AutoCATE: End-to-End, Automated Treatment Effect Estimation
Lists
Only you see this
Abstract
Accurate estimation of heterogeneous treatment effects is critical in domains such as healthcare, economics, and education. While machine learning (ML) has led to significant advances in estimating conditional average treatment effects (CATE), real-world adoption of these methods remains limited due to the complexity of implementing, tuning, and validating them. To this end, we advocate for a more holistic view on the development of ML pipelines for CATE estimation through automated, end-to-end protocols. We formalize the search for an optimal pipeline as a counterfactual Combined Algorithm Selection and Hyperparameter optimization (CASH) problem. We introduce \texttt{AutoCATE}, the first automated solution tailored for CATE estimation that addresses this problem based on protocols for evaluation, estimation, and ensembling. Our experiments show how AutoCATE allows for comparing different protocols, with the final configuration outperforming common strategies. We provide AutoCATE as an open-source software package to help practitioners and researchers develop ML pipelines for CATE estimation.
BibTeX
@inproceedings{vanderschueren2025autocate,
author = {Toon Vanderschueren and Tim Verdonck and Mihaela van der Schaar and Wouter Verbeke},
title = {{AutoCATE: End-to-End, Automated Treatment Effect Estimation}},
booktitle = {International Conference on Machine Learning},
year = {2025},
url = {https://openreview.net/forum?id=QbOoz74GNO},
}