pymc-labs/CausalPy

A Python package for causal inference in quasi-experimental settings

CausalPy – Bayesian‑first causal inference for quasi‑experimental designs

What it does – CausalPy is a Python library that lets analysts estimate causal effects when a true randomized experiment isn’t possible. It implements a suite of classic quasi‑experimental methods (difference‑in‑differences, synthetic control, regression discontinuity, etc.) and runs them as Bayesian models built on PyMC (with an optional OLS fallback). The output includes full posterior summaries, credible intervals, ROPE‑based practical‑significance checks and ready‑to‑publish plots.

Why it matters

  • Uncertainty‑aware: instead of a single point estimate you get a distribution, HDI, tail probabilities, and ROPE analysis.
  • Reproducible workflow: data loading, model specification, fitting and visualisation are all wrapped in high‑level classes (RegressionDiscontinuity, SyntheticControl, …) that produce a single result object with plot() and summary() methods.
  • Built on PyMC: leverages modern probabilistic programming (PyMC 6, PyTensor 3, ArviZ) while still supporting classic OLS via scikit‑learn for quick checks.
  • Consulting‑ready: the team offers paid consulting for high‑stakes lift‑testing or policy‑impact work.

Core features (from the README)

Feature Details
Methods Synthetic control, Geographical lift, ANCOVA, Diff‑in‑Diff, Staggered DiD, Regression discontinuity/kink, Interrupted time series, Instrumental variables, Inverse propensity weighting
Bayesian first Models are written with PyMC; posterior HDI, tail‑probabilities, ROPE. OLS fallback via scikit‑learn.
Decision‑ready outputs effect_summary() returns tables and prose; plot() gives publication‑quality graphics.
Installation pip install CausalPy or conda install -c conda‑forge causalpy. Works on Python 3.12+.
Runtime notes Uses numba (via PyTensor) – first sample incurs compilation cost; native crashes possible; nutpie can silently change NUTS defaults.
AI‑agent skills Small “skills” for AI coding assistants live under causalpy/skills/ and are published to Decision AI Hub.
Documentation & support Read‑the‑Docs site, GitHub Discussions for Q&A, Issues for bugs, and a consulting Calendly link.
License Apache 2.0.

Typical usage (quick‑start example)

import causalpy as cp
import matplotlib.pyplot as plt

# Load a toy dataset bundled with the package
df = (
    cp.load_data("drinking")
    .rename(columns={"agecell": "age"})
    .assign(treated=lambda d: d.age > 21)
)

# Fit a regression‑discontinuity model using a Bayesian linear regression
result = cp.RegressionDiscontinuity(
    df,
    formula="all ~ 1 + age + treated",
    running_variable_name="age",
    model=cp.pymc_models.LinearRegression(),
    treatment_threshold=21,
)

# Visualise the jump at the cutoff
fig, ax = result.plot()

# Print a summary with posterior intervals
result.summary()

The result object encapsulates the fitted model, diagnostics, and helper methods for reporting.

Who should use it?

  • Product/marketing analysts evaluating geo‑lift, policy changes, or staggered roll‑outs.
  • Social scientists needing rigorous DiD or RD analyses with uncertainty quantification.
  • Data scientists who prefer a Bayesian approach but still want the option of fast OLS.

It is not aimed at:

  • Pure causal‑discovery from observational data.
  • Fully automated “black‑box” causal answers.
  • Large‑scale production pipelines (no built‑in orchestration).

Getting started & resources

  • Installation: pip install CausalPy (or conda). Requires Python 3.12+, PyMC 6.
  • Docs: https://causalpy.readthedocs.io – includes tutorials for each method.
  • Community: GitHub Discussions for usage questions; Issues for bugs.
  • Video demo: YouTube link in the README.

TL;DR

CausalPy is a research‑grade, Bayesian‑first Python toolbox for quasi‑experimental causal inference. It wraps classic designs in a PyMC‑powered workflow, delivering uncertainty‑aware estimates, diagnostics, and ready‑to‑share visualisations, while remaining easy to install and use for analysts familiar with Python’s data‑science stack.

Related

  • Project
    py-why/dowhy

    DoWhy is a Python library for end‑to‑end causal inference. It lets you define a causal graph, automatically identifies the causal effect, estimates it with a variety of statistical methods, and runs refutation tests to check robustness. The library also supports graphical causal models for root‑cause analysis, counterfactuals, and interventional sampling. Install via `pip install dowhy` and follow the four‑step workflow (model → identify → estimate → refute).

  • Project
    pgmpy/pgmpy

    A Python library for causal and probabilistic reasoning using graphical models, providing tools for structure learning, parameter estimation, and causal inference.

  • Project
    igerber/diff-diff

    A Python library for Difference-in-Differences causal inference providing an sklearn-like API for a wide range of econometric estimators and rigorous diagnostic tools.

  • Project
    erdogant/bnlearn

    bnlearn is a Python library for learning, parameterising, and querying Bayesian networks. It supports structure discovery (various scores and search methods), CPT estimation, causal inference (do‑calculus), synthetic data generation, and visualisation, all wrapped in a simple API.

  • Project
    process-intelligence-solutions/pm4py

    PM4Py is an open‑source Python library for process mining, offering algorithms to discover, analyse and visualise business process models from event logs. Developed by Process Intelligence Solutions (a Fraunhofer spin‑off), it supports Python 3.9‑3.14, optional ML/LLM extensions, and is AGPL‑3.0 licensed (commercial licence available).