Causal Assistant

This library is a semi-wrapper around Jianqiao Mao's causalBootstrapping package (based upon Little et al.'s Causal Bootstrapping paper) to improve the user-friendliness and performance of the technique when used for general causal de-confounding.

Causal Bootstrapping is a process for improving model robustness in the presence of causal factors (confounders) by iteratively re-weighting samples within a dataset.

For reference on causality, I suggest reading the works of Judea Pearl. To my knowledge, neither of his books on the topic (Causality, The Book of Why) are freely available outside of academia, however Probabilistic and Causal Inference is, and may be a good place to get started.

Usage and Development

• The project has been built with an assumed Python version of 3.12, however earlier and later versions may also work.
• I recommend uv and virtual environments for dependency management.
• Pull requests are welcome for new functionality or fixes!

Basic Usage:

import causal_assistant as ca

# load your feature data
X, y, u = ...

# de-confound the data!
X_dc, y_dc = ca.bootstrap(causal_graph="""X;y;u;X->y;u->X;u->y;""", X=X, y=y, u=u)

Attribution

Some code in this module is derived from the causalBootstrapping library.

If you use this library, please also cite the originating paper:

@article{little2019causal,
  title={Causal bootstrapping},
  author={Little, Max A and Badawy, Reham},
  journal={arXiv preprint arXiv:1910.09648},
  year={2019}
}