Parity with ExPanDaR

A core goal of expdpy is faithful numerical parity with the original R package. Parity is verified at two levels.

Frozen golden values

tests/fixtures/make_goldens.R computes reference values with the base-R functions that ExPanDaR relies on — sd (denominator n-1), quantile (type 7), and cor.test (exact = FALSE) — on a deterministic fixture. The Python test-suite asserts against these, so the fast test run needs no R installation.

Live parity vs ExPanDaR (rpy2)

The against_r-marked tests use rpy2 to call ExPanDaR directly and compare outputs within tolerance. They run in the pixi r environment in CI:

pixi run -e r test-r

Implementation notes

  • Quantiles — numpy method="linear" equals R’s type = 7 (the default). Pass method="averaged_inverted_cdf" to treat_outliers for Stata / R type = 2.
  • Correlations — computed pairwise (per-pair complete observations, p-values and N), Pearson above and Spearman below the diagonal, exactly as ExPanDaR’s cor_mat.
  • Clustered standard errors — prepare_regression_table uses pyfixest with ssc(k_adj=True, G_adj=True), matching lfe::felm(cmethod = "reghdfe").
  • Approximations — the scatter LOESS confidence band is bootstrap-based, and weighted LOESS (loess=2) approximates ggplot’s geom_smooth(weight=...); these are smoke-tested rather than asserted numerically. Logit models are intentionally out of scope (OLS only).