dynamicDML

A Python implementation of dynamic Double Machine Learning (DML) as developed in Bodory, Huber & Laffers (2022) and Bradic, Ji & Zhang (2024). The dynamicDML package allows to flexibly estimate counterfactual outcomes and treatment effects of sequential policies from observational data, where treatment assignment may dynamically depend on time-varying characteristics. For a detailed overview of these methods, see Muny (2025).

Installation

To install the dynamicDML package run

pip install dynamicDML

in the terminal. dynamicDML requires the following dependencies:

• flaml[automl]>=2.3.3
• matplotlib>=3.10.0
• mgzip>=0.2.1
• numpy>=2.2.3
• pandas>=2.2.3
• scikit-learn>=1.6.1
• scipy>=1.15.2
• seaborn>=0.13.2

The implementation has been developed and tested in Python version 3.12.

Examples

The following examples demonstrate the basic usage of the dynamicDML package with default settings.

# load packages
import dynamicDML
import numpy as np
from sklearn.linear_model import LinearRegression, LogisticRegression

# Seed
seed = 999
# Generate data
data = dynamicDML.dyn_data_example(n=2000, random_state=seed)

# Define counterfactual contrasts of interest
all_treat = np.ones_like(data['D1'])
all_control = np.zeros_like(data['D1'])

# Basic setting with Linear and Logistic regression
model = dynamicDML.dml2periods(dynamic_confounding=True, random_state=seed)

# APO treat-treat
model = model.init_sequence(
    d1treat='treat',
    d2treat='treat',
    MLmethod_p1=LogisticRegression(),
    MLmethod_p2=LogisticRegression(),
    MLmethod_mu=LinearRegression(),
    MLmethod_nu=LinearRegression()
    )
model = model.fit_sequence(
    'treat', 'treat', data['Y'], data['D1'], data['D2'], data['X0'],
    data['X1'], g1t=all_treat, g2t=all_treat)
model.sequence_summary()
model = model.compute_APO(d1treat='treat', d2treat='treat')

# APO control-control
model = model.init_sequence(
    d1treat='control',
    d2treat='control',
    MLmethod_p1=LogisticRegression(),
    MLmethod_p2=LogisticRegression(),
    MLmethod_mu=LinearRegression(),
    MLmethod_nu=LinearRegression()
    )
model = model.fit_sequence(
    'control', 'control', data['Y'], data['D1'], data['D2'], data['X0'],
    data['X1'], g1t=all_control, g2t=all_control)
model.sequence_summary()
model = model.compute_APO(d1treat='control', d2treat='control')

# ATE treat-treat vs. control-control
model = model.compute_ATE(
    d1treat='treat', d2treat='treat', d1control='control', d2control='control')

# GATE treat-treat vs. control-control for first covariate
model = model.compute_GATEmATE(
    d1treat='treat', d2treat='treat', d1control='control', d2control='control',
    groupvar=(data['X1'][:, 0] > 0), name_groupvar='X1')

Release Notes

• Version 0.1.0: Unpublished
• Version 0.2.0: Initial release of dynamicDML python package

References

• Bodory, H., Huber, M., & Laffers, L. (2022). Evaluating (weighted) dynamic treatment effects by double machine learning. The Econometrics Journal, 25(3), 648. [1]
• Bradic, J., Ji, W., & Zhang, Y. (2024). High-dimensional inference for dynamic treatment effects. The Annals of Statistics, 52(2), 415-440. [2]
• Muny, F. (2025). Evaluating Program Sequences with Double Machine Learning: An Application to Labor Market Policies. arXiv preprint arXiv:2506.11960. [3]