Efficient Online Estimation of Causal Effects by Deciding What to Observe

Overview

This repository contains the implementation for the paper:
Efficient Online Estimation of Causal Effects by Deciding What to Observe [PDF]
Advances in Neural Information Processing Systems (NeurIPS), 2021
Shantanu Gupta, Zachary Lipton, David Childers

Code and Datasets

The following Jupyter notebooks contain the code for running the experiments in our paper:

• Linear_split_IV_graph.ipynb: Code for the experiment for the instrumental variable (IV) graph (Figures 2a and 3a).
• Two_IVs_graph.ipynb: Code for the experiment for the graph with two IVs (Figures 2b and 3b).
• Confounder_mediator_graph.ipynb: Code for experiment for the confounder-mediator graph (Figures 2c and 3c).
• IHDP_data_graph.ipynb: Code for the experiment with the Infant Health Development Program (IHDP) dataset (Figures 4a and 4b).
• Vietnam_draft_earnings_data_graph.ipynb: Code for the experiment with the Vietnam era draft and earnings dataset (Figure 4c).

The code depends on the Python libraries numpy, scipy, and pandas. In order to execute the iterations in parallel, we use the package ipyparallel.

The datasets and the code used for data preprocessing are located in the data/ folder. The following files inside this folder contain the datasets used

• ihdp.RData: The IHDP dataset taken from Dorie, 2016.
• ihdp_processed.csv: The preprocessed IHDP dataset that we use in our experiments.
• cpi_angrist1990.dta and cwhsc_new.dta: The Vietnam era draft and earnings dataset taken from Angrist, 2009.
• angrist_1951_cohort.csv: The preprocessed Vietnam era draft and earnings dataset that we use in our experiments.

And the following files inside the data/ folder contain the data preprocessing code:

• IHDP_dataset_preprocessing.ipynb: Code for preprocessing the IHDP data (generates ihdp_processed.csv).
• Vietnam_draft_earnings_data_preprocessing.ipynb: Code for preprocessing the Vietnam era draft and earnings dataset (generates angrist_1951_cohort.csv).

Citation

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

@inproceedings{gupta2021efficient,
  title={Efficient Online Estimation of Causal Effects by Deciding What to Observe},
  author={Gupta, Shantanu and Lipton, Zachary C. and Childers, David},
  booktitle={Advances in Neural Information Processing Systems},
  year={2021}
}