CATE (Conditional Average Treatment Effect)

Definition

The Conditional Average Treatment Effect (CATE) is the average treatment effect given covariates $X=x$:

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

where:

• $Y(1)$: potential outcome under treatment
• $Y(0)$: potential outcome under no treatment
• $X$: pre-treatment covariates (feature variables)

Related terms:

• HTE (Heterogeneous Treatment Effect): used synonymously with CATE
• ITE (Individual Treatment Effect): $\tau_i = Y_i(1) - Y_i(0)$ (unobservable)

Intuitive Understanding

Key question:

“How effective is the treatment for a person with specific characteristics?”

ATE vs CATE:

Quantity	Definition	Question
ATE	$E[Y(1) - Y(0)]$	”Is there an effect on average?”
CATE	$E[Y(1) - Y(0) \| X=x]$	”Is there an effect for a person with these characteristics?”

Example:

• The average effect of a new drug is positive (ATE > 0)
• But for patients aged 65 and over it has no or even a negative effect ($\tau(x_{age \geq 65}) \leq 0$)

ATE = E[τ(X)] = ∫ τ(x) dP(x)  (average of CATE)

Key Properties

Fundamental Problem of Causal Inference

At the individual level, $Y(1)$ and $Y(0)$ cannot be observed simultaneously:

• Actual observation: $Y = AY(1) + (1-A)Y(0)$
• The counterfactual is always missing

Identification Assumptions

The standard assumptions for identifying CATE:

1. SUTVA (Stable Unit Treatment Value Assumption)

   • No interference: another unit’s treatment does not affect my outcome
   • Consistency: $Y = Y(A)$

2. Unconfoundedness (Ignorability) $Y(0), Y(1) \perp A | X$

   • Given $X$, treatment assignment is independent of the potential outcomes

3. Positivity (Overlap) $0 < P(A=1|X=x) < 1, \quad \forall x \in \mathcal{X}$

   • At every covariate value the probability of receiving treatment lies strictly between 0 and 1

Structure of CATE

CATE can be decomposed as: $\tau(x) = \mu_1(x) - \mu_0(x)$

where $\mu_a(x) = E[Y|X=x, A=a]$

Estimation Methods

Meta-Learners

Method	Description	Best When
S-Learner	Single model: $\hat{\mu}(x,a)$, then $\hat{\tau}(x) = \hat{\mu}(x,1) - \hat{\mu}(x,0)$	Homogeneous effects
T-Learner	Two models: $\hat{\mu}_1(x)$, $\hat{\mu}_0(x)$ separately	Different response functions
X-Learner	Two-stage imputation with propensity weighting	Unbalanced treatment groups
R-Learner	Residualize then regress: minimize $\sum(Y_i - \hat{\mu}(X_i) - (A_i - \hat{\pi}(X_i))\tau(X_i))^2$	Heterogeneous effects
DR-Learner	Regress doubly robust pseudo-outcome on $X$	Double robustness desired

Tree-Based Methods

• Causal Forest (Wager & Athey): Random forest adapted for CATE
• BART (Bayesian Additive Regression Trees)
• Causal MARS

Deep Learning

• CEVAE (Causal Effect VAE)
• TARNet (Treatment-Agnostic Representation Network)
• DragonNet

Example

Medical scenario:

• $Y$: reduction in blood pressure
• $A$: whether the new drug is administered (0/1)
• $X$: (age, sex, baseline blood pressure, BMI, …)

$\tau(x) = E[\text{BP reduction}|\text{new drug}] - E[\text{BP reduction}|\text{placebo}] \quad \text{given } X=x$

Interpretation:

• $\tau(x) > 0$: the new drug is effective for a patient with these characteristics
• $\tau(x) < 0$: the new drug is harmful for a patient with these characteristics
• $\tau(x) \approx 0$: no effect for a patient with these characteristics

Applications

Treatment Targeting (Policy Learning)

Learning the optimal treatment rule: $d^*(x) = \mathbf{1}[\tau(x) > 0]$

• treat if $\tau(x) > 0$
• don’t treat if $\tau(x) < 0$

Personalized Medicine

• Tailored treatment based on patient characteristics
• Minimize side effects & maximize efficacy

Precision Marketing

• Estimating per-customer marketing effects
• Personalized promotion targeting

Policy Evaluation

• Analyzing policy effects by subgroup
• Exploring heterogeneity

Evaluation Metrics

Evaluating CATE estimates is difficult (the true CATE is unobservable)

When an RCT is Available

• PEHE (Precision in Estimation of HTE): $\sqrt{E[(\hat{\tau}(x) - \tau(x))^2]}$
• ATE Error: $|\hat{\tau}_{ATE} - \tau_{ATE}|$

Observational Data

• AUUC (Area Under Uplift Curve): treatment targeting performance
• Qini Coefficient: uplift modeling evaluation

Related Concepts

• ATE - Average Treatment Effect (the average of CATE)
• ATT - Average Treatment on Treated
• Propensity Score - Treatment assignment probability
• DR-Learner - A doubly robust method for CATE estimation
• Double-Debiased ML - High-dimensional CATE estimation
• Causal Forest - Tree-based CATE estimation

Key Papers

• kunzelMetalearnersEstimatingHeterogeneous2019 - Meta-learners (S, T, X-learner)
• nieQuasiOracleEstimationHeterogeneous2020 - R-learner
• kennedyOptimalDoublyRobust2023 - DR-learner, optimal rates
• Wager & Athey (2018) - Causal Forests
• chernozhukovDoubleDebiasedMachine2018 - DML for treatment effects

Implementation

Python (econml):

from econml.dml import CausalForestDML
from econml.dr import DRLearner

# Causal Forest
cf = CausalForestDML()
cf.fit(Y, T, X=X, W=W)
cate = cf.effect(X_test)

# DR-Learner
dr = DRLearner()
dr.fit(Y, T, X=X, W=W)
cate = dr.effect(X_test)

R (grf):

library(grf)
cf <- causal_forest(X, Y, W)
tau_hat <- predict(cf)$predictions

References

• kunzelMetalearnersEstimatingHeterogeneous2019
• nieQuasiOracleEstimationHeterogeneous2020
• kennedyOptimalDoublyRobust2023
• chernozhukovDoubleDebiasedMachine2018
• Wager & Athey (2018) - “Estimation and Inference of Heterogeneous Treatment Effects using Random Forests”