Method library WORKFLOW / 034

Research workflow/4 steps

Uplift modelling with S-, T-, X- and R-learners (CausalML)

@causalml ✓ P D · Jun 29, 2026 · 992 views

Source CausalML — Uber (Chen, Harinen, Lee, Zhao et al.)

@misc{causalml,
  title        = {CausalML},
  author       = {Chen and Harinen and Lee and Zhao and others},
  howpublished = {\url{https://causalml.readthedocs.io/}},
  note         = {Software / documentation}
}

In the margins · Peer notes@targeting_tara · The uplift gain curve over random is the plot that actually convinces ops. Saved. 2

The method workspace

Follow the analysis.

Select a node to explore the method

Method mapInput → 4 steps → output

START WITHInput data

ARRIVE ATResults

Read step 01

Inside the methodReasoning, equations & implementation

Data prep

Outcome, treatment, features

Step notesField note & discussion1

Data preparation — shapes the raw inputs into what the estimator expects.

What happens here

Assemble Y, the treatment flag W, and features X. The estimand is the conditional uplift τ(x).

Formula

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

Reads fromNot specified Feeds intoNot specified

Key code

# y outcome, w treatment (0/1), X features

Reference / docs ↗

Discussion on this step (0)

StatsOtterDiscussion starter · Editorial note

Before using this step, make its assumptions and its output explicit. A useful peer note explains what you would check, what could go wrong, or what would change your interpretation.

Questions to explore

Which assumption matters most here, and what evidence could challenge it?
What would you check before passing this output to the next step?

Peer discussion

Bring your experience, a question, or a different perspective.

Estimation

Fit a family of meta-learners

Step notesField note & discussion1

The core estimate — where the causal quantity itself is computed.

What happens here

S/T/X-learners build τ̂ from base outcome models; the R-learner targets it directly through a residualised loss (Robinson).

Formula

\hat\tau_R=\arg\min_\tau\textstyle\sum_i\big[(Y_i-\hat m(X_i))-(W_i-\hat e(X_i))\,\tau(X_i)\big]^2

Reads fromNot specified Feeds intoNot specified

Key code

from causalml.inference.meta import (
    BaseSRegressor, BaseTRegressor, BaseXRegressor, BaseRRegressor)

Reference / docs ↗

Discussion on this step (0)

StatsOtterDiscussion starter · Editorial note

Before using this step, make its assumptions and its output explicit. A useful peer note explains what you would check, what could go wrong, or what would change your interpretation.

Questions to explore

Which assumption matters most here, and what evidence could challenge it?
What would you check before passing this output to the next step?

Peer discussion

Bring your experience, a question, or a different perspective.

Heterogeneity

Compare learners; choose by validation

Step notesField note & discussion1

Heterogeneity — who is affected, and by how much, not just on average.

What happens here

There's no single best meta-learner — score CATE error and ATE recovery on held-out data and let the data choose.

Formula

\hat\tau^\star=\arg\min_{\hat\tau\in\{S,T,X,R\}}\ \mathbb{E}_{\text{val}}\big[(\hat\tau(X)-\tau(X))^2\big]

Reads fromNot specified Feeds intoNot specified

Key code

from causalml.metrics import get_cumgain
# compare S/T/X/R on a validation split

Reference / docs ↗

Discussion on this step (0)

StatsOtterDiscussion starter · Editorial note

Before using this step, make its assumptions and its output explicit. A useful peer note explains what you would check, what could go wrong, or what would change your interpretation.

Questions to explore

Which assumption matters most here, and what evidence could challenge it?
What would you check before passing this output to the next step?

Peer discussion

Bring your experience, a question, or a different perspective.

Reporting

Targeting: uplift / Qini gain

Step notesField note & discussion1

Reporting — turn the numbers into a figure or table a reader can act on.

What happens here

Rank units by predicted uplift and read the cumulative gain (AUUC/Qini) versus random — the operational payoff of the model.

Formula

\mathrm{AUUC}=\int_0^1\big(g_{\text{model}}(\phi)-g_{\text{random}}(\phi)\big)\,d\phi

Reads fromNot specified Feeds intoNot specified

Key code

from causalml.metrics import plot_gain, auuc_score
plot_gain(df)

Reference / docs ↗

Discussion on this step (0)

StatsOtterDiscussion starter · Editorial note

Before using this step, make its assumptions and its output explicit. A useful peer note explains what you would check, what could go wrong, or what would change your interpretation.

Questions to explore

Which assumption matters most here, and what evidence could challenge it?
What would you check before passing this output to the next step?

Peer discussion

Bring your experience, a question, or a different perspective.

Data contract

Inputs & data

An outcome Y, a binary treatment W, and features X — typically from an experiment or a de-confounded observational sample.

Show data format & exampleHide example

Format — one row per unit: outcome y, treatment flag w, features X.

from causalml.inference.meta import BaseXRegressor
from xgboost import XGBRegressor
learner = BaseXRegressor(XGBRegressor())
cate = learner.fit_predict(X, treatment=w, y=y)

Read the evidence

Results & interpretation

2 figures Workflow discussion 2

01FigureCausalML benchmarks the S-, T-, X- and R-learners: CATE error (MSE) vs. ATE error, train vs. validation — pick the learner that generalises.

02FigureUplift gain curve: targeting by the model captures far more cumulative effect than random allocation across the population.

Figures reproduced from CausalML — Uber (Chen, Harinen, Lee, Zhao et al.) — unofficial community showcase; all credit to the original authors.

Context & interpretation

⚠️ Unofficial community showcase of causalml. Not affiliated with the authors; all credit to them.

Estimate who responds, not just the average: fit a family of meta-learners for the CATE, pick the best by validation error, then rank and target with an uplift curve.

Peer notes

Discuss the workflow (2)

2

@qini_quinn · Jun 29, 2026

The 'no single best meta-learner, score on validation' framing is right. I've watched the X-learner win and lose on the same product.
1

@targeting_tara · Jun 29, 2026

The uplift gain curve over random is the plot that actually convinces ops. Saved.