rdhte (Python)

Heterogeneous treatment effects in regression discontinuity (RD) designs. Point estimates, robust bias-corrected confidence intervals, and heteroskedastic or cluster-robust standard errors for conditional RD treatment effects.

The theoretical framework is developed in Calonico, Cattaneo, Farrell, Palomba, and Titiunik (2025). arXiv

Install (development)

cd Python/rdhte
pip install -e ".[plot,test]"

Hard dependencies: numpy, scipy, pandas, statsmodels, rdrobust. Plotting (rdhte.plot.plot) additionally requires plotnine.

Quick start

import numpy as np
from rdhte import rdhte, rdbwhte, rdhte_lincom

rng = np.random.default_rng(0)
n = 5000
x = rng.uniform(-1.0, 1.0, n)
W = np.where(rng.random(n) < 0.5, "A", "B")              # categorical W
y = ((x >= 0) * (1.0 + 0.5 * (W == "B"))                 # CATE = 1.0 (A), 1.5 (B)
     + 0.3 * x + rng.normal(0, 0.3, n))

# Per-cell CATEs with rdrobust-style bandwidth selection (default: mserd).
m = rdhte(y=y, x=x, c=0.0, covs_hte=W)
print(m)                 # short header
print(m.summary())       # full table with Nh/h columns
print(m.tidy())          # broom-style DataFrame: term, estimate, std.error, ...
print(m.glance())        # one-row metadata DataFrame

# Data-driven bandwidth per cell, without running the full estimator.
bw = rdbwhte(y=y, x=x, c=0.0, covs_hte=W)
print(bw.h)              # (n_lev, 2) array of left/right bandwidths

# Test H0: CATE(A) == CATE(B), and separately H0: CATE(A) == 0.
res = rdhte_lincom(m, linfct=["A - B = 0", "A = 0"])
print(res["individual"])
print(res["joint"])

Empirical illustration

The package includes a full empirical illustration using the bundled Granzier, Pons, and Tricaud election dataset. From the Python folder:

python rdhte_illustration.py

Command Reference

The companion commands are:

Main estimation options:

bwselect accepts mserd, msetwo, msesum, msecomb1, msecomb2, cerrd, certwo, cersum, cercomb1, and cercomb2. MSE denotes mean square error; CER denotes coverage error rate.

Return Objects

rdhte() returns an RdhteResult with fields for estimates, inference results, bandwidths, labels, and metadata: Estimate, Estimate_bc, coef, coef_bc, se_rb, ci_rb, t_rb, pv_rb, vcov, W_lev, W_names, kernel, bwselect, vce, vce_select, c, h, p, q, N, Nh, level, and rdmodel. Convenience methods include summary(), tidy(), and glance().

rdbwhte() returns an RdbwhteResult with bandwidths h, effective sample sizes Nh, group labels, and the same metadata. rdhte_lincom() returns a dictionary with individual and joint test-result DataFrames.

Multiple heterogeneity variables

For factor-by-factor interactions and formula expressions, supply covs_hte as a list of arrays / DataFrame / formula string:

import pandas as pd
df = pd.DataFrame({"y": y, "x": x, "W1": W1, "W2": W2})

# Cartesian-product factor cells.
m_2d = rdhte(y="y", x="x", c=0.0, covs_hte=[df["W1"], df["W2"]], data=df)

# Formula syntax (linear-in-Wk CATE per term).
m_form = rdhte(y="y", x="x", c=0.0, covs_hte="W1 * W2", data=df)
# Coefs: T, T:W1, T:W2, T:W1.W2.

# factor() is accepted as an alias for patsy C().
m_mix = rdhte(y="y", x="x", c=0.0, covs_hte="factor(W1) * W2", data=df)

Supported Features

vce accepts: nn, hc0, hc1, hc2, hc3, cr1, cr2, cr3. When cluster= is supplied the default switches to cr1; passing an hc* value with a cluster issues a warning and upgrades to the matching cr* form.

References

• Calonico, Cattaneo, Farrell, Palomba, and Titiunik (2025a): Treatment Effect Heterogeneity in Regression Discontinuity Designs.
• Calonico, Cattaneo, Farrell, Palomba, and Titiunik (2025b): rdhte: Conditional Average Treatment Effects in RD Designs.
• Cattaneo and Titiunik (2022): Regression Discontinuity Designs.
• Calonico, Cattaneo, and Farrell (2020): Optimal Bandwidth Choice for Robust Bias Corrected Inference in Regression Discontinuity Designs.
• Calonico, Cattaneo, Farrell, and Titiunik (2019): Regression Discontinuity Designs using Covariates.
• Calonico, Cattaneo, and Titiunik (2014): Robust Nonparametric Confidence Intervals for Regression-Discontinuity Designs.
• Granzier, Pons, and Tricaud (2023): Coordination and Bandwagon Effects: How Past Rankings Shape the Behavior of Voters and Candidates.

Related RD software is available at https://rdpackages.github.io/.

Authors

Sebastian Calonico, Matias D. Cattaneo, Max H. Farrell, Filippo Palomba, Rocio Titiunik.

License

GPL-3; see LICENSE in this package and LICENSE.md at the repository root.