Changelog

0.1.0 (2026-11-06)

First release: matching for data too large for memory, with treatment effects and the standard errors that go with them.

Three functions, one for each kind of matching. Each takes only its own options and returns its own result, with matches(), weights(), counts(), balance() and effect().

• match_nn: nearest-neighbor matching, on a propensity score or on the covariates themselves.

  • The score is fitted for you (logit or probit, with high-dimensional fixed effects absorbed, out of core by hdfe-stream), or it is a column you supply. Or match on Mahalanobis, inverse-variance or Euclidean distance.

  • One match or k per treated unit, with or without replacement, ties kept or broken, and in any of five orders without replacement.

  • Calipers on the score and on covariates, exact strata, common-support trimming, and Mahalanobis distance within a score caliper.

  • The focal units are the treated (the ATT) or the controls (the ATC); effect() can also match both ways for the ATE.

• match_kernel: kernel and ridge matching on the score. Five kernels, ridge weights, a default bandwidth from the 1:1 nearest-neighbor distances (PairDistance) and cross-validated bandwidths (CrossValidation), without forming pairs.

• match_cem: coarsened exact matching, with cells, weights and balance.

Treatment effects with the standard errors the literature uses. effect() gives the ATT, ATC and ATE, for one outcome or several in one pass.

• Abadie-Imbens standard errors for matching with replacement, with the bias adjustment, with heteroskedastic variances (n_variance_matches=), and with the correction for an estimated propensity score (Abadie and Imbens 2016).

• Influence-function standard errors, by unit or clustered on the pair, the matched unit or your own columns, as Abadie and Spiess (2022).

• Results follow hdfe-stream’s conventions, so they print, tabulate and export to pyfixest the same way.

Balance diagnostics. A table of standardized differences, variance ratios and empirical-CDF distances before and after matching, with the sample sizes by arm.

Built for large data.

• Only a few narrow columns are held; the rest of the data stays on disk. Large data are matched a batch of exact strata at a time (memory="by_stratum"), with memory bounded by the largest stratum.

• Matching with replacement and by covariates uses k-d trees, and searches run in parallel. On a 16-core machine, 1:1 matching of 100 million units takes about one to two minutes (see the benchmarks).

• profile=True estimates the memory and disk a match would need, before running it. Threads and memory follow your job’s limits (PBS, Slurm, containers).

Checked against other software. On shared reference data the matches and weights are the same, unit for unit, as R’s MatchIt; the effects and standard errors agree with R’s Matching and Stata’s teffects and kmatch; and the matches agree with psmatch2’s. docs/implementations.md lists where packages differ and which option reproduces each.

Documentation and benchmarks.

• A guide to choosing a matching method, runnable examples, a page on how it works, a function reference and the conventions compared with other packages.

• Benchmarks against MatchIt and R Matching, pymatchit, fastmatch, bare nearest-neighbor searches and Polars, from 10,000 to 100 million units.

• Memory tests (pytest -m resource) check that matching a few narrow columns stays under a limit that holding the data would exceed.

Requires hdfe-stream 0.9.0 or later.