Treatment Effects ATE Cate Ite

Treatment Effects: ATE, CATE, and ITE

The one idea: "The effect" is not one number — it's a distribution of individual effects that you can only ever summarize. Average it over everyone (ATE) and you get a headline; average it within a segment (CATE) and you get a decision; ask for one person's exact effect (ITE) and you're asking for something that can never be observed.

⚠️ The trap: "The average effect is basically zero, so this treatment doesn't do anything." That feels right and it's wrong. A near-zero ATE can be two large, opposite effects canceling out — a treatment that meaningfully helps one segment and meaningfully hurts another averages out to "no effect" on the topline, and the topline is exactly what hides that.

The last lesson gave you Y1 and Y0 — a unit's two potential outcomes — and named their difference, Y1 − Y0, the causal effect for that one unit. This lesson gives that quantity a name and shows why the level you average it at is one of the most consequential choices in applied causal work.

Three effects, three altitudes

ITE — the Individual Treatment Effect, written Y1ᵢ − Y0ᵢ for unit i. This is the effect of treatment on one specific customer, patient, or account. It is the most precise possible answer to "did this work for this person" — and it is also, per the fundamental problem of causal inference from the last lesson, permanently unobservable. You never see both Y1ᵢ and Y0ᵢ for the same i. The ITE is a real number that exists in principle and that no dataset, however large, will ever hand you directly.