Changelog

0.1.0

The first version.

  • Estimators. Arellano–Bond difference GMM and Blundell–Bond system GMM; one-step, two-step or iterated; first differences or forward orthogonal deviations. Where conventions differ, the design follows Stata’s xtabond2 (3.7.2), the reference for the instrument layout, the generalized inverse (including invsym()’s treatment of negative pivots), the weights’ normalizations and the tests.

  • Instruments. Lag limits, collapse, eq, split, passthru, backward orthogonal deviations, IV-style mz, h() and pca, written in the formula (gmm(), iv()) or passed as Gmm / Iv objects.

  • Out of core. Polars reads the data once, from a Parquet file or glob or a Lazy- or DataFrame, into blocks of whole panels. Three passes over the blocks, in numba, give two-step estimates with Windmeijer’s correction and the Arellano–Bond tests; everything else comes from the sums. No instrument matrix is ever built.

  • Inference. Windmeijer’s (2005) correction; one- and multi-way clustering; small; aweights, pweights and fweights; the Arellano–Bond AR(m), Sargan, Hansen and difference-in-Hansen tests; Andrews and Lu’s MMSC; delta-method long-run effects (long_run).

  • Identification tests. Kleibergen–Paap rk LM and Wald and Sanderson–Windmeijer tests (Windmeijer 2025), robust to any clustering, with weights, PCA and absorbed effects; first-stage partial and Shea partial R², the Anderson–Rubin test (and, with one endogenous regressor, its confidence set) and Montiel Olea and Pflueger’s effective F. They are ivreg2’s forms, not the LIML-based ones of Stata’s underid, so the digits differ (see the identification page). The algebra is hdfe-stream’s core.ivtests.

  • Ahn–Schmidt nonlinear moments (nl="noserial" or "iid"), with xtdpdgmm’s first step, one-step fits and Windmeijer’s correction in its general form. The nonlinear fits’ standard errors are the exact ones for the moments and differ by a few percent from xtdpdgmm’s (see the nonlinear page).

  • Absorbed effects. Time effects (| year) and, in difference GMM, period-by-group effects (| region^year), with the estimates, variances and tests of their dummies, under first differences or orthogonal deviations, weights, any clustering and iterated GMM.

  • DPDResult. A hdfe-stream RegressionResult: summary(), tidy(), confint(), wald_test(), etable, summary_json(), with_vcov(), the one-step results of a two-step fit (onestep), and lazy predict(), resid() and sample() over the source.

  • dpd_stream.simulate. simulate_dynamic and truth_dynamic: dynamic panels with predetermined, endogenous or exogenous regressors, stationary or not, unbalanced, with gaps, clusters and weights.

  • Tested against a dense implementation of the algebra, Roodman’s (2009) published output, pydynpd 0.2.2 and plm 2.6.7’s pgmm (on a simulated panel shipped with the tests; Arellano and Bond’s data are not distributed), and Stata’s xtabond2 and xtdpdgmm on about fifty specifications, whose bundle is in benchmarks/external/stata.