Causal Machine Learning - Index

Routing Summary

The ML-meets-causal-inference bridge: how to use flexible learners for nuisance functions and effect heterogeneity while keeping valid frequentist inference. Anchored by Chernozhukov et al. (2018) on Double/Debiased ML, Wager & Athey (2018) on causal forests, Athey, Tibshirani & Wager (2019) on generalized random forests, and Nie & Wager (2021) on the R-learner.

• Need the big picture, a DML-vs-forest-vs-R-learner comparison, or marketing relevance? → Causal Machine Learning - Overview
• Need to know why plugging ML into a regression breaks inference? → Regularization Bias and the Partially Linear Model
• Need the formal definition of an orthogonal score, how to construct one, or the $o(N^{-1/4})$ rate condition? → Neyman Orthogonality
• Need the DML1/DML2 algorithms, choice of $K$, variance estimator, or median-over-splits? → Cross-Fitting and Sample Splitting
• Need the AIPW / ATTE / LATE scores and the theorem for ATE inference with ML nuisances? → DML Estimators for ATE and the Interactive Model
• Need what a causal tree is, what honesty means, or double-sample vs propensity trees? → Honest Trees and Causal Forests
• Need forests for arbitrary moment conditions (quantiles, partial effects, IV), gradient splitting, or local centering? → Generalized Random Forests - Local Moment Equations
• Need the CLT for forests, the subsample-size condition, infinitesimal jackknife or bootstrap of little bags? → Asymptotic Normality and Inference for Forests
• Need a loss function for CATE usable with any learner, or a way to cross-validate / stack CATE models? → R-Learner and Orthogonal CATE Estimation
• Which estimator do I use? → average effect with CI: cross-fit AIPW; constant/“lift” coefficient with continuous treatment: DML-PLR; $\tau (x)$ with pointwise CIs: centered causal forest; $\tau (\cdot )$ with a preferred learner: R-learner.

Concept Map

Concept	Note	Type	Depends On	Key Result
Framing and comparison	Causal Machine Learning - Overview	overview	Potential Outcomes; CIA; Frequentist Causal Estimation	Orthogonalize + split + re-use; DML for scalars, forests for functions, R-learner as bridge
Regularization bias in PLR	Regularization Bias and the Partially Linear Model	concept	Overview; CIA; OVB	Naive plug-in bias $\sqrt{n}{n}^{-{\phi }_g}\to \infty$; double residualization gives product bias $\sqrt{n}{n}^{-({\phi }_m+{\phi }_g)}$; Thm 4.1
Neyman orthogonality	Neyman Orthogonality	theorem	PLR note	${\partial }_{\eta }\mathbb{E}\psi (W;{\theta }_0,{\eta }_0)[\eta -{\eta }_0]=0$; constructions (Lemma 2.1, IF adjustment); nuisance rate $o(N^{-1/4})$, or $o(1)$ when ${\lambda }_N^{\prime }=0$
Cross-fitting	Cross-Fitting and Sample Splitting	method	Orthogonality	DML1/DML2; Chebyshev replaces Donsker; $K=4$–5; $s_g{s}_m\ll N$ vs $s_g^2+s_m^2\ll N$; median over $S$ splits
DML for ATE/ATTE/LATE	DML Estimators for ATE and the Interactive Model	method / theorem	Orthogonality; Cross-fitting; DR estimation	AIPW score; Thm 5.1 $\sqrt{N}$-normal, efficient (Hahn bound) under product-rate condition; 401(k) and bonus examples
Honest causal trees and forests	Honest Trees and Causal Forests	method	Potential Outcomes; CIA; Overlap	Leaf difference-in-means; honesty; double-sample and propensity trees; variance-of-$\hat{\tau }$ splitting; CF vs $k$-NN simulations
Generalized random forests	Generalized Random Forests - Local Moment Equations	method	Causal forests; Orthogonality	Forest weights ${\alpha }_i(x)$ solve local moment equation; $\Delta$-criterion and gradient tree; quantile, CAPE/causal, IV forests; local centering
Forest asymptotics and CIs	Asymptotic Normality and Inference for Forests	theorem	Causal forests; GRF	W&A Thm 1/11 with $s\asymp n^{\beta }$, $\beta >{\beta }_{\min }$; Hájek projection + incrementality; ${\hat{V}}_{IJ}$; ATW Thm 5 via pseudo-forest; bootstrap of little bags
R-learner	R-Learner and Orthogonal CATE Estimation	method	PLR note; Orthogonality; Cross-fitting; Metalearners	R-loss from Robinson’s transformation; quasi-oracle regret bound (Thm 3); X-learner counterexample; R-stacking

Notes

• Causal Machine Learning - Overview — CONTAINS: two target types (scalar vs function), the common orthogonalize/split/re-use recipe, DML vs GRF vs R-learner comparison table, relevance to marketing measurement (lift, geo experiments, uplift, MMM diagnostics), Pennsylvania bonus and 401(k) headline numbers.
• Regularization Bias and the Partially Linear Model — CONTAINS: PLR model definition, naive estimator decomposition $a+b$, orthogonalized estimator decomposition $a^{\ast }+b^{\ast }+c^{\ast }$, Robinson partialling-out score, Theorem 4.1 with variance ${\sigma }^2$, efficiency and sparsity-tightness remarks, Figure 1 simulation, Python cross-fit PLR sketch.
• Neyman Orthogonality — CONTAINS: Gateaux derivative and Definitions 2.1–2.2 (exact and near-orthogonality), PLR orthogonality check, Neyman’s ${\mu }_0=J_{\theta \beta }{J}_{\beta \beta }^{-1}$ construction, GMM/concentrating-out/conditional-moment constructions, influence-function adjustment, Assumptions 3.1–3.2 rate conditions, Theorem 3.1 uniform normality, catalogue of orthogonal scores.
• Cross-Fitting and Sample Splitting — CONTAINS: overfitting-bias example ($N^{ϵ}$ blow-up), Chebyshev argument, Donsker failure in high dimension, DML1 and DML2 definitions, Remark 3.1 recommendations, variance estimator and uniform CIs, median/mean aggregation over splits, cross-fitting vs cross-validation, empirical 2-fold vs 5-fold table, pseudocode.
• DML Estimators for ATE and the Interactive Model — CONTAINS: interactive regression model, AIPW (ATE) and ATTE scores, orthogonality verification, closed-form cross-fit estimator, Assumption 5.1 and Theorem 5.1, rate double-robustness, LATE score and Theorem 5.2 condition, overlap/trimming cautions, 401(k) ATE and LATE tables, Python AIPW sketch.
• Honest Trees and Causal Forests — CONTAINS: setup (unconfoundedness, overlap), causal tree/forest definitions, honesty definition, Procedure 1 (double-sample) and Procedure 2 (propensity trees), splitting-rule rationale, why honesty removes bias, random-split/$\alpha$-regular/symmetric conditions, simulation Tables 1–3, grf usage.
• Generalized Random Forests - Local Moment Equations — CONTAINS: local moment condition, forest weights, Proposition 1 $\Delta$-criterion, gradient tree labeling/regression steps, Algorithm 1, quantile forests vs Meinshausen, CAPE/causal forest estimator and pseudo-outcomes, local centering, instrumental forest and Angrist–Evans application, Table 1 MSE comparison, grf usage.
• Asymptotic Normality and Inference for Forests — CONTAINS: forest as U-statistic, Theorems 1 and 11, ${\beta }_{\min }$ formula and worked values, leaf-diameter and bias bounds, Hájek projection and $\nu$-incrementality, $k$-PNN predictors, Lemma 7/Theorems 8–9, infinitesimal jackknife formula, GRF pseudo-forest coupling and Theorem 5, delta-method variance, bootstrap of little bags, scope caveats.
• R-Learner and Orthogonal CATE Estimation — CONTAINS: Robinson’s transformation, two-step R-learner algorithm, weighted-regression implementation trick, U-learner instability, quasi-oracle Theorem 3 with RKHS assumptions, X-learner counterexample, R-stacking, relation to centered causal forests, voting-study and simulation results, Python sketch.

External / Cross-Folder Links

• Potential Outcomes Framework, Causal Estimands, Conditional Independence Assumption, Common Support and Overlap — identification assumptions maintained throughout.
• Frequentist Causal Estimation, Doubly-Robust Estimands for ATT(g,t) — doubly-robust estimation, which DML generalizes and justifies with ML nuisances.
• Metalearners for CATE, S-Learner, T-Learner and Minimax Rate, X-Learner, Künzel 2019 - Overview — the existing CATE metalearner cluster.
• Nonparametric Causal Inference — BART, the Bayesian tree-ensemble counterpart.
• Propensity Score Matching - Overview, Propensity Score and the Balancing Property, Matching Methods and Distance Measures — propensity-score and matching methods.
• Instrumental Variables, Local Average Treatment Effects, GMM Estimation and Instruments for Price Endogeneity — IV and moment-condition estimation.
• Quantile Regression — parametric counterpart of quantile forests.
• Omitted Variables Bias, Table 2 Fallacy, Horseshoe and Regularized Horseshoe Priors — regularization and nuisance-coefficient interpretation.
• Geo-Experiment Methodology - Overview, Bayesian Media Mix Modeling - Overview — applied marketing-measurement contexts.

Sources

• Chernozhukov 2018 - Double Debiased Machine Learning — Chernozhukov, V., Chetverikov, D., Demirer, M., Duflo, E., Hansen, C., Newey, W. & Robins, J. (2018), “Double/Debiased Machine Learning for Treatment and Structural Parameters,” The Econometrics Journal 21(1). arXiv:1608.00060.
• Wager Athey 2018 - Heterogeneous Treatment Effects using Random Forests — Wager, S. & Athey, S. (2018), “Estimation and Inference of Heterogeneous Treatment Effects using Random Forests,” Journal of the American Statistical Association 113(523). arXiv:1510.04342.
• Athey Tibshirani Wager 2019 - Generalized Random Forests — Athey, S., Tibshirani, J. & Wager, S. (2019), “Generalized Random Forests,” Annals of Statistics 47(2). arXiv:1610.01271.
• Nie Wager 2021 - Quasi-Oracle Estimation of Heterogeneous Treatment Effects — Nie, X. & Wager, S. (2021), “Quasi-Oracle Estimation of Heterogeneous Treatment Effects,” Biometrika 108(2). arXiv:1712.04912.