This code is the official implementation of Counterfactual Prediction Under Outcome Measurement Error, published at FAccT 23'.

Requirements

To install dependencies, run

! pip3 install -r requirements.txt

For licensing and privacy reasons, we omit raw data for semi-synthetic experiments from this repository. The raw data for the semi-synthetic evaluations can be downloaded from:

• JOBS Dataset: train, test
• Oregon Health Insurance Experiment (OHIE) Dataset: https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/SJG1ED

After accessing the data, place the files in the relevant directory data/JOBS/ and data/OHIE/. The OHIE data can be merged and pre-processed by executing data/OHIE/preprocess.py after placing the download package in the folder.

Running experiments

To run the synthetic experiment, execute:

! python3 drivers.py erm syn_convergence_learned
! python3 drivers.py erm syn_convergence_oracle

To run the semi-synthetic experiment, execute:

! python3 drivers.py erm_semisyn semisyn_oracle_exp
! python3 drivers.py erm_semisyn_assumption semisyn_learned_exp

Bibtex Citation

@article{guerdan2023counterfactual,
  title={Counterfactual Prediction Under Outcome Measurement Error},
  author={Guerdan, Luke and Coston, Amanda and Holstein, Kenneth and Wu, Zhiwei Steven},
  booktitle={Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency},
  year={2023}
}