Replication files for Chernozhukov, Newey, Quintas-Martínez and Syrgkanis (2021) "RieszNet and ForestRiesz: Automatic Debiased Machine Learning with Neural Nets and Random Forests"

Main Files

1. RieszNet_IHDP.ipynb runs the ATE MAE and coverage experiments based on IHDP semi-synthetic data using RieszNet. It takes the files in "data/IHDP" as inputs, and outputs in "results/IHDP/RieszNet." It produces a table "IHDP_MAE_NN.tex" and a histogram "IHDP_coverage_NN.pdf," which correspond to Table 1 Panel (a) and Figure 2 panel (a) in the paper. It also produces the ablation study results in Table 3.
2. ForestRiesz_IHDP.ipynb runs the ATE MAE and coverage experiments based on IHDP semi-synthetic data using ForestRiesz. It takes the files in "data/IHDP" as inputs, and outputs in "results/IHDP/ForestRiesz." It produces a table "IHDP_MAE_RF.tex" and a histogram "IHDP_coverage_RF.pdf," which correspond to Table 1 Panel (b) and Figure 2 panel (b) in the paper.
3. RieszNet_BHP.ipynb runs the average derivative experiments based on BHP semi-synthetic data using RieszNet. It takes the file in "data/BHP" as input, and outputs in "results/BHP/RieszNet." It produces a table collecting all results "res_avg_der_NN.tex" and, for each design, a histogram "all.pdf" in the corresponding sub-folder. Outputs "res_avg_der_NN.tex" and "true_f_compl_nonlin_conf/all.pdf" correspond to Table A1 and Figure 3 panel (a) of the paper, respectively. Table 2 panel (a) is the last row of Table A1.
4. ForestRiesz_BHP.ipynb runs the average derivative experiments based on BHP semi-synthetic data using ForestRiesz. It takes the file in "data/BHP" as input, and outputs in "results/BHP/ForestRiesz." It produces a table collecting all results "res_avg_der_RF.tex" and, for each design, "(method)_all.pdf" in the corresponding sub-folder, where (method) is a string detailing the type of cross-fitting (0 = none, 1 = simple, 2 = double) and whether multitasking is used (0 = No, 1 = Yes). Outputs "res_avg_der_NN.tex" and "true_f_compl_nonlin_conf/all.pdf" correspond to Table A2 and Figure 3 panel (b) of the paper, respectively. Table 2 panel (b) is the last row of Table A2, and the ablation results in Table 4 are the last five rows of Table A2. This file also produces the RF Plugin benchmark results in Table A3, which are stored in "plugin.tex".

Utils Folder

1. riesznet.py contains the main class for RieszNet.
2. forestriesz.py contains the main class for ForestRiesz.
3. moments.py defines some moment functions to use with RieszNet (currently only ATE and avg_small_diff are used).
4. NN_avgmom_sim defines functions for experiments with semi-synthetic data that are used in RieszNet_BHP.ipynb.
5. RF_avgmom_sim defines functions for experiments with semi-synthetic data that are used in ForestRiesz_BHP.ipynb.
6. ihdp_data.py contains utils to load and format the IHDP data, and it is largely drawn from Shi et el. (2019)'s replication code.

Data Folder

1. IHDP has two subfolders: sim_data, which contains 1000 semi-synthetic datasets based on IHDP that were generated using Dorie (2016)'s NPCI R Package under setting "A", and sim_data_redraw_T, hich contains 100 semi-synthetic datasets based on IHDP, also generated using NPCI under setting "A", but redrawing the treatment variable according to the propensity score setting "True". These are used for the ATE MAE and coverage experiments, respectively.
2. BHP contains the gasoline demand data from Blundell et al. (2017)'s replication files, converted into csv. These are used to generate semi-synthetic data for the average derivative experiments.