🦞LobsterNet: Heterogeneous Treatment Assignment Effect Estimation Under Non-adherance

🔍 Overview

• LobsterNet is a multitask neural network designed for heterogeneous treatment assignment effect estimation under treatment non-adherence.

• It leverages the conditional front-door adjustment (CFD), which theoretically gaurantees lower variance estimates than the commonly used standard backdoor adjustment (CFD) when the true treatment effect is small.

• This repository contains reproducible codes for experimental results in "Heterogeneous Treatment Assignment Effect Estimation Under Non-adherance with Conditional Front-door Adjustment" (CHIL 2025).

---

▶️ Quick Start

1: obtain required datasets

• Synthetic datasets will be automatically generated in the following steps
• IHDP dataset is provided under data/IHDP/ihdp.RData, originally downloaded from the npci repository
• AMR-UTI dataset need to be obtained from PhysioNet, and place the all_prescriptions.csv, all_uti_features.csv, and all_uti_resist_labels.csv files in data/AMR-UTI folder.

2: Install required packages

    pip install -r requirements.txt

3: train and evaluate SBD and CFD estimators:

Path	Description
src/run_batch_main_sim_A.sh	Run synthetic dataset A experiments method
src/run_batch_main_sim_B.sh	Run synthetic dataset B experiments usage
src/run_batch_main_ihdp.sh	Run IHDP experiments
src/run_batch_main_amruti.sh	Run AMR-UTI experiments

4: analyze results

Path	Description
src/analysis/variance_viz.ipynb	Analyze asymptotic variance
src/analysis/sim_experiment_results.ipynb	Analyze synthetic datasets experiment results
src/analysis/ihdp_experiment_results.ipynb	Analyze IHDP experiments results
src/analysis/amruti_experiment_results.ipynb	AMR-UTI experiments results

---

📝 Citation

If you use LobsterNet in your research, please cite:

Chen W, Chang T, Wiens J. Conditional Front-door Adjustment for Heterogeneous Treatment Assignment Effect Estimation Under Non-adherence. Conference on Health, Inference and Learning (2025).

---

🛠️ License

This project is licensed under the Apache 2.0 License.