Uplift Modeling

S-learner, T-learner, X-learner, R-learner — heterogeneous treatment effects and targeting

A marketing team has a $1M budget for promotional emails. They can send to 1M users, but only 200K will benefit from the discount — people who would not have purchased without it. Sending to users who would purchase anyway wastes the discount cost. Sending to users who actively dislike being contacted is counterproductive. A simple propensity-to-purchase model predicts who will buy — but that is not uplift. Uplift is: who changes their behavior because of the treatment?

Four user types define the targeting problem. Persuadables will buy with treatment and not without — your target. Sure Things will buy regardless — wasting budget on them yields zero incremental revenue. Lost Causes will not buy regardless. Sleeping Dogs will buy without treatment but not with it — negative uplift, worse than doing nothing. A model maximizing purchase propensity concentrates budget on Sure Things. An uplift model maximizes incremental effect.

Uplift estimation targets τ(x) = E[Y(1) − Y(0) | X = x] — the Conditional Average Treatment Effect (CATE). You cannot observe τ(x) directly because you never observe both Y(1) and Y(0) for the same person. Meta-learners estimate it from RCT data. Two-model (T-learner): fit Y ~ X separately for treated and control, subtract predictions. S-learner: include T as a feature, fit one model. X-learner: imputes individual treatment effects and then builds a CATE model by borrowing strength across groups — better for imbalanced treatment/control splits. R-learner: fit nuisance models m(x) = E[Y|X] and e(x) = P(T=1|X) on one data fold, then solve Y − m(X) = τ(X)(T − e(X)) + ε by regressing the outcome residual on the treatment residual on a held-out fold (cross-fitting). This residual-on-residual step is what makes it orthogonalized: Neyman orthogonality means the loss's gradient with respect to τ vanishes at the true m and e, so small errors in the nuisance models don't bias τ̂ to first order — R-learner is often preferred when nuisance estimation is noisy or hard, such as with imbalanced treatment/control splits. Causal forests: non-parametric CATE estimator with valid confidence intervals.

Worked T-learner example: suppose the RCT gives a high-spend segment a treated purchase rate μ̂₁ = 0.52 and control rate μ̂₀ = 0.48, so τ̂(high-spend) = 0.52 − 0.48 = 0.04. A low-spend segment gives μ̂₁ = 0.31 and μ̂₀ = 0.09, so τ̂(low-spend) = 0.31 − 0.09 = 0.22 — 5.5× the incremental lift of the high-spend segment, even though its raw purchase rate (0.31) is lower than high-spend's (0.52). A response model, which ranks purely on μ̂₁, would rank high-spend above low-spend and get the targeting backwards; the T-learner's subtraction recovers the correct ranking.

Evaluation: Qini curves rank users by predicted uplift descending. At each percentile of targeted users, compute cumulative incremental outcome in the treatment arm versus the control arm of a held-out RCT split — because both arms are real, this is a direct measurement of incremental effect, not merely a ranking-quality proxy. Summing that gap across percentile bins gives the AUUC; when treatment and control groups are equal-sized this equals the plain unnormalized uplift-curve area, but a genuine Qini curve additionally rescales the control arm's cumulative outcome by the treatment/control size ratio, so the two areas diverge once the groups are imbalanced. What one offline Qini curve does NOT catch: tuning a model repeatedly against the same historical RCT split can overfit to that split's noise, and a historical split can miss distribution shift after deployment (new users, new season, a competitor's promo). A fresh, post-deployment no-contact holdout — held out from all tuning — is what catches those two failure modes; it is not required because offline Qini fails to measure real treatment effect, which it does measure when built from a genuine RCT split.

What a response model is not: an uplift model. A response model predicts P(purchase) — dominated by Sure Things. An uplift model predicts P(purchase | treatment) − P(purchase | no treatment) — targets Persuadables. These are orthogonal quantities. Deploying a response model as an uplift model wastes marketing budget and misses the actual causal effect of the treatment.

Key points

• Keep a fresh, post-deployment no-contact holdout that was never used for tuning. An offline Qini curve built from a real held-out RCT split already measures actual incremental effect — it compares treated vs. control outcomes at each percentile, so it is not merely a ranking-quality proxy. The real risk is that repeatedly tuning against the same historical split can overfit to its noise, and that split can miss distribution shift after deployment. A never-tuned-on post-deployment holdout is what catches both of those failure modes, which is why it's required in addition to (not instead of) offline Qini.
• Trap: using the T-learner on highly imbalanced treatment/control designs (95% treated, 5% control). The control model has very little data and its predictions are noisy. Errors from both models compound in the subtraction — variance of τ̂(x) is dominated by noise in the smaller group. Use X-learner or causal forests, which are specifically designed for imbalanced designs and borrow strength across groups.
• Diagnostic: if your model assigns high uplift to users with high baseline purchase rates — high overlap between uplift deciles and response deciles — the uplift model is collapsing to a response model. Check the correlation between CATE estimates and baseline propensity. Worked example: five baseline-propensity quintiles of [0.09, 0.20, 0.35, 0.48, 0.61] paired with τ̂ of [0.22, 0.15, 0.09, 0.04, 0.01] give corr(τ̂, propensity) ≈ −0.98 — healthy, since uplift falls as baseline propensity rises (Persuadables cluster at low baseline propensity here). If instead τ̂ tracked propensity directly, e.g. τ̂ = [0.03, 0.07, 0.13, 0.17, 0.22] rising alongside those same quintiles, corr(τ̂, propensity) ≈ +0.99 — over the 0.7 threshold, meaning the model is not capturing incremental effects, it's just re-deriving baseline propensity. Segment users by baseline propensity quintile and verify that τ̂(x) varies within each quintile (not just across quintiles) before trusting the targeting.

Recap

• Uplift is not propensity: a response model predicts *who will convert* (dominated by people who'd buy anyway); an uplift model predicts *who converts because of the treatment* — orthogonal quantities, and confusing them wastes budget on people who'd have bought regardless.
• Four user types define the targeting problem: Persuadables (buy only if treated — your target), Sure Things (buy regardless — wasted discount), Lost Causes (never buy — wasted), Sleeping Dogs (buy *without* treatment but not with it — negative uplift, actively backfires).
• Target τ(x) = E[Y(1) − Y(0) | X=x], the CATE: you can never observe it directly (never both potential outcomes per person), so you estimate it from RCT data with a meta-learner.
• The meta-learners: S-learner (include T as a feature, one model), T-learner (fit treated and control separately, subtract), X-learner (impute individual effects and borrow strength across groups), R-learner (residualize outcome and treatment against nuisance models m(x), e(x) via cross-fitting, then regress residual-on-residual — Neyman orthogonality keeps τ̂ robust to small nuisance errors), causal forests (non-parametric, with valid confidence intervals).
• Use X-learner (or causal forests) for imbalanced splits (e.g. 95% treated / 5% control): the T-learner's small-group model is noisy and its error compounds in the subtraction, whereas X-learner borrows strength from the larger group to stabilise.
• Evaluate with a Qini curve / AUUC — rank users by predicted uplift and measure cumulative incremental outcome in the treatment arm vs. the control arm of a held-out RCT split; because both arms are real, this already measures actual incremental effect, not just ranking quality. The real risk is overfitting the model to that one historical split, or a post-deployment distribution shift it can't see — guard against both with a fresh, never-tuned-on, post-deployment no-contact holdout.
• Collapse warning: if corr(τ̂, baseline purchase propensity) exceeds ~0.7, the uplift model has degenerated into a plain response model — it's ranking Sure Things, not Persuadables.

Check your understanding