did-had 0.2.0

Difference-in-Differences with Heterogeneous Adoption Design (HAD) estimator

did-had

A Python implementation of the Heterogeneous Adoption Design (HAD) estimator for difference-in-differences analysis from de Chaisemartin et al. (2025).

Overview

The did-had package implements the HAD estimator for settings where:

• All groups receive treatment but with different intensities (no pure control group)
• Treatment adoption occurs at the same time for all groups, but doses vary
• Some groups have treatment doses close to zero, serving as "quasi-untreated" groups

This is a Python port of the official Stata package did_had, producing numerically identical results.

Installation

pip install did-had

For plotting support:

pip install did-had[plot]

Quick Start

import pandas as pd
from did_had import DidHad

# Load your panel data
df = pd.read_stata("tutorial_data.dta")

# Create and fit the model
model = DidHad(kernel="tri")
results = model.fit(
    df=df,
    outcome="y",        # outcome variable
    group="g",          # group identifier
    time="t",           # time period
    treatment="d",      # treatment dose
    effects=5,          # post-treatment periods
    placebo=4           # pre-treatment placebos
)

# View results
print(results)

# Get ATT
print(f"Average Treatment Effect: {results.att():.4f}")

# Save results
model.save_results("results.csv", format="csv")

Example Output

===========================================================================
DID-HAD Estimation Results
===========================================================================
Number of groups: 1,000
Number of periods: 10
Adoption period (F): 6.0
Kernel: tri
Confidence level: 95%

---------------------------------------------------------------------------
                          Effect Estimates                      QUG* Test
         --------------------------------------------------- ---------------
          Estimate       SE     LB.CI     UB.CI     N      BW    N.BW        T    p.val
Effect_1   4.28198  0.55814   2.71751   4.90538 1,000 0.36133    371  3.96182  0.20154
Effect_2   3.59260  0.66675   2.02563   4.63925 1,000 0.27104    282  3.96182  0.20154
Effect_3   4.25466  0.67544   2.71354   5.36123 1,000 0.31407    324  3.96182  0.20154
...

Features

• Exact replication of Stata's did_had command
• Bias-corrected inference using local polynomial regression (lprobust-style)
• Quasi-untreated group tests to validate the estimation strategy
• Event-study plots for visualization
• Multiple output formats: CSV, Stata, Excel, pickle

API Reference

DidHad Class

DidHad(kernel="epa", alpha=0.05, nnmatch=3)

Parameters:

• kernel: Kernel function - "epa" (Epanechnikov), "tri" (triangular), "uni" (uniform), "gau" (Gaussian)
• alpha: Significance level for confidence intervals (default: 0.05 for 95% CI)
• nnmatch: Number of nearest neighbors for variance estimation

fit() Method

results = model.fit(
    df,                      # Panel DataFrame
    outcome,                 # Outcome variable name
    group,                   # Group identifier name
    time,                    # Time period name
    treatment,               # Treatment dose name
    effects=1,               # Number of post-treatment periods
    placebo=0,               # Number of pre-treatment placebos
    dynamic=False,           # Use cumulative treatment dose
    bandwidth=None,          # Global bandwidth (or Silverman if None)
    bandwidth_effect=None,   # Effect-specific bandwidths (dict or scalar)
    bandwidth_placebo=None   # Placebo-specific bandwidths (dict or scalar)
)

DidHadResults Object

results.summary()           # Formatted summary string
results.to_dataframe()      # Full results as DataFrame
results.att()               # Average treatment effect on treated
results.effects             # DataFrame of effect estimates
results.placebos            # DataFrame of placebo estimates

Plotting

# Basic plot
model.plot()

# Customized plot
model.plot(
    figsize=(10, 6),
    title="My Event Study",
    xlabel="Periods since treatment",
    ylabel="Treatment Effect",
    show_ci=True
)

Saving Results

model.save_results("results.csv", format="csv")
model.save_results("results.dta", format="stata")
model.save_results("results.xlsx", format="excel")
model.save_results("results.pkl", format="pickle")

Comparison with Stata

This package produces numerically identical results to the official Stata did_had command when using the same bandwidths:

Estimate	Python	Stata	Match
Effect_1	4.28198	4.28198	Yes
Effect_2	3.59260	3.59260	Yes
Effect_3	4.25466	4.25466	Yes
...	...	...	...

Requirements

• Python >= 3.8
• NumPy >= 1.20.0
• Pandas >= 1.3.0
• Matplotlib >= 3.4.0 (optional, for plotting)

References

• de Chaisemartin, C., D'Haultfoeuille, X., Pasquier, F., & Vazquez-Bare, G. (2025). "Difference-in-Differences Estimators for Treatments Continuously Distributed at Every Period". SSRN

• Calonico, S., Cattaneo, M. D., & Farrell, M. H. (2019). "nprobust: Nonparametric Kernel-Based Estimation and Robust Bias-Corrected Inference". Journal of Statistical Software.

License

MIT License. See LICENSE for details.

Contributing

Contributions are welcome! Please open an issue or submit a pull request on GitHub.