EconML

0.16.0 · active · verified Mon Apr 13

EconML is a Python library for estimating Conditional Average Treatment Effects (CATEs) from observational or experimental data. It provides a suite of advanced machine learning methods, including Double Machine Learning (DML) and Causal Forests, to infer causal relationships and individual-level treatment effects. The current version is 0.16.0, and it maintains an active development pace with major updates and bugfix releases every few months.

Warnings

breaking The `DynamicDML` estimator was moved from `econml.dml` to `econml.panel`. Direct imports from the old path will raise an `ImportError`.
Fix: Update import statements from `from econml.dml import DynamicDML` to `from econml.panel import DynamicDML`.
breaking The default `alpha` value for confidence intervals in methods like `effect_interval` changed from `None` (user-defined or estimator-specific) to `0.05` (representing 95% confidence).
Fix: Explicitly set the `alpha` parameter in `effect_interval` or similar methods if you rely on a specific alpha value or wish to restore prior behavior.
gotcha EconML v0.16.0 requires `shap` versions `>=0.40.0` and `<0.44.0`. Using `shap` version 0.44.0 or higher will lead to installation conflicts or runtime errors when calling `shap_values`.
Fix: Ensure your installed `shap` version is within the compatible range. Downgrade `shap` to `<0.44.0` if you encounter issues (`pip install 'shap<0.44.0'`).
gotcha The `deepiv` module, containing the `DeepIV` estimator, requires `tensorflow` and `keras` to be installed separately. These libraries can have strict Python and other dependency version requirements that may cause conflicts.
Fix: Install `econml` with the `[deepiv]` extra (`pip install econml[deepiv]`) to ensure the correct `tensorflow`/`keras` versions are pulled, or manually manage these dependencies carefully to avoid conflicts with other parts of your environment.
gotcha Distributed training features (e.g., for scaling `OrthoLearner`s) rely on the `ray` library, which is an optional dependency and not installed by default.
Fix: Install `econml` with the `[ray]` extra (`pip install econml[ray]`) to enable distributed training capabilities.

Install

pip install econml Base installation
pip install econml[ray] # For distributed training with Ray pip install econml[deepiv] # For DeepIV estimator with TensorFlow/Keras pip install econml[all] # For all optional dependencies With optional dependencies

Imports

CausalForestDML
```
from econml.dml import CausalForestDML
```
CausalForestDML, like other DML estimators, resides in `econml.dml`.
LinearDML
```
from econml.dml import LinearDML
```
LinearDML is part of the Double Machine Learning (DML) module.
DynamicDML
```
from econml.panel import DynamicDML
```
As of v0.14.0, DynamicDML was moved to the `econml.panel` subpackage.

Quickstart

This quickstart demonstrates how to use `CausalForestDML` to estimate Conditional Average Treatment Effects (CATEs) using synthetic data. It involves defining confounders (W), features for heterogeneity (X), treatment (T), and outcome (Y), then fitting the model and predicting CATEs for new feature values. `RandomForestRegressor` and `RandomForestClassifier` are used as base learners for Y and T models respectively.

import numpy as np
import pandas as pd
from econml.dml import CausalForestDML
from sklearn.ensemble import RandomForestRegressor, RandomForestClassifier

# Simulate data
np.random.seed(42)
n_samples = 1000
W = np.random.normal(0, 1, size=(n_samples, 3)) # Confounders
X = np.random.normal(0, 1, size=(n_samples, 2)) # Features for heterogeneity
T = (W[:, 0] + W[:, 1] + np.random.normal(0, 1, n_samples) > 0).astype(float) # Treatment
Y = W[:, 0] + W[:, 2] + T * (X[:, 0] + np.random.normal(0, 0.1, n_samples)) + np.random.normal(0, 1, n_samples) # Outcome

# Initialize and fit the CausalForestDML model
est = CausalForestDML(
    model_y=RandomForestRegressor(min_samples_leaf=5, n_estimators=100, random_state=42),
    model_t=RandomForestClassifier(min_samples_leaf=5, n_estimators=100, random_state=42),
    cv=5,
    random_state=42
)
est.fit(Y, T, X=X, W=W)

# Estimate CATE for new data (or original X)
X_test = np.array([[0.5, 0.5], [-0.5, -0.5]])
cate_estimates = est.effect(X_test)
print(f"CATE estimates for X_test: {cate_estimates}")

# Expected output: CATE estimates for X_test: [0.67204641 0.44976767]

view raw JSON →