Introduction to S/T/X/R Meta-Learners for CATE Estimation

A tutorial from zhichai.net on estimating individual-level causal effects — the principles, quirks, and selection criteria of the four most common meta-learners. Full translation below (source is a long-form article; structure preserved).

The Problem

Given 20,000 users and a budget to send 10,000 promotional SMS messages, how do you decide *for each individual* how much extra revenue the message would generate? Standard approaches fail:

• Classification of "will treated users convert" is meaningless — untreated users never had the chance.
• A/B testing only gives an average, masking heterogeneity (some users respond strongly, others negatively).
• Naive regression with treatment as a feature captures correlation, not causation.

What we want is the CATE:

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

Key points

• Potential outcomes: Every individual has two outcomes, Y(0) and Y(1), but only one is ever observed (the Fundamental Problem of Causal Inference, formalized in the Rubin Causal Model). All meta-learners are strategies for inferring the unobserved half.
• Three adversaries: confounding (e.g., user activity drives both purchases and receiving promotions), selection bias in treatment assignment, and heterogeneous treatment effects (HTE).
• Shared assumptions: all four learners assume unconfoundedness (all confounders observed in X), SUTVA (no spillover between units), and overlap (both groups present for any covariate profile).

The Four Learners

S-Learner (Single)

Fit one model μ(x, t) with treatment as an ordinary feature; estimate τ̂(x) = μ̂(x, 1) − μ̂(x, 0).

• *Pros*: dead simple; one model, two queries.
• *Cons*: regularized models (XGBoost/Lasso) tend to zero out the low-importance binary treatment feature, collapsing τ̂(x) ≈ 0; extrapolation is unreliable since each x is only observed at one t.

T-Learner (Two)

Fit separate models μ̂₀(x) and μ̂₁(x) on control and treated data; estimate τ̂(x) = μ̂₁(x) − μ̂₀(x).

• *Pros*: no regularization bias against T; each arm can use its own model class.
• *Cons*: each model sees only half the data (higher variance); if the two groups' covariate distributions differ, both models extrapolate independently with poor tail accuracy.

X-Learner (Künzel, 2017)

Designed for severely imbalanced arms (e.g., 100 treated vs 900 control). Three steps:

1. Fit μ̂₀ and μ̂₁ as in T-Learner. 2. Compute imputed individual effects: D̃₁ᵢ = Y(1)ᵢ − μ̂₀(Xᵢ) for treated units, D̃₀ⱼ = μ̂₁(Xⱼ) − Y(0)ⱼ for controls — each group "borrows" the other's model to imagine its counterfactual. 3. Fit effect models τ̂₁(x), τ̂₀(x) on D̃₁, D̃₀ and combine: τ̂(x) = g(x)·τ̂₁(x) + (1 − g(x))·τ̂₀(x), where g(x) is the propensity score weighting toward the larger group.

• *Pros*: the go-to method when treated samples are scarce (medical trials, rare-event interventions, marketing canary rollouts).
• *Cons*: dependent on the quality of the first-stage models; degrades under poor overlap.

R-Learner (Robinson, 1988)

Starts from the partially linear model Y = m(X) + τ(X)·T + ε. Robinson's residualization gives:

where e(x) = P(T=1 | X=x) is the propensity score. First estimate outcome residuals (Y − m̂(X)) and treatment residuals (T − ê(X)) with any ML models, then solve for τ as the slope between the two residual sets, typically minimizing a Robinson-style loss. This makes the effect estimation robust to errors in modeling m(x).

Which one to use

| Situation | Recommendation | |---|---| | Huge data, treatment correlated with X, non-regularized base model (e.g., deep nets) | S-Learner | | Both arms well-sampled, distributions overlap | T-Learner | | Treated arm ≤ ~10% of control, reasonable overlap | X-Learner | | Want robustness to nuisance-model error; can estimate propensity | R-Learner |

Implementation

All four are available in Uber's causalml Python package (causalml.inference.meta: BaseSClassifier/BaseSRegressor, BaseT*, BaseX*, BaseR*), with a scikit-learn-style fit/estimate_ate API.

Limitations

None of the meta-learners fix unobserved confounding (e.g., user mood not captured in data). If ignorability fails, switch to design-based methods: instrumental variables (IV), difference-in-differences (DiD), or regression discontinuity design (RDD).