rd2d 1.0.0

Local polynomial methods for boundary discontinuity designs

rd2d

rd2d provides local polynomial estimation, robust bias-corrected inference, and bandwidth helpers for boundary discontinuity designs with bivariate running variables. The package includes location-based and distance-based methods, sharp and fuzzy designs, pointwise confidence intervals, covariance-backed summary inference, uniform confidence bands, and aggregate boundary effect summaries. The default fitting method is the joint treatment-interacted regression; use fitmethod="separate" to reproduce the earlier two-sample fitting path.

from rd2d import rd2d, rdbw2d, rd2d_dist, rdbw2d_dist, summary

Main Functions

• rd2d(): location-based estimation and inference.
• rdbw2d(): location-based bandwidth selection.
• rd2d_dist() and rd2d_distance(): distance-based estimation and inference.
• rdbw2d_dist() and rdbw2d_distance(): distance-based bandwidth selection.
• summary(): summary tables with optional uniform bands, WBATE, and LBATE.

In the source repository, the sibling scripts ../rd2d_illustration.py and ../rd2d_plot.py read ../rd2d_data.csv and run the same dataset-based illustration workflow as the R and Stata examples. They do not create an output/ folder or save fitted objects, tables, or plot files.

Installation

python -m pip install rd2d

For local development:

python -m pip install -e .

Optional plotting and testing dependencies can be installed with:

python -m pip install -e ".[plots,test]"

Basic Usage

import numpy as np
from rd2d import rd2d

rng = np.random.default_rng(123)
n = 800
x1 = rng.normal(size=n)
x2 = rng.normal(size=n)
assignment = (x1 >= 0).astype(float)
y = 3 + 2 * x1 + 1.5 * x2 + assignment + rng.normal(size=n)
x = np.column_stack([x1, x2])
b = np.array([[0.0, 0.0], [0.0, 1.0]])

fit = rd2d(y, x, assignment, b, h=0.9, params_cov="main")
fit.main
fit.summary(cbands="main", repp=999).tables["main"]

Pre-intervention covariates can be supplied for efficiency adjustment. The same covariate coefficient is used across treatment sides, matching the R package.

z = np.column_stack([x1 + rng.normal(size=n), x2 + rng.normal(size=n)])
fit_cov = rd2d(y, x, assignment, b, h=0.9, covs_eff=z, fitmethod="joint")

For distance-based designs, pass one signed-distance column per evaluation point. Nonnegative distances identify observations on the treated side.

from rd2d import rd2d_dist

distance = x1.reshape(-1, 1)
fit_dist = rd2d_dist(y, distance, h=0.5, b=np.array([[0.0, 0.0]]))
fit_dist.main

Development

From this directory:

python -m pytest

Publishing From GitHub

The workflow .github/workflows/python-publish.yml builds, checks, tests, and publishes the Python package to PyPI with trusted publishing. Configure:

• PyPI project: rd2d.
• Trusted publisher owner: rdpackages.
• Trusted publisher repository: rd2d.
• Trusted publisher workflow: python-publish.yml.
• Trusted publisher environment: pypi.

In GitHub, create an environment named pypi. Add required reviewers there if you want each PyPI upload to require manual approval. Publish by creating a GitHub Release or by running the Publish Python package workflow manually.