EconML/prototypes/dynamic_dml

Introduction

Code for recreating figures and numbers from paper: "Double/Debiased Machine Learning for Dynamic Treatment Effects"

Generating paper figures

To generate the paper figures run

./all_coverage.sh

This will produce a set of .png files that contain distributions of point estimates, distribution of standard error estimates, coverage probabilities of confidence intervals (for both the estimated parameters and for a set of target counterfactual policies).

The results for constant effects will be in the newly created folder:

results/long_range_constant

and the results for heterogeneous effects in the folder:

results/long_range_hetero

The code assumes Python 3 and requires the standard packages: statsmodels, numpy, scipy, scikit-learn, matplotlib. All can be pip installed.

To create the figure that contains the benchmark performance run the jupyter notebook: high_dim_state_any_m_panel.ipynb.

The jupyter notebook high_dim_state_any_panel_hetero.ipynb compares performance with benchmarks when there is effect heterogeneity.

Files

Estimator Classes

• panel_dynamic_dml.py : Contains the estimator DynamicPanelDML that estimates dynamic treatment effects without heterogeneity.
• hetero_panel_dynamic_dml.py: Contains the estimator HeteroDynamicPanelDML that estimates heterogeneous dynamic treatment effects.

Expository Notebooks

• high_dim_state_any_m_panel.ipynb: examples of runnign the estimators with constant effects
• high_dim_state_any_m_panel_hetero.ipynb: examples of running the estiamtors with heterogeneous effects

Data Generating Processes

• dynamic_panel_dgp.py: Contains several data generating processes for dynamic treatment effect estimation

Coverage Experiments

• coverage_panel.py: Runs coverage experiments for contant dynamic effects
• coverage_panel_hetero.py: Runs coverage experiments for heterogeneous dynamic effects
• all_coverage.sh: Shell script that runs all coverage experiments
• postprocess_panel.ipynb: Post-processing notebook that reads the coverage results for constant effects so as to produce more plots later on
• postprocess_panel_hetero.ipynb: Post-processing notebook that reads the coverage results for heterogeneous effects so as to produce more plots later on