Design-based estimation

The estimator layer delegates regression parsing and fitting to PyFixest and Formulaic. This supplies fixest-style fixed effects, instrumental variables, categorical expansions, interactions, stepwise model operators, and mature covariance estimators. declaredesign only adds a thin Polars-input adapter and returns the native PyFixest model, preserving its complete post-estimation API.

import numpy as np
import polars as pl
import declaredesign as dd

rng = np.random.default_rng(37)
n = 600
x = rng.normal(size=n)
z = dd.complete_ra(n, m=n // 2, rng=rng)
y = 1 + 1.5 * z + 0.8 * x + rng.normal(size=n)
data = pl.DataFrame({"Y": y, "Z": z, "X": x})

fit = dd.lm_robust("Y ~ Z + X", data, se_type="HC2")
fit.tidy()

	Estimate	Std. Error	t value	Pr(>\|t\|)	2.5%	97.5%
Coefficient
Intercept	0.936765	0.060521	15.478443	0.0	0.817906	1.055624
Z	1.467508	0.085443	17.175276	0.0	1.299703	1.635314
X	0.743967	0.044191	16.835382	0.0	0.657179	0.830755

Direct wrappers feols(), fepois(), feglm(), and quantreg() also accept Polars frames and return native PyFixest models.

fixed_effect_data = data.with_columns(
    pl.Series("group", np.repeat(np.arange(30), 20))
)
fixed_effect_fit = dd.feols(
    "Y ~ Z + X | group",
    fixed_effect_data,
    vcov="HC1",
)
fixed_effect_fit.tidy()

	Estimate	Std. Error	t value	Pr(>\|t\|)	2.5%	97.5%
Coefficient
Z	1.459250	0.088801	16.432811	0.0	1.284831	1.633668
X	0.734447	0.045965	15.978447	0.0	0.644165	0.824729

Experimental estimators

unadjusted = dd.difference_in_means("Y ~ Z", data)
lin = dd.lm_lin("Y ~ Z", data, covariates=["X"])
pl.concat([
    unadjusted.tidy().with_columns(pl.lit("Difference in means").alias("estimator")),
    lin.tidy().with_columns(pl.lit("Lin adjustment").alias("estimator")),
]).select("estimator", "estimate", "std_error", "conf_low", "conf_high")

shape: (2, 5)

estimator	estimate	std_error	conf_low	conf_high
str	f64	f64	f64	f64
"Difference in means"	1.413853	0.104229	1.209154	1.618553
"Lin adjustment"	1.467512	0.085514	1.299566	1.635458

difference_in_means() uses the Neyman standard error in an unblocked two-arm experiment and supports block-weighted contrasts. lm_lin() centers baseline covariates and fully interacts them with treatment. horvitz_thompson() accepts known condition probabilities. iv_robust() uses PyFixest’s IV formula and solver. lm_robust() maps the familiar classical and HC0–HC3 names to PyFixest and supports CRV1, CRV3, Newey–West, and Driscoll–Kraay through the same backend. Its HC2 and HC3 settings reproduce estimatr’s small-sample scaling.

PyFixest supplies CRV1 and leave-one-cluster-out CRV3. The prototype still does not relabel either as estimatr’s Bell–McCaffrey CR2 adjustment; CR2 remains a distinct unimplemented covariance estimator.