Two-stage framework that combines observational and IV data to reliably estimate conditional average treatment effects (CATEs), addressing both unobserved confounding and low compliance issues.

Replication code for Estimating Heterogeneous Treatment Effects by Combining Weak Instruments and Observational Data.

• scikit-learn
• pytorch
• doubleml

Use the following commands to replicate the figures from the "Estimating Heterogeneous Treatment Effects by Combining Weak Instruments and Observational Data" paper:

• For Figure 2, top row, and Table 1: python para_sims.py
• For Figure 2, bottom row: python rep_sims.py
• For Figure 3 & 4, Table 4 & 5: python 401k.py

For interactive experimentation, we provide three Jupyter notebooks in the notebooks folder. Each notebook demonstrates the results of one simulation/run for the three considered settings: parametric extrapolation simulation, representation learning simulation, and the 401(k) participation treatment effect.