Meta-learners for estimating multi-valued treatments heterogeneuous effects.

Code Author: Naoufal Acharki (naoufal.acharki@gmail.com)

This repositery multipleT-MetaLearners contains the code/simulations in R (see Appendices A-D) and the semi-synthetic dataset (described in Appendix E) as detailed in the Supplementary Materials of our the paper Comparison of meta-learners for estimating multi-valued treatment heterogeneous effects (Acharki et al., 2023)

Here, we implement meta-learners (T-, S-, M-, DR-, R-) and different versions (X- and Naive X-) learners for multi-valued treatments to estimate heterogenous treatments. These implementations can be found in MetaLearners_tools.R in Scripts and the analytical tests presented in the paper (linear model and Hazard rate) are written in each scripts of Appendix_D1 to Appendix_D4.

This software is currently in beta, and we expect to make continual improvements to its performance and usability.

Experiments and simulations

All experiments in Acharki et al. (2023) can be replicated using this repository in the fold Scripts. The necessary code for each table in the main paper or Supplementary Materials can be reproduced by running to script associated to the experiment (e.g. Lin_Rand_Case1.R in Appendix D1 for Table 5).

Semi-synthetic datasets:

In the fold Datasets, you can find and upload the following datasets in the zip file Semi-Synthetic-EGS.zip :

• "Single_Fracture_Simulation_Cases_16200.csv"
• "Fracture_Efficency.csv"
• "Main_Dataset.csv"

We refer the reader to Appendix E for more details about the physical model used to generate this dataset and how it can be useful for further use/application in Causal Inference. An example of the use of this semi-synthetic dataset, the creation of a non-randomized biased dataset, is described in Appendix E and can be found in Scripts/Appendix_EGS/EGS_CATE.

Citation

If you use this software or the datasets please cite the corresponding paper(s):

@misc{acharki2023comparison,
      title={Comparison of meta-learners for estimating multi-valued treatment heterogeneous effects}, 
      author={Naoufal Acharki and Ramiro Lugo and Antoine Bertoncello and Josselin Garnier},
      year={2023},
      eprint={2205.14714},
      archivePrefix={arXiv},
      primaryClass={stat.ML}
}