Fast matching estimators for causal inference under unconfoundedness. Interpretable, fast alternative to doubleML under unconfoundedness and overlap.
Supports
- Estimand: ATT, ATE
- Estimator: vanilla KNN, bias-corrected KNN (the latter lets you use standard errors using Rai/Otsu 2017)
- Nearest Neighbours Solvers: scikit, faiss, voyager
- the latter two are very fast and can be used with GPU support, and can also be exported to a production environment where adding new units and future nearest-neighbours queries are very fast
Uses modern vectorDBs for knn search
To install, run
pip install git+https://github.com/py-econometrics/fastmatch.git
Install with gpu support: pip install fastmatch[gpu]