Transfer Learning for Individual Treatment Effects

This repository contains the code accompanying the paper:

“Advantages and Limitations in the Use of Transfer Learning for Individual Treatment Effects in Causal Machine Learning”
by Aydin & Brandt.

---

Data

• Simulation (Data/Datasets/Simulation):
  Includes simulated source datasets. Using the functions in the Generation file, target datasets (randomized or non randomized) can be created.

• Empirical Example (Data/Datasets/Empirical):
  Datasets from the IHDS-II household survey dataset for empirical analysis. Code under Generation file can be used to create datasets.

• Generation (Data/Generation): Includes functions to create empirical or simulation datasets.

Functions (Model & Training Procedure)

• TARNet Model (TARNet.py):
  Defines the shared representation and two potential outcome heads.

• Phase 1 – Distribution Alignment (Optimize_IPM.py):
  Trains the representation to align source and target treatment/control distributions using an Integral Probability Metric (IPM).

• Phase 2 – Factual Loss Training (Optimize_Loss.py):
  Trains the treatment and control outcome heads on the target dataset.

Distance Measures

• Distribution Distances (Distances/Distance.py):
  Implements IPM-based metrics for quantifying dataset distribution differences.

---

Repository Structure

TL_TARNet
|-- Data
|   |-- Datasets
|   |   |-- Empirical
|   |   |   |-- biased_subsample.csv
|   |   |   |-- random_subsample.csv
|   |   |   |-- punjab.csv
|   |   |   `-- uttar_pradesh.csv
|   |   `-- Simulation
|   |       |-- source_1000.csv
|   |       |-- source_5000.csv
|   |       |-- source_10000.csv
|   |       |-- source_20000.csv
|    |      `-- source_30000.csv
|   `-- Generation
|
|-- Distances
|   `-- Distance.py
|
|-- Functions
|   |-- TARNet.py
|   |-- Optimize_IPM.py
|   `-- Optimize_loss.py
|
`-- Results
    |-- simulation
    |-- empirical
    `-- plots

Contact

For questions or discussion, feel free to contact the authors.