Uplift Modeling & Causal Treatment EffectsOpen in Colab

Uplift Modeling & Causal Treatment Effects

Standard recommendation predicts who will convert. Uplift modeling predicts who will convert because of an intervention — a discount, an email, a product recommendation. The target variable is the Individual Treatment Effect (ITE):

τi=Yi(1)−Yi(0)\tau_i = Y_i(1) - Y_i(0)

where Yi(1)Y_i(1) is the outcome if user ii is treated and Yi(0)Y_i(0) is the outcome if not. The fundamental problem: you can only observe one potential outcome per user. Uplift models estimate τi\tau_i from A/B experiment data where treatment is randomly assigned.

The four uplift segments

Users fall into four segments based on their behavior with and without treatment:

SegmentNo treatmentTreatmentUplift
PersuadablesWould not convertWould convert✅ Positive
Sure thingsWould convertWould convert0
Lost causesWould not convertWould not convert0
Sleeping dogsWould convertWould not convert❌ Negative

The goal is to target Persuadables and avoid Sleeping dogs — users who would convert anyway (Sure things) waste budget; Sleeping dogs are made worse by the intervention (e.g., a discount that insults a loyal customer).

S-Learner (Single Model)

Train a single model that takes both features XX and treatment indicator TT as inputs:

μ(x,t)=E[Y∣X=x,T=t]\mu(x, t) = \mathbb{E}[Y \mid X = x, T = t]

Estimate ITE by computing the model's prediction with and without treatment:

τ^(x)=μ(x,1)−μ(x,0)\hat{\tau}(x) = \mu(x, 1) - \mu(x, 0)

Pros: Simple, uses all data jointly, no need for two separate models.
Cons: If the treatment feature TT has weak main effects but strong interaction effects, regularized models may shrink its coefficient toward zero, underestimating uplift.

T-Learner (Two Models)

Train separate models for treated and control groups:

μ0(x)=E[Y∣X=x,T=0],μ1(x)=E[Y∣X=x,T=1]\mu_0(x) = \mathbb{E}[Y \mid X = x, T = 0], \qquad \mu_1(x) = \mathbb{E}[Y \mid X = x, T = 1] τ^(x)=μ1(x)−μ0(x)\hat{\tau}(x) = \mu_1(x) - \mu_0(x)

Pros: Each model is optimized for its own group, captures group-specific patterns.
Cons: High variance when treatment/control group sizes are imbalanced; the two models may have different implicit regularization that introduces bias into τ^\hat{\tau}.

X-Learner (Cross-fitted)

Proposed by Künzel et al. (2019). Three-stage approach:

Stage 1: Train μ0\mu_0 on control, μ1\mu_1 on treated (same as T-Learner).

Stage 2: Compute imputed treatment effects using cross-group predictions:

τ~i(1)=Yi−μ^0(Xi)for treated units\tilde{\tau}_i^{(1)} = Y_i - \hat{\mu}_0(X_i) \quad \text{for treated units} τ~i(0)=μ^1(Xi)−Yifor control units\tilde{\tau}_i^{(0)} = \hat{\mu}_1(X_i) - Y_i \quad \text{for control units}

Stage 3: Train two uplift models:

  • τ1(x)\tau_1(x): regress τ~(1)\tilde{\tau}^{(1)} on features XX for treated units.
  • τ0(x)\tau_0(x): regress τ~(0)\tilde{\tau}^{(0)} on features XX for control units.

Combine using propensity score e(x)=P(T=1∣X=x)e(x) = P(T=1 \mid X=x):

τ^(x)=e(x)⋅τ0(x)+(1−e(x))⋅τ1(x)\hat{\tau}(x) = e(x) \cdot \tau_0(x) + (1 - e(x)) \cdot \tau_1(x)

Pros: Better performance when treatment/control sizes are very imbalanced (X-Learner upweights the larger group's estimate).
Cons: Requires propensity score estimation; 3-stage pipeline is complex.

Evaluation: Qini curve and AUUC

Standard accuracy metrics are meaningless for uplift models — the true ITE is never observed. Use ranking-based metrics:

Qini curve: Sort users by predicted uplift τ^i\hat{\tau}_i (descending). Plot cumulative incremental gains vs. fraction of population treated. A perfect model concentrates all Persuadables at the top.

Qini gain at decile k=Y(k)TNT−Y(k)CNC\text{Qini gain at decile } k = \frac{Y^T_{(k)}}{N^T} - \frac{Y^C_{(k)}}{N^C}

AUUC (Area Under the Uplift Curve): normalize the Qini curve area. Higher AUUC = better uplift model.

Real-world applications from case studies

CompanyUse caseEstimated uplift
NetflixPersonalize "push notification to re-engage" targeting+8% conversion vs. random targeting
LyftIncentive targeting (discount only high-uplift riders)−12% promotion cost, flat revenue
GojekTargeted special offers for dormant users+15% reactivation vs. blanket discount
InstacartBudget-constrained promotion allocation+9% ROAS
NYTPaywall presentation personalization+6% subscription conversion

Causal forests

For non-parametric uplift estimation, Causal Forests (Wager & Athey, 2018) build an ensemble of causal trees that partition the feature space to find heterogeneous treatment effects:

τ^(x)=LocalWeightedEstimator({(Yi,Ti,Xi)}i:Xi≈x)\hat{\tau}(x) = \text{LocalWeightedEstimator}(\{(Y_i, T_i, X_i)\}_{i: X_i \approx x})

The causalml and EconML Python packages implement causal forests with confidence intervals, enabling statistical significance testing of uplift estimates.

Related concepts

Concept neighbourhood

How this concept connects — click any node to jump there, or explore the whole concept map.

RelatedReferenced by
Uplift Modeling & Causal Treatmen…Decile Analysis: Gains, L…Interventions & Potential…Correlation, Causation & …Bagging & Random Forests