HTE (Heterogeneous Treatment Effects)

Definition

The phenomenon in which the treatment effect varies with an individual’s characteristics

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

If $\tau(x)$ varies with $x$, then heterogeneous treatment effects (HTE) are present.

---

Intuitive Understanding

Homogeneous vs Heterogeneous

Type	Definition	Example
Homogeneous	$\tau(x) = \tau$ (constant)	The same drug effect for all patients
Heterogeneous	$\tau(x)$ depends on $x$	A larger drug effect for younger patients

Why Does It Matter?

1. Personalized decision-making: deciding whom to treat
2. Policy targeting: prioritizing groups with large effects
3. Resource optimization: efficient allocation of limited resources
4. Scientific understanding: insight into mechanisms

---

HTE vs Related Concepts

Concept	Definition	Relationship
ITE	$Y_i(1) - Y_i(0)$	Individual-level effect
CATE	$E[Y(1)-Y(0) \mid X=x]$	Functional representation of HTE
HTE	Variability of CATE	If CATE varies, HTE is present
ATE	$E[Y(1)-Y(0)]$	The average of CATE

---

HTE Estimation Methods

1. Meta-learners

CATE estimation leveraging existing ML algorithms:

Method	Approach	Characteristics
S-Learner	Single model	Simple; suitable when HTE is negligible
T-Learner	Separate model per treatment	Flexible; no information sharing
X-Learner	Two-stage imputation	Suitable for imbalanced data
R-Learner	Residualized regression	Theoretical guarantees

2. Tree-based Methods

Method	Characteristics
Causal Forest	Heterogeneity-based splitting, provides confidence intervals
BART	Bayesian uncertainty quantification

3. Representation Learning

Method	Characteristics
CFR	Balanced representation learning
CEVAE	VAE-based latent confounder inference
BNN	Balancing Neural Network
GANITE	GAN-based ITE estimation

4. Multi-task Learning

Method	Characteristics
CMGP	Gaussian Process multi-task
Dragonnet	Targeted regularization

---

HTE Analysis Workflow

1. Exploring Effect Heterogeneity

# Estimate CATE with a Causal Forest
from econml.dml import CausalForestDML

cf = CausalForestDML()
cf.fit(Y, T, X)
cate = cf.effect(X)

# Check heterogeneity
print(f"CATE std: {cate.std():.3f}")
print(f"CATE range: [{cate.min():.3f}, {cate.max():.3f}]")

2. Subgroup Analysis

# Effect analysis by characteristic
for group in ['young', 'old']:
    idx = X['age_group'] == group
    print(f"{group}: CATE = {cate[idx].mean():.3f}")

3. Variable Importance

# Which variables contribute to heterogeneity?
importance = cf.feature_importances_

---

Cautions in Interpreting HTE

1. The Subgroup Hunting Problem

• Finding high-effect groups via post-hoc analysis → multiple-comparisons problem
• Separate validation data required

2. Checking Confidence Intervals

• Do not draw conclusions from point estimates alone
• Account for uncertainty

3. Causal Interpretation

• CATE estimation ≠ optimal treatment decision
• A causal interpretation of the interventional effect is needed

---

Applications

1. Personalized Medicine

• “Which treatment is optimal for this patient?”
• Side-effect risk vs. effect trade-off

2. Marketing Targeting

• “For which customers is the promotion effective?”
• Uplift modeling

3. Policy Design

• “To which group should the policy be applied first?”
• Optimization under resource constraints

4. Recommender Systems

• “For which users is this content effective?”
• Connected to MOC-LLM4RecSys, etc.

---

Evaluation Metrics

Metric	Definition	Use
PEHE	$\sqrt{E[(\tau(X) - \hat{\tau}(X))^2]}$	CATE estimation accuracy
ATE Estimation Error	$	\text{ATE} - \widehat{\text{ATE}}
Qini coefficient	Based on the uplift curve	Targeting efficiency

---

Related Concepts

• Treatment Effects Overview - integrated overview of treatment effects
• CATE - functional definition of CATE
• ITE - individual-level effect
• Meta-learners - HTE estimation framework
• Causal Forest - tree-based HTE
• CFR - representation-learning-based HTE
• PEHE - HTE evaluation metric

---

References

• yaoSurveyCausalInference2021 - Section 2.2, 3.4, 3.5
• kunzelMetalearnersEstimatingHeterogeneous2019 - Meta-learners
• Athey, S., & Imbens, G. (2016). Recursive partitioning for heterogeneous causal effects
• Wager, S., & Athey, S. (2018). Estimation and inference of HTE using random forests