open-econs 1.4.3

Empirical economics and causal inference in Python - a scikit-learn-style unified API with Stata/R-grade numerical parity.

open-econs

open-econs is a Python library for empirical economics and causal inference that reproduces Stata and R results to a verified numerical tolerance — built for researchers migrating off Stata/R and for Python users who need causal inference, it is the econometrics toolkit a Stata/R researcher would reach for, but native in Python.

Unlike statsmodels, linearmodels, or fixest, open-econs is validated against Stata and R across the whole causal-inference stack: 550+ parity tests (330+ vs Stata, 220+ vs R) run in CI on every release, and a numerical mismatch fails the build before it ships. New methods are checked to ≤1e-6; IV, Arellano-Bond, and synthetic control reproduce reference results to machine precision.

It covers 40+ estimators in one consistent API: OLS, fixed effects, IV/2SLS, GMM & Arellano-Bond, logit/probit/mlogit, Oaxaca-Blinder, nonlinear least squares, the full difference-in-differences family (Callaway-Sant'Anna, Sun-Abraham, Gardner DID2S, event studies), regression discontinuity, propensity-score & coarsened-exact matching, synthetic control with permutation inference, and a time-series module (ARIMA, VAR/VECM, GARCH, ARDL/UECM with the Pesaran-Shin-Smith bounds test, unit-root & cointegration).

Install

pip install open-econs                      # core estimators
pip install open-econs[plot]                # + matplotlib for .plot()
pip install open-econs[nls]                 # + sympy for nls()
pip install open-econs[dev,lint]            # + development & linting tools

Requires Python ≥ 3.10.

Quick start

import open_econs as oe
import pandas as pd

df = pd.DataFrame({
    "income":    [30, 45, 55, 70, 85, 40, 60, 95],
    "education": [10, 12, 14, 16, 18, 11, 15, 20],
    "age":       [25, 30, 35, 40, 45, 28, 38, 50],
    "female":    [0,  0,  0,  0,  1,  1,  1,  1],
    "province":  ["A","A","B","B","C","C","A","B"],
})

r = oe.ols("income ~ education + age", data=df, cluster="province")
print(r.tidy())        # coefficient table (named pd.DataFrame)
print(r.summary())     # full OLS results

# Causal estimators use the same API:
oe.did_cs("y ~ x1 + x2 | group + time", data=panel_df)   # Callaway-Sant'Anna DiD
oe.psm("treat ~ x1 + x2", data=df)                        # propensity-score matching
oe.synth(...)                                            # synthetic control

Every result returns named pd.Series / pd.DataFrame outputs and exports to JSON, CSV, LaTeX, or HTML with one call — .export(), .to_latex(), .to_html(). Results are immutable after estimation.

Applied-micro example (limited dependent variables)

open-econs reproduces the robust and cluster standard errors that applied microeconometricians rely on daily — validated to ≤1e-6 against Stata:

import open_econs as oe

# Poisson PPML — non-clustered robust SE matches Stata ppmlhdfe's
# Correia-Guimaraes-Zylkin adjustment; cluster-robust matches to <=1e-6.
r = oe.poisson("y ~ x1 + x2", data=df, vcov_backend="stata")
r = oe.poisson("y ~ x1 + x2", data=df, vcov_backend="stata", cluster="id")

# Tobit with robust SEs (matches Stata tobit, vce(robust))
r = oe.tobit("y ~ x1 + x2", data=df, cov_type="HC1")

# Ordered logit with heteroskedasticity-robust SEs (matches Stata ologit, vce(robust))
r = oe.ologit("y_ordered ~ x1 + x2", data=df, cov_type="HC1")

Stata / R → open-econs (top mappings)

Stata / R	open-econs
regress (Stata)	oe.ols()
xtreg, fe (Stata)	oe.fe()
ivregress 2sls / AER::ivreg	oe.iv()
xtabond2 (Stata)	oe.abond()
csdid (Stata) / did (R)	oe.did_cs()
eventstudyinteract (Stata) / fixest::sunab (R)	oe.did_sa()
did2s (Stata) / fixest::did2s (R)	oe.did_gardner()
rdrobust (Stata)	oe.rdd()
teffects psmatch (Stata) / MatchIt (R)	oe.psm()
synth_runner (Stata) / synth (R)	oe.synth() / oe.placebo_space() / oe.placebo_time()

Full mapping: docs/stata-r-mapping.md. Migration guides: Stata / R.

Parity coverage (validated to ≤1e-6)

Every estimator below is checked against its Stata and/or R reference in CI; a numerical mismatch fails the build. The limited-DV family's robust and cluster standard errors were hardened this release to match Stata exactly.

Family	Stata anchor	R anchor	Robust / cluster SE
OLS / WLS	reg, newey	lm	✓ robust, cluster, HC0–HC3, HAC
Fixed effects (2-way)	xtreg, fe	plm	✓
IV / 2SLS	ivregress 2sls	AER::ivreg	✓ robust, cluster, HC0–HC3 (HAC = R-reference)
Arellano-Bond (Diff + System GMM)	xtabond2	—	✓ AR(1)/AR(2), Windmeijer 2-step
Poisson (PPML)	ppmlhdfe	fixest::fepois	✓ cluster and non-clustered robust (CGZ)
Tobit (censored)	tobit	AER::tobit	✓ OIM, robust, cluster
Ordered logit / probit	ologit / oprobit	MASS::polr	✓ OIM and robust (HC1)
Negative binomial	nbreg	glm.nb / fenegbin	✓ cluster; Stata constant dispersion toggle
Logit / probit / mlogit	logit / probit / mlogit	glm / nnet	✓
DiD (CS2021, Gardner, Sun-Abr, ES)	csdid / did2s / eventstudyinteract	did / fixest	✓
RDD	rdrobust	rdrobust	✓
PSM / CEM	teffects psmatch / cem	MatchIt / cem	✓
Synthetic control	synth_runner	synth	✓ + permutation inference
ARDL/UECM, VAR, GARCH, ARIMA, unit-root	ardl / var / arch	ARDL / vars / rugarch	✓ (TS CV vintage labelled)
Robust regression (M/MM)	rreg	MASS::rlm	✓ bisquare M-estimator

Performance

v1.0.3 hardens hot loops with bit-identical vectorization/parallelization — no parity tolerance loosened: psm is ~4× faster (k-NN + variance loops batched), _hac_S Newey-West accumulation is a single einsum, and did_cs bootstrap runs through an opt-in parallel= pool. GPU offload was evaluated and deliberately declined (BLAS is already CPU-threaded). Full detail: docs/performance.md.

Compare

open-econs is the only Python library covering the full causal-inference stack and enforcing Stata/R parity. The side-by-side matrix against statsmodels, linearmodels, and fixest is here: docs/comparison.md.

Documentation

• Methodology — per-estimator math, assumptions, inference, and Stata/R equivalents.
• Tutorials — OLS, FE, IV, DiD, RDD, PSM, synthetic control walkthroughs.
• Roadmap · Changelog · v1.1.0 release notes · v1.0.3 release notes

Development

pip install -e ".[dev]"
python -m pytest tests/

License

MIT