Heterogeneous Treatment Effect (HTE)

The foundational estimand in HTE analysis. CATE estimates the expected causal effect of a treatment for a subgroup defined by a specific set of covariates X = x.

• Formula: τ(x) = E[Y(1) - Y(0) | X = x]
• Distinct from ATE (Average Treatment Effect), which collapses all heterogeneity into a single population mean.
• Estimation requires methods that can model complex, non-linear interactions between treatment assignment and covariates.
• Serves as the input to personalized decision rules: treat if τ(x) > cost.

HTE estimation is constrained by the fact that we can never observe both potential outcomes for the same individual simultaneously.

• For any unit i, we observe either Y_i(1) under treatment or Y_i(0) under control, never both.
• The individual treatment effect Y_i(1) - Y_i(0) is fundamentally unobservable.
• All HTE methods must therefore estimate missing counterfactuals using statistical assumptions and observed covariates.
• This missing data problem drives the need for randomized experiments or strong ignorability assumptions in observational studies.

Variation in causal impact arises from multiple sources that must be distinguished analytically.

• Observed Heterogeneity: Effect modification captured by measured covariates such as age, prior purchase frequency, or channel preference.
• Unobserved Heterogeneity: Latent variation not captured by available features, requiring random effects or latent class models.
• Essential Heterogeneity: Variation correlated with treatment selection itself, where individuals select into treatment based on private, unobserved anticipated gains.
• Failing to model the correct source leads to biased CATE estimates and suboptimal targeting.

A direct modeling paradigm that predicts the causal effect rather than outcomes, segmenting individuals into four behavioral types.

• Persuadables: Respond positively only when treated — the target segment for intervention.
• Sure Things: Convert regardless of treatment — wasteful to target.
• Lost Causes: Never convert — no intervention helps.
• Sleeping Dogs: Respond negatively to treatment — targeting causes harm.
• Uplift models optimize resource allocation by focusing exclusively on persuadables, maximizing incremental ROI.

A family of modular algorithms that decompose CATE estimation into subproblems solvable by any supervised learning base model.

• S-Learner: A single model trained on features plus treatment indicator. CATE = μ(x, T=1) - μ(x, T=0). Simple but can bias treatment effect toward zero.
• T-Learner: Two separate models for treated and control groups. CATE = μ₁(x) - μ₀(x). Captures heterogeneity well but suffers from high variance with imbalanced data.
• X-Learner: Cross-estimates counterfactuals and propensity scores, then models imputed individual effects. Robust to imbalanced treatment assignment.
• R-Learner: Directly optimizes CATE via a loss function derived from Robinson's partialling-out transformation, offering strong theoretical guarantees.

Non-parametric ensemble methods that adaptively partition the covariate space to discover heterogeneous subgroups without pre-specification.

• Causal Trees: Recursively split data to maximize the difference in treatment effects between child nodes, using modified splitting criteria distinct from predictive trees.
• Honest Estimation: Uses one subsample to determine the tree structure and a separate, independent subsample to estimate leaf-level effects, preventing overfitting.
• Generalized Random Forests: Extend the random forest framework to estimate any local parameter identified by a moment condition, including CATE, conditional quantile effects, and instrumental variable effects.
• Provide asymptotic normality and valid confidence intervals for estimated treatment effects.