zhaozhiqian/single-cause-perturbation-neurips-2021

Code for Estimating Multi-cause Treatment Effects via Single-cause Perturbation (NeurIPS 2021)

Estimating Multi-cause Treatment Effects via Single-cause Perturbation (NeurIPS 2021)

Single-cause Perturbation (SCP) is a framework to estimate the multi-cause conditional average treatment effect (CATE) from observational data.

Most existing CATE estimation methods are designed for single cause interventions, i.e. only one variable can be intervened on at one time. However, many applications involve simultaneous intervention on multiple variables. This is the multi-cause estimation problem addressed by SCP.

SCP leverages the connection between single and multi-cause interventions and overcomes the confounding bias via data augmentation. Compared with existing works, SCP does not make assumptions about the distributional or functional form of the DGP.

Usage

To run the code locally, make sure to first install the required python packages specified in requirements.txt. Python 3.6+ is recommended (the code has been tested with Python 3.7).

All the experiments scripts are located under experiments/. The reproduce_all.sh shell script contains commands to reproduce all synthetic data tables and figures in the paper. The reproduce_real.sh shell script contains commands to reproduce all real data tables and figures in the paper. The Fig[x].sh or Tab[x].sh shell script contain commands to generate results for individual figures or tables. The Fig[x].ipynb notebooks contain commands to create the visualizations.

An implementation of SCP is provided in the file run_simulation_scp.py. Note that SCP is a general framework agnostic to the exact choice of step one and step two estimators. In this implementation, we use DR-CFR in step one and neural network regression in step two. The benchmarks are implemented in the files benchmarks/run_simulation_[x].py.

Citation

If you find the software useful, please consider citing the following paper:

@inproceedings{scp2021,
  title={Estimating Multi-cause Treatment Effects via Single-cause Perturbation},
  author={Qian, Zhaozhi and Curth, Alicia and van der Schaar, Mihaela},
  booktitle={Advances in neural information processing systems},
  year={2021}
}

License

Copyright 2021, Zhaozhi Qian.

This software is released under the MIT license.

single-cause-perturbation-neurips-2021's Issues

Inconsistentence between code and simulation.

Dear Zhiqian,

Thank you so much for presenting this nice paper.

However, I noticed that there may exist some name inconsistence between the code and simulation. For example, for Figure 3, you mentioned that 8 methods were included in the benchmark, including NN, NN-IPW, OP, DEC, VSR, SCP, CFR and DR-CFR. However, I could not find related name in the code (run_one_simulation.sh). In this file, there were eight methods, dor, scp, propensity, tarnet, vsr, drcrn, bmc overlap. Would you mind sharing some hints that which refers to which here?

Thank you!