Method library WORKFLOW / 033

Research workflow/4 steps

Heterogeneous effects with causal-forest double ML (EconML)

@econml ✓ P D · Jun 29, 2026 · 1k views

Source EconML — Microsoft Research (ALICE)

@misc{econml,
  title        = {EconML},
  author       = {ALICE},
  howpublished = {\url{https://econml.azurewebsites.net/}},
  note         = {Software / documentation}
}

In the margins · Peer notes@nuisance_nina · Residual-on-residual plus forest weights — clean. How are you choosing the number of trees in practice? 2

The method workspace

Follow the analysis.

Select a node to explore the method

Method mapInput → 4 steps → output

START WITHInput data

ARRIVE ATResults

Read step 01

Inside the methodReasoning, equations & implementation

Data prep

Split features: effect-modifiers X vs controls W

Step notesField note & discussion1

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

What happens here

Decide what the effect is allowed to vary over (X) and what is merely confounding to adjust for (W). The target is the CATE.

Formula

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

Reads fromNot specified Feeds intoNot specified

Key code

# Y outcome, T treatment, X effect modifiers, W controls

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

Partial out nuisance (Neyman-orthogonal DML)

Step notesField note & discussion1

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

What happens here

Cross-fit flexible models for E[Y|X,W] and E[T|X,W]; regress the residuals so first-stage error doesn't bias the effect.

Formula

\tilde Y=Y-\hat\ell(X,W),\quad \tilde T=T-\hat e(X,W),\quad \tilde Y=\theta(X)\,\tilde T+\varepsilon

Reads fromNot specified Feeds intoNot specified

Key code

from econml.dml import CausalForestDML
est = CausalForestDML().fit(Y, T, X=X, W=W)

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.

Heterogeneity

Forest-weighted local effect τ(x)

Step notesField note & discussion1

Heterogeneity — who is affected, and by how much, not just on average.

What happens here

A causal/orthogonal forest defines adaptive weights α_i(x); the effect at x solves a locally-weighted residual moment.

Formula

\hat\theta(x)=\Big(\textstyle\sum_i\alpha_i(x)\tilde T_i^2\Big)^{-1}\textstyle\sum_i\alpha_i(x)\tilde T_i\tilde Y_i

Reads fromNot specified Feeds intoNot specified

Key code

tau = est.effect(X_test)
est.feature_importances_

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

Confidence intervals for τ(x)

Step notesField note & discussion1

Uncertainty quantification — standard errors, intervals, and aggregation.

What happens here

Honest forests / bootstrap-of-little-bags give pointwise intervals — heterogeneity you can actually defend, not just a colourful plot.

Formula

\hat\theta(x)\;\pm\;z_{1-\alpha/2}\,\hat\sigma(x)

Reads fromNot specified Feeds intoNot specified

Key code

lb, ub = est.effect_interval(X_test, alpha=0.05)

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.

Data contract

Inputs & data

Outcome Y, treatment T, the effect-modifiers X you want to vary the effect over, and controls W to adjust for.

Show data format & exampleHide example

Format — outcome Y, treatment T, effect-modifiers X, controls W (one row per unit).

from econml.orf import DMLOrthoForest
est = DMLOrthoForest(n_trees=1000)
est.fit(Y, T, X=X, W=W)
tau = est.effect(X_test)        # τ(x) at new points

Read the evidence

Results & interpretation

1 figure Workflow discussion 2

01FigureEconML's Orthogonal Random Forest recovers the heterogeneous effect τ(x) as a function of a covariate — the ORF estimate tracks the true effect closely.

Figures reproduced from EconML — Microsoft Research (ALICE) — unofficial community showcase; all credit to the original authors.

Context & interpretation

⚠️ Unofficial community showcase of econml. Not affiliated with the authors; all credit to them.

Double machine learning with a forest final stage: partial out nuisance with flexible learners, then read the conditional effect τ(x) — with valid confidence intervals.

Peer notes

Discuss the workflow (2)

2

@hte_hannah · Jun 29, 2026

CausalForestDML with honest intervals is my default for CATE now. The orthogonalization makes the first-stage choice forgiving.
0

@nuisance_nina · Jun 29, 2026

Residual-on-residual plus forest weights — clean. How are you choosing the number of trees in practice?

1

@forest_fox · Jun 29, 2026

Crank it until the intervals stabilise; 1–2k trees is usually plenty, and bootstrap-of-little-bags handles the inference.