did_multiplegt_stat

Python implementation of the did_multiplegt_stat Stata package by de Chaisemartin, D'Haultfœuille, Pasquier, Sow, and Vazquez-Bare (2024) — heterogeneity-robust difference-in-differences estimators with stayers for binary, discrete, or continuous treatments (and instruments).

The package estimates the Average Slope (AS), Weighted Average Slope (WAS), and IV-WAS parameters in static designs where parallel trends is assumed conditional on the baseline treatment.

Installation

pip install did-multiplegt-stat

Optional extras:

pip install "did-multiplegt-stat[linearmodels]"   # alternative IV backend
pip install "did-multiplegt-stat[docs]"           # build docs locally
pip install "did-multiplegt-stat[dev]"            # tests + tooling

Quick start: default scikit-learn regressions

By default, DIDMultiplegtStat uses scikit-learn's LinearRegression for the outcome-change nuisance function and LogisticRegression for the stayer probability nuisance functions. This is equivalent to setting asinstata=False. The option is written explicitly below so that the backend used by each example is unambiguous.

Example 1: AS and WAS of gasoline taxes on log consumption

Original Stata command:

did_multiplegt_stat lngca id year tau, or(1) estimator(as was) placebo(3) as_vs_was

Python with the default scikit-learn OLS and logit regressions:

# Uncomment this line when running the example in a Jupyter notebook:
# %pip install pandas statsmodels scikit-learn did-multiplegt-stat

import pandas as pd
from did_multiplegt_stat import DIDMultiplegtStat

data_url = (
    "https://raw.githubusercontent.com/Credible-Answers/"
    "py_did_multiplegt_stat/main/tests/data/gazoline_did_multiplegt_stat.dta"
)
df = pd.read_stata(data_url)

model = DIDMultiplegtStat(
    estimator=["as", "was"],
    order=1,
    placebo=3,
    as_vs_was=True,
    asinstata=False,  # scikit-learn LinearRegression and LogisticRegression
)
model.fit(df, Y="lngca", ID="id", Time="year", D="tau")
model.summary()
model.plot()

Example 2: AS and WAS with no extrapolation

Original Stata command:

did_multiplegt_stat lngpinc id year tau, or(1) estimator(as was) estimation_method(dr) placebo(3) noextra as_vs_was

Python with the default scikit-learn OLS and logit regressions:

# Uncomment this line when running the example in a Jupyter notebook:
# %pip install pandas statsmodels scikit-learn did-multiplegt-stat

import pandas as pd
from did_multiplegt_stat import DIDMultiplegtStat

data_url = (
    "https://raw.githubusercontent.com/Credible-Answers/"
    "py_did_multiplegt_stat/main/tests/data/gazoline_did_multiplegt_stat.dta"
)
df = pd.read_stata(data_url)

model = DIDMultiplegtStat(
    estimator=["as", "was"],
    order=1,
    placebo=3,
    noextrapolation=True,  # Stata: noextra
    as_vs_was=True,
    asinstata=False,  # scikit-learn LinearRegression and LogisticRegression
)
model.fit(df, Y="lngpinc", ID="id", Time="year", D="tau")
model.summary()
model.plot()

Reproducing Stata's OLS and logit results

Set asinstata=True to replace the default scikit-learn nuisance regressions with the Stata-faithful implementations: statsmodels OLS and the package's Newton-Raphson logit matching Stata's logit, asis behavior. The following cell runs both examples with the Stata-faithful backend:

# Uncomment this line when running the example in a Jupyter notebook:
# %pip install pandas statsmodels scikit-learn did-multiplegt-stat

import pandas as pd
from did_multiplegt_stat import DIDMultiplegtStat

data_url = (
    "https://raw.githubusercontent.com/Credible-Answers/"
    "py_did_multiplegt_stat/main/tests/data/gazoline_did_multiplegt_stat.dta"
)
df = pd.read_stata(data_url)

# Stata-faithful version of Example 1
stata_model_1 = DIDMultiplegtStat(
    estimator=["as", "was"],
    order=1,
    placebo=3,
    as_vs_was=True,
    asinstata=True,
)
stata_model_1.fit(df, Y="lngca", ID="id", Time="year", D="tau")
stata_model_1.summary()
stata_model_1.plot()

# Stata-faithful version of Example 2
stata_model_2 = DIDMultiplegtStat(
    estimator=["as", "was"],
    order=1,
    placebo=3,
    noextrapolation=True,
    as_vs_was=True,
    asinstata=True,
)
stata_model_2.fit(df, Y="lngpinc", ID="id", Time="year", D="tau")
stata_model_2.summary()
stata_model_2.plot()

Using machine learning to estimate the nuisance functions

DIDMultiplegtStat accepts custom scikit-learn-style estimators for its nuisance functions. The examples below use a random forest regressor for the outcome-change model and a random forest classifier for the stayer model. Supplying these models overrides the built-in OLS and logit regressions, regardless of the value of asinstata:

# Uncomment this line when running the example in a Jupyter notebook:
# %pip install pandas statsmodels scikit-learn did-multiplegt-stat

import pandas as pd
from sklearn.ensemble import RandomForestClassifier, RandomForestRegressor

from did_multiplegt_stat import DIDMultiplegtStat

data_url = (
    "https://raw.githubusercontent.com/Credible-Answers/"
    "py_did_multiplegt_stat/main/tests/data/gazoline_did_multiplegt_stat.dta"
)
df = pd.read_stata(data_url)

# Random-forest version of Example 1
ml_model_1 = DIDMultiplegtStat(
    estimator=["as", "was"],
    order=1,
    placebo=3,
    as_vs_was=True,
    model_deltay=RandomForestRegressor(
        n_estimators=100,
        random_state=42,
    ),
    model_stayer=RandomForestClassifier(
        n_estimators=100,
        random_state=42,
    ),
)
ml_model_1.fit(df, Y="lngca", ID="id", Time="year", D="tau")
ml_model_1.summary()
ml_model_1.plot()

# Random-forest version of Example 2
ml_model_2 = DIDMultiplegtStat(
    estimator=["as", "was"],
    order=1,
    placebo=3,
    noextrapolation=True,
    as_vs_was=True,
    model_deltay=RandomForestRegressor(
        n_estimators=100,
        random_state=42,
    ),
    model_stayer=RandomForestClassifier(
        n_estimators=100,
        random_state=42,
    ),
)
ml_model_2.fit(df, Y="lngpinc", ID="id", Time="year", D="tau")
ml_model_2.summary()
ml_model_2.plot()

Any compatible estimators may be supplied: model_deltay must implement fit and predict, while model_stayer must implement fit and predict_proba.

Backends: asinstata

Two regression backends are bundled:

Stata parity uses statsmodels OLS + a from-scratch Newton-Raphson logit that matches Stata's logit, asis defaults; results agree to ~1e-7 relative error.

Main options

The Python API follows the terminology used in the paper and Stata package:

• estimator (as / was / iv-was)
• order (scalar, 4-tuple, or 8-tuple for IV)
• placebo(N) (multi-period placebos)
• exact_match, noextrapolation
• switchers (up / down)
• as_vs_was
• by, by_fd, by_baseline
• controls, weight, cluster
• other_treatments
• cross_fitting, trimming, on_placebo_sample
• bootstrap + seed
• twfe (with same_sample, full_sample, percentile)
• cross_validation (k-fold CV for polynomial order)

The package uses doubly robust estimation by default. When exact_match=True, it uses regression adjustment internally; users do not select RA or PS as a separate estimation method.

See the full Python documentation for the help-file style reference.

Citation

If you use this software in academic work, please cite the underlying paper:

de Chaisemartin, C., D'Haultfœuille, X., Pasquier, F., Sow, D., Vazquez-Bare, G. (2024). Difference-in-Differences for Continuous Treatments and Instruments with Stayers. arXiv:2201.06898.

A CITATION.cff is bundled for tooling integration.

License

GPL-3.0-or-later — see the license.

Authors

Originally authored by the team behind the Stata package:

• Clément de Chaisemartin (Sciences Po)
• Diego Ciccia (Sciences Po)
• Xavier D'Haultfœuille (CREST-ENSAE)
• Felix Knau (Sciences Po)
• Felix Pasquier (CREST-ENSAE)
• Doulo Sow (Sciences Po)
• Gonzalo Vazquez-Bare (UCSB)

Python port: Anzony Quispe.

Contact: chaisemartin.packages@gmail.com