Causal inference for when randomization is impossible

What it does

CausalPy provides a Python toolkit for analyzing natural experiments—policy rollouts, geographic lifts, threshold-based eligibility rules, and staggered adoptions. It implements methods like difference-in-differences, synthetic control, regression discontinuity, and instrumental variables, estimating effects through PyMC’s Bayesian machinery or traditional OLS via scikit-learn. Every analysis returns effect summaries with uncertainty intervals and optional ROPE testing for practical significance.

The interesting bit

Most causal libraries treat uncertainty as an afterthought; CausalPy makes it the main event by reporting HDIs and tail probabilities by default. The library also bakes in diagnostics-first design—effect summaries come in both tabular and prose formats, which is the kind of reproducible artifact you actually want to paste into a decision memo.

Key highlights

• Covers ten quasi-experimental methods including staggered DiD, regression kink, and inverse propensity weighting
• Bayesian-first via PyMC with Highest Density Intervals, plus fallback OLS through scikit-learn
• Built-in ROPE analysis and direction testing (e.g., P(effect > 0)) for practical significance
• Outputs publication-quality plots and structured effect summaries meant for direct interpretation
• Includes “AI agent skills” to teach coding assistants how to use the library

Caveats

• Explicitly not for black-box automated causal discovery or weakly identified observational data without a plausible design
• Production workflow features like multi-user governance and pipelines are out of scope
• Zenodo DOI for stable releases is still planned, so citation currently relies on the GitHub repository

Verdict

Worth a look if you run geo-lifts, policy evaluations, or any quasi-experiment where defensible uncertainty quantification matters. Skip it if you are hunting for automated causal discovery in messy observational data or need enterprise pipeline orchestration.