Method library WORKFLOW / 008

Research workflow/5 steps

Policy learning via optimal decision trees

@grf ✓ P D · Jun 29, 2026 · 771 views

@misc{grf,
  title        = {grf},
  author       = {Athey and Tibshirani and Wager},
  howpublished = {\url{https://grf-labs.github.io/grf/}},
  note         = {Software / documentation}
}

In the margins · Peer notes@aipw_amir · Doubly-robust scores as the policytree reward is the key link. Garbage scores → garbage policy. 2

The method workspace

Follow the analysis.

Select a node to explore the method

Method mapInput → 5 steps → output

START WITHInput data

ARRIVE ATResults

Read step 01

Inside the methodReasoning, equations & implementation

Estimation

[GRF] Causal forest

Step notesField note & discussion1

The core estimate — where the causal quantity itself is computed.

What happens here

Estimate CATEs and the AIPW score for each unit.

Formula

\tau(x)=\mathbb{E}\!\left[\,Y(1)-Y(0)\mid X=x\,\right]

The estimator

Causal forest — Honest random forest for heterogeneous treatment effects — CATE for a binary treatment via GRF moment conditions.

CATE Code @ v2.6.1 ↗ Paper ↗

Reads fromNot specified Feeds into#2

Key code

cf <- causal_forest(X, Y, W)          # Y.hat, W.hat cross-fit
tau.hat <- predict(cf)$predictions    # OOB CATEs

Discussion on this step (0)

StatsOtterDiscussion starter · Editorial note

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?

Peer discussion

Bring your experience, a question, or a different perspective.

Data prep

double_robust_scores()

Step notesField note & discussion1

Data preparation — shapes the raw inputs into what the estimator expects.

What happens here

Build the doubly-robust reward matrix policytree maximizes over.

Formula

\Gamma_i=\hat\mu_1(X_i)-\hat\mu_0(X_i)+\tfrac{W_i(Y_i-\hat\mu_1)}{\hat e(X_i)}-\tfrac{(1-W_i)(Y_i-\hat\mu_0)}{1-\hat e(X_i)}

Reads from#1 Feeds into#3

Key code

dr.scores <- double_robust_scores(cf)   # n × K reward matrix

Reference / docs ↗

Discussion on this step (0)

StatsOtterDiscussion starter · Editorial note

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?

Peer discussion

Bring your experience, a question, or a different perspective.

Estimation

policytree: depth-2 optimal tree

Step notesField note & discussion1

The core estimate — where the causal quantity itself is computed.

What happens here

Learn an interpretable assignment rule that maximizes expected welfare.

Formula

\hat\pi=\arg\max_{\pi\in\Pi}\ \tfrac1n\textstyle\sum_i\Gamma_i\big(\pi(X_i)\big)

Reads from#2 Feeds into#4#5

Key code

library(policytree)
tree <- policy_tree(X, dr.scores, depth = 2)
plot(tree)

Reference / docs ↗

Discussion on this step (0)

StatsOtterDiscussion starter · Editorial note

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?

Peer discussion

Bring your experience, a question, or a different perspective.

Inference

Evaluate policy value (held-out)

Step notesField note & discussion1

Uncertainty quantification — standard errors, intervals, and aggregation.

What happens here

Estimate the value of the learned rule vs treat-all / treat-none on held-out data.

Formula

V(\pi)=\mathbb{E}\!\left[\,\Gamma(\pi(X))\,\right]

Reads from#3 Feeds into#5

Key code

# value of the learned rule vs treat-all / treat-none on held-out data
mean(dr.scores[cbind(seq_len(n), predict(tree, X))])

Discussion on this step (0)

StatsOtterDiscussion starter · Editorial note

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?

Peer discussion

Bring your experience, a question, or a different perspective.

Data contract

Inputs & data

Per-unit CATEs / doubly-robust scores from a causal forest, plus a cost.

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

Read the evidence

Results & interpretation

1 figure Workflow discussion 2

01FigureThe learned treatment rule: a shallow decision boundary splitting treat vs do-not-treat in covariate space.

Figures reproduced from grf — Athey, Tibshirani & Wager — unofficial community showcase; all credit to the original authors.

Context & interpretation

The GRF + policytree tutorial: turn CATEs into an interpretable, near-optimal treatment rule and honestly evaluate its value. Unofficial summary; policytree is a separate grf-labs package.

Peer notes

Discuss the workflow (2)

4

@policy_pia · Jun 29, 2026

Depth-2 trees hit the sweet spot: near-optimal but you can actually explain the rule to ops. The held-out value evaluation keeps you honest.

5

@targeting_tara · Jun 29, 2026

The 'evaluate vs treat-all/treat-none' baseline is what convinces stakeholders. Always include it.
5

@aipw_amir · Jun 29, 2026

Doubly-robust scores as the policytree reward is the key link. Garbage scores → garbage policy.