canbiplotme

Canonical biplot for grouped estimates measured with error — Python port of the R package gcbiplot (Gravity Canonical Biplot, formerly canbiplotME; Freire-Cadme, 2026, doctoral thesis Biplot canónico y modelo gravitacional de comercio: regímenes comerciales y exportadores medidos con error).

The individuals of a canonical (MANOVA) biplot are estimated coefficient vectors, each with its own measurement-error covariance, correlated within groups and with components that may be undefined for some individuals.

pip install canbiplotme            # core
pip install canbiplotme[all]       # + pyfixest (first stage) and matplotlib

Reproducing the thesis

import canbiplotme as cb
x = cb.load_gcb_trade()                               # 115 exporters, 6 regimes, 5 frictions
fit = cb.gcb(x, S_vars=["log_dist", "comlang_off"])   # trace 18.579
t = cb.gcb_test(fit, B=1999, seed=20251025 + 17000000)
print(t)                                              # p(trace) = 0.096; 2 of 15 pairs (Holm)

With an integer seed, gcb_test draws replicate b with cb.RRNG(seed + b), a generator that reproduces R's set.seed() and rnorm(): the replicates are those of gcbiplot::gcb_test() and of the thesis (differences below 1e-11). The gcb* names are aliases of the cbme* ones.

Minimal example

import canbiplotme as cb
x = cb.load_sim_me()
fit = cb.cbme(x, S_vars=["z1", "z2"])        # second stage
tt = cb.cbme_test(fit, B=1999, seed=1)       # G0 bootstrap: trace, largest root, pairs (Holm), S
rg = cb.cbme_regions(fit, seed=1)            # regions R1-R4
fit.plot(rg, region="R3")
cb.cbme_influence(fit, B_ref=200)            # jackknife + reference
cb.cbme_simulate_tests(fit, nsim=1000, B=199)  # validate the size in your own design

First stage from trade flows (needs pyfixest):

d = cb.load_sim_trade()
s1 = cb.gravity_by_individual(d, "flow", "orig", "dest", "year", ["ldist", "lang", "fta"], "regime", controls=["contig"])
bt = cb.pairs_bootstrap(s1, d, R=199)
cnt = cb.support_counts(d, "orig", "dest", ["lang", "fta"])
idn = cb.identify_d6(s1.beta, bt["V"], s1.id, cnt, ["lang", "fta"], components=s1.spec["frictions"])
fit = cb.cbme(cb.as_me_data(s1, bt, idn))

Validation against R

tests/test_vs_r.py compares with fixtures produced by the R package (tests/fixtures/make_fixtures.R):

• second stage (heterogeneity, means and covariances, metric, eigenvalues, markers, structure correlations, statistics, D²) in four configurations: max. difference ~1e-14;
• regions R1–R3 and jackknife ρ: < 1e-10;
• first stage with pyfixest vs fixest: coefficients < 1e-7, also on a resampled bootstrap dataset and in the dynamic extension. Clustered covariances differ by ~0.1 % because of the small-sample degrees-of-freedom convention of each library (the analysis uses the bootstrap covariance).

Random draws (bootstrap, G0) use NumPy's generator, so p-values agree with R in distribution, not replicate by replicate. The REML comparator cbme_reml reproduces metafor::rma.mv (T < 6e-7, means < 2e-6, LRT < 1e-8). cbme_permutation_test is the permutation comparator.

mvmeta is a general multivariate random-effects meta-analysis and meta-regression (unstructured between-study covariance, missing outcomes, REML / ML / fixed), validated against mixmeta 1.2.2: coefficients, Psi, covariance of the coefficients, log-likelihood and Cochran's Q agree to < 1e-6 (REML with a badly scaled moderator, where the surface is flat: 4e-5 in the coefficients at the same log-likelihood).