone because "it works on average," and you've made things worse for half your users while the topline looks fine.

Hold the ATE fixed and slide heterogeneity up: the average never moves, but the CATE curve τ(X) tilts and a growing slice of the population falls below zero — groups for which the treatment's average effect is negative. The dots are individual effects, which exist here only because the world is simulated; the line is the conditional average, which is the part a real study can estimate. Two campaigns with an identical headline ATE can mean "helps everyone a little" or "helps half a lot and hurts the other half," and the ATE can't tell them apart.

Surfacing τ(X) is exactly what the modern causal-ML toolkit is for:

Uplift modelling — rank people by their predicted conditional effect, then treat only the segments the model says will actually respond. In marketing the vocabulary is persuadables, sure things (who'd convert anyway) and sleeping dogs (whom contact annoys into leaving) — useful categories, though no model can place a given person in one of them with certainty.
Causal forests / meta-learners (T-, S-, X-learners, DR-learner) — machine-learning estimators built to output τ(X) rather than a prediction of y. Note the difference: a good predictive model tells you who is likely to convert; a good causal model tells you whom the treatment will change. Confusing the two is the classic mistake.

The practical punchline: when effects vary, targeting on τ(X) can beat treating everyone — most clearly when the treatment costs something, when the budget is finite, or when some groups sit below zero. It is not automatic: if every group benefits and the treatment is free, treating everyone is fine, and a targeting rule built on a badly estimated τ(X) can do worse than no targeting at all.

The full version — ATE vs CATE, uplift, double ML, causal forests, and the predictive-vs-causal trap — is in the Causal ML Beyond Econometrics note.