SyncTwin: Treatment Effect Estimation with Longitudinal Outcomes (NeurIPS 2021)

SyncTwin is a treatment effect estimation method tailored for observational studies with longitudinal data. Specifically, it applies to the LIP setting: Longitudinal, Irregular and Point treatment. In these studies, the covariates are observed at irregular intervals leading to treatment allocation; the outcomes are measured longitudinally before and after the treatment; the treatment is assigned at a specific time point and stays unchanged during the study.

The key insight of SyncTwin is to fully leverage the pre-treatment outcomes. It uses the temporal structure in the outcome time series to improve the accuracy of counterfactual prediction. It further uses the pre-treatment outcomes to control the estimation error on the individual level. Finally, the method enables interpretability by example: the user can examine the key contributing examples that leads to the estimate.

Installation

To run the code locally, make sure to first install the required python packages specified in requirements.txt. Python 3.7 is recommended for best compatibility. Note that tensorflow and GPy are only needed for running the benchmarks. The directory clairvoyance contains a streamlined version of the clairvoyance library. It is used to run the benchmarks CRN and RMSN.

For some benchmarks (SC, MC-NNM, 1NN), we use their public implementations in the R language. To run these benchmarks, please install R and the dependencies listed in requirements_R.txt.

For coda users, an environment YAML file environment.yml is provided, which includes both Python and R dependencies.

Usage

Scripts for reproducing paper experiments are provided under the directory experiments/.

The reproduce_all.sh shell script contains commands to reproduce all 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. The results will be written in the results folder. For instance, Tab2_C1_MAE.txt corresponds to the first Column of Table 2.

An implementation of SyncTwin is provided in the file SyncTwin.py. Note that SyncTwin is a general framework agnostic to the exact architectural choice of encoder and decoder. In this implementation, we use attentive GRU-D encoder and time-LSTM decoder. In the simulations, SyncTwin is trained in pkpd_sim3_model_training.py.

Citation

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

@inproceedings{synctwin2021,
  title={SyncTwin: Treatment Effect Estimation with Longitudinal Outcomes},
  author={Qian, Zhaozhi and Zhang, Yao and Bica, Ioana and Wood, Angela 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.