Cross-fold validation of heterogeneity
@misc{grf,
title = {grf},
author = {Athey and Tibshirani and Wager},
howpublished = {\url{https://grf-labs.github.io/grf/}},
note = {Software / documentation}
}Follow the analysis.
Select a node to explore the method
K-fold split (e.g. K = 5)
Data preparation — shapes the raw inputs into what the estimator expects.
Partition units into K folds; each unit's CATE will be predicted from a forest trained without it.
folds <- sample(rep(1:5, length.out = n)) # cross-fit by fold
Before using this step, make its assumptions and its output explicit. A useful peer note explains what you would check, what could go wrong, or what would change your interpretation.
Questions to explore- Which assumption matters most here, and what evidence could challenge it?
- What would you check before passing this output to the next step?
Choose a question to start your comment.
Peer discussion
- Bring your experience, a question, or a different perspective.
Log in to comment on this step.
[GRF] Regression forest
The core estimate — where the causal quantity itself is computed.
Cross-fit nuisances Y.hat, W.hat across folds.
Regression forest — Honest non-parametric regression for E[Y|X], with out-of-bag predictions and pointwise CIs.
rf <- regression_forest(X, Y)
Y.hat <- predict(rf)$predictions
Before using this step, make its assumptions and its output explicit. A useful peer note explains what you would check, what could go wrong, or what would change your interpretation.
Questions to explore- Which assumption matters most here, and what evidence could challenge it?
- What would you check before passing this output to the next step?
Choose a question to start your comment.
Peer discussion
- Bring your experience, a question, or a different perspective.
Log in to comment on this step.
[GRF] Causal forest
The core estimate — where the causal quantity itself is computed.
For each fold, fit on the other K−1 and predict τ̂(x) on the held-out fold.
Causal forest — Honest random forest for heterogeneous treatment effects — CATE for a binary treatment via GRF moment conditions.
cf <- causal_forest(X, Y, W) # Y.hat, W.hat cross-fit
tau.hat <- predict(cf)$predictions # OOB CATEs
Before using this step, make its assumptions and its output explicit. A useful peer note explains what you would check, what could go wrong, or what would change your interpretation.
Questions to explore- Which assumption matters most here, and what evidence could challenge it?
- What would you check before passing this output to the next step?
Choose a question to start your comment.
Peer discussion
- Bring your experience, a question, or a different perspective.
Log in to comment on this step.
[GRF] Rank-weighted ATE — RATE / AUTOC / Qini
Heterogeneity — who is affected, and by how much, not just on average.
RATE / AUTOC with the cross-fitted τ̂ as priorities — uses all the data without double-dipping.
Rank-weighted ATE — RATE / AUTOC / Qini — Evaluate how well a CATE estimate prioritizes treatment: TOC curve, AUTOC and Qini with confidence intervals.
rate <- rank_average_treatment_effect(eval.forest, priorities = tau.hat)
plot(rate) # TOC curve
rate$estimate / rate$std.err # AUTOC z-stat
Before using this step, make its assumptions and its output explicit. A useful peer note explains what you would check, what could go wrong, or what would change your interpretation.
Questions to explore- Which assumption matters most here, and what evidence could challenge it?
- What would you check before passing this output to the next step?
Choose a question to start your comment.
Peer discussion
- Bring your experience, a question, or a different perspective.
Log in to comment on this step.
TOC curve + bootstrap CI
Reporting — turn the numbers into a figure or table a reader can act on.
Plot the curve; report whether AUTOC is bounded away from zero.
Before using this step, make its assumptions and its output explicit. A useful peer note explains what you would check, what could go wrong, or what would change your interpretation.
Questions to explore- Which assumption matters most here, and what evidence could challenge it?
- What would you check before passing this output to the next step?
Choose a question to start your comment.
Peer discussion
- Bring your experience, a question, or a different perspective.
Log in to comment on this step.
Inputs & data
A single dataset (X, W, Y) — no separate evaluation sample needed.
Show data format & exampleHide example
Format — one row per unit. A covariate matrix X (numeric), a binary treatment W ∈ {0,1}, and an outcome Y.
X1 X2 X3 W Y
0.42 -1.1 0 1 3.10
-0.07 0.6 1 0 1.85
1.20 0.3 0 1 4.02
Figures reproduced from grf — Athey, Tibshirani & Wager — unofficial community showcase; all credit to the original authors.
The GRF 'Cross-fold validation of heterogeneity' tutorial. Sample-splitting RATE wastes half the data; K-fold cross-fitting fixes that while keeping the priorities out-of-sample at every unit. Unofficial summary.
Discuss the workflow (0)
Log in to join the discussion.