End-to-End Balancing for Causal Continuous Treatment-Effect Estimation

Authors: Taha Bahadori, Eric Tchetgen Tchetgen, David Heckerman

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Using synthetic and real-world data, we show that our proposed algorithm outperforms the entropy balancing in accuracy of treatment effect estimation.
Researcher Affiliation Collaboration 1Amazon.com, Inc. 2Wharton School of the University of Pennsylvania. Correspondence to: Mohammad Taha Bahadori <bahadorm@amazon.com>.
Pseudocode Yes Algorithm 1 Stochastic Training of ℓθ for End-to-End Balancing
Open Source Code No The paper does not provide an explicit statement or link for open-source code for the methodology.
Open Datasets Yes We study the impact of PM2.5 particle level on the cardiovascular mortality rate (CMR) in 2132 counties in the US using the data provided by the National Studies on Air Pollution and Health (Rappold, 2020). The data is publicly available under U.S. Public Domain license.
Dataset Splits Yes All neural networks are trained using Adam (Kingma & Ba, 2014) with early stopping based on validation error. The learning rate and architectural parameters of the neural networks are tuned via hyperparameter search on the validation data.
Hardware Specification Yes We performed our experiments on a CPU machine with 16 cores from a cloud provider that uses hydroelectric power.
Software Dependencies No The paper mentions using PyTorch but does not specify its version number or any other software dependencies with version numbers.
Experiment Setup Yes All neural networks are trained using Adam (Kingma & Ba, 2014) with early stopping based on validation error. The learning rate and architectural parameters of the neural networks are tuned via hyperparameter search on the validation data. Learning algorithm: Adam with learning rate 0.001, no AMSGrad. Batch size: 100 Max epochs: 400 Weight decay: 2.5e-5.