This code accompanies the paper Undersmoothing Causal Estimators with Generative Trees [1].
The experimental setup is based on the CATE benchmark and extended to test our proposed DeGeTs framework.
The easiest way to replicate the running environment is through Anaconda. Once installed, follow the steps below.
- Download the repo.
- Enter the directory (i.e.
cd cate-benchmark). - Run the following command to recreate the 'cate-bench' conda environment:
conda env create -f environment.yml
- Download datasets from here. Once downloaded, extract them to 'datasets' directory.
main.py- main script to run the experiments.models/tree_balancing.py- our proposed DeGeTs framework (DeGeDTs and DeGeF).utils.py- helper functions.
Head to `experiments' folder to run the main script in a more automated and convenient way.
experiments/full_run.sh - Replicates 100% of the experiments as in the paper. Note this full setup takes days/weeks to complete, even on fairly strong machines.
experiments/demo.sh - A small experimental run for demonstration purposes. It involves the IHDP data set (10/1000 iterations) and only selected estimators.
Note the intermediate results printed to the console while running the scripts show mean +- standard error. The final results in the paper were computed separately from outputted CSV files in order to show mean +- 95% confidence intervals.
For more details about how to use the framework and analyse the results, please visit the CATE benchmark website.
[1] D. Machlanski, S. Samothrakis, and P. Clarke, ‘Undersmoothing Causal Estimators with Generative Trees’, arXiv:2203.08570 [cs, stat], Mar. 2022, doi: 10.48550/arXiv.2203.08570.