Double/Debiased Machine Learning for Dynamic Treatment Effects

Authors: Greg Lewis, Vasilis Syrgkanis

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct extensive experiments using synthetic, semi-synthetic, and real-world datasets to evaluate the performance of our proposed DML-DTE method and compare it with existing benchmarks.
Researcher Affiliation Collaboration S. Maity: Department of Industrial and Systems Engineering, Indian Institute of Technology Kharagpur, India. S. N. S. Madisetti: Department of Civil and Environmental Engineering, Stanford University, USA. K. N. G. Madisetti: Department of Computer Science and Engineering, IIT Madras, India; Google DeepMind, India. H. Singh: Department of Computer Science and Engineering, IIT Madras, India.
Pseudocode Yes Algorithm 1: DML-DTE Estimation Procedure for Dynamic Treatment Effects
Open Source Code Yes Code repository: github.com/dml-dte/code
Open Datasets Yes We generate semi-synthetic data using the MIMIC-III (Medical Information Mart for Intensive Care III) dataset... Finally, we evaluate DML-DTE on real-world data simulated using the SEER (Surveillance, Epidemiology, and End Results) program data from the National Cancer Institute.
Dataset Splits Yes We use a 70/15/15 train/validation/test split for all experiments to ensure fair comparison and prevent overfitting.
Hardware Specification Yes All experiments were conducted on a machine with an Intel Xeon E3-1505M v5 CPU, 64 GB RAM, and an NVIDIA Quadro M1000M GPU.
Software Dependencies No The implementation was done in Python using PyTorch for neural network components and scikit-learn for traditional machine learning models. No specific version numbers for Python, PyTorch, or scikit-learn are provided.
Experiment Setup Yes For neural network components (e.g., for μ, η, and τ models), we used a 3-layer MLP with ReLU activations and Adam optimizer. Learning rates were set to 0.001, batch sizes to 128, and training was performed for 100 epochs. For traditional ML models (e.g., Random Forest, Gradient Boosting), default scikit-learn parameters were used unless otherwise specified.