Causal Forest GRF & CATE Estimation

Causal Forests & Generalized Random Forests (GRF, Athey & Wager, Stanford) establishes rigorous non-parametric estimation of Conditional Average Treatment Effects (CATE: $\tau(x) = \mathbb{E}[Y(1) - Y(0) | X=x]$) with asymptotic Gaussian normality; Unlike standard Random Forests maximizing label purity, Causal Forests introduce: 1) Causal Splitting Rules: maximizing the heterogeneity variance of treatment effects across child nodes $\Delta \tau^2$ to recursively isolate treatment-sensitive subgroups; 2) Honest Tree Estimation: partitioning training data such that tree structure splits are determined on sample set $S^{\text{tr}}$ while leaf treatment effects are evaluated on an independent holdout sample set $S^{\text{est}}$, eliminating adaptive overfitting bias; 3) Pointwise Confidence Intervals: providing closed-form asymptotic standard errors for individual causal estimates.

Personalized dynamic pricing, heterogeneous policy targeting, and non-linear causal feature interaction discovery.

Classical OLS interaction terms enforce rigid linear assumptions; Causal Forests adaptively partition high-dimensional covariate space to discover non-linear treatment heterogeneity.

Contrast standard Random Forest splitting criteria (MSE minimization) vs Causal Tree splitting criteria (treatment effect variance maximization)?

Explain why Honest Estimation (splitting structure selection from leaf evaluation) eliminates adaptive overfitting and restores asymptotic normality?

Derive how GRF transforms trees into adaptive kernel weighting functions $\alpha_i(x)$ solving local Generalized Method of Moments (GMM) equations?

How does Causal Forest compute variable importance metrics ranking which covariates drive the strongest treatment heterogeneity?

Explain Double Machine Learning (DML) residualization orthogonalizing outcomes and treatments to accelerate Causal Forest convergence?