Beyond the Average: Distributional Causal Inference under Imperfect Compliance

This repository contains the implementation code for reproducing results from the paper "Beyond the Average: Distributional Causal Inference under Imperfect Compliance" (NeurIPS 2025) by Undral Byambadalai, Tomu Hirata, Tatsushi Oka, and Shota Yasui.

Simulation with Python

Environment

• Python >= 3.12
• uv >= 0.7.0

Installation

# Clone the repository
git clone https://github.com/CyberAgentAILab/ldte.git
cd ldte

# Setup Python environment
# See https://docs.astral.sh/uv/ for other installation
curl -LsSf https://astral.sh/uv/install.sh | sh
uv python install 3.12
uv python pin 3.12
uv venv
source .venv/bin/activate

# For running DTE estimation
uv sync

Run Simulation

You can run the simulation using simulation.py. The available arguments for this script are described below.

Arguments:

• n: Specifies the number of samples for each simulation iteration. This is a required argument and must be a positive integer.
• iterations: The number of simulation iterations.
• discrete: Indicates whether the Data Generating Process (DGP) is discrete outcome or continuous outcome. By default, the DGP is continuous. Use this flag to switch to discrete mode.

Example:

• For 1000 samples with a continuous DGP: --n 1000
• For 5000 samples with a discrete DGP: --n 5000 --discrete true

python simulation.py --n 5000 --discrete true

Empirical Application

Data preparation

Download Oregon Insurance Experiment Data from https://www.nber.org/research/data/oregon-health-insurance-experiment-data.

python oregon.py --data_dir path/to/OHIE_Data --n_bootstrap 50

Contributors

• Undral Byambadalai (undral21@gmail.com)
• Tomu Hirata (tomu.hirata@databricks.com)