PyCausalML

Classical and deep causal inference in Python

PyCausalML

pycausalml

Version: 0.2.0 · Requires: Python ≥ 3.9 · License: MIT

Classical and deep causal inference in one installable package. Trees, uplift, meta-learners, double machine learning, treatment policies, Bayesian quasi-experiments, and the neural models from PyDeepCausalML share one import.

pip install https://github.com/zia207/PyCausalML/releases/download/v0.2.0/pycausalml-0.2.0-py3-none-any.whl
Package Release v0.2.0 · GitHub · Homepage
Docs Tutorial · Glossary
Author Zia U. Ahmed · Upatta Analytics

What is it?

PyCausalML 0.2.0 extends PyDeepCausalML 0.2.0 and keeps every public estimator from that release (TARNet, DragonNet, NOTEARS, TCDF, and the rest). On top of those neural models it adds tree and forest estimators, uplift learners, meta-learners, double machine learning, treatment policies, and Bayesian quasi-experiments.

import pycausalml and import pydeepcausalml both work after installation. The deep-learning release on its own is still PyDeepCausalML v0.2.0.

Use it when you need to:

  • Estimate conditional average treatment effects with honest trees, forests, meta-learners, or neural nets
  • Score uplift on a binary outcome and choose who should be treated
  • Estimate a constant effect with a confidence interval (double machine learning, neural DML)
  • Run quasi-experiments: difference-in-differences, synthetic control, regression discontinuity, instrumental variables
  • Put Bayesian uncertainty on a causal network, a structural model, or a quasi-experiment
  • Discover causal graphs from static or time-series data
  • Train neural models on CPU, CUDA, or Apple MPS with the same scikit-learn style API

Installation

Requires Python ≥ 3.9. PyTorch ≥ 2.0 is included because the deep models are part of this distribution.

Install release v0.2.0:

pip install https://github.com/zia207/PyCausalML/releases/download/v0.2.0/pycausalml-0.2.0-py3-none-any.whl

From source:

git clone https://github.com/zia207/PyCausalML.git
cd PyCausalML
pip install -e ".[dev]"
pytest

Plotting (matplotlib / seaborn) and the optional libraries:

pip install -e ".[plot]"
pip install -e ".[dowhy,econml,causallearn,uber,doubleml,causalpy,skgrf,shap,boost]"

Main features

Module What it estimates
pycausalml.tree_based_models Honest trees, doubly robust forests, survival and instrumental forests, boosted causal models
pycausalml.uplift_learners Uplift trees and forests, interaction trees, uplift XGBoost, Qini curves
pycausalml.meta_learners S, T, X, R, and DR learners, plus domain adaptation
pycausalml.double_machine_learning Partially linear, linear, sparse, nonparametric, kernel, and interactive DML
pycausalml.policy_learners Doubly robust policy learner, policy tree, and policy forest
pycausalml.bayesian_causalML Bayesian DiD, RDD, ITS, IV, synthetic control, causal networks, structural models
pydeepcausalml TARNet, CFRNet, DragonNet, GANITE, CausalVAE, NOTEARS, TCDF, and the rest of the neural catalog

Optional backends in pycausalml.integrations wrap DoWhy, EconML, causal-learn, Uber CausalML, DoubleML, CausalPy, and skgrf when those libraries are installed.

Quickstart

Classical estimators share fit(X, treatment, y) and return NumPy arrays of conditional average treatment effects.

from pycausalml import DRLearner, PartiallyLinearDML, synthetic_data
from pycausalml.data import make_linear_treatment

data = synthetic_data(n=1000, p=6, sigma=0.8, mode=2, random_state=0)
dr = DRLearner(learner="gbr", n_estimators=40, random_state=0).fit(
    data["X"], data["w"], data["y"]
)
print("CATE:", dr.predict(data["X"][:5]))
print("ATE:", dr.average_treatment_effect())

linear = make_linear_treatment(n=800, tau=2.0, random_state=0)
dml = PartiallyLinearDML(model_y="ridge", random_state=0).fit(
    linear["X"], linear["w"], linear["y"]
)
print("DML effect:", dml.effect(), "95% CI:", dml.confidence_interval())

Neural estimators use the same top-level import:

from pycausalml import DragonNet
from pydeepcausalml.datasets import make_confounded_data

df = make_confounded_data(n=2000, random_state=0)
X = df[["age", "education", "prior_income"]].values
est = DragonNet(epochs=50, random_state=0).fit(
    X, df["treatment"].values, df["outcome"].values
)
print("ATE:", est.predict_ate(X))

Choosing an estimator

Question Reach for
Honest tree or forest CATE, with a doubly robust score CausalTree, CausalForest, DRLearner
Uplift on a binary outcome UpliftRandomForest, UpliftXGBoost
Constant effect with a confidence interval PartiallyLinearDML, NeuralDML
Who should be treated? PolicyLearner, DRPolicyTree
Threshold, before/after, or a single treated unit RegressionDiscontinuity, DifferenceInDifferences, SyntheticControl
Bayesian uncertainty for a network or a quasi-experiment bayesian_causalML
Neural effect on i.i.d. data TARNet, CFRNet, DragonNet
Which variables cause which? NOTEARSLinear, DAGGNN, TCDF

Documentation

The sidebar is the tutorial series for this package: tree-based models, meta-learners, uplift, double machine learning, deep causal learning, Bayesian causal ML, policy learners, and quasi-experiments.

Start here:

Source, issues, and the full README live on GitHub.

License

MIT — see LICENSE.

Verified details

License MIT
Author Zia U. Ahmed
Requires Python ≥ 3.9
Framework PyTorch ≥ 2.0, scikit-learn ≥ 1.2

Optional extras

plot, boost, shap, dowhy, econml, causallearn, uber, doubleml, causalpy, skgrf, dev

Classifiers

  • Development Status :: Beta
  • Intended Audience :: Science/Research
  • Topic :: Scientific/Engineering :: Artificial Intelligence
Back to top