Debiased Front-Door Learners for Heterogeneous Effects

This repository contains the public release for Debiased Front-Door Learners for Heterogeneous Effects.

It includes only:

• reproducible experiment code,
• a short usage guide,
• the paper PDF.

Contents

Path	Description
paper/debiased-front-door-learners-for-heterogeneous-effects.pdf	Paper PDF.
code/FDCATE.py	Synthetic experiments and estimator implementations for FD-PI, FD-DR, and FD-R.
code/analyze_fars_2000_fd.py	FARS 2000 front-door case-study script.
requirements.txt	Minimal Python dependencies.

Setup

Use Python 3.9 or newer.

python -m venv .venv
source .venv/bin/activate
python -m pip install -U pip
python -m pip install -r requirements.txt

Synthetic Experiments

Run the help command:

python code/FDCATE.py --help

Fast smoke run:

python code/FDCATE.py \
  --ns 100 200 \
  --dim 2 \
  --rounds 1 \
  --delta-grid-fixed-n 0 \
  --fixed-n-for-sweep 100 \
  --skip-weak-overlap \
  --skip-mediator-confound \
  --no-save \
  --no-plots \
  --quiet

Paper-scale runs use larger sample-size grids and more Monte Carlo replications. See the CLI defaults in code/FDCATE.py.

FARS Case Study

The FARS script expects raw public FARS/NHTSA input files under data/fars/:

ACCIDENT.CSV
PERSON.CSV
state_primary_2000.csv

Run:

python code/analyze_fars_2000_fd.py --dir data/fars

The public repository does not include raw FARS data.

Citation

If you use this code, please cite:

@article{jung2025debiasedfrontdoor,
  title={Debiased Front-Door Learners for Heterogeneous Effects},
  author={Jung, Yonghan},
  journal={arXiv preprint arXiv:2509.22531},
  year={2025}
}