Transfer Learning for Treatment Effect Estimation in Small Samples

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/          # empirical datasets (IHDS subsamples, Punjab, Uttar Pradesh)
│   │   │   ├── biased_subsample.csv
│   │   │   ├── punjab.csv
│   │   │   ├── random_subsample.csv
│   │   │   └── uttar_pradesh.csv
│   │   ├── Simulation_1/       # source datasets at varying sample sizes
│   │   │   ├── source_1000.csv
│   │   │   ├── source_5000.csv
│   │   │   ├── source_10000.csv
│   │   │   ├── source_20000.csv
│   │   │   └── source_30000.csv
│   │   └── Simulation_2/       # 93 simulated datasets (.csv)
│   └── Generation/             # data-generation scripts
│       ├── IHDS.R
│       ├── Simulation_1.R
│       └── Simulation_2.py
├── Distances/
│   └── Distance.py
├── Functions/                  # TARNet model + optimization code
│   ├── Optimize_IPM.py
│   ├── Optimize_loss.py
│   └── TARNet.py
├── Results/
│   ├── Empirical/              # empirical results, with/without transfer learning
│   │   ├── biased_and_source_w_TL.csv
│   │   ├── biased_wo_TL.csv
│   │   ├── random_and_source_w_TL.csv
│   │   └── random_wo_TL.csv
│   ├── Plots/
│   │   ├── Empirical/          # 8 plot images (.png)
│   │   ├── Simulation_1/       # 30 plot images (.png)
│   │   └── Simulation_2/       # 74 plot images (.png)
│   ├── Simulation_1/
│   │   ├── with_TL.csv
│   │   └── without_TL.csv
│   └── Simulation_2/
│       ├── results_all/        # 200 per-run result files (.csv)
│       ├── results.csv
│       └── summary.csv
└── README.md