qte — Quantile Treatment Effects in Python

Alpha software: the API may change without notice between releases.

This is a Python implementation of some of the estimators in the qte R package by Brantly Callaway (link). See the project website for the API reference and benchmarks.

The main features are:

• Cross-sectional quantile treatment effects estimators (unadjusted, inverse probability weighted, outcome regression, doubly robust) and a nonlinear difference-in-differences estimator (changes-in-changes and quantile difference-in-differences).
• Fast: py-qte uses more efficient algorithms and parallelizes the bootstrap standard errors effectively. On larger datasets it is more than 20× faster than qte for the AIPW estimator of the quantile treatment effect (see the benchmark timings).
• Beautiful: Graphs and tables for the console, the web, and LaTeX powered by Altair, Great Tables and Rich.

Installation

Requires Python 3.12 or newer.

uv add py-qte # or pip install py-qte

On Linux and macOS the Fortran solver for quantile regression should work and be used in the computations. On Windows a fallback (based on the statsmodels library) will be used.

Examples

Cross-Sectional Data

We can estimate quantile treatment (and average) treatment effects using an augmented inverse probability weighting (AIPW) estimator.

from qte.cross_sectional import estimate_aipw_effects
from qte.datasets import load_lalonde

ds = load_lalonde()

res = estimate_aipw_effects(
    ds=ds,
    outcome="re78",
    treatment="treat",
    outcome_regression_formula="age + education",
    propensity_score_formula="age + education",
)

res.plot()  # Vega-Altair plot (see below)
res.tabulate()  # Great Tables output (see below)

Nonlinear Difference-in-Differences

We can estimate quantile treatment (and average) treatment effects with the changes-in-changes estimator.

from qte.nonlinear_did import (
    CounterfactualModel,
    TrtGroupConfig,
    estimate_nonlinear_did_for_panel,
)
from qte.datasets import load_mpdta

ds = load_mpdta()
res = estimate_nonlinear_did_for_panel(
    ds,
    "lemp",
    TrtGroupConfig("first.treat", 0),  # never-treated group is 0
    "year",
    "countyreal",
    qs=[0.25, 0.5, 0.75],
    counterfactual_model=CounterfactualModel.CIC,
)

Development

Development is somewhat complicated due to the inclusion of the Fortran code for solving quantile regression. A working way to set up the project for development is:

git clone git@github.com:mohelm/py-qte.git
cd py-qte
uv sync --all-groups --no-install-project && uv sync --all-groups

Website

See docs/README.md for how to build and preview the Quarto website and regenerate the API reference.

Regenerating the README assets

Regenerate the figures above after changing the estimators or their presentation:

uv run python -m docs.readme

README_PYPI.md (what project.readme points at) is generated from this file with relative links made absolute, since PyPI can't resolve repo-relative assets. Regenerate it after editing this file:

uv run python -m docs.readme --pypi-readme-only

Releasing

See RELEASING.md for how to cut a release.