Local Causal Discovery for Estimating Causal Effects

Overview

This repository contains the implementation for the paper:
Local Causal Discovery for Estimating Causal Effects
Conference on Causal Learning and Reasoning (CLeaR), 2023
Shantanu Gupta, David Childers, Zachary Lipton

Code and Datasets

The causal_discovery folder contains the code for the various causal discovery algorithms used in our work:

• pc_alg.py: An implementation of the PC algorithm (code adapted from the pcalg library).
• sd_alg.py: An implementation of the SD algorithm which recursively runs the PC algorithm locally starting from the treatment, its neighbors, and so on.
• mb_by_mb.py: An implementation of the MB-by-MB algorithm (Wang et. al., 2014) which performs local causal discovery by sequentially finding Markov blankets and local structures within the Markov blankets.
• ldecc.py: An implementation of the LDECC algorithm proposed in our work.

The following Jupyter notebooks contain example usages of the various algorithms and code for running the experiments in our paper:

• Results_on_synthetic_linear_graphs.ipynb: Code for experiments on synthetic linear Gaussian graphs (Fig. 10).
• Results_on_semi_synthetic_linear_Gaussian_graphs.ipynb: Code for experiments on semi-synthetic linear Gaussian graphs from bnlearn (Figs. 11, 17).
• Results_on_synthetic_linear_Erdos_Renyi_graphs.ipynb: Code for experiments on synthetic Erdos-Renyi linear Gaussian graphs (Fig. 16).
• Results_on_semi_synthetic_discrete_graphs.ipynb: Code for experiments on three semi-synthetic discrete graphs from bnlearn (Fig. 18).

The data folder contains the bnlearn graphs used in our experiments.

Citation

If you find this work useful, please consider citing our work:

@inproceedings{gupta2023local,
  title={Local Causal Discovery for Estimating Causal Effects},
  author={Gupta, Shantanu and Childers, David and Lipton, Zachary C.},
  booktitle={Conference on Causal Learning and Reasoning},
  year={2023}
}