Examples

Each page mirrors a script in the repository’s examples/ directory and shows its output. “Matching” means different things: most pages use nearest-neighbor matching on a propensity score (a logit of the treatment on the covariates), and the titles say when one uses another method (Mahalanobis, kernel and ridge, coarsened exact). To choose among them, see Choosing a matching method. Run the scripts from the repository’s root:

python examples/quickstart.py

Page	What it shows
Quickstart	1:1 matching on a fitted score with replacement, exact strata, a caliper; balance; the ATT with Abadie-Imbens SEs
Matches only	Greedy 1:1 nearest-neighbor matching on the propensity score, without replacement, calipers on the score and covariates, a score model with absorbed fixed effects, pairs to Parquet
k nearest neighbors	k = 3 on the propensity score, with ties and calipers; min_matches; rounds or one pass without replacement
Mahalanobis matching	Nearest-neighbor matching on the covariates, with no propensity score, within strata
Kernel and ridge matching	The default bandwidth, cross-validation, influence-function SEs
Coarsened exact matching	Cells, weights, balance and the effect
Effects and standard errors	When to use which standard error: Abadie-Imbens SEs, n_variance_matches, bias adjustment, the ATE, influence-function SEs clustered on the pair or unit
Out of core	A million units in Parquet files, matched a stratum batch at a time

The simulated data

simulate_matching draws units in 20 exact strata (r x i) with three normal covariates x1, x2, x3, a logistic treatment d whose probability (ps, the true score) depends on the covariates and the stratum, and an outcome

\[y = 1 + d + x_1 + 0.5\,x_2 + 0.3\,x_3 + \text{stratum effect} + \varepsilon,\]

so the treatment effect is 1 for everyone (effect= changes it). With its defaults it reproduces tests/data/ref_data.csv, the data the test suite’s reference values were computed on.