Cross-language parity matrix¶
Auto-generated — do not hand-edit. Regenerate with
python scripts/build_parity_index.py. Every row traces to a committed test artifact; nothing here is asserted from memory.
StatsPAI's promise is that every number it reports is either aligned with an external reference implementation, recovered against a known truth, or honestly marked as neither. This page makes that promise auditable function-by-function, and keeps the three cases apart — a method with no Stata or R sibling can reach the second and never the first, and saying so is the point. Query any function programmatically:
import statspai as sp
sp.parity_status("feols") # one function
sp.parity_matrix() # the whole matrix
sp.parity_summary() # honest coverage counts
Taxonomy¶
| grade | meaning |
|---|---|
bit-exact |
matches a named R/Stata reference on identical input bytes within the strict pre-registered tolerance (headline estimate and SE relative error ≤ 1e-6; most modules land at 1e-9 to 1e-15). A numerical-tolerance grade, not IEEE bitwise equality, and a statement about the compared configuration only — see sp.validation_scope for which options and outputs were compared |
aligned |
matches a named reference within a documented, pre-registered looser tolerance (cross-fit / convention disagreement) |
analytical-only |
recovers a known population parameter on a deterministic DGP, or a closed-form identity (no cross-package reference) |
external-replication |
reproduces published-paper numbers on a calibrated replica |
unverified |
registered public API, no qualifying numerical-parity evidence attached yet — the honest gap |
Coverage at a glance¶
Read the two evidence kinds separately. Only the first answers "does StatsPAI agree with Stata/R"; the second answers "does StatsPAI recover the right answer", which is a different — and for methods with no Stata/R sibling, the only available — question. Summing them into one "verified" figure would let the smaller claim borrow the authority of the larger one, so this page does not print that total.
| evidence kind | grade | functions |
|---|---|---|
| Compared against R/Stata (T2) | bit-exact | 529 |
| aligned | 77 | |
| subtotal | 606 | |
| No external software reference | analytical-only (T1) | 373 |
| external-replication (published numbers) | 54 | |
| subtotal | 427 | |
| No numerical evidence yet | unverified | 654 |
Honest denominators¶
The all-registered denominator understates coverage: it counts result and exception classes, which can never carry a parity grade, and infrastructure functions that render tables, draw plots, build agent schemas or load data. Infrastructure is the registry's infrastructure categories plus the plots, bundled datasets, exporters, language-model helpers and catalogue listings that live in estimator modules, listed by name with the reason in statspai._parity_taxonomy.NON_NUMERIC_CALLABLES. The estimator denominator is the number to drive release over release.
| denominator | cross-language | any evidence | total | cross-lang share |
|---|---|---|---|---|
| estimator callables | 606 | 1013 | 1013 | 59.8% |
| infrastructure (parity N/A) | 0 | 14 | 236 | 0.0% |
| result / exception classes | 0 | 6 | 438 | 0.0% |
| all registered | 606 | 1033 | 1687 | 35.9% |
Coverage by estimator family¶
Families with zero cross-language rows are the highest-leverage targets when a reference implementation exists, and the honest ceiling when one does not — a method with no Stata/R sibling can reach analytical-only and no further. This table is generated from the same records as the rest of the page, so it cannot drift from them.
| family | cross-language | any evidence | estimator callables |
|---|---|---|---|
| causal | 200 | 363 | 363 |
| regression | 55 | 65 | 65 |
| inference | 34 | 60 | 60 |
| timeseries | 24 | 54 | 54 |
| bayes | 13 | 50 | 50 |
| panel | 38 | 42 | 42 |
| diagnostics | 24 | 39 | 39 |
| spatial | 28 | 35 | 35 |
| experimental | 16 | 30 | 30 |
| decomposition | 21 | 25 | 25 |
| network | 23 | 24 | 24 |
| mendelian | 18 | 20 | 20 |
| structural | 7 | 19 | 19 |
| epi | 17 | 18 | 18 |
| postestimation | 14 | 15 | 15 |
| power | 8 | 13 | 13 |
| survival | 8 | 11 | 11 |
| nonparametric | 10 | 10 | 10 |
| survey | 6 | 9 | 9 |
| robustness | 4 | 9 | 9 |
| dag | 6 | 9 | 9 |
| conformal_causal | 3 | 9 | 9 |
| frontier | 5 | 9 | 9 |
| causal_discovery | 3 | 7 | 7 |
| target_trial | 2 | 6 | 6 |
| transport | 3 | 6 | 6 |
| interference | 3 | 6 | 6 |
| fairness | 0 | 6 | 6 |
| neural_causal | 0 | 5 | 5 |
| longitudinal | 2 | 5 | 5 |
| bartik | 4 | 4 | 4 |
| other | 2 | 4 | 4 |
| surrogate | 0 | 3 | 3 |
| causal_rl | 0 | 3 | 3 |
| assimilation | 0 | 3 | 3 |
| missing | 1 | 3 | 3 |
| mediation | 2 | 3 | 3 |
| multivariate | 0 | 2 | 2 |
| gformula | 1 | 2 | 2 |
| ope | 0 | 2 | 2 |
| causal_text | 0 | 2 | 2 |
| forest | 0 | 1 | 1 |
| censoring | 1 | 1 | 1 |
| synth | 0 | 1 | 1 |
bit-exact — 529 functions¶
Machine-tolerance agreement with a named R/Stata reference.
| function | reference | versions | tolerance | rel err (R / Stata) | test |
|---|---|---|---|---|---|
abnormal_returns |
Stata 18 estudy (Pacicco, Vena and Venegoni): abnormal returns and CARs of four models; Patell, BMP and Kolari-Pynnonen statistics; Stata signrank for the signed-rank test | Stata 18 | abnormal returns and CARs 5e-6 rel (estudy works in single precision); test statistics 1e-9 rel given the same standardised CARs, 5e-5 end to end for two models | — / — | test_event_study_returns_parity.py (+1) |
absorb_ols |
fixest::feols 0.14.0 and Stata reghdfe (Track A 03_hdfe / 15_hdfe_cluster goldens); Stata reghdfe with aweights, singleton dropping, two-way clustering | R 4.5.2; fixest 0.14.0; Stata 18 MP | coefficients rtol 1e-12 (observed 2e-15), iid SEs 1e-12 (observed 8e-15), clustered SEs 1e-10 (observed 5.6e-11) | — / — | test_panel_absorb_ols_parity.py |
adjust_pvalues |
base R stats::p.adjust (bonferroni/holm/BH) | R 4.5.2 | exact (atol 1e-15; observed 0) | — / — | test_mht_parity.py (+1) |
aipw |
Stata teffects aipw (ATE, POmeans); R AIPW::AIPW 0.6.9.3 stratified_fit(k_split = 1) | R 4.5.2; AIPW 0.6.9.3; SuperLearner 2.0.40; Stata 18 | 1e-10 rel on ATE, potential-outcome means and SEs (observed <= 1.4e-14) | — / — | test_teffects_R_parity.py (+2) |
ancova |
Stata regress, vce(robust) / vce(cluster) | Stata 18 | 1e-9 rel | — / — | test_misc_sens_stata_parity.py |
anderson_rubin_ci |
R ivmodel::AR.test confidence set | ivmodel 1.9.1; car 3.1.5; metafor 5.0.1 | Both endpoints at 5e-15 after the boundary bisection replaced the grid-point endpoints (previously 8.1e-3 / 4.9e-3). A reference-free test also asserts the two AR entry points agree with each other and that neither endpoint lands exactly on a grid node. | — / — | test_weakiv_meta_parity.py |
anderson_rubin_test |
R ivmodel::AR.test | ivmodel 1.9.1; car 3.1.5; metafor 5.0.1 | Statistic 1.1e-15, p-value 1.2e-13, degrees of freedom exact, and the analytic AR confidence set 1.2e-14. | — / — | test_weakiv_meta_parity.py |
arima |
stats::arima | R 4.5.2; stats 4.5.2 | rel_est<=1e-06, rel_se<=1e-06 | 7.6e-07 / 1.4e-08 | 39_arima.py (+2) |
assortativity |
R igraph::assortativity_degree | igraph 2.3.3; sna 2.8; ergm 4.12.0; dyadRobust 0.0.1.0001 | 1.2e-16 on karate. | — / — | test_network_parity.py |
attributable_risk |
base-R closed form (attributable fraction exposed + PAF) | R 4.5.2 | AFE + PAF point estimates 1e-12 abs (observed 0); CI not pinned | — / — | test_epi_extra_parity.py (+1) |
attrition_bounds |
Stata leebounds 1.5 as shipped (upper bound, tie branch) | leebounds 1.5; Stata 18 | 1e-12 rel | — / — | test_misc_sens_stata_parity.py (+1) |
attrition_test |
Stata tabulate, chi2; regress | Stata 18 | 1e-10 rel | — / — | test_misc_sens_stata_parity.py |
auc |
R pROC::auc; Stata roctab | R 4.5.2; pROC 1.19.0.1; Stata 18 MP | 1e-10 rel (observed 2e-16) | — / — | test_survival_epi_R_parity.py (+2) |
average_treatment_effect |
grf::average_treatment_effect 2.6.1 (target.sample all, treated, control, overlap), forest outputs held fixed | R 4.5.2; grf 2.6.1 | estimate and std.err 1e-10 rel (observed 2.0e-15) | — / — | test_ml_causal_R_parity.py (+1) |
bacon_decomposition |
bacondecomp::bacon | R 4.5.2; bacondecomp 0.1.1 | rel_est<=1e-06, rel_se<=1e-06 | 5.6e-16 / 9.6e-09 | 20_bacon.py (+2) |
bal_tab |
cobalt::bal.tab (mean differences, variance ratios, KS, correlations, effective sample sizes; pairwise and focal multi-category summaries) | cobalt 4.6.2; WeightIt 1.7.0; R 4.5.2 | every statistic 1e-9 relative (observed about 1e-15) | — / — | test_weightit_cobalt_parity.py (+3) |
balance_check |
Stata iebaltab (ietoolkit) | ietoolkit 7.5; Stata 18 | 1e-9 rel | — / — | test_misc_sens_stata_parity.py |
balance_diagnostics |
cobalt::bal.tab and cobalt::col_w_smd / col_w_vr / col_w_ks / col_w_mean / col_w_sd; halfmoon::bal_energy and bal_vr | cobalt 4.6.2; halfmoon 0.2.0; propensity 0.1.0; R 4.5.2 | 1e-9 relative, 1e-12 absolute where a statistic is zero | — / — | test_matching_balance_r_parity.py (+2) |
balance_panel |
base R counts == n_periods | R 4.5.2 | rel_est<=1e-06, rel_se<=1e-06 | 0 / 0 | 69_balance_panel.py (+2) |
bartik |
2SLS: R AER::ivreg + sandwich (HC1 / classical), Stata ivregress 2sls, vce(robust) small / small; Rotemberg weights: R bartik.weight::bw and Stata bartik_weight (Goldsmith-Pinkham, Sorkin & Swift) | R 4.5.2; AER 1.2.16; sandwich 3.1.1; bartik.weight 0.1.0 (GitHub paulgp/bartik-weight@722ceb85484d6a2bf77985edf2403515eacd1770, R-code/pkg); Stata 18 MP; bartik_weight GitHub paulgp/bartik-weight@722ceb85484d6a2bf77985edf2403515eacd1770 code/bartik_weight.ado | 1e-9 rel on all coefficients and SEs (observed <= 5.4e-15) and on Rotemberg alpha_k / beta_k (observed 2.5e-12) | — / — | test_did_synth_shiftshare_parity.py (+2) |
bauer_sinning |
Stata mvdcmp (Powers, Yoshioka & Yun), logit | Stata 18 MP; mvdcmp SSC | Logit: explained, unexplained, gap, Yun-weighted detailed explained and unexplained terms, their 6x6 delta-method covariance and the aggregate SEs at 1e-9. Probit aligned at 2e-6 (reference probit not fully converged; see note). | — / — | test_decomp_R_parity.py |
bayes_net |
bnlearn::bn.fit(method = 'mle') for the conditional probability tables; base R table arithmetic for a Dirichlet-smoothed table, an interventional distribution written as the truncated factorisation, and a conditional query by full enumeration | R 4.5.2; bnlearn 5.2.1 | 1e-10 rel on every table cell and every probability | — / — | test_ness_causal_ai_parity.py (+1) |
benjamini_hochberg |
base R stats::p.adjust(method='BH') | R 4.5.2 | exact (atol 1e-15; observed 0) | — / — | test_mht_parity.py (+1) |
best_linear_projection |
grf 2.6.1 best_linear_projection (HC0 to HC3, clusters and equalized cluster weights; instrumental and causal survival scores), forest outputs held fixed | R 4.5.2; grf 2.6.1; sandwich 3.1.1 | coefficients and SEs 1e-9 relative (observed <= 8.5e-15) with clusters; 1e-12 absolute on the family operators | — / — | test_grf_cluster_operator_parity.py (+3) |
best_subset |
leaps::regsubsets(method = 'exhaustive') | R 4.5.2; leaps 3.2 | RSS and adjusted R2 of the best model of each size 1e-9 rel; Cp 1e-8 abs; the same variables at every size | — / — | test_linear_model_extensions_parity.py (+1) |
betareg |
betareg::betareg(link.phi="log") | R 4.5.2; betareg 3.2.4 | rel_est<=1e-06, rel_se<=1e-06 | 3.1e-08 / 2.1e-08 | 61_betareg.py (+2) |
betweenness_centrality |
R igraph::betweenness | igraph 2.3.3 | Normalised and raw on karate, raw on the directed graph, all at 1e-10. | — / — | test_network_parity.py (+1) |
bias_factor |
R EValue::multi_bound(confounding(), RRAUc, RRUcY) | R 4.5.2; EValue 4.1.4 | 1e-14 rel (observed 0) | — / — | test_inference_sens_R_parity.py (+1) |
binned_residuals |
arm::binned.resids (xbar, ybar, n, range, 2 se; default and chosen bin counts, binning on a regressor, tied values) | R 4.5.2; arm 1.15.3 | 1e-8 abs | — / — | test_regression_stories_r_parity.py (+1) |
biprobit |
GJRM::gjrm(margins = c("probit", "probit"), model = "B", copula = "N"); Stata side is the built-in biprobit (canonical). GJRM's bivariate normal CDF (VGAM::pbinorm) is accurate to about 1e-7, which is the size of the R gap | R 4.5.2; GJRM 0.2.6.9; VGAM 1.1.14 | rel_est<=1e-06, rel_se<=1e-06 | 1.4e-07 / 3.1e-16 | 91_biprobit.py (+2) |
bjs |
didimputation::did_imputation | R 4.5.2; didimputation 0.5.1 | rel_est<=1e-06, rel_se<=1e-06 | 4.8e-08 / 3.5e-07 | 16_bjs.py (+3) |
block_weights |
spdep::nb2blocknb(NULL, ID) + nb2listw(style = 'W') | R 4.5.2; spdep 1.4.2 | neighbour sets exact; weights 1e-12 rel (observed 0) | — / — | test_spatial_survey_R_parity.py (+1) |
blp_test |
GenericML::BLP 0.2.3 (vcovHC const and HC1; with and without the baseline proxy) | R 4.5.2; GenericML 0.2.3; sandwich 3.1.1 | beta and SE 1e-10 rel (observed 2.3e-15); one-sided p 1e-8 rel | — / — | test_ml_causal_R_parity.py (+1) |
bma |
BMS::bms (g-prior, enumeration); BMA::bic.glm; BMA::bicreg | R 4.5.2; BMS 0.3.5; BMA 3.18.21 | vs BMS: inclusion probabilities, posterior means and sds 1e-9 abs for g = UIP, BRIC, RIC; vs bic.glm: model probabilities 1e-9 (logit, Poisson), 1e-5 (gamma), standard deviations 1e-4 (R's IRLS tolerance); vs bicreg: same models, BIC 1e-9 after rounding R-squared to bicreg's five decimals | — / — | test_bayes_mcmc_parity.py (+1) |
bonacich_power |
R igraph::power_centrality and sna::bonpow | igraph 2.3.3; sna 2.8; ergm 4.12.0; dyadRobust 0.0.1.0001 | 3.9e-16 against igraph and 5.8e-16 against sna at beta = 0.1. | — / — | test_network_parity.py |
bonferroni |
base R stats::p.adjust(method='bonferroni') | R 4.5.2 | exact (atol 1e-15; observed 0) | — / — | test_mht_parity.py (+1) |
borusyak_jaravel_spiess |
didimputation::did_imputation | R 4.5.2; didimputation 0.5.1 | rel_est<=1e-06, rel_se<=1e-06 | 4.8e-08 / 3.5e-07 | 16_bjs.py (+3) |
boxcox |
MASS::boxcox (profile); Stata 18 boxcox, model(lhsonly) | R 4.5.2; MASS 7.3.65; Stata 18 | profile log likelihood equal to MASS up to a constant (spread < 1e-9); lambda 1e-7 abs, log likelihoods 1e-8 rel and LR statistics 1e-4 rel vs Stata | — / — | test_linear_model_extensions_parity.py (+2) |
breslow_day_test |
R DescTools::BreslowDayTest (correct=FALSE/TRUE); Stata cc, by() bd tarone | R 4.5.2; DescTools 0.99.60; Stata 18 MP | 1e-10 rel (observed 1.3e-14) | — / — | test_survival_epi_R_parity.py (+2) |
calibrate_confounding_strength |
R sensemakr::ovb_bounds (kd = ky = k) | sensemakr 0.1.6 | 1e-12 rel (CI 1e-10) | — / — | test_misc_sens_R_parity.py |
calibration_test |
grf::test_calibration 2.6.1 (vcov.type HC3 default and HC1), forest outputs held fixed | R 4.5.2; grf 2.6.1; sandwich 3.1.1 | coef, se, t 1e-10 rel (observed 1.1e-15); one-sided p-value 1e-8 rel (observed 1.6e-13) | — / — | test_ml_causal_R_parity.py (+1) |
callaway_santanna |
did::att_gt + aggte | R 4.5.2; did 2.5.1 | rel_est<=1e-06, rel_se<=1e-09 | 1.3e-15 / 1.3e-15 | 04_csdid.py (+2) |
cate_gain_curve |
fklearn.causal.validation.curves.effect_curves and effect_by_segment (linear_effect) | python 3.11.15; fklearn 4.3.0; pandas 3.1.1 | 1e-10 relative on the cumulative effect, the cumulative and relative gain at each of 100 steps, the area and the ten quantile effects (observed 4e-14) | — / — | test_cate_gain_curve_fklearn_parity.py (+1) |
cgs_continuous_did |
contdid::cont_did | R 4.5.2; contdid 0.1.1 | rel_est<=1e-06 | 2.4e-14 / — | 80_contdid.py (+1) |
clogit |
survival::clogit | R 4.5.2; survival 3.8.3 | rel_est<=1e-06, rel_se<=1e-06 | 2.0e-13 / 1.2e-10 | 46_clogit.py (+2) |
clone_censor_weight |
ipw::ipwtm(type = 'cens', family = 'binomial', link = 'logit') on the at-risk rows of each strategy, stabilised (numerator ~1) and unstabilised | R 4.5.2; ipw 1.3.0; ltmle 1.3.0; sandwich 3.1.1; SuperLearner 2.0.40 | 1e-9 relative on every weight (observed 6.6e-15); uncensored (id, time) rows identical | — / — | test_tte_longitudinal_r_parity.py (+1) |
closeness_centrality |
Wasserman-Faust closeness from R igraph::distances | igraph 2.3.3; sna 2.8; ergm 4.12.0; dyadRobust 0.0.1.0001 | Exact (0.0) on a disconnected graph with three blocks and three isolates -- the case the correction exists for; the connected-graph values also match igraph::closeness(normalized = TRUE) to 1e-10. | — / — | test_network_parity.py |
cluster_cross_interference |
stats::lm on cluster means and Stata collapse + regress (direct and spillover coefficients, classical SE) | R 4.5.2; ivreg 0.6.7; sandwich 3.1.1; Stata Stata 18 MP (07 Jun 2023) | 1e-9 relative on estimates and SEs (observed 6.0e-14 R, 6.9e-14 Stata) | — / — | test_interference_r_parity.py (+2) |
cluster_matched_pair |
stats::t.test(paired = TRUE) estimate and stderr, Stata ttest on the collapsed cluster means | R 4.5.2; ivreg 0.6.7; sandwich 3.1.1; Stata Stata 18 MP (07 Jun 2023) | 1e-9 relative on estimate and SE (observed 6.5e-16 R, 5.8e-16 Stata) | — / — | test_interference_r_parity.py (+2) |
cluster_robust_se |
R sandwich::vcovCL (HC1, cadjust; HC0; two-way multi0=FALSE); Stata regress, vce(cluster) | R 4.5.2; sandwich 3.1.1; Stata 18 | SE 1e-10 rel (observed 2.1e-15 R, 9.8e-16 Stata) | — / — | test_inference_sens_R_parity.py (+3) |
clustering |
R igraph::transitivity(type = 'local', isolates = 'zero') | igraph 2.3.3 | Exact on karate and on a disconnected graph whose isolates and degree-1 nodes score 0 on both sides. | — / — | test_network_parity.py (+1) |
cmtest |
R micsr 0.1-5 cmtest (Croissant 2025, Microeconometrics with R) |
micsr 0.1-5 | Tobit: normality, heteroskedasticity, skewness and kurtosis statistics, Hessian form and outer-product form, at 1e-9 (observed 4e-11 to 4e-16). Probit: 1e-6 (observed 8e-7 and 4e-8); micsr::binomreg stops 1e-8 short of the probit optimum and the statistic inherits it, the coefficients of both sides are compared in the same test. | — / — | test_cmtest_micsr_parity.py |
cnsreg |
Stata cnsreg with two linear constraints (default, vce(robust), vce(cluster)) | Stata 18 | coefficients, standard errors, root MSE and F 1e-9 relative | — / — | test_hansen_methods_stata_parity.py (+1) |
cohen_kappa |
base-R closed form (Cohen's kappa point estimate) | R 4.5.2 | kappa + agreements 1e-12 abs (observed ~1e-16); SE not pinned | — / — | test_epi_extra_parity.py (+1) |
compare_metalearners |
R DoubleML::DoubleMLIRM(score = 'ATE') on supplied folds for the average effect; econml TLearner / XLearner for the spread of the effects | R 4.5.2; DoubleML 1.0.2; mlr3 1.6.0; mlr3learners 0.14.0; econml 0.16.0 | 1e-9 relative on the average effect and its standard error (observed 1e-15); 1e-10 absolute on cate_std and cate_iqr | — / — | test_metalearner_tools_parity.py (+2) |
conformal_counterfactual |
cfcausal conformalCf_split steps with estimand='unconditional' (cfcausal::conformal per arm, weights 1/e(x) and 1/(1-e(x))), split supplied | R 4.5.2; cfcausal 0.2.1 (GitHub lihualei71/cfcausal e4ca670); conformalInference 1.1 (GitHub ryantibs/conformal 15c51c6) | 1e-12 relative on both arms' lower, upper and centre (observed <= 2.1e-15); infinite band where overlap fails | — / — | test_conformal_cfcausal_r_parity.py (+1) |
conformal_synth |
scinference (Chernozhukov-Wuthrich-Zhu authors' package, GitHub kwuthrich/scinference 567c688): estimation_method='sc', permutation_method='mb' | R 4.5.2; scinference 0.0.0.9000 567c6889ce0a1d269a62d415b88aa6baf723a3fe; limSolve 2.0.3; quadprog 1.5.8 | p-values exact (rank statistics); ATT 1e-9 rel (observed 9e-16); CI end points exact on the grid | — / — | test_synth_rest_R_parity.py (+1) |
conley |
Stata acreg (Colella, Lalive, Sakalli & Thoenig) | Stata 18 MP; acreg 1.1.0 | SE 1e-9 rel (observed ~5e-15); absorbed-IV spatial 1e-10 | — / — | test_conley_acreg_spacetime_parity.py (+1) |
conley_sensitivity |
Stata acreg spatial HAC at 25, 75, 150 and 300 km, OLS and 2SLS, uniform and Bartlett kernels | Stata 18; acreg 1.1.0 | standard errors 1e-9 relative (observed 6e-13) | — / — | test_ivreport_stata_parity.py (+1) |
continuous_did |
fixest::feols(y ~ dose:post | id + time) (method='twfe') | R 4.5.2; fixest 0.14.0 | slope & SE 1e-9 rel (observed 2.2e-15), iid / ~id / ~region, balanced + unbalanced | — / — |
contrast |
Stata 18 margins r.g / rb3.g / ar.g / gw.g | Stata 18 | 1e-10 rel (observed <= 7.9e-15 est / 4.9e-15 SE) | — / — | test_r2_postest_parity.py (+1) |
cox |
survival::coxph | R 4.5.2; survival 3.8.3 | rel_est<=1e-06, rel_se<=1e-06 | 8.4e-16 / 2.1e-10 | 24_coxph.py (+2) |
cr2_se |
clubSandwich::vcovCR(type="CR2"/"CR3") | R 4.5.2; clubSandwich 0.6.2 | rel_est<=1e-06, rel_se<=1e-06 | 1.8e-08 / 2.2e-08 | 53_cr2.py (+2) |
cr3_jackknife_vcov |
R sandwich::vcovJK(center='estimate'), summclust (CV3); Stata regress, vce(jackknife, cluster() mse double) | R 4.5.2; sandwich 3.1.1; summclust 0.7.0; Stata 18 | vcov 1e-10 rel (observed 6.5e-14 R, SE 4.3e-15 Stata) | — / — | test_inference_sens_R_parity.py (+3) |
ctree |
partykit::ctree (quadratic test statistic, Bonferroni and univariate tests) | R 4.5.2; partykit 1.3.0 | tree shape, split variables, cut points and level sets, terminal node and prediction of every row exact; test statistics 1e-9 relative; adjusted p-values 1e-7 relative | — / — | test_ctree_partykit_parity.py (+1) |
cuminc |
R cmprsk::cuminc (estimate, var, Tests); Stata stcompet (ci, se, hi, lo) | R 4.5.2; cmprsk 2.2.12; Stata 18 MP; stcompet 1.0.7 (06nov2012) | 1e-10 rel (observed: CIF 6e-16, Gray variance 3e-13, delta SE 2e-16, Gray test 9e-15) | — / — | test_survival_epi_R_parity.py (+2) |
cumulative_periodogram_test |
Stata 18 wntestb | Stata 18 | statistic 1e-12 rel; p-value 2e-8 abs | — / — | test_spectral_parity.py (+1) |
cusum_test |
strucchange::efp(type = 'Rec-CUSUM') + sctest 1.5.4; Stata 18 estat sbcusum | R 4.5.2; strucchange 1.5.4; Stata 18 | process, statistic and p-value 1e-10 rel (observed 7.1e-14); Stata boundary constants 3e-6 | — / — | test_timeseries_R_parity.py (+2) |
dag |
dagitty::adjustmentSets(type = 'minimal'), impliedConditionalIndependencies, dseparated, instrumentalVariables, paths | R 4.5.2; dagitty 0.3.4 | exact: sets and truth values, on 120 random graphs (1,440 d-separation statements) | — / — | test_dag_dagitty_parity.py (+3) |
das_gupta |
R DasGuptR::dgnpop (product rate function, summed over strata) | DasGuptR 2.2.0; ddecompose 1.0.0; cdgd 1.0.1 | Das Gupta's Table 2.1 (two factors) and Table 6.5 (four factors x six age groups), every factor effect and both crude rates at 1e-10. | — / — | test_decomp_R_parity.py |
ddd |
Stata 18 MP regress [aw=w], robust (aweight HC1) | Stata 18 MP | b / se 1e-12 abs (observed <= 3e-15) | — / — | test_did2x2_ddd_weighted_robust_parity.py |
ddd_heterogeneous |
triplediff::ddd + agg_ddd | R 4.5.2; triplediff 0.2.4 | rel_est<=1e-06, rel_se<=1e-06 | 5.7e-15 / — | 77_ddd.py (+1) |
decompose |
oaxaca::oaxaca | R 4.5.2; oaxaca 0.1.5 | rel_est<=1e-06, rel_se<=1e-06 | 2.6e-15 / 2.1e-15 | 30_oaxaca.py (+2) |
degree_centrality |
R igraph::degree | igraph 2.3.3 | Normalised on karate and raw in / out / all modes on a 40-node directed graph, exact. | — / — | test_network_parity.py (+1) |
demean |
textbook mean-within (algorithmic) | R 4.5.2 | rel_est<=1e-06, rel_se<=1e-06 | 3.5e-15 / 8.8e-15 | 68_demean_within.py (+2) |
design_aberration |
DoE.base 1.2-5 GWLP (generalized word length pattern) of given designs: replicated L9, columns of L18, an unbalanced mixed-level design | R 4.5.2; DoE.base 1.2.5 | 1e-10 absolute | — / — | test_joseph_doe_parity.py (+1) |
design_augment |
MaxPro 4.1-2 MaxProAugment (greedy MaxPro augmentation from given candidates) | R 4.5.2; MaxPro 4.1.2 | 1e-14 absolute on the chosen runs (the same six candidates in the same order) | — / — | test_joseph_doe_parity.py (+1) |
design_criteria |
SFDesign 0.1.5 maxpro.crit, maximin.crit (plain and surrogate, r = 4), uniform.crit on a given design | R 4.5.2; SFDesign 0.1.5 | 1e-12 relative | — / — | test_joseph_doe_parity.py (+1) |
dfl_decompose |
ddecompose::dfl_decompose | R 4.5.2; ddecompose 1.0.0 | rel_est<=1e-06, rel_se<=1e-06 | 1.2e-09 / 1.8e-13 | 31_dfl.py (+3) |
diagnostic_test |
R epiR::epi.tests; Stata diagti | R 4.5.2; epiR 2.0.94; Stata 18 MP; diagt 2.032 (diagti 2.053) | 1e-12 rel (observed 2e-15) | — / — | test_survival_epi_R_parity.py (+2) |
did_2stage |
did2s::did2s | R 4.5.2; did2s 1.2.1 | rel_est<=1e-06, rel_se<=1e-06 | 4.8e-08 / 2.4e-12 | 73_did2s.py (+3) |
did_2x2 |
Stata 18 MP regress [aw=w], robust (aweight HC1) | Stata 18 MP | b / se 1e-12 abs (observed <= 3e-16) | — / — | test_did2x2_ddd_weighted_robust_parity.py |
did_estimate |
synthdid::did_estimate 0.0.9 | R 4.5.2; synthdid 0.0.9 | estimate and jackknife SE at 1e-9 on the Prop. 99 replica and a five-treated panel (observed est 3.1e-15, jackknife SE 5.7e-16); R's 40 placebo and 40 bootstrap replications replayed at 1e-9 (observed 1.9e-13) | — / — | test_did_synth_R_parity.py (+1) |
did_imputation |
didimputation::did_imputation | R 4.5.2; didimputation 0.5.1 | rel_est<=1e-06, rel_se<=1e-06 | 4.8e-08 / 3.5e-07 | 16_bjs.py (+2) |
did_multiplegt |
DIDmultiplegt::did_multiplegt (archived 0.1.4) | R 4.5.2; DIDmultiplegt 0.1.4 | rel_est<=1e-06 | 3.9e-15 / 3.2e-15 | 81_didm.py (+2) |
did_multiplegt_dyn |
DIDmultiplegtDYN::did_multiplegt_dyn | R 4.5.2; DIDmultiplegtDYN 2.3.4 | rel_est<=1e-06 | 3.3e-15 / 2.1e-15 | 78_multiplegt_dyn.py (+2) |
did_timevarying_covariates |
ptetools::pte_default(d_outcome=TRUE, est_method='reg') and did::att_gt(est_method='reg') | R 4.5.2; ptetools 1.0.1; did 2.5.1; DRDID 1.3.0 | ATT(g,t) and overall ATT 1e-9 rel (observed 3.8e-15) | — / — | test_did_synth_didvar_parity.py (+1) |
direct_method |
Open Bandit Pipeline (obp) 0.5.7 DirectMethod (same reward-model matrix via q_hat=) | obp 0.5.7 | value 1e-12 rel (observed 0.0); SE identity 1e-10 | — / — | test_ml_causal_obp_parity.py (+1) |
direct_standardize |
R epitools::ageadjust.direct; Stata dstdize | R 4.5.2; epitools 0.5.10.1; Stata 18 MP | 1e-10 rel (observed 3e-13) | — / — | test_survival_epi_R_parity.py (+2) |
discos |
DiSCos::DiSCo 0.1.4 (Gunsilius distributional synthetic controls), mixture = FALSE; mixture = TRUE vs GLPK on DiSCo's LP | R 4.5.2; DiSCos 0.1.4; pracma 2.4.6; quadprog 1.5.8; CVXR 1.8.2; Rglpk 0.6.5.1 | quantile weights 1e-11 abs (observed 4.9e-13), counterfactual quantile functions 1e-10 rel, quantile effects 1e-10 abs (observed 1.9e-12); mixture LP weights vs GLPK 1e-12 abs, vs DiSCo's SCS solution 5e-6 abs (observed 7.1e-7) | — / — | test_did_synth_synthvar_parity.py (+1) |
distance_band |
R spdep::dnearneigh | spdep 1.4.2; spatialreg 1.4.3 | Neighbour sets identical for all 120 points at a 0.25 radius. | — / — | test_spdep_parity.py |
distributional_did |
didFF::distDD 0.1.0 (Roth & Sant'Anna) | R 4.5.2; didFF 0.1.0; did 2.5.1 | per-bin effect & SE 1e-9 (observed est 2.6e-12 abs, SE 2.5e-10 rel) | — / — | test_did_synth_didvar_parity.py (+3) |
dlm |
dlm::dlmFilter, dlmSmooth, dlmLL, dlmMLE | R 4.5.2; dlm 1.1.6.1 | filtered and smoothed means, filtered variances, one-step forecasts and the log likelihood 1e-9 rel at fixed variances; smoothed variances 1e-5 under the diffuse prior; maximum-likelihood variances 1e-5 with a likelihood at least as high | — / — | test_dlm_parity.py (+1) |
dml |
DoubleML::DoubleMLPLR | R 4.5.2; DoubleML 1.0.2 | rel_est<=1e-10, rel_se<=1e-10 | 0 / 3.7e-15 | 08_dml.py (+2) |
dml_did |
doubleml (Python) DoubleMLDID and DoubleMLDIDCS | doubleml 0.11.3; scikit-learn 1.6.1 | estimate and standard error 1e-8 on a shared fold column, linear and unpenalised-logit learners; two layouts x two scores x with and without in-sample normalisation | — / — | test_dml_did_doubleml_parity.py (+1) |
dml_panel |
fixest::demean 0.14.0 (unit and unit+time absorption, balanced and unbalanced) and ddml::ddml_plm 0.3.1 (OLS learner, unit clusters, shared folds) | R 4.5.2; fixest 0.14.0; ddml 0.3.1; sandwich 3.1.1; doubleml 0.11.3; scikit-learn 1.6.1 | within transform 1e-10 abs (observed 2.6e-14); estimate 1e-10 rel vs ddml and DoubleML (observed 4.1e-16); SE 1e-10 rel vs DoubleML (observed 1.4e-15) and vs ddml after the CR1 factor | — / — | test_ml_causal_dml_parity.py (+2) |
dml_sensitivity |
doubleml (Python) DoubleML.sensitivity_analysis | — | bias_bound and adjusted theta bounds 1e-12 (observed 2.5e-15); RV 1e-6 (observed 9.2e-8); RVa is a documented convention gap (<5e-3, observed 1.4e-3) because StatsPAI exhausts | theta | -z*se with the unadjusted SE while doubleml lets the SE move with the confounding scenario |
do_calculus_apply |
dagitty::dseparated on the corresponding mutilated graph, built in R from the definition | R 4.5.2; dagitty 0.3.4 | exact: a truth value, on 720 queries (Y, X, Z, W) | — / — | test_dag_dagitty_parity.py (+1) |
do_rule1 |
dagitty::dseparated on the graph without the edges into X, built in R from the definition | R 4.5.2; dagitty 0.3.4 | exact: a truth value, on 720 queries (Y, X, Z, W) | — / — | test_dag_dagitty_parity.py (+1) |
do_rule2 |
dagitty::dseparated on the graph without the edges into X and out of Z, built in R from the definition | R 4.5.2; dagitty 0.3.4 | exact: a truth value, on 720 queries (Y, X, Z, W) | — / — | test_dag_dagitty_parity.py (+1) |
do_rule3 |
dagitty::dseparated on the graph without the edges into X and into the members of Z that are not ancestors of W, built in R from the definition | R 4.5.2; dagitty 0.3.4 | exact: a truth value, on 720 queries (Y, X, Z, W) | — / — | test_dag_dagitty_parity.py (+1) |
dose_response |
Stata doseresponse / gpscore (Hirano-Imbens normal GPS, quadratic T and GPS with interaction) | Stata 18; doseresponse SSC | 1e-9 rel on the dose-response function at 5 doses (observed 5.1e-10) | — / — | test_teffects_R_parity.py (+2) |
double_negative_control |
ivreg::ivreg + sandwich::vcovHC(type = 'HC1') and Stata ivregress 2sls, vce(robust) small (y on d, nco, X with d, nce, X) | R 4.5.2; ivreg 0.6.7; sandwich 3.1.1; Stata Stata 18 MP (07 Jun 2023) | 1e-9 relative on estimate and SE (observed 4.9e-15 R, 6.0e-15 Stata) | — / — | test_proximal_r_parity.py (+2) |
doubly_robust |
Open Bandit Pipeline (obp) 0.5.7 DoublyRobust (lambda_ inf and 2; same reward model via q_hat=) | obp 0.5.7 | value 1e-12 rel (observed 1.8e-16); SE identity 1e-10 | — / — | test_ml_causal_obp_parity.py (+1) |
drdid |
DRDID::drdid_imp_panel | R 4.5.2; DRDID 1.3.0 | rel_est<=1e-06, rel_se<=1e-06 | 2.6e-15 / 2.2e-16 | 38_drdid.py (+2) |
dyadic_regression |
R dyadRobust (Aronow-Samii-Assenova dyadic-robust variance) | igraph 2.3.3; sna 2.8; ergm 4.12.0; dyadRobust 0.0.1.0001 | Coefficients 7e-16 and standard errors 2e-15 on undirected and directed dyads, after the 1.28.0 fix to the shared-member weighting; also asserted against a brute-force construction of the definition. | — / — | test_network_parity.py |
ebalance |
ebal::ebalance 0.2.1 (Hainmueller 2012) | — | ATT rel <= 1e-5 (observed 3.2e-7); moment gap <= 1e-10 | — / — | test_matching_r_parity.py (+1) |
effective_f_test |
Stata weakivtest (Montiel Olea & Pflueger) after ivreg2; R ivDiag::eff_F 1.0.6 | Stata 18 MP; weakivtest 10/28/2020; ivreg2 4.1.12; R 4.5.2; ivDiag 1.0.6 | F_eff rel 1e-9 on 6 designs (observed 3.7e-14 vs Stata, 1.3e-12 vs ivDiag for k = 1) | — / — | test_rd_iv_R_parity.py (+2) |
eigenvector_centrality |
R sna::evcent (unit L2 norm, as here) | sna 2.8; igraph 2.3.3 | 5e-11 undirected, 8e-11 directed -- power-iteration tolerance on both sides. igraph::eigen_centrality max-scales instead (and 2.x ignores scale = FALSE), so it agrees only up to one scalar; the earlier note claiming the igraph convention was wrong. | — / — | test_network_parity.py (+1) |
eivreg |
eivtools::eivreg 0.1-8 (a port of Stata's eivreg; archived on CRAN, installed from the archive into a private library, see R_ENVIRONMENT.md); Stata side is the built-in eivreg, the canonical implementation | R 4.5.2; eivtools 0.1.8 | rel_est<=1e-06, rel_se<=1e-06 | 3.2e-15 / 4.7e-15 | 90_eivreg.py (+2) |
energy_distance |
halfmoon::bal_energy (unweighted and weighted) | halfmoon 0.2.0; R 4.5.2 | 1e-9 relative (observed 1e-14) | — / — | test_matching_balance_r_parity.py (+2) |
engle_granger |
egranger 1.0.6 (Stata SSC); urca::ur.df on lm residuals; aTSA::coint.test | R 4.5.2; urca 1.3.4; aTSA 3.1.2.1; Stata 18; egranger 1.0.6 | Z(t) and step-1 coefficients 1e-10 rel (observed 2.1e-13); MacKinnon (2010) critical values 1e-12 vs egranger | — / — | test_timeseries_R_parity.py (+2) |
ergm |
R ergm::ergm(estimate = 'MPLE') | igraph 2.3.3; sna 2.8; ergm 4.12.0; dyadRobust 0.0.1.0001 | Coefficients 2e-16 to 7e-13 for edges + triangle + nodematch + nodecov + absdiff (undirected) and edges + mutual (directed); standard errors 2e-8 (directed) and <= 3.2e-7 (undirected), inside the 1e-6 budget. | — / — | test_network_parity.py |
esize |
Stata 18 MP esize twosample, all (with and without unequal) | Stata 18 MP | estimates and interval limits 1e-8 rel (observed <= 1e-12) | — / — | test_effect_size_power_ttest_parity.py |
ess |
WeightIt::ESS and halfmoon::ess, overall and by group | WeightIt 1.7.0; halfmoon 0.2.0; R 4.5.2 | 1e-9 relative (observed 2e-14) | — / — | test_matching_balance_r_parity.py (+2) |
etpoisson |
Stata 18 MP official etpoisson and margins |
Stata 18 MP | Every block at 1e-6; observed 3e-15 on coefficients, 8e-11 on standard errors, with 24 and 64 Gauss-Hermite points, under vce(oim), vce(robust) and vce(cluster). Log-likelihood to 1e-8, the Wald test of independent equations at 1e-6. The ATE on the count scale and the two potential-outcome means agree with margins r.d and margins d to 1e-9, the delta-method standard error of the ATE to 1e-6. |
— / — | test_etpoisson_stata_parity.py |
etregress |
Stata 18 MP official etregress (Maddala 1983 model) |
Stata 18 MP | Two-step: 5e-9 on every coefficient and every standard error, including the Heckman correction for the estimated first stage. ML: the likelihood, score and observed information are pinned at 9e-11 -- our Hessian reproduces Stata's reported standard errors when evaluated at Stata's own parameter vector, which is independent of either optimiser. At our own optimum the parameters sit within 2e-5 of Stata's; that gap is the two optimisers' stopping points, not a formula difference, and StatsPAI's stops at the HIGHER log-likelihood with a gradient ~300x smaller (asserted, so a regression that makes our optimum worse fails even though the 1e-4 parity assertions would still pass). vce(robust) carries Stata's N/(N-1) meat factor and vce(cluster) its g/(g-1). | — / — | test_etregress_stata_parity.py |
etwfe |
etwfe::etwfe + emfx | R 4.5.2; etwfe 0.6.2 | rel_est<=1e-06, rel_se<=1e-06 | 1.8e-13 / 3.9e-14 | 17_etwfe.py (+2) |
etwfe_emfx |
etwfe::etwfe + emfx | R 4.5.2; etwfe 0.6.2 | rel_est<=1e-06, rel_se<=1e-06 | 1.8e-13 / 3.9e-14 | 17_etwfe.py (+2) |
evalue |
EValue::evalues.RR | R 4.2.3; EValue 4.1.4 | rel_est<=1e-06, rel_se<=1e-06 | 5.8e-14 / 1.2e-16 | 23_evalue.py (+2) |
evalue_from_result |
R EValue::twoXtwoRR -> evalues.RR | R 4.5.2; EValue 4.1.4 | 1e-13 rel (observed 3.3e-16) | — / — | test_inference_sens_R_parity.py (+1) |
evalue_rd |
R EValue::evalues.RD | R 4.5.2; EValue 4.1.4 | 1e-12 rel (observed 5.8e-14: grid built by seq vs np.arange) | — / — | test_inference_sens_R_parity.py (+1) |
evalue_rr |
R EValue::evalues.RR | EValue 4.1.4 | Point and CI E-values at 1e-12 across ten cases, including RR < 1 and CIs crossing the null. | — / — | test_evalue_rr_parity.py |
event_study |
fixest::feols(y ~ i(rel, treat, ref=-1) | R 4.5.2; fixest 0.14.0 | rel_est<=1e-09, rel_se<=1e-09 | 3.2e-13 / 1.5e-14 | 85_twfe_event_study.py (+2) |
event_study_vcov |
did::aggte(type='dynamic') 2.5.1 influence-function covariance; fixest::sunab 0.14.0 (A V A', cohort shares fixed); fixest::feols i(rel, ref=-1) 0.14.0; did2s::did2s 1.2.1; etwfe::emfx(type='event') 0.6.2; Stata 18 did_imputation (Borusyak, SSC, 2023-11-22) | R 4.5.2; did 2.5.1; fixest 0.14.0; did2s 1.2.1; etwfe 0.6.2; marginaleffects 0.32.0; Stata 18 | full joint covariance (every off-diagonal entry) 1e-9 rel (observed 3.4e-15 cs / 5.1e-13 twfe / 8.9e-11 sunab / 6.7e-11 did2s), 1e-8 did_imputation (observed 1.2e-9), 1e-6 etwfe (observed 3.4e-9 not-yet / 8.1e-8 never); coefficients 1e-9 (observed <= 3.3e-10) | — / — | test_event_study_vcov_R_parity.py (+3) |
factor_importance |
first 2.1 first (FIRST factor importance: forward selection with early dropping, backward elimination) | R 4.5.2; first 2.1 | 1e-10 absolute on every importance | — / — | test_joseph_doe_parity.py (+1) |
factorial_effects |
unrepx 1.0-2 (effects and Lenth's pseudo standard error) and stats::lm on a half fraction (aliased coefficients) | R 4.5.2; unrepx 1.0.2 | effects and coefficients 1e-10 absolute; pseudo standard error 1e-12 relative | — / — | test_joseph_doe_parity.py (+1) |
fairlie |
Stata fairlie 1.0.7 (Jann, SSC) | Stata 18.0 MP; fairlie 1.0.7 16jun2008 (SSC) | tightly converged logit/probit: contributions and SEs 1e-12 rel (observed 4.5e-15 / 5.4e-14); at Stata's default logit tolerance contributions 1e-8 and SEs 1e-5 (observed 1.6e-10 / 1.6e-6) | — / — | test_decomp_qte_parity.py (+1) |
fama_macbeth |
plm::pmg; sandwich::NeweyWest on the coefficient series; Stata 18 xtfmb (with and without lag()) | R 4.5.2; plm 2.6.7; sandwich 3.1.1; Stata 18 | coefficients, covariance and Newey-West SE 1e-9 rel (R); 1e-6 rel (xtfmb keeps the per-period estimates as floats) | — / — | test_accounting_research_parity.py (+2) |
fci |
pcalg::fci(skel.method = 'stable') with gaussCItest | R 4.5.2; pcalg 2.7.12 | partial ancestral graphs compared for equality, every edge with the mark at each end, on 39 data sets | — / — | test_fci_pcalg_parity.py (+3) |
fect |
fect::fect(Y ~ D + X1 + X2, method=, force="two-way", se=FALSE, CV=FALSE, tol=1e-12, max.iteration=20000); Stata side uses the authors' fect_stata (GitHub, installed into a local ado path) | R 4.5.2; fect 2.4.1 | rel_est<=1e-06, rel_se<=1e-06 | 1.8e-13 / 9.8e-10 | 86_fect.py (+2) |
feglm |
fixest::feglm (family="logit") / fixest::fepois | R 4.5.2; fixest 0.14.0 | rel_est<=1e-06, rel_se<=1e-06 | 3.8e-15 / 1.8e-09 | 67_panel_glm.py (+2) |
feols |
fixest::feols | R 4.5.2; fixest 0.14.0 | rel_est<=1e-06, rel_se<=1e-06 | 5.2e-15 / 2.9e-15 | 03_hdfe.py (+3) |
fepois |
fixest::feglm (family="logit") / fixest::fepois | R 4.5.2; fixest 0.14.0 | rel_est<=1e-06, rel_se<=1e-06 | 3.8e-15 / 1.8e-09 | 67_panel_glm.py (+2) |
ffl_decompose |
R ddecompose::ob_decompose(reweighting = TRUE) | rifreg 1.1.0; dineq 0.1.0; ddecompose 1.0.0 | Observed difference, composition, structure, specification and reweighting errors at 1e-9 (1e-12 absolute floor; logit MLE in the path), both reference directions, for the variance, Gini (exact RIF supplied as custom_rif_function) and 10th/50th/90th percentiles. | — / — | test_decomp_R_parity.py |
finegray |
R cmprsk::crr (coef, var, invinf, loglik); Stata stcrreg | R 4.5.2; cmprsk 2.2.12; Stata 18 MP | R 1e-10 rel (observed 6e-15); Stata coefficients 1e-7 and SEs 1e-8 (Stata's ml stops 5e-9 away), log likelihood 1e-10 | — / — | test_survival_epi_R_parity.py (+2) |
fisher_exact |
R ri2::conduct_ri (randomizr full enumeration); Stata ritest over the full assignment set | R 4.5.2; ri2 0.5.0; Stata 18; ritest 1.1.7 | p exact (k / N_assignments); statistic 1e-12 rel vs R, 2^-23 vs Stata | — / — | test_inference_sens_R_parity.py (+3) |
florentine_families |
R ergm flomarriage | igraph 2.3.3; sna 2.8; ergm 4.12.0; dyadRobust 0.0.1.0001 | The 20 marriage ties are identical edge for edge; the Pucci isolate is omitted (15 nodes against 16), which is documented. | — / — | test_network_parity.py |
four_way_decomposition |
Stata med4way (yreg/mreg linear); CMAverse::cmest 0.1.0 (rb, paramfunc, delta) | R 4.5.2; CMAverse 0.1.0; Stata 18 | components 1e-12 rel; delta SEs 1e-12 rel vs CMAverse (vcov='ols'); variances 5e-9 rel vs med4way (vcov='ml') | — / — | test_teffects_R_parity.py (+2) |
fracreg |
stats::glm(quasibinomial('logit')) [fractional response] | R 4.5.2 | coefficients 1e-10 abs (observed ~8e-15) | — / — | test_glm_ext_parity.py (+1) |
frontier |
sfaR::sfacross | R 4.5.2; sfaR 1.0.1 | rel_est<=1e-06, rel_se<=1e-06 | 1.6e-09 / 1.1e-09 | 28_frontier.py (+2) |
functional_form_test |
didFF::didFF | R 4.5.2; didFF 0.1.0 | rel_est<=1e-06 | 1.3e-14 / — | 79_didff.py (+1) |
g_computation |
base R stats::lm g-formula standardization (Robins 1986) | — | psi 1e-8 (observed <= 7e-16; bootstrap SE pinned loosely +/-25%) | — / — | test_gformula_parity.py (+1) |
g_estimation |
DTRreg::DTRreg 2.4 (method = 'gest', treat.type = 'bin', weight = 'none') | R 4.5.2; DTRreg 2.4 | 1e-10 rel on each stage psi (observed 4.6e-15) | — / — | test_teffects_R_parity.py (+2) |
gam |
mgcv::gam with s(x, bs = 'ps' | 'tp' | 're'), te(x, z), methods GCV.Cp and REML; summary.gam for the term tests | R 4.5.2; mgcv 1.9.3 | at given smoothing parameters, Gaussian, binomial and Poisson: coefficients, SEs, edf, GCV / UBRE, scale, deviance and fitted values 1e-9 rel, predictions and term curves with SEs 1e-8; selected smoothing parameters 1e-4 rel, the resulting fit 1e-5; tensor, thin plate and random-effect terms 1e-9 at given parameters, 1e-5 under REML (2e-3 for the random-effect model); term tests: ref.df 1e-9, p-values 5e-6 abs |
gap_closing |
R ddecompose::dfl_decompose (method='ipw') and ob_decompose (method='regression') | DasGuptR 2.2.0; ddecompose 1.0.0; cdgd 1.0.1 | Observed, counterfactual and closed gaps at 1e-9 for IPW in both directions (logit MLE in the path) and 1e-10 for regression. method='aipw' has no reference and is checked for double robustness on a known-truth DGP (T1). | — / — | test_decomp_R_parity.py |
gardner_did |
did2s::did2s | R 4.5.2; did2s 1.2.1 | rel_est<=1e-06, rel_se<=1e-06 | 4.8e-08 / 2.4e-12 | 73_did2s.py (+2) |
gate_test |
GenericML::GATES 0.2.3 (monotonize = FALSE) with GenericML::quantile_group membership | R 4.5.2; GenericML 0.2.3; sandwich 3.1.1 | gamma_k, SE and gamma_K - gamma_1 1e-10 rel (observed 1.9e-15); group membership identical | — / — | test_ml_causal_R_parity.py (+1) |
geary |
R spdep::geary.test | spdep 1.4.2; spatialreg 1.4.3 | C 2.0e-15. The closed-form variance and z are new in 1.27.0 (they were NaN whenever permutations=0) and match both spdep nulls at 1.5e-14: randomisation (with the m4/m2^2 term) and normality. | — / — | test_spdep_parity.py |
gee |
gee::gee (independence, exchangeable, AR-M); Stata 18 xtgee (independent, exchangeable, ar 1; with and without nmp) | R 4.5.2; gee 4.13.30; Stata 18 | vs gee 1e-9 rel on coefficients, naive and robust SE, scale and alpha; vs xtgee 1e-6 rel (its convergence tolerance) | — / — | test_linear_model_extensions_parity.py (+2) |
gelbach |
Stata b1x2 (Gelbach's own command), robust and homoskedastic | Stata 18 MP; b1x2 SSC | Per-variable contributions at 1e-10; their full covariance, SEs and the total-change variance at 1e-9, robust and homoskedastic. | — / — | test_decomp_R_parity.py |
gelman_rubin |
coda::gelman.diag(transform = FALSE) | R 4.5.2; coda 0.19.4.1 | point estimate and upper limit 1e-9 rel, with and without autoburnin | — / — | test_bayes_mcmc_parity.py (+1) |
geographic_rd |
rdmulti::rdms; Stata side uses the rdms ado from the rdpackages GitHub mirror (rdmulti is not on SSC: ssc describe rdmulti returns r(601)), installed into a local gitignored ado path | R 4.5.2; rdmulti 2.0.0 | rel_est<=1e-06, rel_se<=1e-06 | 1.2e-12 / 5.4e-10 | 89_rdms.py (+3) |
ges |
pcalg::ges(GaussL0penObsScore, phase = c('forward', 'backward'), iterate = FALSE) | R 4.5.2; pcalg 2.7.12 | essential graphs compared for equality, every edge with its orientation, on 30 data sets | — / — | test_ges_pcalg_parity.py (+1) |
get_scores |
grf 2.6.1 get_scores (causal, instrumental, multi-arm and causal survival forests), forest outputs held fixed | R 4.5.2; grf 2.6.1 | 1e-11 absolute elementwise (observed <= 2.3e-14) | — / — | test_grf_family_operator_parity.py (+3) |
getis_ord_g |
R spdep::globalG.test (binary weights) | spdep 1.4.2; spatialreg 1.4.3 | G 5.4e-16. Binary weights, which is what spdep recommends for this statistic. | — / — | test_spdep_parity.py |
getis_ord_local |
R spdep::localG | spdep 1.4.2; spatialreg 1.4.3 | Gi* 1.7e-14; Gi 4.3e-13 after the star=False branch stopped borrowing Gi*'s standardisation. | — / — | test_spdep_parity.py |
geweke_diag |
coda::geweke.diag | R 4.5.2; coda 0.19.4.1 | z-scores 1e-9 rel at (0.1, 0.5) and (0.2, 0.3) | — / — | test_bayes_mcmc_parity.py (+1) |
gformula_ice_fn |
ltmle::ltmle 1.3.0 (gcomp = TRUE, SL.library = list(Q = 'SL.lm')) point estimate; base-R lm() ICE with geex::m_estimate 1.1.1 sandwich SE | R 4.5.2; ltmle 1.3.0; SuperLearner 2.0.40; geex 1.1.1 | point 1e-10 rel vs ltmle (observed 2.1e-11), 1e-12 vs hand lm (observed 2.4e-15); sandwich SE 1e-9 rel (observed 7.1e-12) | — / — | test_r2_teffects_parity.py (+3) |
glm |
base R stats::glm (binomial logit + Poisson log) | R 4.5.2 | coef / logLik / AIC 1e-8 abs (observed <= 5e-13); SE ~1e-3 rel | — / — | test_glm_parity.py (+1) |
glmnet |
glmnet::glmnet and glmnet::cv.glmnet (gaussian, binomial; thresh = 1e-14) | R 4.5.2; glmnet 4.1.10 | penalty path 1e-10 and equal length; coefficients 1e-6 absolute (5e-5 with more predictors than rows); deviance explained 1e-7; cross-validated error and its standard error 1e-4; lambda.min and lambda.1se the same grid point | — / — | test_glmnet_r_parity.py (+1) |
gmm |
Stata 18 gmm (linear and exponential-mean IV; twostep, igmm, onestep); R gmm::gmm | Stata 18 MP; R 4.5.2; gmm 1.9.1 | vs Stata estimates rtol 1e-10 (observed 1.2e-12), SEs 1e-9 linear / 1e-6 nonlinear (Stata's numerical Jacobian; observed 2.5e-7), J rtol 1e-10; vs R rtol 1e-6 (observed 1.8e-7) | — / — | test_panel_gmm_stata_parity.py (+1) |
gologit |
gologit2 (all, no and partial parallel lines; logit and probit links; vce(robust), vce(cluster)); R VGAM::vglm(cumulative(reverse = TRUE)) coefficients | Stata 18; gologit2 3.2.8; R 4.5.2; VGAM 1.1.14 | coefficients and standard errors 1e-6 relative (observed 4.4e-7 against gologit2, 1.7e-7 against VGAM); log likelihood 1e-12 | — / — | test_regcmd_stata_r_parity.py (+2) |
granger_causality |
Stata 18 vargranger; vars::causality 1.6.1 | R 4.5.2; vars 1.6.1; Stata 18 | chi2 and F 1e-10 rel (observed 4.1e-15), p-values 1e-8 | — / — | test_timeseries_R_parity.py (+2) |
grf_calibration |
grf::test_calibration 2.6.1 (vcov.type HC3 default and HC1; with clusters, with and without equalize.cluster.weights), forest outputs held fixed | R 4.5.2; grf 2.6.1; sandwich 3.1.1 | coef, se, t 1e-10 rel without clusters, 1e-9 rel with; one-sided p-value 1e-8 rel | — / — | test_ml_causal_R_parity.py (+1) |
group_sequential |
Stata 18 gsbounds (intpointsscale(60) infotol(1e-12)): 117 designs over every boundary family, one- and two-sided, efficacy, futility, binding and nonbinding | Stata Stata 18 MP | 3e-7 relative on every bound and on the maximum information ratio (test bound 1e-6) | — / — | test_group_sequential_stata_parity.py (+2) |
gsynth |
gsynth::gsynth | R 4.5.2; gsynth 1.4.0 | rel_est<=1e-06, rel_se<=1e-06 | 7.7e-14 / 2.2e-15 | 19_gsynth.py (+2) |
gwr |
GWmodel::gwr.basic 2.4.1 | R 4.5.2; GWmodel 2.4.1 | local betas and SEs 1e-9 rel (observed 1.3e-11 / 3.5e-14); RSS, AIC, AICc, BIC, enp, edf 1e-10 rel | — / — | test_spatial_survey_R_parity.py (+1) |
gwr_bandwidth |
GWmodel::bw.gwr 2.4.1 (golden section gold(); gwr.aic / gwr.cv) | R 4.5.2; GWmodel 2.4.1 | selected bandwidth 1e-10 rel (observed 0 on 8 configurations); criterion values 1e-11 rel (observed 7.0e-15) | — / — | test_spatial_survey_R_parity.py (+1) |
harvest_did |
R did::att_gt(control_group='notyettreated', base_period='universal') for every (cohort, horizon) cell; did::aggte(type='dynamic') for the event study under weighting='n_treated' | did 2.5.1; DRDID 1.3.0 | Every 2x2 cell (ATT and SE) and every event-study horizon (ATT and SE) at 1e-9 relative; observed agreement 1.7e-14. The inverse-variance aggregate over horizons (the default headline) has no reference and is checked by identity only. | — / — | test_did_synth_misc_parity.py |
hdfe_ols |
fixest::feols | R 4.5.2; fixest 0.14.0 | rel_est<=1e-06, rel_se<=1e-06 | 5.2e-15 / 2.9e-15 | 03_hdfe.py (+3) |
heckman |
sampleSelection::heckit | R 4.5.2; sampleSelection 1.2.14 | rel_est<=1e-06, rel_se<=1e-06 | 1.0e-11 / 1.0e-11 | 43_heckman.py (+2) |
heckprobit |
Stata heckprobit (oim, vce(robust), vce(cluster), [pw=], [pw=] with vce(cluster)); R sampleSelection::selection(method = 'ml') with a binary outcome, Newton-Raphson and BHHH | Stata 18; R 4.5.2; sampleSelection 1.2.14 | against Stata: coefficients and standard errors 1e-9 relative (observed 4e-11); against sampleSelection's Newton-Raphson fit 1e-8 (observed 2e-9); log likelihood 1e-12 | — / — | test_regcmd_stata_r_parity.py (+2) |
heidel_diag |
coda::heidel.diag | R 4.5.2; coda 0.19.4.1 | stationarity decision and start equal; p-value, mean and half-width 1e-9 rel | — / — | test_bayes_mcmc_parity.py (+1) |
het_test |
lmtest::bptest (studentized Breusch-Pagan) | R 4.5.2; lmtest 0.9.40 | statistic & p-value 1e-10 rel (observed ~1e-13) | — / — | test_diagnostics_parity.py (+1) |
heterogeneity_of_effect |
R metafor::rma(method = 'DL') | metafor 5.0.1 | 1e-12 rel | — / — | test_misc_sens_R_parity.py |
hill_climb |
R bnlearn::score(type = "bic" / "bic-g") and hc() | R 4.5.2; bnlearn 5.2.1 | BIC of eight fixed graphs 1e-12 rel; the score hc() ends at 1e-10 rel on the two fixture data sets | — / — | test_hill_climb_bnlearn_parity.py (+1) |
holm |
base R stats::p.adjust(method='holm') | R 4.5.2 | exact (atol 1e-15; observed 0) | — / — | test_mht_parity.py (+1) |
honest_did |
HonestDiD::createSensitivityResults (exact quantile) + closed form at M = 0 | R 4.5.2; HonestDiD 0.2.8 | abs_est<=1e-06, abs_se<=1e-06 | 1.4e-08 / 9.4e-09 | 10_honest_did.py (+2) |
horowitz_manski |
Stata tebounds 1.8 worst-case bounds (identical estimand for ATE worst-case bounds) | Stata 18; tebounds 1.8 (SJ15-2 st0386) | 1e-12 rel / 1e-15 abs on both bounds | — / — | test_teffects_R_parity.py (+2) |
hpd_interval |
coda::HPDinterval | R 4.5.2; coda 0.19.4.1 | 1e-9 rel at 95 and 80 percent | — / — | test_bayes_mcmc_parity.py (+1) |
hurdle |
pscl::hurdle(dist='poisson', zero.dist='binomial') | R 4.5.2; pscl 1.5.9 | count + zero coefficients 1e-6 abs (observed ~2e-8) | — / — | test_glm_ext_parity.py (+1) |
icc |
Stata 18 estat icc after mixed (ML, REML) and melogit; performance::icc; psych::ICC (balanced ANOVA identity) | Stata 18 MP; R 4.5.2; performance 0.16.0; psych 2.6.5 | estimate / SE rtol 1e-6, logit-scale CI rtol 2e-6 (observed <= 2.4e-7 / 7e-7 / 1.2e-6); balanced REML ICC = ANOVA ICC(1) rtol 1e-9 | — / — | test_panel_icc_lrtest_parity.py |
icp |
InvariantCausalPrediction::ICP(test = 'normal', selection = 'all') | R 4.5.2; InvariantCausalPrediction 0.8 | accepted subsets and parents exact on 32 fits; 1e-9 absolute on the parents' confidence intervals (observed 5e-15) | — / — | test_icp_R_parity.py (+1) |
ida |
pcalg::ida(method = 'local' / 'global') | R 4.5.2; pcalg 2.7.12 | 1e-10 on the sorted multiset of possible effects | — / — | test_ida_pcalg_parity.py (+1) |
identify |
causaleffect::causal.effect (Shpitser-Pearl ID) for identifiability; base R table arithmetic for the value of the estimand | R 4.5.2; causaleffect 1.3.15 | identifiable or not: equal on seven graphs (backdoor, bow, front door, instrument, napkin, M-bias, broken front door); .estimate(data): 1e-10 rel against the truncated factorisation | — / — | test_ness_causal_ai_parity.py (+1) |
impacts |
R spatialreg::impacts on a lagsarlm fit | spdep 1.4.2; spatialreg 1.4.3 | Direct, indirect and total at 1e-6, inheriting the SAR rho's own agreement. | — / — | test_spdep_parity.py |
implied_weights |
lmw::lmw (URI; MRI for ATE, ATT and ATC) | lmw 0.0.2; R 4.5.2 | 1e-9 relative (observed 7e-13) | — / — | test_matching_balance_r_parity.py (+2) |
incidence_rate_ratio |
base-R closed form (rate ratio + conditional-binomial exact CI) | R 4.5.2 | estimate 1e-12; exact CI 1e-10 abs (observed ~3e-15) | — / — | test_epi_parity.py (+1) |
indirect_standardize |
Stata istdize (exact CI); R epitools::ageadjust.indirect (log-normal CI) | R 4.5.2; epitools 0.5.10.1; Stata 18 MP | 1e-10 rel (observed 6e-16) | — / — | test_survival_epi_R_parity.py (+2) |
inequality_index |
base-R closed form (Gini/Theil-T/Theil-L/Atkinson; = ineq) | R 4.5.2 | all indices 1e-12 abs (observed ~2e-16) | — / — | test_inequality_parity.py (+1) |
influence_measures |
Stata 18 predict after regress (rstandard, rstudent, cooksd, dfits, welsch, covratio, stdp, stdf, stdr) and dfbeta | Stata 18 | mean and sd of each statistic over the sample 1e-9 rel; DFBETAs 1e-5 rel | — / — | test_kohler_kreuter_stata_parity.py (+1) |
interactive_fe |
Stata regife (SSC, Gomez) ..., noconstant; R phtt::Eup(additive.effects = 'none') | Stata 18 MP; regife 2026-03-30 SSC; R 4.5.2; phtt 3.1.2 | slopes rtol 1e-9 vs both (observed <= 2e-11); SEs rtol 1e-9 vs regife with dof='regife' (homoskedastic and cluster), phtt SE reconstructed rtol 1e-9 | — / — | test_panel_ife_parity.py |
interflex |
interflex::interflex(vartype="delta", vcov.type="robust", neval=5, nbins=3, bw=1); Stata side uses the SSC interflex command | R 4.5.2; interflex 1.4.0 | rel_est<=1e-06, rel_se<=1e-06 | 4.0e-15 / 1.5e-14 | 87_interflex.py (+2) |
intreg |
Stata intreg (oim, vce(robust), vce(cluster), [pw=], [pw=] with vce(cluster)); R survival::survreg(Surv(type = 'interval2'), dist = 'gaussian') with its observed-information covariance | Stata 18; R 4.5.2; survival 3.8.3 | coefficients and standard errors 1e-9 relative (observed 5e-11); log likelihood 1e-12 | — / — | test_regcmd_stata_r_parity.py (+2) |
inward_outward_spillover |
stats::lm classical covariance and Stata regress (inward and outward exposure coefficients) | R 4.5.2; ivreg 0.6.7; sandwich 3.1.1; Stata Stata 18 MP (07 Jun 2023) | 1e-9 relative on estimates and SEs (observed 6.7e-15 R, 9.7e-15 Stata) | — / — | test_interference_r_parity.py (+2) |
ipcw |
survival::coxph(ties = 'breslow') + basehaz(centered = FALSE) | R 4.5.2; survival 3.8.3 | 1e-9 rel on every weight (observed 2.0e-11) | — / — | test_teffects_R_parity.py (+2) |
ips |
Open Bandit Pipeline (obp) 0.5.7 InverseProbabilityWeighting (lambda_ inf and 2) | obp 0.5.7 | value 1e-12 rel (observed 0.0); SE identity 1e-10 | — / — | test_ml_causal_obp_parity.py (+1) |
ipw |
base R stats::glm(binomial) + hand-rolled Hajek weighted means | — | Hajek ATE/ATT estimate 1e-9 (observed <= 2e-15; SE not pinned) | — / — | test_ipw_parity.py (+1) |
irf |
vars::irf 1.6.1; Stata 18 irf create | R 4.5.2; vars 1.6.1; Stata 18 | 1e-10 rel vs vars, 1e-9 vs Stata irf file (observed 2.6e-14) | — / — | test_timeseries_R_parity.py (+2) |
itcv |
Stata 18 pkonfound (ITCV and RIR); the formula of Frank (2000) evaluated in R | R 4.5.2; Stata 18 | threshold, correlations and benchmark impacts 1e-9 rel (R); 1e-6 rel (pkonfound, fed rounded inputs) | — / — | test_accounting_research_parity.py (+2) |
its |
lm + sandwich::NeweyWest 3.1.1; Stata 18 newey; itsa 1.0.0 (SSC) | R 4.5.2; sandwich 3.1.1; Stata 18; itsa 1.0.0 | coefficients and Newey-West SE 1e-10 rel (observed 7.9e-14); itsa 1e-6 (glm2 IRLS) | — / — | test_timeseries_R_parity.py (+2) |
iv |
AER::ivreg | R 4.5.2; AER 1.2.16 | rel_est<=1e-06, rel_se<=1e-06 | 1.1e-11 / 1.1e-11 | 02_iv.py (+3) |
iv_compare |
R ivmodel KClass(k = 1) / LIML / Fuller(b = 1); Stata ivregress 2sls / liml, ivreg2 fuller(1), jive ujive1, estat firststage, weakivtest | R 4.5.2; ivmodel 1.9.1; Stata 18 MP; ivreg2 4.1.12; jive 1.0.2 | 1e-9 relative on every estimate, standard error and F statistic (observed <= 1e-11) | — / — | test_iv_compare_parity.py (+3) |
iv_diag |
R ivDiag::ivDiag 1.0.6 (analytic block) | R 4.5.2; ivDiag 1.0.6; lfe 3.1.1 | 2SLS / OLS coefficients and SEs, classical first-stage F, effective F, tF critical value and interval: rel 1e-9 on 6 designs (observed <= 6e-12) | — / — | test_rd_iv_R_parity.py (+1) |
iv_mediate |
Stata ivmediate 1.0.4 (Dippel, Ferrara and Heblich), SSC | Stata 18.0 MP; ivmediate 1.0.4 | 1e-9 rel (coefficients and F 1e-14, 3x3 V 1.6e-11) | — / — | test_iv_mediate_stata_parity.py (+1) |
ivpoisson |
Stata 18 MP official ivpoisson gmm |
Stata 18 MP | Every block at 1e-6 on coefficients, standard errors and Hansen's J. Observed 5e-15 to 2e-7 across additive and multiplicative errors, one-step, two-step and iterated GMM, exact and over-identification, and robust, cluster and unadjusted weight matrices, including wmatrix() set apart from vce(). The residual is Stata's stopping rule for the GMM criterion. | — / — | test_ivpoisson_stata_parity.py |
ivprobit |
Stata 18 MP official ivprobit (ML and Newey two-step) |
Stata 18 MP | Every block at 1e-6. Observed: Newey two-step coefficients and standard errors 3e-10 to 2e-8 with one and with two endogenous regressors; maximum likelihood coefficients 9e-9 or better, standard errors 4e-8 or better under vce(oim), vce(robust) and vce(cluster), log-likelihood to 1e-8; the Wald test of exogeneity and the model Wald chi2 at the same level. The do-file tightens Stata's ml stopping rule (nrtolerance 1e-13): at Stata's default the ML blocks differ by up to 3e-6, which is the stopping rule and not the likelihood. | — / — | test_ivprobit_ivtobit_stata_parity.py |
ivreg |
AER::ivreg | R 4.5.2; AER 1.2.16 | rel_est<=1e-06, rel_se<=1e-06 | 1.1e-11 / 1.1e-11 | 02_iv.py (+2) |
ivtobit |
Stata 18 MP official ivtobit (ML and Newey two-step) |
Stata 18 MP | Every block at 1e-6. Observed: Newey two-step 4e-13 to 1e-11 (left-censored, two-limit, two endogenous regressors); maximum likelihood coefficients 8e-14 or better and standard errors 8e-8 or better under vce(oim), vce(robust) and vce(cluster), log-likelihood to 1e-8; both Wald statistics at the same level. Stata's ml stopping rule is tightened in the do-file (see ivprobit). | — / — | test_ivprobit_ivtobit_stata_parity.py |
jackknife_se |
R sandwich::vcovJK(center='mean'); Stata regress, vce(jackknife, cluster() double) | R 4.5.2; sandwich 3.1.1; Stata 18 | SE / CI 1e-10 rel, p 1e-9 (observed SE 2.0e-15, p 1.4e-14, CI 9.2e-12) | — / — | test_inference_sens_R_parity.py (+3) |
jive |
Stata jive 1.0.2 (Stata Journal st0108) ujive1 / ujive2 | Stata 18 MP; jive 1.0.2 | coefficients and SEs (default and robust) rel 1e-9 (observed 3.0e-13 / 4.4e-13) | — / — | test_rd_iv_R_parity.py (+1) |
johansen |
urca::ca.jo 1.3.4; Stata 18 vecrank | R 4.5.2; urca 1.3.4; Stata 18 | eigenvalues, trace & max-eigenvalue statistics 1e-11 rel vs ca.jo, 1e-10 vs vecrank (observed 8.2e-14); Osterwald-Lenum table equal to Stata _vecgetcv cell by cell | — / — | test_timeseries_R_parity.py (+2) |
johansen_lrtest |
urca::blrtest, bh5lrtest, alrtest | R 4.5.2; urca 1.3.4 | LR statistic 1e-8 rel, degrees of freedom equal, restricted vectors 1e-6 | — / — | test_johansen_lrtest_parity.py (+1) |
join_counts |
R spdep::joincount.multi (binary weights) | spdep 1.4.2; spatialreg 1.4.3 | BB, WW and BW all exact. A reference-free guard also asserts BB + WW + BW = S0/2, the identity the BW defect violated (70.75 against 50). | — / — | test_spdep_parity.py |
jwdid |
Stata 18 jwdid ..., method(ppmlhdfe) + estat simple | Stata 18 MP; jwdid 2.2; ppmlhdfe 2.3.3 | Link-scale simple ATT 1e-9 and its SE 1e-7 relative (observed <= 3e-14 and <= 3e-11); response-scale ATT and its default (response_se='margins') SE 1e-6 (Stata margins takes numerical derivatives; observed <= 2e-7). Sixteen specifications: every hettype(), never, i.xcat / xc covariates, over(xcat), and exovar(i.year#i.xcat), never exovar(i.year#i.xcat), exovar(c.xc#i.xcat). | — / — | test_jwdid_Stata_parity.py |
kalman_filter |
KFAS::KFS, logLik; statsmodels KalmanSmoother | R 4.5.2; KFAS 1.6.0; statsmodels 0.14.6 | log likelihood, predicted, filtered and smoothed states and covariances 1e-9 rel | — / — | test_statespace_parity.py (+1) |
kaplan_meier |
survival::survfit | R 4.5.2; survival 3.8.3 | S(t) at every event time 1e-12 (observed ~3e-17); median exact | — / — | test_survival_km_parity.py (+1) |
karate_club |
R igraph::make_graph('Zachary') | igraph 2.3.3; sna 2.8; ergm 4.12.0; dyadRobust 0.0.1.0001 | Adjacency matrix identical. | — / — | test_network_parity.py |
katz_centrality |
R igraph::alpha_centrality | igraph 2.3.3; sna 2.8; ergm 4.12.0; dyadRobust 0.0.1.0001 | 7.1e-16 with normalized = False (normalized = True L2-scales the same vector). | — / — | test_network_parity.py |
kdensity |
Stata kdensity (at(), bwidth(), kernel()); R bw.nrd0 / bw.SJ | R 4.5.2; Stata 18 MP | density and default widths 1e-10 rel (observed 9e-16); Sheather-Jones 1e-6 vs bw.SJ(nb=1e7, tol=1e-14) (observed 2e-7, R's pair-count binning) | — / — | test_survival_epi_R_parity.py (+2) |
kernel_weights |
spdep::nb2listwdist(type = 'dpd', alpha = 2) on dnearneigh(0, h); GWmodel::gw.weight | R 4.5.2; spdep 1.4.2; GWmodel 2.4.1 | weights 1e-12 rel / 1e-15 abs (observed 6.9e-16 rel, 3.3e-16 abs) | — / — | test_spatial_survey_R_parity.py (+1) |
kitagawa_decompose |
R DasGuptR::dgnpop with ratefunction sum(size*rate)/sum(size) | DasGuptR 2.2.0; ddecompose 1.0.0; cdgd 1.0.1 | Rate and composition effects on Das Gupta's Table 5.1 at 1e-10; interaction exactly 0. | — / — | test_decomp_R_parity.py |
knn_weights |
R spdep::knearneigh + knn2nb | spdep 1.4.2; spatialreg 1.4.3 | Neighbour sets identical for all 120 points, k=4, on a random point set chosen so no distance ties make the answer non-unique. | — / — | test_spdep_parity.py |
krippendorff_alpha |
Stata kalpha 1.3.0 and kappaetc 2.1.0; R irr::kripp.alpha 0.85 | Stata 18.0 MP; kalpha 1.3.0; kappaetc 2.1.0; R 4.5.2; irr 0.85 | 1e-10 rel (observed 2e-16), four levels with missing codings | — / — | test_krippendorff_alpha_parity.py (+2) |
ksmirnov |
Stata 18 ksmirnov | Stata 18 | D statistics and p-values 1e-9 rel | — / — | test_kohler_kreuter_stata_parity.py (+1) |
ktau |
Stata 18 ktau | Stata 18 | tau-a, tau-b, score, its standard error and p-value 1e-9 rel | — / — | test_kohler_kreuter_stata_parity.py (+1) |
kwallis |
Stata 18 kwallis | Stata 18 | chi-squared with and without ties 1e-5 rel | — / — | test_kohler_kreuter_stata_parity.py (+1) |
l1_imbalance |
cem::imbalance(breaks =) global L1, unweighted and weighted | cem 1.1.31; R 4.5.2 | 1e-12 relative | — / — | test_matching_balance_r_parity.py (+1) |
lee_bounds |
Stata leebounds 1.5 (Tauchmann), vce(analytic) | Stata 18; leebounds 1.5 (2013-07-17, Tauchmann) | 1e-10 rel on bounds and analytic variances | — / — | test_teffects_R_parity.py (+3) |
liml |
ivmodel::LIML | R 4.5.2; ivmodel 1.9.1 | rel_est<=1e-06, rel_se<=1e-06 | 1.7e-15 / 9.7e-16 | 59_liml.py (+2) |
linear_calibration |
survey::calibrate(calfun='linear', unbounded); Stata svycal regress | R 4.5.2; survey 4.5; Stata 18 MP | calibrated weights 1e-12 rel (observed 5.5e-15) | — / — | test_survey_calib_R_parity.py (+2) |
lingam |
lingam.DirectLiNGAM (Python, measure 'pwling'): causal order | python 3.11.15; lingam 1.13.0; numpy 2.2.6 | exact: the causal order is a permutation, equal on 40 data sets | — / — | test_lingam_package_parity.py (+1) |
lm_tests |
R spdep::lm.RStests | spdep 1.4.2; spatialreg 1.4.3 | All five statistics and their p-values at 1e-9. Before the fix: LM_err 39.47 against 19.58, and Robust_LM_err 20.49 (p=6e-6) against 0.0397 (p=0.84) -- the Anselin lag-vs-error decision rule, reversed. | — / — | test_spdep_parity.py |
lmtp |
lmtp::lmtp_tmle and lmtp_sdr (SL.glm, folds = 1) | R 4.5.2; lmtp 1.5.4 | 3e-7 relative on estimate and SE (observed <= 1.9e-7 on the binary TMLE rows); density ratios 1e-11 | — / — | test_lmtp_parity.py (+1) |
local_projections |
lpirfs::lp_lin | R 4.5.2; lpirfs 0.2.5 | rel_est<=1e-06, rel_se<=1e-06 | 5.0e-15 / 4.4e-15 | 34_lp.py (+2) |
logit |
stats::glm(family=binomial("logit")) | R 4.5.2; stats 4.5.2 | rel_est<=1e-06, rel_se<=1e-06 | 2.7e-11 / 2.7e-11 | 57_logit.py (+2) |
logit_gof |
Stata 18 estat gof and estat gof, group() | Stata 18 | chi-squared and degrees of freedom 1e-5 rel | — / — | test_kohler_kreuter_stata_parity.py (+1) |
logit_influence |
Stata 18 predict after logit (residuals, hat, rstandard, deviance, dx2, ddeviance, dbeta) | Stata 18 | mean and sd of each statistic over the sample 1e-5 rel | — / — | test_kohler_kreuter_stata_parity.py (+1) |
logrank_test |
survival::survdiff | R 4.5.2; survival 3.8.3 | chi-square 1e-10 rel (observed ~8e-16); p-value 1e-10 abs | — / — | test_survival_km_parity.py (+1) |
longitudinal_analyze |
IPW path: ipw::ipwpoint weights + lm(Y ~ 1, weights) + sandwich HC0; g-formula path: ltmle::ltmle(gcomp = TRUE) with SL.lm (iterated conditional expectations with linear regressions on the full history) | R 4.5.2; ipw 1.3.0; ltmle 1.3.0; sandwich 3.1.1; SuperLearner 2.0.40 | IPW path 1e-9 relative on estimate and SE (observed 2.6e-15); g-formula path 1e-9 on the estimate (observed 1.9e-11) | — / — | test_tte_longitudinal_r_parity.py (+1) |
longitudinal_contrast |
IPW path: lm(Y ~ A, weights = ipwpoint weights) + sandwich HC0; g-formula path: ATE of ltmle::ltmle(gcomp = TRUE, SL.lm) | R 4.5.2; ipw 1.3.0; ltmle 1.3.0; sandwich 3.1.1; SuperLearner 2.0.40 | IPW path 1e-9 relative on contrast and SE (observed 2e-15); g-formula path 1e-9 on the contrast (observed 8.4e-12) | — / — | test_tte_longitudinal_r_parity.py (+1) |
loo |
loo::loo on a log-likelihood matrix (pointwise elpd, p, Pareto k, n_eff, smoothed weights, estimates) | R 4.5.2; loo 2.9.0 | 1e-9 abs on pointwise elpd, p, Pareto k and log weights; 1e-8 on the estimates; 1e-9 rel on n_eff | — / — | test_regression_stories_r_parity.py (+1) |
loo_compare |
loo::loo_compare (elpd_diff, se_diff, ordering) | R 4.5.2; loo 2.9.0 | 1e-8 abs | — / — | test_regression_stories_r_parity.py (+1) |
lp_did |
direct transcription (no LP-DiD R package installed); Stata side uses the authors' lpdid | R 4.5.2 | rel_est<=1e-10, rel_se<=1e-10 | 5.0e-15 / 2.5e-15 | 83_lpdid.py (+2) |
lpbwselect_ce_rot |
R nprobust::lpbwselect(bwselect = 'ce-rot') (black box); Stata nprobust lpbwselect as a cross-check | R 4.5.2; nprobust 1.0.0; Stata 18 MP | 1e-9 relative on h and b | — / — | test_lprobust_cleanroom.py (+2) |
lpbwselect_imse_dpi |
R nprobust::lpbwselect(bwselect = 'imse-dpi') (black box); Stata nprobust lpbwselect as a cross-check | R 4.5.2; nprobust 1.0.0; Stata 18 MP | 1e-9 relative on h and b | — / — | test_lprobust_cleanroom.py (+2) |
lpbwselect_imse_rot |
R nprobust::lpbwselect(bwselect = 'imse-rot') (black box); Stata nprobust lpbwselect as a cross-check | R 4.5.2; nprobust 1.0.0; Stata 18 MP | 1e-9 relative on h (1e-8 at p = 3, where the reference's own degree-6 global fit rounds at 4e-9) | — / — | test_lprobust_cleanroom.py (+2) |
lpbwselect_mse_dpi |
R nprobust::lpbwselect(bwselect = 'mse-dpi') (black box); Stata nprobust lpbwselect as a cross-check | R 4.5.2; nprobust 1.0.0; Stata 18 MP | 1e-9 relative on h and b | — / — | test_lprobust_cleanroom.py (+2) |
lpbwselect_mse_rot |
R nprobust::lpbwselect(bwselect = 'mse-rot') (black box); Stata nprobust lpbwselect as a cross-check | R 4.5.2; nprobust 1.0.0; Stata 18 MP | 1e-9 relative on h (1e-8 at p = 3, where the reference's own degree-6 global fit rounds at 4e-9) | — / — | test_lprobust_cleanroom.py (+2) |
lpoly |
Stata lpoly (at(), bwidth(), degree(), kernel(), se(), pwidth()) | Stata 18 MP | 1e-10 rel (observed 1.6e-12) | — / — | test_survival_epi_R_parity.py (+2) |
lprobust_at_point |
R nprobust::lprobust and Stata nprobust lprobust at supplied bandwidths (vce nn) | R 4.5.2; nprobust 1.0.0; Stata 18 MP | 1e-9 relative on tau_us, tau_bc, se_us, se_rb against R (192 cases); 1e-9 absolute against Stata (10 boundary cases); effective sample size exact | — / — | test_lprobust_cleanroom.py (+3) |
lrtest |
Stata 18 lrtest; R anova() on lme4 ML fits | Stata 18 MP; R 4.5.2; lme4 2.0.1 | chi2 rtol 1e-6, df exact, p rtol 1e-5 (observed chi2 <= 1e-9) | — / — | test_panel_icc_lrtest_parity.py |
lrvar |
sandwich::kernHAC, bwAndrews, bwNeweyWest, NeweyWest, lrvar | R 4.5.2; sandwich 3.1.1 | long-run variance and bandwidth 1e-9 rel over 104 kernel x bandwidth x prewhitening cases | — / — | test_lrvar_parity.py (+1) |
ltmle |
ltmle::ltmle 1.3-0 (glm, gbounds c(0.01, 1), variance.method = 'ic') | R 4.5.2; ltmle 1.3.0 | psi1 / psi0 / ATE / SE 1e-9 rel without censoring (observed 5e-13); 1e-8 with censoring nodes (observed 5e-10) | — / — | test_teffects_R_parity.py (+2) |
ltmle_msm |
ltmle::ltmleMSM (glm nuisance fits, variance.method = 'ic') | R 4.5.2; ltmle 1.3.0 | 1e-8 absolute on TMLE coefficients, 1e-8 relative on their SEs; IPTW coefficients 1e-11 | — / — | test_ltmle_msm_parity.py (+1) |
ltmle_survival |
ltmle::ltmle(survivalOutcome = TRUE, variance.method = 'ic'), one call per horizon | R 4.5.2; ltmle 1.3.0 | 1e-8 absolute on the counterfactual risks (observed 1.6e-9), 1e-8 relative on their SEs and on the SE of the risk difference | — / — | test_ltmle_survival_R_parity.py (+1) |
manski_bounds |
Stata tebounds 1.8 (SJ15-2 st0386), erates(0) | Stata 18; tebounds 1.8 (SJ15-2 st0386) | 1e-12 rel / 1e-15 abs on both bounds | — / — | test_teffects_R_parity.py (+2) |
mantel_haenszel |
base-R closed form (Robins-Breslow-Greenland MH; = epiR) | R 4.5.2 | estimate, se_log, CI 1e-12 abs (observed 0) | — / — | test_epi_parity.py (+1) |
margins_at |
Stata 18 margins, at(...); R marginaleffects::avg_predictions | Stata 18; R 4.5.2; marginaleffects 0.32.0 | 1e-10 rel (observed <= 3.5e-15 est / SE; marginaleffects SE 1.8e-10, numerical Jacobian) | — / — | test_r2_postest_parity.py (+2) |
markup |
Stata markupest 1.0.1 (Rovigatti), method(dlw) pmethod(lp) | Stata 18; markupest 1.0.1 10May2020 | 1e-10 rel (observed 8.1e-13 corrected, 3.4e-15 uncorrected) | — / — | test_frontier_struct_R_parity.py (+1) |
match |
MatchIt::matchit 4.7.2 (nearest, glm/logit distance) | — | 1:1 and 2:1 PS matching without replacement: rel <= 1e-9. Mahalanobis metric pinned against MatchIt:::mahalanobis_dist (rel <= 1e-10); greedy m_order='data'/'closest' rel <= 1e-9. | — / — | test_matching_r_parity.py (+1) |
mc_panel |
MCPanel::mcnnm_fit (Athey, Bayati, Doudchenko, Imbens & Khosravi; github.com/susanathey/MCPanel) and fect::fect(method = "mc") | R 4.5.2; MCPanel 0.0 @ 6b2706fd7c35f3266048ceb22a7e9a61ae1774da; fect 2.4.1; gsynth 1.4.0 | ATT and fitted untreated matrix 1e-9 rel at fixed lambda (observed <= 2.6e-13), four fixed-effect modes x two lambdas | — / — | test_did_synth_mc_parity.py (+1) |
mc_synth |
MCPanel::mcnnm_fit (Athey, Bayati, Doudchenko, Imbens & Khosravi; github.com/susanathey/MCPanel) and fect::fect(method = "mc") | R 4.5.2; MCPanel 0.0 @ 6b2706fd7c35f3266048ceb22a7e9a61ae1774da; fect 2.4.1 | ATT and fitted untreated matrix 1e-9 rel at fixed lambda (observed <= 6.0e-13), two-way and no-FE x two lambdas | — / — | test_did_synth_mc_parity.py (+1) |
mccrary_test |
R rdd::DCdensity 0.57 (CRAN archive) | R 4.5.2; rdd 0.57 | theta, se, z rel 1e-9; p rel 1e-8; bin width 1e-12 (observed 6.5e-12) | — / — | test_rd_iv_rd_R_parity.py (+1) |
mcmc_ess |
coda::effectiveSize, coda::spectrum0.ar | R 4.5.2; coda 0.19.4.1 | 1e-9 rel | — / — | test_bayes_mcmc_parity.py (+1) |
mcmc_summary |
summary(coda::mcmc(x)) | R 4.5.2; coda 0.19.4.1 | mean, sd, naive and time-series SE, quantiles 1e-9 rel | — / — | test_bayes_mcmc_parity.py (+1) |
mde |
base-R closed form (RCT minimum detectable effect) | R 4.5.2 | effect size 1e-6 abs (output rounded to 6 dp; observed ~2e-8) | — / — | test_power_extra_parity.py (+1) |
median_test |
Stata 18 median | Stata 18 | chi-squared, its continuity-corrected form and p-value 1e-9 rel | — / — | test_kohler_kreuter_stata_parity.py (+1) |
mediate |
mediation::mediate | R 4.5.2; mediation 4.5.1 | rel_est<=1e-06, rel_se<=1e-06 | 6.7e-15 / 3.6e-15 | 36_mediation.py (+3) |
mediate_interventional |
CMAverse::cmest 0.1.0 (gformula with postc: rpnde / rpnie / te; rb paramfunc without) | R 4.5.2; CMAverse 0.1.0 | IIE / IDE / total 1e-9 rel (observed 2.3e-15) | — / — | test_teffects_R_parity.py (+2) |
mediate_sensitivity |
R mediation::medsens (lm/lm, rho.by = 0.1) | mediation 4.5.1 | 1e-10 rel | — / — | test_misc_sens_R_parity.py |
mediation |
mediation::mediate | R 4.5.2; mediation 4.5.1 | rel_est<=1e-06, rel_se<=1e-06 | 6.7e-15 / 3.6e-15 | 36_mediation.py (+2) |
mediation_decompose |
Stata paramed (Liu & Emsley, SSC), yreg(linear) mreg(linear) with interaction | Stata 18.0 MP; paramed SSC | CDE/NDE/NIE/total and delta-method SEs 1e-12 rel (observed 9.5e-15) | — / — | test_decomp_qte_parity.py (+1) |
megamma |
glmmTMB 1.1.14 Gamma(link = 'log') (Laplace, AD Hessian); Stata 18 meglm, family(gamma) link(log) intmethod(laplace) | R 4.5.2; glmmTMB 1.1.14; TMB 1.9.25; Stata 18 MP | vs glmmTMB estimates / SEs rtol 1e-6 (observed 6.9e-11 / 3.8e-8); vs Stata Laplace estimates rtol 1e-6 (observed 2.2e-7); objective identity rtol 1e-11 | — / — | test_panel_glmm_parity.py |
meglm |
Stata 18 meglm (gaussian; binomial with binomial()); lme4::lmer(REML = FALSE); lme4::glmer | Stata 18 MP; R 4.5.2; lme4 2.0.1 | gaussian vs Stata estimates / SEs rtol 1e-6 (observed 4.2e-9 / 2.8e-7), vs lmer ML estimates 6.1e-11; binomial-trials vs Stata 2.4e-9 / 6.6e-8 | — / — | test_panel_glmm_parity.py |
melly_decompose |
Stata cdeco, method(qr) 1.0.2 (Chernozhukov, Fernandez-Val & Melly; bmelly/Stata counterfactual) | Stata 18.0 MP; cdeco 1.0.2 01mar2023 (bmelly/Stata counterfactual, Distribution-Date 20220803) | fitted and counterfactual quantiles 1e-10 rel (observed 6.2e-16); QR process coefficients 1e-12 abs (observed 7.5e-15) | — / — | test_decomp_qte_parity.py (+1) |
menbreg |
glmmTMB 1.1.14 nbinom2 (Laplace, AD Hessian); Stata 18 menbreg intmethod(laplace) / mcaghermite(7) | R 4.5.2; glmmTMB 1.1.14; TMB 1.9.25; lme4 2.0.1; Stata 18 MP | vs glmmTMB estimates / SEs rtol 1e-6 (observed 1.7e-11 / 2.3e-7); vs Stata Laplace estimates rtol 1e-6 (observed 7.5e-8); objective identity at Stata's and glmmTMB's estimates rtol 1e-11 | — / — | test_panel_glmm_parity.py |
mendelian_randomization |
R MendelianRandomization::mr_allmethods(method = 'main') | MendelianRandomization 0.10.0 | 1e-12 rel | — / — | test_misc_sens_R_parity.py |
meologit |
Stata 18 meologit intmethod(mcaghermite) intpoints(7) / intmethod(laplace); ordinal::clmm | Stata 18 MP; R 4.5.2; ordinal 2025.12.29 | AGHQ-7 vs Stata estimates rtol 1e-6 (observed 3.8e-7), SEs incl. cutpoints rtol 2e-6 (observed 1.3e-6; 5.7e-7 at Stata's estimates); objective identity vs Stata and clmm rtol 1e-11 (observed <= 1.4e-11) | — / — | test_panel_glmm_parity.py |
mepoisson |
Stata 18 mepoisson, intmethod(laplace) / intmethod(mcaghermite) intpoints(7); lme4::glmer(nAGQ = 1 / 7) | Stata 18 MP; R 4.5.2; lme4 2.0.1 | objective identity at Stata's estimates rtol 1e-11 (observed 6.6e-12); Laplace estimates / SEs vs Stata rtol 1e-6 (observed 9.0e-11 / 1.2e-7); AGHQ-7 vs lme4 rtol 1e-6 (observed 4.4e-8 / 9.3e-8) | — / — | test_panel_glmm_parity.py |
meta_analysis |
R metafor::rma (method='FE' and 'DL') | ivmodel 1.9.1; car 3.1.5; metafor 5.0.1 | 6.2e-16 across all nine reported quantities: fixed and random pooled effect and standard error, tau^2, Cochran Q and its p-value, I^2 and H^2. REML is not implemented here, which is a capability gap rather than a disagreement. | — / — | test_weakiv_meta_parity.py |
metalearner |
econml.metalearners SLearner / TLearner / XLearner | — | S / T / X conditional-average-treatment-effect vectors match econml elementwise to 1e-12 absolute (observed <= 1.1e-15) when both fitting stages use the same base learner | — / — | test_metalearner_econml_parity.py |
mgwr |
PySAL mgwr 2.2.1 Sel_BW(multi=True).search() + MGWR.fit() (authors' implementation); fixed-bandwidth back-fitting also GWmodel::gwr.multiscale 2.4.1 | python 3.13.9; mgwr 2.2.1; numpy 2.2.6; R 4.5.2; GWmodel 2.4.1 | bandwidths, initial bandwidth, bandwidth history and iteration count exact; betas 1e-9 rel (observed 3.0e-12); SEs, ENP_j, tr(S), sigma2, AICc/AIC/BIC 1e-10 rel (observed <= 1e-15); SOC path 1e-9 (observed 8e-13) | — / — | test_r2_spatial_parity.py (+3) |
mi_estimate |
R mice::pool; Stata mi estimate: regress | mice 3.19.0; Stata 18 | 1e-9 rel (p, CI via t quantiles at fractional df); 1e-12 est/SE/df | — / — | test_misc_sens_R_parity.py (+1) |
mixed |
lme4::lmer | R 4.5.2; lme4 2.0.1 | rel_est<=1e-06, rel_se<=1e-06 | 2.6e-12 / 8.3e-11 | 25_lmm.py (+2) |
mixlogit |
Stata mixlogit 1.4.0 (SSC, Hole), nrep(50) burn(15), on identical Halton draws | Stata 18 MP; mixlogit 1.4.0 | means / SDs / Sigma / SEs rtol 1e-6 (observed <= 2.2e-7), log-likelihood rtol 1e-10 (observed 5e-13) | — / — | test_panel_mixlogit_parity.py |
mixture_design |
mixexp 1.2.7 SLD (simplex lattice, 3 to 5 components, degree 1 to 4) and SCD (simplex centroid, 3 to 6 components) | R 4.5.2; mixexp 1.2.7 | 1e-12 absolute on every share, the set of runs compared after sorting | — / — | test_mixture_design_r_parity.py (+1) |
mlogit |
nnet::multinom | R 4.5.2; nnet 7.3.20 | rel_est<=1e-06, rel_se<=1e-06 | 2.6e-07 / 1.5e-11 | 44_mlogit.py (+2) |
mmpc |
R bnlearn::mmpc and si.hiton.pc (tests "cor" / "mi", alpha 0.05) | R 4.5.2; bnlearn 5.2.1 | exact: the skeleton is a set of edges | — / — | test_mmhc_bnlearn_parity.py (+1) |
moran |
R spdep::moran.test (randomisation null) | spdep 1.4.2; spatialreg 1.4.3 | I 1.9e-15, expectation, variance and z all at 1e-15 on the row-standardised lattice. | — / — | test_spdep_parity.py |
moran_local |
R spdep::localmoran | spdep 1.4.2; spatialreg 1.4.3 | Every Ii at 8.1e-15. | — / — | test_spdep_parity.py |
moran_residuals |
R spdep::lm.morantest | spdep 1.4.2; spatialreg 1.4.3 | Statistic 5e-16; the p-value at 1e-7 once X is supplied so the Cliff-Ord regression-residual null can be formed. Both spdep alternatives are recorded because lm.morantest defaults to one-sided. | — / — | test_spdep_parity.py |
morris_screening |
sensitivity 1.31.0 morris statistics (mu, mu*, sigma) from the same trajectories and responses | R 4.5.2; sensitivity 1.31.0 | 1e-10 absolute on effects of order 100 | — / — | test_joseph_doe_parity.py (+1) |
mprobit |
Stata mprobit (independent errors of variance one, intpoints(120)): base outcome 1 and 3, noconstant | Stata 18 | coefficients 1e-8 relative (observed 1.2e-9), standard errors 1e-6 (observed 2.8e-7), log likelihood 1e-10 | — / — | test_regcmd_stata_r_parity.py (+1) |
mr |
R MendelianRandomization::mr_ivw through the sp.mr dispatcher | MendelianRandomization 0.10.0; TwoSampleMR 0.7.9; RadialMR 1.2.4; MRPRESSO 1.0; mr.raps 0.4.3 | IVW estimate and default random-effects SE, 1e-10. | — / — | test_mr_R_parity.py |
mr_bma |
Zuber et al. summary_mvMR_BF (GitHub verena-zuber/demo_AMD 4981b5a) | summary_mvMR_BF.R demo_AMD@4981b5a | 1e-12 rel | — / — | test_misc_sens_R_parity.py |
mr_cml |
R MendelianRandomization::mr_cML (DP = FALSE, n = 17723) | MendelianRandomization 0.10.0; TwoSampleMR 0.7.9; RadialMR 1.2.4; MRPRESSO 1.0; mr.raps 0.4.3 | Estimate and SE for every K = 0..6, the BIC-selected fit with its invalid set {12, 14}, and the MA-BIC average, all at 1e-9 (both sides iterate to | d theta | <= 1e-7). |
mr_egger |
R MendelianRandomization::mr_egger, TwoSampleMR::mr_egger_regression | MendelianRandomization 0.10.0; TwoSampleMR 0.7.9; RadialMR 1.2.4; MRPRESSO 1.0; mr.raps 0.4.3 | Slope, intercept and both SEs at 1e-10. | — / — | test_mr_R_parity.py |
mr_f_statistic |
R MendelianRandomization::mr_ivw @Fstat | MendelianRandomization 0.10.0; TwoSampleMR 0.7.9; RadialMR 1.2.4; MRPRESSO 1.0; mr.raps 0.4.3 | Mean F statistic at 1e-10. | — / — | test_mr_R_parity.py |
mr_heterogeneity |
R TwoSampleMR::mr_ivw / mr_egger_regression (Q, Q_df, Q_pval) | MendelianRandomization 0.10.0; TwoSampleMR 0.7.9; RadialMR 1.2.4; MRPRESSO 1.0; mr.raps 0.4.3 | IVW and Egger (Ruecker) Q, degrees of freedom and p-values at 1e-10. | — / — | test_mr_R_parity.py |
mr_ivw |
R MendelianRandomization::mr_ivw (default / fixed / random), TwoSampleMR::mr_ivw | MendelianRandomization 0.10.0; TwoSampleMR 0.7.9; RadialMR 1.2.4; MRPRESSO 1.0; mr.raps 0.4.3 | Estimate, SE under all three models, RSE and Cochran's Q at 1e-10 (observed <= 1e-15). | — / — | test_mr_R_parity.py |
mr_leave_one_out |
R MendelianRandomization::mr_ivw on each leave-one-out subset | MendelianRandomization 0.10.0; TwoSampleMR 0.7.9; RadialMR 1.2.4; MRPRESSO 1.0; mr.raps 0.4.3 | All 28 leave-one-out estimates and default-model SEs at 1e-10. | — / — | test_mr_R_parity.py |
mr_median |
R MendelianRandomization::mr_median (weighted / simple / penalized) | MendelianRandomization 0.10.0; TwoSampleMR 0.7.9; RadialMR 1.2.4; MRPRESSO 1.0; mr.raps 0.4.3 | Point estimates for all three weightings at 1e-10. The bootstrap SE is Monte Carlo on both sides and is not compared (T3). | — / — | test_mr_R_parity.py |
mr_mediation |
R MendelianRandomization::mr_ivw + mr_mvivw | MendelianRandomization 0.10.0 | 1e-12 rel | — / — | test_misc_sens_R_parity.py |
mr_mode |
R MendelianRandomization::mr_mbe (weighted / unweighted, stderror = simple) | MendelianRandomization 0.10.0; TwoSampleMR 0.7.9; RadialMR 1.2.4; MRPRESSO 1.0; mr.raps 0.4.3 | Point estimates for both weightings at 1e-10 -- the same point of the same 512-point density grid. The bootstrap SE is Monte Carlo on both sides and is not compared (T3). | — / — | test_mr_R_parity.py |
mr_multivariable |
R MendelianRandomization::mr_mvivw (default random effects) | MendelianRandomization 0.10.0; TwoSampleMR 0.7.9; RadialMR 1.2.4; MRPRESSO 1.0; mr.raps 0.4.3 | Three direct effects and SEs at 1e-10. | — / — | test_mr_R_parity.py |
mr_pleiotropy_egger |
R TwoSampleMR::mr_egger_regression (intercept test) | MendelianRandomization 0.10.0; TwoSampleMR 0.7.9; RadialMR 1.2.4; MRPRESSO 1.0; mr.raps 0.4.3 | Intercept, SE and t(n - 2) p-value at 1e-10. | — / — | test_mr_R_parity.py |
mr_presso |
R MRPRESSO::mr_presso (NbDistribution = 2000) | MendelianRandomization 0.10.0; TwoSampleMR 0.7.9; RadialMR 1.2.4; MRPRESSO 1.0; mr.raps 0.4.3 | Raw estimate and SE, observed RSS, and the outlier-corrected estimate and SE at 1e-10; outlier set {12, 14} identical. The simulated p-values are Monte Carlo on both sides (T3) and follow the reference's k / B convention. | — / — | test_mr_R_parity.py |
mr_radial |
R RadialMR::ivw_radial (alpha = 0.05, no Bonferroni) | MendelianRandomization 0.10.0; TwoSampleMR 0.7.9; RadialMR 1.2.4; MRPRESSO 1.0; mr.raps 0.4.3 | Square-root weights, per-variant Q contributions and total Q at 1e-10; the outlier set is identical with bonferroni=False (StatsPAI's default applies Bonferroni). | — / — | test_mr_R_parity.py |
mr_steiger |
R TwoSampleMR::mr_steiger with r from get_r_from_bsen | MendelianRandomization 0.10.0; TwoSampleMR 0.7.9; RadialMR 1.2.4; MRPRESSO 1.0; mr.raps 0.4.3 | R^2 on both traits and the direction at 1e-10; the p-value (1.8e-73) at 1e-12. | — / — | test_mr_R_parity.py |
msm |
ipw::ipwtm + lm / glm(quasibinomial) + sandwich::vcovCL(type = 'HC1') | R 4.5.2; ipw 1.3.0; sandwich 3.1.1 | coefficients and cluster SEs 1e-9 rel (observed 1.5e-13); Stata regress / logit [pw] 1e-7 | — / — | test_teffects_R_parity.py (+2) |
mswitch |
Stata 18 mswitch dr / ar, estat transition, estat duration, predict; statsmodels MarkovAutoregression | Stata 18; statsmodels 0.14.6 | log likelihood, filtered and smoothed probabilities at Stata's estimates 1e-11; covariance 5e-6; estimates within 5e-5 standard errors | — / — | test_mswitch_parity.py (+1) |
multi_cutoff_rd |
rdmulti::rdmc 2.0.0 (Cattaneo, Titiunik, Vazquez-Bare & Keele) | R 4.5.2; rdmulti 2.0.0 | identical to sp.rdmc on the fixture (exact); per-cutoff and pooled estimates vs R 1e-9 rel | — / — | test_rdmulti_parity.py (+1) |
multi_outcome_synth |
augsynth::augsynth_multiout 0.2.0 (progfunc='None', scm=TRUE, combine_method 'concat' / 'avg'); synth_qp re-run at OSQP eps 1e-12 | R 4.5.2; augsynth 0.2.0; osqp 1.0.0 | weights 1e-10 abs (observed <= 9e-15); per-outcome ATT 1e-9 rel | — / — | test_synth_rest_R_parity.py (+1) |
multi_treatment |
Stata teffects aipw with a multivalued treatment (mlogit propensity) | Stata 18 | 1e-10 rel on both contrasts, potential-outcome means and sandwich SEs | — / — | test_teffects_R_parity.py (+2) |
multiway_cluster_vcov |
sandwich::vcovCL(cluster=~g1+g2+g3) | R 4.5.2; sandwich 3.1.1 | rel_est<=1e-06, rel_se<=1e-06 | 2.1e-15 / 2.1e-15 | 56_multiway_cluster.py (+2) |
nbreg |
MASS::glm.nb | R 4.5.2; MASS 7.3.65 | rel_est<=1e-06, rel_se<=1e-06 | 1.0e-13 / 1.0e-11 | 42_nbreg.py (+2) |
ndcg |
farr::ndcg, evaluated in R from its definition | R 4.5.2 | 1e-9 rel | — / — | test_accounting_research_parity.py (+1) |
negative_control_exposure |
stats::lm + sandwich::vcovHC(type = 'HC1') and Stata regress, vce(robust) | R 4.5.2; ivreg 0.6.7; sandwich 3.1.1; Stata Stata 18 MP (07 Jun 2023) | 1e-9 relative on estimate and SE (observed 2.1e-16 R, 6.7e-16 Stata) | — / — | test_proximal_r_parity.py (+2) |
negative_control_outcome |
stats::lm + sandwich::vcovHC(type = 'HC1') and Stata regress, vce(robust) | R 4.5.2; ivreg 0.6.7; sandwich 3.1.1; Stata Stata 18 MP (07 Jun 2023) | 1e-9 relative on estimate and SE (observed 1.2e-15 R, 2.2e-15 Stata) | — / — | test_proximal_r_parity.py (+2) |
negd |
Stata regress, vce(robust) | Stata 18 | 1e-9 rel | — / — | test_misc_sens_stata_parity.py |
netlm |
R sna::netlm | igraph 2.3.3; sna 2.8; ergm 4.12.0; dyadRobust 0.0.1.0001 | Coefficients 2e-15 directed and undirected. QAP p-values are permutation draws and are not compared. | — / — | test_network_parity.py |
netlogit |
R sna::netlogit | igraph 2.3.3; sna 2.8; ergm 4.12.0; dyadRobust 0.0.1.0001 | Coefficients 1.4e-9 (IRLS on both sides). QAP p-values are permutation draws and are not compared. | — / — | test_network_parity.py |
network_components |
R igraph::components | igraph 2.3.3; sna 2.8; ergm 4.12.0; dyadRobust 0.0.1.0001 | Counts and sizes exact on a disconnected graph; weak and strong counts on the directed graph. | — / — | test_network_parity.py |
network_modularity |
R igraph::modularity | igraph 2.3.3; sna 2.8; ergm 4.12.0; dyadRobust 0.0.1.0001 | Exact for a fixed split and for igraph's own fast-greedy partition. | — / — | test_network_parity.py |
network_summary |
R igraph edge_density / diameter / mean_distance / transitivity | igraph 2.3.3; sna 2.8; ergm 4.12.0; dyadRobust 0.0.1.0001 | Density, diameter, mean path length, transitivity and assortativity exact; average clustering matches igraph::transitivity(type = 'average', isolates = 'zero'), the convention used here. | — / — | test_network_parity.py |
nlogit |
Stata 18 MP official nlogit (RUM-consistent); R mlogit mlogit(nests = ...) |
Stata 18 MP; mlogit 2.0.0 | Stata, every block at 1e-6: coefficients 1e-10, standard errors 4e-8 under vce(oim) and 8e-8 under vce(robust) and vce(cluster), log-likelihood 1e-11, LR test of IIA 1e-12. R: coefficients at 1e-6 and log-likelihoods at 1e-8 with separate dissimilarity parameters, a common one, and no constants. | — / — | test_nlogit_parity.py |
nls |
R stats::nls with the analytic gradient from deriv() and sandwich::sandwich / vcovCL (gnr, HC1, cluster); Stata nl (gnr, robust, cluster, hc2, hc3) | R 4.5.2; sandwich 3.1.1; Stata 18 | against R: estimates and standard errors 1e-8 relative at tol=1e-12 (observed 3e-9), 1e-6 at the default tol=1e-10 (estimates 2e-7); against Stata nl: estimates 1e-7, standard errors 1e-6 (observed 4.2e-7) | — / — | test_regcmd_stata_r_parity.py (+2) |
noisy_covariates |
Agarwal and Singh replication package (Zenodo 10.5281/zenodo.19866751, CC-BY-4.0), src/wt/stage2.R::rrr_iv, the method authors' own code | R 4.5.2 | estimate and SE 1e-12 relative on standardised covariates (observed 1e-15), 1e-8 on raw covariates (observed 5e-12 to 2.6e-10; the R reference moves 1.8e-11 across platforms) | — / — | test_noisy_covariates_authors_parity.py (+1) |
number_needed_to_treat |
base-R closed form (NNT = 1/risk difference) | R 4.5.2 | estimate 1e-12 abs (observed 0); CI not pinned | — / — | test_epi_parity.py (+1) |
oaxaca |
oaxaca::oaxaca | R 4.5.2; oaxaca 0.1.5 | rel_est<=1e-06, rel_se<=1e-06 | 2.6e-15 / 2.1e-15 | 30_oaxaca.py (+3) |
odds_ratio |
base-R closed form (Woolf logit; = epiR::epi.2by2) | R 4.5.2 | estimate, se_log, CI 1e-12 abs (observed 0) | — / — | test_epi_parity.py (+1) |
ologit |
MASS::polr(method="logistic") | R 4.5.2; MASS 7.3.65 | rel_est<=1e-06, rel_se<=1e-06 | 1.5e-07 / 3.4e-07 | 45_ologit.py (+2) |
oneway |
Stata 18 oneway | Stata 18 | F, sums of squares and Bartlett's test 1e-9 rel | — / — | test_kohler_kreuter_stata_parity.py (+1) |
oprobit |
MASS::polr(method="probit") | R 4.5.2; MASS 7.3.65 | rel_est<=1e-06, rel_se<=1e-06 | 3.7e-07 / 2.1e-10 | 49_oprobit.py (+2) |
oster_bounds |
Stata psacalc (Oster); R robomit::o_delta / o_beta | Stata 18; psacalc 2.1; R 4.5.2; robomit 1.0.7 | 1e-12 rel vs psacalc (observed 4.4e-14); 5e-7 abs vs robomit (it rounds to 6 dp) | — / — | test_inference_sens_R_parity.py (+3) |
oster_delta |
Stata psacalc (Oster); R robomit::o_delta / o_beta | Stata 18; psacalc 2.1; R 4.5.2; robomit 1.0.7 | 1e-12 rel vs psacalc (observed 4.4e-14); 5e-7 abs vs robomit | — / — | test_inference_sens_R_parity.py (+3) |
overlap_weights |
WeightIt::weightit 1.7.0 (method='glm'), R 4.5.2 | R 4.5.2; WeightIt 1.7.0 | All four estimands of the shared-propensity family (Li, Li & Li 2019 Table 1) relative to WeightIt: ATO 2.5e-14, ATE 4.1e-14, ATT 2.3e-14, ATC 4.1e-14. The propensity score itself matches R glm(family=binomial) to 2.6e-14 absolute. | — / — | test_overlap_weights_r_parity.py (+1) |
pagerank |
R igraph::page_rank (damping 0.85) | igraph 2.3.3; sna 2.8; ergm 4.12.0; dyadRobust 0.0.1.0001 | 5e-12 undirected, 1e-12 directed. | — / — | test_network_parity.py |
panel |
plm::plm + plm::phtest | R 4.5.2; plm 2.6.7 | rel_est<=1e-06, rel_se<=1e-06 | 4.7e-14 / 1.5e-15 | 35_panel.py (+2) |
panel_fgls |
Stata 18 xtgls, panels(hetero) | Stata 18 MP | 6.3e-16 on every coefficient and 7.0e-16 on every standard error against the two-step default, and 4.7e-08 against xtgls, igls for the iterated variant. A reference-free test also asserts the two are distinct estimators, so a silent return to iterating fails. |
— / — | test_panel_stata_parity.py |
panel_logit |
Stata 18 xtlogit, re | Stata 18 MP | Graded by CONVERGENCE rather than a fixed tolerance: Stata integrates adaptively and StatsPAI does not, so the honest claim is that agreement improves as the Gauss-Hermite rule is refined. Observed 3.8e-04 at 12 points and 2.4e-07 at 60, with the log-likelihood at 2.1e-08 -- the sharpest single check, since it is the same objective evaluated at the same optimum. sigma_u and rho at 1e-4. | — / — | test_panel_stata_parity.py |
panel_match |
PanelMatch::PanelMatch / PanelEstimate / get_covariate_balance / placebo_test | R 4.5.2; PanelMatch 3.1.5 | matched sets exact; refinement weights, estimates, conditional and unconditional SEs, balance and placebo estimates 1e-10 absolute | — / — | test_panel_match_R_parity.py (+1) |
panel_probit |
Stata 18 xtprobit, re | Stata 18 MP | Same convergence grading: 8.0e-05 at 12 quadrature points and 4.0e-08 at 60, log-likelihood at 1.7e-09, sigma_u and rho at 1e-4. | — / — | test_panel_stata_parity.py |
panel_qtet |
qte::panel.qtet 1.3.1 (Callaway & Li 2019) | — | all 19 quantiles: abs < 1e-8 (observed 6.8e-12); ATT abs < 1e-6. panel.qtet composes ordinary ecdf evaluations and type-7 quantiles, both of which have exact numpy equivalents, so this is machine-precision agreement rather than a tolerance band. | — / — | test_panel_qtet_parity.py (+1) |
panel_unitroot |
plm::purtest 2.6.7; Stata 18 xtunitroot | R 4.5.2; plm 2.6.7; Stata 18 | all statistics 1e-10 rel (observed 9.1e-15) | — / — | test_timeseries_R_parity.py (+2) |
partial_corr_pvalue |
pcalg::gaussCItest (Fisher z test of a zero partial correlation) | R 4.5.2; pcalg 2.7.12 | 1e-9 relative on the p-value (observed 1.5e-12), 224 tests with p from 3e-144 to 1 | — / — | test_partial_corr_pcalg_parity.py (+1) |
path_analysis |
R lavaan::sem / growth (ML, expected information, fixed.x) and estimator = "MLM" | R 4.5.2; lavaan 0.6.21 | observed-variable models: estimates / SEs / standardised solution 2e-5 rel (lavaan's optimiser; 1e-12 on the closed-form models), fit measures and the scaled test 1e-6; latent-variable models: 5e-6 on everything, bounded by lavaan's stopping rule | — / — | test_path_analysis_lavaan_parity.py (+5) |
pc_algorithm |
pcalg::pc(skel.method = 'stable') with gaussCItest, and with bnlearn::ci.test(test = 'x2-adf') as its independence test | R 4.5.2; pcalg 2.7.12; bnlearn 5.2.1 | graphs compared for equality: Gaussian data, every directed and undirected edge of the CPDAG and every separating set; categorical data, skeleton and every separating set | — / — | test_ness_causal_ai_parity.py (+1) |
pcmci |
tigramite PCMCI.run_pcmci(tau_min = 1) with ParCorr(significance = 'analytic') | python 3.11.15; tigramite 5.2.10.1; numpy 2.2.6 | 1e-10 absolute on the MCI p-values and partial correlations of every lagged link (observed 1e-15); link decisions exact | — / — | test_pcmci_tigramite_parity.py (+1) |
peer_effects |
ivreg::ivreg + sandwich::vcovHC(type = 'HC0') and Stata ivregress 2sls, vce(robust): y on 1, Wy, X[, WX] with instruments 1, X, WX, W^2X, W^3X | R 4.5.2; ivreg 0.6.7; sandwich 3.1.1; Stata Stata 18 MP (07 Jun 2023) | 1e-9 relative on every coefficient and SE (observed 3.8e-14 R, 2.7e-14 Stata) | — / — | test_interference_r_parity.py (+2) |
periodogram |
stats::spec.pgram, stats::spec.ar; Stata 18 pergram | R 4.5.2; Stata 18 | frequencies, spectrum, degrees of freedom, bandwidth and band 1e-9 rel | — / — | test_spectral_parity.py (+2) |
poisson |
stats::glm(family=poisson()) | R 4.5.2; stats 4.5.2 | rel_est<=1e-06, rel_se<=1e-06 | 9.2e-15 / 8.7e-12 | 58_poisson.py (+2) |
policy_tree |
policytree::policy_tree | R 4.5.2; policytree 1.2.4 | rel_est<=1e-06, rel_se<=1e-06 | 9.6e-16 / 1.4e-16 | 70_policy_tree.py (+2) |
policy_value |
grf::average_treatment_effect(subset = policy == 1) x mean(policy) on grf get_scores; policytree::double_robust_scores reward contrast | R 4.5.2; grf 2.6.1; policytree 1.2.4 | 1e-12 rel (observed 6.7e-16) | — / — | test_r2_teffects_parity.py (+1) |
positive_correlation_test |
Stata 18 biprobit (method='biprobit') | Stata 18 | log likelihood and LR statistic 1e-12 relative; rho and athrho 1e-8 (observed 1.6e-9); athrho standard error and Wald statistic 1e-8 (observed 2e-10) | — / — | test_positive_correlation_test_stata_parity.py (+1) |
power_case_control |
Stata power twoproportions; R stats::power.prop.test(strict=TRUE) | R 4.5.2; Stata 18 MP | 1e-11 rel vs Stata (observed 4e-13, Stata's normal CDF); 1e-10 vs R (observed 6e-16) | — / — | test_survival_epi_R_parity.py (+2) |
power_cluster_rct |
base-R closed form (design-effect-inflated z-approx power) | R 4.5.2 | power 1e-12 abs (observed ~2e-16) | — / — | test_power_extra_parity.py (+1) |
power_logrank |
base-R closed form (Schoenfeld log-rank power) | R 4.5.2 | power 1e-12 abs (observed ~2e-16) | — / — | test_power_parity.py (+1) |
power_rct |
base-R closed form (two-sample pooled-sigma z-approx power) | R 4.5.2 | power 1e-12 abs (observed ~2e-16) | — / — | test_power_parity.py (+1) |
power_ttest |
Stata 18 MP power onemean / twomeans / pairedmeans; R pwr::pwr.t.test | Stata 18 MP; R 4.5.2 | power / delta 1e-8 rel vs Stata (observed <= 1e-11); fractional n 1e-6 rel vs pwr (its uniroot tolerance) | — / — | test_effect_size_power_ttest_parity.py |
power_two_proportions |
base-R closed form (unpooled Wald two-proportion z-approx, test='wald') | R 4.5.2 | power 1e-12 abs (observed ~2e-16) | — / — | test_power_parity.py (+1) |
ppmlhdfe |
fixest::fepois | R 4.5.2; fixest 0.14.0 | rel_est<=1e-06, rel_se<=2e-06 | 4.9e-13 / 2.2e-15 | 37_ppmlhdfe.py (+2) |
prais |
Stata prais: Prais-Winsten and Cochrane-Orcutt, rhotype regress / freg / tscorr / dw / theil / nagar, iterated and twostep, vce(robust) / vce(hc2) / vce(hc3) | Stata 18 | coefficients, standard errors, rho, R-squared, F, root MSE and both Durbin-Watson statistics 1e-9 relative (observed 2.5e-15) | — / — | test_regcmd_stata_r_parity.py (+1) |
predict_cate |
econml.metalearners SLearner / TLearner / XLearner .effect() on new rows | python 3.11.0; econml 0.16.0; scikit-learn 1.6.1 | 1e-10 absolute on the predicted effect at 200 new rows (observed 1e-15) | — / — | test_metalearner_tools_parity.py (+1) |
pretrends_power |
pretrends::pretrends / pretrends::slope_for_power (GitHub, not CRAN) | R 4.5.2; pretrends 0.1.0 | rel_est<=1e-06 | 1.6e-08 / — | 76_pretrends.py (+2) |
pretrends_slope_for_power |
pretrends::pretrends / pretrends::slope_for_power (GitHub, not CRAN) | R 4.5.2; pretrends 0.1.0 | rel_est<=1e-06 | 1.6e-08 / — | 76_pretrends.py (+2) |
prevalence_ratio |
base-R closed form (Katz-log; = epiR::epi.2by2) | R 4.5.2 | estimate, se_log, CI 1e-12 abs (observed ~2e-16) | — / — | test_epi_parity.py (+1) |
principal_strat |
AER::ivreg 1.2.16 (Wald LATE); SACE bounds via Stata leebounds | R 4.5.2; AER 1.2.16 | 1e-10 rel on the monotonicity complier LATE | — / — | test_teffects_R_parity.py (+2) |
private_info_bound |
Hendren's own makem.do (Econometrica supplement 10931) in Stata 18, the method author's code | Stata 18 | bound and Wald statistic 1e-9 relative (observed 2e-11, 1.3e-11), per-row m_Z 1e-9 (observed 1e-10), Wald p 1e-8 on a p of 4e-35 | — / — | test_private_info_bound_stata_parity.py (+1) |
probit |
stats::glm(family=binomial("probit")) | R 4.5.2; stats 4.5.2 | rel_est<=1e-06, rel_se<=1e-06 | 9.8e-10 / 1.6e-08 | 48_probit.py (+2) |
propensity_score |
stats::glm(binomial('logit' / 'probit'), epsilon = 1e-14) | R 4.5.2 | logit 1e-12 relative; probit 1e-8 (glm stops with a score of order 1e-7; ours is the maximum, score below 1e-10, asserted) | — / — | test_matching_balance_r_parity.py (+2) |
proximal |
ivreg::ivreg default covariance and Stata ivregress 2sls, small (outcome-bridge 2SLS: y on 1, d, X, W with instruments 1, d, X, Z) | R 4.5.2; ivreg 0.6.7; sandwich 3.1.1; Stata Stata 18 MP (07 Jun 2023) | 1e-9 relative on estimate and SE (observed 2.3e-14 R, 2.8e-15 Stata) | — / — | test_proximal_r_parity.py (+2) |
ps_balance |
cobalt::col_w_smd / col_w_vr / col_w_ks with propensity::wt_* weights | cobalt 4.6.2; propensity 0.1.0; R 4.5.2 | 1e-9 relative, 1e-12 absolute where a statistic is zero | — / — | test_matching_balance_r_parity.py (+1) |
ps_weights |
propensity::wt_ate / wt_att / wt_atu / wt_atm / wt_ato, with stabilize = TRUE and after ps_trunc (fixed and percentile) | propensity 0.1.0; R 4.5.2 | 1e-9 relative (observed 2e-14) | — / — | test_matching_balance_r_parity.py (+2) |
psis |
loo::psis through loo::loo (smoothed log weights, Pareto k, n_eff) | R 4.5.2; loo 2.9.0 | 1e-9 abs | — / — | test_regression_stories_r_parity.py (+1) |
psm |
MatchIt::matchit | R 4.5.2; MatchIt 4.7.2 | rel_est<=1e-06, rel_se<=1e-06 | 1.2e-15 / 2.0e-16 | 11_psm.py (+2) |
psmatch2 |
Stata 18 MP + psmatch2 4.0.12 / pstest 4.2.2 (Leuven & Sianesi 2003) | Stata 18 MP; psmatch2 4.0.12 | Observed py<->Stata relative gaps on the committed fixtures: nearest-neighbour ATT 1.2e-16 and its analytic SE exactly 0 (as is _weight per row); radius ATT 2.8e-16; Abadie-Imbens ai(1)/ai(2) SE 3.0e-15/1.7e-15; PSM-DID across all five weight regimes 2.0e-14; pstest per-covariate rows 1.3e-14; Mahalanobis ATT 1.2e-13; llr ATT 2.1e-10 (worst of four kernels); kernel ATT 9.1e-10; pstest summary block 1.3e-9. The two loosest rows are bounded by fixture precision, not by the estimator: pstest accumulates MeanBias/MedBias in a Stata float (2.6e-9) and r(seatt) for the llr reroute was captured at 8 significant digits (9.9e-9). | — / — | test_psmatch2_parity.py (+3) |
pwcompare |
Stata 18 margins g, pwcompare(effects) mcompare(noadjust | bonferroni | sidak); R emmeans pairs(adjust=) | Stata 18; R 4.5.2; emmeans 2.0.3 | 1e-10 rel (observed <= 7.9e-15 diff / 4.9e-15 SE) |
qreg |
quantreg::rq | R 4.5.2; quantreg 6.1 | rel_est<=1e-06, rel_se<=1e-06 | 1.0e-15 / 5.6e-16 | 40_qreg.py (+2) |
queen_weights |
spdep::poly2nb(queen = TRUE) + nb2listw(style = 'W' / 'S' / 'U') | R 4.5.2; spdep 1.4.2; sf 1.1.1; spData 2.3.5 | neighbour sets exact; weights 1e-12 rel (observed 0) | — / — | test_spatial_survey_R_parity.py (+1) |
raftery_diag |
coda::raftery.diag | R 4.5.2; coda 0.19.4.1 | burn-in, total and minimum length equal as integers; dependence factor equal after rounding to coda's three significant digits | — / — | test_bayes_mcmc_parity.py (+1) |
rake |
survey::rake (to its fixed point) and survey::calibrate(calfun='raking'); Stata svycal rake | R 4.5.2; survey 4.5; Stata 18 MP | calibrated weight shares 1e-12 rel at tol=1e-14 (observed 1.5e-15); 1e-8 at the default tol=1e-10 (observed 1.0e-10) | — / — | test_survey_calib_R_parity.py (+2) |
ranksum |
Stata 18 ranksum | Stata 18 | z, rank sum and adjusted variance 1e-9 rel | — / — | test_kohler_kreuter_stata_parity.py (+1) |
rd2d |
R rd2d::rd2d / rd2d.distance 1.0.0 (Cattaneo, Titiunik & Yu) | R; rd2d 1.0.0; sandwich 3.1.1 | estimate.p/q, std.err.p/q, t, CI, cross-point covariance rel 1e-9 (observed 1.6e-11); bandwidths rel 1e-8 (observed 1.2e-12); N exact | — / — | test_rd_open_R_parity.py (+1) |
rd2d_bw |
R rd2d::rdbw2d / rdbw2d.distance 1.0.0 | R; rd2d 1.0.0 | per-point bandwidths rel 1e-8 (observed 5.0e-13) | — / — | test_rd_open_R_parity.py (+1) |
rd_honest |
RDHonest::RDHonest 1.0.1.9000 (Armstrong & Kolesar) | R 4.5.2; RDHonest 1.0.1.9000 | estimate / std.error / maximum.bias / conf.low / conf.high 1e-9 rel at fixed bandwidth; 1e-6 rel when the bandwidth and M are selected | — / — | test_rdhonest_parity.py (+1) |
rdbwhte |
R rdhte::rdbwhte 0.2.0 | R 4.5.2; rdhte 0.2.0; rdrobust 4.0.0 | bandwidths rel 1e-8 (continuous and per-subgroup) | — / — | test_rd_iv_rd_R_parity.py (+1) |
rdbwselect |
rdrobust::rdbwselect; Stata side uses the authors' rdbwselect ado. certwo is R-only: Stata rdbwselect 10.0.0 exits r(3200) on it, including on the package's own rdrobust_senate.dta | R 4.5.2; rdrobust 4.0.0 | rel_est<=1e-06 | 9.4e-13 / 3.5e-09 | 88_rdbwselect.py (+2) |
rdd |
rdrobust::rdrobust | R 4.5.2; rdrobust 4.0.0 | rel_est<=1e-06, rel_se<=1e-06 | 5.7e-14 / 9.4e-11 | 06_rd.py (+3) |
rddensity |
rddensity::rddensity | R 4.5.2; rddensity 2.6 | rel_est<=1e-06, rel_se<=1e-06 | 9.3e-12 / 1.8e-11 | 09_rddensity.py (+2) |
rdhte |
R rdhte::rdhte 0.2.0 (sandwich 3.1.1) | R 4.5.2; rdhte 0.2.0; sandwich 3.1.1; rdrobust 4.0.0 | coef, coef.bc, se.rb, vcov rel 1e-9 (observed 2.2e-12); bandwidths 1e-8 | — / — | test_rd_iv_rd_R_parity.py (+1) |
rdhte_lincom |
R rdhte::rdhte_lincom 0.2.0 | R 4.5.2; rdhte 0.2.0 | estimate, z, CI, joint chi-square rel 1e-9; p 1e-8 | — / — | test_rd_iv_rd_R_parity.py (+1) |
rdmc |
rdmulti::rdmc 2.0.0 (Cattaneo, Titiunik, Vazquez-Bare & Keele) | R 4.5.2; rdmulti 2.0.0 | per-cutoff coefficients, robust coefficients, robust SEs and the pooled weighted estimate 1e-9 rel; selected bandwidths 1e-5 rel | — / — | test_rdmulti_parity.py (+1) |
rdms |
rdmulti::rdms; Stata side uses the rdms ado from the rdpackages GitHub mirror (rdmulti is not on SSC: ssc describe rdmulti returns r(601)), installed into a local gitignored ado path | R 4.5.2; rdmulti 2.0.0 | rel_est<=1e-06, rel_se<=1e-06 | 1.2e-12 / 5.4e-10 | 89_rdms.py (+2) |
rdplot |
R rdrobust::rdplot 4.0.0 | R 4.5.2; rdrobust 4.0.0 | J / J_IMSE / J_MV and bin counts exact; bin means, SEs, t intervals and polynomial values rel 1e-9 (observed 4.0e-12); p = 4 global polynomial coefficients rel 1e-8 (observed 1.6e-10, raw-scale conditioning) | — / — | test_rd_iv_rd_R_parity.py (+1) |
rdplotdensity |
R rddensity::rdplotdensity 2.6 (lpdensity 2.5) | R 4.5.2; rddensity 2.6; lpdensity 2.5 | f_p, f_q, se_p, se_q at every grid point rel 1e-8 (observed 8.0e-12); nh exact | — / — | test_rd_iv_rd_R_parity.py (+1) |
rdpower |
rdpower::rdpower 3.0 (Cattaneo, Titiunik & Vazquez-Bare) | R 4.5.2; rdpower 3.0 | robust bias-corrected SE & power 1e-8 rel (observed 4.2e-14) | — / — | test_rdlocrand_parity.py (+1) |
rdrandinf |
rdlocrand::rdrandinf 2.0 (Cattaneo, Titiunik & Vazquez-Bare) | R 4.5.2; rdlocrand 2.0 | observed statistic & asymptotic p-value 1e-8 rel (observed 2.3e-15) | — / — | test_rdlocrand_parity.py (+1) |
rdrobust |
rdrobust::rdrobust | R 4.5.2; rdrobust 4.0.0 | rel_est<=1e-06, rel_se<=1e-06 | 5.7e-14 / 9.4e-11 | 06_rd.py (+2) |
rdsampsi |
rdpower::rdsampsi 3.0 (Cattaneo, Titiunik & Vazquez-Bare) | R 4.5.2; rdpower 3.0 | required sample sizes n_left / n_right / n_total asserted as exact integer equality (no tolerance) | — / — | test_rdlocrand_parity.py (+1) |
rdwinselect |
rdlocrand::rdwinselect 2.0 (Cattaneo, Titiunik & Vazquez-Bare) | R 4.5.2; rdlocrand 2.0 | window grid 1e-12 rel (observed 0); per-window counts Nl / Nr asserted as exact integer equality | — / — | test_rdlocrand_parity.py (+1) |
recentered_iv |
Stata ivreghdfe on the recentred instrument (method 'recenter' and 'control') and ritest on 200 fixed permutations | Stata 18 MP; ivreghdfe 1.1.4; ivreg2 4.1.12; reghdfe 6.13.1; ritest 1.1.7 | 1e-8 relative on coefficient and SE (observed <= 3e-15); randomisation-inference count exact and observed statistic 1e-9 relative | — / — | test_recentered_iv_stata.py (+1) |
reciprocity |
R igraph::reciprocity | igraph 2.3.3; sna 2.8; ergm 4.12.0; dyadRobust 0.0.1.0001 | Exact on a 40-node directed graph. | — / — | test_network_parity.py |
regress |
lm + sandwich::vcovHC | R 4.5.2; sandwich 3.1.1 | rel_est<=1e-06, rel_se<=1e-06 | 1.1e-12 / 1.3e-12 | 01_ols.py (+2) |
relative_risk |
base-R closed form (Katz-log; = epiR::epi.2by2 / Stata epitab) | R 4.5.2 | estimate, se_log, CI 1e-12 abs (observed 0) | — / — | test_epi_parity.py (+1) |
reset_test |
lmtest::resettest(power=2:3, type='fitted') | R 4.5.2; lmtest 0.9.40 | F-statistic & p-value 1e-10 rel (observed ~1e-13) | — / — | test_diagnostics_parity.py (+1) |
retrodesign |
retrodesign::retro_design_closed_form (power, type S, type M); retrodesign::retrodesign power and type S with 20 degrees of freedom | R 4.5.2; retrodesign 0.2.2 | 1e-9 rel | — / — | test_regression_stories_r_parity.py (+1) |
ri_test |
R ri2::conduct_ri (randomizr full enumeration); Stata ritest over the full assignment set | R 4.5.2; ri2 0.5.0; Stata 18; ritest 1.1.7 | p exact (k / N_assignments); statistic 1e-12 rel vs R, 2^-23 vs Stata | — / — | test_inference_sens_R_parity.py (+3) |
ridge |
MASS::lm.ridge | R 4.5.2; MASS 7.3.65 | coefficients, GCV, kHKB and kLW 1e-9 rel | — / — | test_linear_model_extensions_parity.py (+1) |
rif_decomposition |
dineq::rif + manual OLS | R 4.5.2; dineq 0.1.0 | rel_est<=1e-06, rel_se<=1e-06 | 2.2e-15 / 1.4e-16 | 32_rif.py (+2) |
rifreg |
R rifreg::rifreg (variance, quantiles) and dineq::rif + lm (Gini) | rifreg 1.1.0; dineq 0.1.0; ddecompose 1.0.0 | Coefficients at 1e-10 for the variance and the 10th/50th/90th percentiles (quantile_convention='rifreg') and for the Gini against dineq's exact RIF; the stock rifreg Gini, which integrates the Lorenz curve numerically, agrees to 1e-4. | — / — | test_decomp_R_parity.py |
risk_difference |
base-R closed form (Wald; = epiR::epi.2by2 / Stata epitab) | R 4.5.2 | estimate, se, CI 1e-12 abs (observed 0) | — / — | test_epi_parity.py (+1) |
rkd |
R rdrobust::rdrobust(deriv = 1, vce = 'hc1') 4.0.0; Stata rdrobust, deriv(1) 11.1.0 | R 4.5.2; rdrobust 4.0.0; Stata 18 MP; rdrobust (Stata) 11.1.0 | conventional and robust estimate / SE, bandwidth: rel 1e-9 (observed 5.2e-12 vs R, 6.0e-11 vs Stata) | — / — | test_rd_iv_rd_R_parity.py (+2) |
rlasso |
R hdm::rlasso 0.3.2 | R 4.5.2; hdm 0.3.2 | support exact; beta / sigma / loadings / residuals atol 1e-6, lambda0 rtol 1e-8 (observed rel <= 1.8e-12) | — / — | test_rlasso_parity.py (+1) |
rlasso_effect |
R hdm::rlassoEffect 0.3.2 | R 4.5.2; hdm 0.3.2 | alpha / se atol 1e-6 (observed rel <= 3.3e-15), incl. hdm's GrowthData vignette | — / — | test_rlasso_parity.py (+2) |
rlasso_effects |
R hdm::rlassoEffects 0.3.2 | R 4.5.2; hdm 0.3.2 | alpha / se rtol 1e-9 on cps2012 (observed 2.8e-12); atol 1e-6 on the synthetic fixture | — / — | test_rlasso_parity.py (+2) |
rlasso_iv |
R hdm::rlassoIV 0.3.2 | R 4.5.2; hdm 0.3.2 | coef / se atol 1e-6 (observed <= 3e-15); EminentDomain select_Z-only atol 1e-8 (observed 6.3e-13, against AER::ivreg) | — / — | test_rlasso_parity.py (+2) |
rlassologit_effect |
R hdm::rlassologitEffect 0.3.2 | R 4.5.2; hdm 0.3.2 | alpha / se atol 1e-6 (observed rel 1.1e-15 / 5.3e-14 post; se 3.2e-7 with post=False) | — / — | test_rlassologit_effect_parity.py (+1) |
rlassologit_effects |
R hdm::rlassologitEffects 0.3.2 | R 4.5.2; hdm 0.3.2 | coef / se atol 1e-6 (observed rel 8.6e-16 / 2.5e-14) | — / — | test_rlassologit_effect_parity.py (+1) |
robreg |
Stata 18 robreg m / s / mm (Jann); robustbase::lmrob (coefficients, scale, weights, covariance); MASS::rlm (M estimates and Huber-formula SE) | R 4.5.2; robustbase 0.99.7; MASS 7.3.65; Stata 18 | coefficients, scale and covariance 1e-6 rel (the references iterate to their own tolerance); lmrob scale 1e-9 rel | — / — | test_accounting_research_parity.py (+2) |
robust_synth |
scpi::scest(w.constr = list(name = 'ols')) with scdata(constant = TRUE) and stats::lm (unconstrained SC with intercept); glmnet (ridge / lasso / elastic net, unpenalised intercept) | R 4.5.2; scpi 4.0.1; glmnet 4.1.10 | OLS weights / intercept / fitted path 1e-10 rel (observed 8.4e-13); penalised paths 1e-8 rel, weights atol 1e-10 (observed 3.4e-11 abs, glmnet's coordinate-descent stop) | — / — | test_did_synth_synthvar_parity.py (+1) |
robvar |
Stata 18 robvar | Stata 18 | W0, W50 and W10 1e-9 rel | — / — | test_kohler_kreuter_stata_parity.py (+1) |
roc_curve |
R pROC::roc/auc/var/ci.auc (DeLong); Stata roctab (default and hanley) | R 4.5.2; pROC 1.19.0.1; Stata 18 MP | 1e-10 rel (observed 6e-15) | — / — | test_survival_epi_R_parity.py (+2) |
rook_weights |
spdep::poly2nb(queen = FALSE) + nb2listw(style = 'W') | R 4.5.2; spdep 1.4.2; sf 1.1.1; spData 2.3.5 | neighbour sets exact; weights 1e-12 rel (observed 0) | — / — | test_spatial_survey_R_parity.py (+1) |
rosenbaum_bounds |
R DOS2::senWilcox (Rosenbaum); Stata rbounds; R stats::binom.test (sign test); R rbounds::psens (zero_method='wilcox', 4-dp) | R 4.5.2; DOS2 0.5.2; rbounds 2.2; Stata 18; rbounds (Stata) 1.1.6 | bounding p-values 1e-12 rel (observed 4.6e-15); Stata sig- 1e-15 abs; psens at its own 4-dp rounding | — / — | test_inference_sens_R_parity.py (+3) |
rosenbaum_gamma |
R DOS2::senWilcox (Rosenbaum); Stata rbounds | R 4.5.2; DOS2 0.5.2; Stata 18; rbounds (Stata) 1.1.6 | bounding p-values 1e-12 rel (observed 4.6e-15) | — / — | test_inference_sens_R_parity.py (+1) |
sac |
R spatialreg::sacsarlm | spdep 1.4.2; spatialreg 1.4.3 | rho and lambda at 1e-5, slope coefficients at 1e-6 -- a bounded two-parameter ML line search on both sides. | — / — | test_spdep_parity.py |
sar |
spatialreg::lagsarlm / spatialreg::errorsarlm / spatialreg::lagsarlm(Durbin=TRUE) | R 4.5.2; spatialreg 1.4.3 | rel_est<=1e-06, rel_se<=1e-06 | 4.1e-08 / 3.6e-08 | 65_spatial.py (+2) |
sar_gmm |
spatialreg::stsls(W2X=FALSE) / spatialreg::GMerrorsar | R 4.5.2; spatialreg 1.4.3 | rel_est<=1e-06, rel_se<=1e-06 | 4.6e-08 / 7.3e-16 | 66_spatial_gmm.py (+2) |
sbw |
sbw::sbw 1.2 (Zubizarreta 2015), quadprog solver | — | ATT rel <= 1e-8 (observed <= 4e-10) under both standardisation conventions: tolerance_scale='target' == bal_std='target' and 'group' == bal_std='group', at bal_tol 0.05 and 0.02. | — / — | test_matching_r_parity.py (+1) |
sc_estimate |
synthdid::sc_estimate 0.0.9 (Arkhangelsky, Athey, Hirshberg, Imbens & Wager) | R 4.5.2; synthdid 0.0.9 | estimate, unit/time weights and jackknife SE at 1e-9 on the Prop. 99 replica and a five-treated panel (observed est 8.1e-15, jackknife SE 4.8e-15); each of R's 40 placebo and 40 bootstrap replications replayed at 1e-9 (observed 2.1e-11) | — / — | test_did_synth_R_parity.py (+1) |
scdata |
R scpi::scdata (features = outcome, no cov.adj, constant = FALSE) | scpi 4.0.1; CVXR 1.9.2; ECOSolveR 0.6.1; Qtools 1.6.0; quantreg 6.1 | A, B, P matrices identical (0 difference) on scpi_germany; donor order = R sort(B.names). | — / — | test_did_synth_scpi_parity.py |
sdid |
synthdid::synthdid_estimate | R 4.5.2; synthdid 0.0.9 | rel_est<=1e-06, rel_se<=1e-06 | 7.8e-16 / 1.1e-15 | 12_sdid.py (+2) |
sdm |
spatialreg::lagsarlm / spatialreg::errorsarlm / spatialreg::lagsarlm(Durbin=TRUE) | R 4.5.2; spatialreg 1.4.3 | rel_est<=1e-06, rel_se<=1e-06 | 4.1e-08 / 3.6e-08 | 65_spatial.py (+2) |
sem |
spatialreg::lagsarlm / spatialreg::errorsarlm / spatialreg::lagsarlm(Durbin=TRUE) | R 4.5.2; spatialreg 1.4.3 | rel_est<=1e-06, rel_se<=1e-06 | 4.1e-08 / 3.6e-08 | 65_spatial.py (+2) |
sem_gmm |
spatialreg::stsls(W2X=FALSE) / spatialreg::GMerrorsar | R 4.5.2; spatialreg 1.4.3 | rel_est<=1e-06, rel_se<=1e-06 | 4.6e-08 / 7.3e-16 | 66_spatial_gmm.py (+2) |
sensemakr |
sensemakr::sensemakr | R 4.5.2; sensemakr 0.1.6 | rel_est<=1e-06, rel_se<=1e-06 | 5.8e-13 / 6.6e-13 | 22_sensemakr.py (+2) |
sensitivity_specificity |
R epiR::epi.tests (method wilson / exact); Stata diagti | R 4.5.2; epiR 2.0.94; Stata 18 MP; diagt 2.032 (diagti 2.053) | 1e-12 rel on intervals, 1e-10 on ratios (observed 2e-15) | — / — | test_survival_epi_R_parity.py (+2) |
series |
R lm on poly(raw = TRUE) and splines::bs with sandwich::vcovHC / vcovCL | R 4.5.2; sandwich 3.1.1 | 1e-8 relative on the cross-validation criterion, the fitted function, its standard errors and the linear coefficient (observed <= 4e-9); 1e-7 on the derivative | — / — | test_series_R_parity.py (+1) |
sharpened_qvalues |
Anderson (2008) fdr_sharpened_qvalues.do (last modified 11/20/07; Wayback copy, sha256 in the .do header), run unchanged | Stata 18.0 MP | exact equality after float rounding (16 p-vectors, 2,474 p-values, M = 1 to 2,000) | — / — | test_sharpened_qvalues_stata_parity.py (+1) |
shift_share_political |
AER::ivreg + sandwich HC1; ShiftShareSE::ivreg_ss (EHW / AKM / AKM0); bartik.weight::bw; anova(lm) share balance | R 4.5.2; AER 1.2.16; sandwich 3.1.1; ShiftShareSE 1.1.0; bartik.weight 0.1.0 | estimate / SEs / Rotemberg / F 1e-9 rel (observed <= 5e-15); AKM p-value 1e-7 rel (ShiftShareSE uses 2*(1-pnorm), cancellation at p ~ 1e-9) | — / — | test_synth_rest_R_parity.py (+1) |
shift_share_political_panel |
fixest::feols 0.14.0 (ssc adj=FALSE, cluster.adj=FALSE; unit / time / two-way clusters; unit, time, two-way FE; unbalanced panel); ShiftShareSE::ivreg_ss with FE dummies (AKM); bartik.weight::bw with FE dummies; AER + HC0 per period | R 4.5.2; fixest 0.14.0; ShiftShareSE 1.1.0; bartik.weight 0.1.0; AER 1.2.16; sandwich 3.1.1 | estimate / SEs / Rotemberg / first-stage F 1e-9 rel (observed <= 1.1e-14) | — / — | test_synth_rest_R_parity.py (+1) |
shift_share_se |
R ShiftShareSE::ivreg_ss (Adao, Kolesar & Morales), AKM row; Stata SSC ivreg_ss | R 4.5.2; ShiftShareSE 1.1.0; Stata 18 MP; ivreg_ss SSC 20241116 | 1e-9 rel on beta and the AKM / AKM0 / EHW / Homoscedastic SEs; observed <= 6e-14 | — / — | test_did_synth_shiftshare_parity.py (+2) |
signrank |
Stata 18 signrank | Stata 18 | z, positive rank sum and adjusted variance 1e-9 rel | — / — | test_kohler_kreuter_stata_parity.py (+1) |
simex |
R simex 1.8 simex(asymptotic=TRUE) and mcsimex, fed the same pseudo-errors and re-misclassified vectors as StatsPAI | R 4.5.2; simex 1.8 | estimates and asymptotic SEs 1e-12 relative (observed 8e-15); jackknife SEs 1e-8 (observed 1.7e-9, R glm covariance at the pre-update weights); rational extrapolant 1e-6 (observed 2.9e-7, R nls stopping rule) | — / — | test_simex_r_parity.py (+1) |
slx |
R spatialreg::lmSLX | spdep 1.4.2; spatialreg 1.4.3 | Every coefficient at 1e-10. | — / — | test_spdep_parity.py |
snips |
Open Bandit Pipeline (obp) 0.5.7 SelfNormalizedInverseProbabilityWeighting | obp 0.5.7 | value 1e-12 rel (observed 0.0); delta-method SE identity 1e-10 and equal to sp.ope.snips | — / — | test_ml_causal_obp_parity.py (+1) |
source_decompose |
Stata descogini (Lerman-Yitzhaki) | Stata 18 MP; descogini SSC | With gini='population': total Gini, and each source's S_k, G_k, R_k and share of the total at 1e-12. The default n/(n-1)-corrected Gini differs by exactly that factor; S_k, R_k and shares are identical. | — / — | test_decomp_R_parity.py |
spatial_iv |
sphet::spreg(model = 'lag', het = TRUE) 2.1.1 | R 4.5.2; sphet 2.1.1; spdep 1.4.2 | coefficients and HC0 SEs 1e-10 rel (observed 9.5e-14 / 6.8e-14) | — / — | test_spatial_survey_R_parity.py (+1) |
spatial_treatment |
Pollmann's spatialTreat-example reference code (R + Rcpp, MIT, commit a73bca9), the method author's own implementation | R 4.5.2; Rcpp 1.1.1.1.1; data.table 1.18.2.1 | estimate, both means, standard error and variance components 1e-9 relative | — / — | test_spatial_treatment_R_parity.py (+1) |
spearman |
Stata 18 spearman | Stata 18 | rho and p-value 1e-9 rel | — / — | test_kohler_kreuter_stata_parity.py (+1) |
spillover_did |
R did::att_gt(control_group='nevertreated') + did::aggte(type='simple') per group (direct / ring r, ring cohort = exposure onset); single cohort also fixest::feols(dbar ~ treat + ring1 + ring2, vcov='hetero', ssc(adj=FALSE)) | did 2.5.1; DRDID 1.3.0; fixest 0.14.0 | Direct and ring effects, their SEs and every (group, onset cohort, period) cell at 1e-9 relative on a single-cohort and a staggered spatial panel; observed agreement 1e-14. The ring construction is recomputed independently in R from the coordinates. | — / — | test_did_synth_misc_parity.py |
split_data |
twinning 1.1 twin (method='twinning': the test rows from a given start, numeric and factor columns, rows not divisible by r) and SPlit 1.3 subsample (nearest-row rule) | R 4.5.2; twinning 1.1; SPlit 1.3 | identical row indices | — / — | test_joseph_doe_parity.py (+1) |
sqreg |
R quantreg::rq (Barrodale-Roberts), Koenker 2005 | quantreg see sqreg_R.json provenance | Coefficients 3.5e-14 against quantreg::rq at tau = 0.25 / 0.50 / 0.75 -- both sides minimise the same pinball loss with the same simplex. Standard errors differ from R's se='iid' by ONE SCALAR PER QUANTILE, constant across coefficients to 6e-16: the sandwich is identical and only the sparsity estimate 1/f(0) differs (Powell kernel here, Koenker-Bassett with a Siddiqui/Hall-Sheather bandwidth there). The test asserts the ratio's constancy rather than a numerical band, which a structural difference could not satisfy. R's default se='nid' (Hendricks-Koenker, also Stata qreg's) is a third convention and is recorded as one. | — / — | test_sqreg_parity.py |
ssaggregate |
R ShiftShareSE::ivreg_ss / reg_ss (Adao, Kolesar & Morales) and ssaggregate (Borusyak, Hull & Jaravel; R kylebutts/ssaggregate + AER::ivreg/sandwich HC0); Stata SSC ivreg_ss / reg_ss and ssaggregate + ivreg2, robust | R 4.5.2; ShiftShareSE 1.1.0; ssaggregate (R) 0.0.0.9000 (GitHub kylebutts/ssaggregate@22df93980250891a0cc247f6020136cd33c65ba2); AER 1.2.16; sandwich 3.1.1; Stata 18 MP; reg_ss / ivreg_ss SSC 20241116; ssaggregate (Stata) SSC 1.2.2 (20200826) | 1e-9 rel on beta, every SE row (Homoscedastic, EHW, Reg. cluster, AKM, AKM0), AKM/AKM0 CIs and the shock-level frame; p-values also atol 1e-15 (references use 2*(1-Phi)); observed <= 6e-14 (frame 2.9e-13) | — / — | test_did_synth_shiftshare_parity.py (+2) |
stabilized_weights |
ipw::ipwtm 1.3.0 (type = 'all'; binomial logit; gaussian via geepack::geeglm) | R 4.5.2; ipw 1.3.0; geepack 1.3.13 | 1e-10 rel on every weight (observed 1.5e-12) | — / — | test_teffects_R_parity.py (+2) |
stacked_did |
hand-written stack + fixest::feols | R 4.5.2; fixest 0.14.0 | rel_est<=1e-06 | 3.6e-13 / 7.4e-13 | 75_stacked.py (+2) |
staggered_cs |
staggered::staggered_cs 1.2.2 (Roth & Sant'Anna) | R 4.5.2; staggered 1.2.2 | estimate, Neyman SE and adjusted SE at abs 1e-9 on mpdta, a randomised rollout and a null panel x simple/cohort/calendar (observed rel 6.6e-14) | — / — | test_staggered_extended_parity.py (+1) |
staggered_rollout |
staggered::staggered / staggered_cs / staggered_sa (1.2.2) | R 4.5.2; staggered 1.2.2 | rel_est<=1e-10 | 9.1e-16 / 5.7e-16 | 82_staggered.py (+2) |
staggered_sa |
staggered::staggered_sa 1.2.2 (Roth & Sant'Anna) | R 4.5.2; staggered 1.2.2 | estimate, Neyman SE and adjusted SE at abs 1e-9 on mpdta, a randomised rollout and a null panel x simple/cohort/calendar (observed rel 6.6e-14) | — / — | test_staggered_extended_parity.py (+1) |
staggered_synth |
augsynth::multisynth 0.2.0 (Ben-Michael, Feller & Rothstein; partially pooled SCM for staggered adoption) | R 4.5.2; augsynth 0.2.0; osqp 1.0.0 | ATT, per-unit / per-cohort ATT, event-time ATT and jackknife SE 1e-8 rel (observed <= 2.6e-11); weights 1e-8 abs (observed 2.1e-10); nu 1e-8, imbalance norms 1e-6 rel | — / — | test_did_synth_synthvar_parity.py (+1) |
standardize |
arm::rescale (all four binary rules) and coefficients of arm::standardize(lm) refitted on sp.standardize(df, formula=) | R 4.5.2; arm 1.15.3 | 1e-12 abs on rescaled columns, 1e-9 rel on refitted coefficients | — / — | test_regression_stories_r_parity.py (+1) |
standardize_coefs |
Stata regress, beta (e(beta), with and without [aw=]) and regress / ivreg2 / reghdfe / ivreghdfe on z-scored or x-only z-scored variables | Stata 18; ivreg2 4.1.12; reghdfe 6.13.1; ivreghdfe 1.1.4 | e(beta) 1e-10 relative; coefficients 1e-10 absolute and standard errors 1e-9 relative on the refits (observed 1e-13) | — / — | test_ivreport_stata_parity.py (+3) |
statespace |
statsmodels MLEModel; KFAS likelihood | R 4.5.2; KFAS 1.6.0; statsmodels 0.14.6 | likelihood 1e-9 rel at common parameters; estimates 1e-6; standard errors 1e-5 | — / — | test_statespace_parity.py (+1) |
structural_break |
strucchange::Fstats + sctest(type = 'supF') / breakpoints; mbreaks::dosequa (Bai-Perron sequential); Stata estat sbsingle | R 4.5.2; strucchange 1.5.4; mbreaks 1.0.1; Stata 18 | statistics 1e-10 rel (observed <= 1e-14); sup-F p-values 1e-10 rel for p > 1e-6, 1e-13 abs below (R uses 1 - pchisq); break dates exact | — / — | test_r2_ts_parity.py (+3) |
subcluster_wild_bootstrap |
Stata boottest, bootcluster() (CRVE clustered at g6, signs flipped at s12) | Stata 18; boottest 4.5.3 | p exact (multiple of 1/4096); t 1e-10 rel (observed 1.8e-14) | — / — | test_inference_sens_stata_parity.py (+1) |
subgroup_analysis |
Stata regress + testparm | Stata 18 | 1e-9 rel | — / — | test_misc_sens_stata_parity.py |
subgroup_decompose |
Stata ineqdeco, bygroup() | Stata 18 MP; ineqdeco SSC | GE(0), GE(1), GE(2) totals and within components at 1e-12, between components at 1e-11. The Gini path (Dagum) has no ineqdeco counterpart and keeps its analytical checks. | — / — | test_decomp_R_parity.py |
suest |
Stata suest after regress (one-way); stacked reghdfe 6.13.1 for multi-way clusters and absorbed effects | Stata 18.0 MP; reghdfe 6.13.1 | suest 1e-10 rel (observed 7e-15 on b and V); stacked identities 1e-9 after reghdfe's (N-1)/(N-K) factor | — / — | test_suest_multiway_fe_parity.py (+1) |
summary_index |
Stata swindex (SSC, 2021-02-24) for the Anderson index; summarize / egen / Mata by hand for KLK, weights, imputation | Stata 18.0 MP; swindex 2021-02-24 (SSC) | 1e-10 rel / 1e-12 abs (observed 2.8e-15 under set type double) | — / — | test_summary_index_parity.py (+1) |
sun_abraham |
fixest::sunab | R 4.5.2; fixest 0.14.0 | rel_est<=1e-06, rel_se<=1e-06 | 4.0e-11 / 3.9e-11 | 05_sunab.py (+2) |
super_learner |
SuperLearner::SuperLearner(method = 'method.CC_LS') with validRows supplied | R 4.5.2; SuperLearner 2.0.40 | cross-validated predictions 1e-9 absolute, cross-validated risks 1e-9 relative, weights and ensemble predictions 1e-6 absolute | — / — | test_super_learner_R_parity.py (+1) |
sureg |
systemfit::systemfit(method="SUR", noDfCor) | R 4.5.2; systemfit 1.1.30 | rel_est<=1e-06, rel_se<=1e-06 | 1.5e-14 / 1.5e-15 | 60_sureg.py (+2) |
survivor_average_causal_effect |
Stata leebounds 1.5 (Zhang-Rubin SACE bounds = Lee bounds under monotonicity) | Stata 18; leebounds 1.5 (2013-07-17, Tauchmann) | 1e-10 rel (upper bound; both bounds equal sp.lee_bounds) | — / — | test_teffects_R_parity.py (+2) |
survreg |
Stata 18 MP streg ..., time (four distributions, vce(oim), vce(robust), vce(cluster), frailty(gamma)); survival::survreg (Weibull AFT) |
Stata 18 MP; R 4.5.2; survival 3.8.3 | Stata: every block at 1e-6; observed 3e-10 or better on coefficients, standard errors and the frailty variance, 5e-12 on the log-likelihood, with ml's stopping rule tightened in the do-file. R: coefficients and log-scale 5e-5 abs against a frozen fixture. | — / — | test_survreg_streg_stata_parity.py (+2) |
svydesign |
survey::svydesign + svymean/svytotal/svyglm/degf (strata, nested PSUs, fpc, survey.lonely.psu); Stata svyset + svy: mean/total/regress/logit/poisson + estat effects | R 4.5.2; survey 4.5; Stata 18 MP | estimates, SEs, DEFF, CI bounds 1e-10 rel, p-values 1e-9 rel (observed <= 4e-14; p 6e-13) | — / — | test_survey_design_R_parity.py (+2) |
svyglm |
survey::svyglm (design-based GLM + linearization SE) | R 4.5.2 | coefficients + SE 1e-10 abs (observed ~2e-15 / 6e-15) | — / — | test_survey_parity.py (+1) |
svymean |
survey::svymean (Horvitz-Thompson/Hajek + Taylor SE) | R 4.5.2 | estimate + SE 1e-10 abs (observed ~5e-15 / 8e-17) | — / — | test_survey_parity.py (+1) |
svytotal |
survey::svytotal (Horvitz-Thompson + Taylor SE) | R 4.5.2 | estimate 1e-12 rel; SE 1e-10 rel (observed ~2e-12 / 1e-14) | — / — | test_survey_parity.py (+1) |
synth |
Synth::synth | R 4.5.2; Synth 1.1.10 | rel_est<=1e-06, rel_se<=1e-06 | 4.1e-10 / 2.0e-15 | 52_scm_unique.py (+2) |
synth_donor_sensitivity |
Synth::synth 1.1.10 on the replayed numpy donor subsets (custom.v matched; quadprog exact) | R 4.5.2; Synth 1.1.10; quadprog 1.5.8 | ATT and pre-RMSPE 1e-9 rel (observed <= 1e-13) | — / — | test_synth_rest_R_parity.py (+1) |
synth_loo |
Synth::synth 1.1.10 (per fit, custom.v matched; QP solved exactly by quadprog::solve.QP) | R 4.5.2; Synth 1.1.10; quadprog 1.5.8; kernlab 0.9.33 | ATT and pre-RMSPE 1e-9 rel (observed <= 3e-14) | — / — | test_synth_rest_R_parity.py (+1) |
synth_rmspe_filter |
SCtools::mspe.test 0.3.3.1 (discard.extreme, mspe.limit 2/5/20) on Synth fits | R 4.5.2; SCtools 0.3.3.1; Synth 1.1.10; quadprog 1.5.8 | placebo pre-RMSPE and ratios 1e-9 rel (observed <= 1e-13); p-values and kept counts exact | — / — | test_synth_rest_R_parity.py (+1) |
synth_time_placebo |
Synth::synth 1.1.10 (per placebo time, custom.v matched; quadprog exact) | R 4.5.2; Synth 1.1.10; quadprog 1.5.8 | placebo ATT 1e-9 rel (observed <= 1e-13) on identified placebo times | — / — | test_synth_rest_R_parity.py (+1) |
synthdid_estimate |
synthdid::synthdid_estimate | R 4.5.2; synthdid 0.0.9 | rel_est<=1e-06, rel_se<=1e-06 | 7.8e-16 / 1.1e-15 | 12_sdid.py (+3) |
synthesise_evidence |
R metafor::rma(method = 'FE') | metafor 5.0.1 | 1e-12 rel | — / — | test_misc_sens_R_parity.py |
tF_adjustment |
R ivDiag::tF 1.0.6 (LMMP 2022 tF table) | R 4.5.2; ivDiag 1.0.6 | rel 1e-12 on the same 26 F values (observed 0) | — / — | test_rd_iv_R_parity.py (+1) |
tF_critical_value |
R ivDiag::tF 1.0.6 (LMMP 2022 tF table) | R 4.5.2; ivDiag 1.0.6 | critical value at 26 F values from 4 to 1e4: rel 1e-12 (observed 0) | — / — | test_rd_iv_R_parity.py (+1) |
target_trial_emulate |
stats::lm(Y ~ A, weights = w) + sandwich::vcovHC(type = 'HC0') on the eligible subset; weights from ipw::ipwpoint (unweighted, unstabilised, stabilised) | R 4.5.2; ipw 1.3.0; ltmle 1.3.0; sandwich 3.1.1; SuperLearner 2.0.40 | 1e-9 relative on estimate and SE (observed 1.2e-15) | — / — | test_tte_longitudinal_r_parity.py (+1) |
test_calibration |
grf::test_calibration 2.6.1 (vcov.type HC3 default and HC1), forest outputs held fixed | R 4.5.2; grf 2.6.1; sandwich 3.1.1 | coef, se, t 1e-10 rel (observed 1.1e-15); one-sided p-value 1e-8 rel (observed 1.6e-13) | — / — | test_ml_causal_R_parity.py (+1) |
three_sls |
R systemfit::systemfit(method='3SLS') | R 4.5.2; systemfit 1.1.30 | coef 1e-9 abs (observed <= 1e-15); SE ~5e-3 rel | — / — | test_threesls_parity.py (+1) |
tmle |
tmle::tmle | R 4.5.2; tmle 2.1.1 | rel_est<=1e-06, rel_se<=1e-06 | 1.9e-09 / 1.9e-09 | 72_tmle.py (+2) |
tobit |
censReg::censReg | R 4.5.2; censReg 0.5.38 | rel_est<=1e-06, rel_se<=1e-06 | 1.1e-13 / 4.6e-12 | 41_tobit.py (+2) |
transitivity |
R igraph::transitivity(type = 'global') | igraph 2.3.3; sna 2.8; ergm 4.12.0; dyadRobust 0.0.1.0001 | Exact on karate. | — / — | test_network_parity.py |
transport_weights_fn |
R glm + quantile(type = 7) + sandwich::vcovHC(HC0) | sandwich 3.1.1 | 1e-12 rel | — / — | test_misc_sens_R_parity.py |
truncreg |
truncreg::truncreg(method="NR") | R 4.5.2; truncreg 0.2.5 | rel_est<=1e-06, rel_se<=1e-06 | 3.5e-10 / 7.9e-08 | 62_truncreg.py (+2) |
tsfilter |
Stata 18 tsfilter hp, bk, cf, bw; statsmodels hpfilter, bkfilter, cffilter | Stata 18; statsmodels 0.14.6 | trend and cycle 1e-9 rel | — / — | test_tsfilter_parity.py (+1) |
tvp_var |
KFAS::KFS, logLik (method='kalman'); sp.dlm per equation | R 4.5.2; KFAS 1.6.0 | proper prior: log likelihood, filtered and smoothed coefficients and variances 1e-11; diffuse prior: 1e-5 after the first k dates | — / — | test_tvp_var_parity.py (+1) |
twoway_cluster |
sandwich::vcovCL(cluster=~g1+g2) | R 4.5.2; sandwich 3.1.1 | rel_est<=1e-06, rel_se<=1e-06 | 7.8e-16 / 7.8e-16 | 54_twoway_cluster.py (+2) |
unified_sensitivity |
R EValue::evalues.OLS; sensemakr::sensemakr (rv_q, rv_qa) | EValue 4.1.4; sensemakr 0.1.6 | 1e-10 rel | — / — | test_misc_sens_R_parity.py |
var |
vars::VAR | R 4.5.2; vars 1.6.1 | rel_est<=1e-06, rel_se<=1e-06 | 3.1e-15 / 6.6e-15 | 33_var.py (+2) |
variable_importance |
grf 2.6.1 variable_importance (decay 2, depth 4) from grf's split frequencies, for nine forest kinds | R 4.5.2; grf 2.6.1 | 1e-12 absolute | — / — | test_grf_family_operator_parity.py (+1) |
vif |
R car::vif | ivmodel 1.9.1; car 3.1.5; metafor 5.0.1 | 2.0e-16 on every variance inflation factor, once the returned values stopped being rounded to two decimals. | — / — | test_weakiv_meta_parity.py |
vuong |
R pscl 1.5.9 vuong, on glm, MASS::glm.nb, zeroinfl and hurdle fits; Stata 18 MP zip / zinb, forcevuong |
pscl 1.5.9; Stata 18 MP | Raw, AIC-corrected and BIC-corrected statistics at 1e-6 on five model pairs (observed 6e-8 at worst, which is where R's optimisers stopped); the per-observation log-likelihoods under them at 1e-7 against R's own densities. pscl::vuong only prints, so the statistic is rebuilt in the R script and checked against the printed lines. Stata's forcevuong statistic for the two zero-inflated pairs agrees at 1e-6 (observed 2e-10 and 1e-9). | — / — | test_vuong_pscl_parity.py |
waic |
loo::waic (estimates, pointwise elpd and p) | R 4.5.2; loo 2.9.0 | 1e-9 abs pointwise, 1e-8 on the estimates | — / — | test_regression_stories_r_parity.py (+1) |
weighted_conformal_prediction |
cfcausal::conformal + predict.conformalSplit (type='mean', side='two'), unweighted and with density-ratio weights including the test point's own weight; conformalInference::conformal.pred.split(w=) agrees | R 4.5.2; cfcausal 0.2.1 (GitHub lihualei71/cfcausal e4ca670); conformalInference 1.1 (GitHub ryantibs/conformal 15c51c6) | 1e-12 relative on lower, upper and centre (observed <= 1e-15); infinite bands at the same test points | — / — | test_conformal_cfcausal_r_parity.py (+1) |
weightit |
WeightIt::weightit (glm, cbps, ebal, energy; binary, multi-category and continuous treatments) | WeightIt 1.7.0; cobalt 4.6.2 | weights 1e-9 relative, except multinomial glm 2e-7 (WeightIt's optimiser stops at a score of order 1e-6; ours is at 1e-11, asserted) and CBPS ATT 1e-8 (WeightIt's own CBPS ATT is 2.7e-9 from its entropy-balancing ATT, which it equals exactly); energy weights 1e-9 on 14 settings | — / — | test_weightit_cobalt_parity.py (+3) |
wild_cluster_boot |
R fwildclusterboot::boottest; Stata boottest (WCR, Rademacher, full enumeration) | R 4.5.2; fwildclusterboot 0.14.3; Stata 18; boottest 4.5.3 | p exact (multiple of 1/4096); t 1e-10 rel (observed 1.9e-14) | — / — | test_inference_sens_R_parity.py (+3) |
wild_cluster_bootstrap |
R fwildclusterboot::boottest; Stata boottest (WCR, Rademacher, full enumeration) | R 4.5.2; fwildclusterboot 0.14.3; Stata 18; boottest 4.5.3 | p exact (multiple of 1/4096); t 1e-10 rel (observed 1.9e-14) | — / — | test_inference_sens_R_parity.py (+3) |
wild_cluster_ci_inv |
R fwildclusterboot::boottest confidence interval (uniroot tol 1e-13) | R 4.5.2; fwildclusterboot 0.14.3 | CI endpoints 1e-9 rel (observed 4.8e-12) | — / — | test_inference_sens_R_parity.py (+2) |
wooldridge_did |
etwfe::etwfe + emfx | R 4.5.2; etwfe 0.6.2 | rel_est<=1e-06, rel_se<=1e-06 | 1.8e-13 / 3.9e-14 | 17_etwfe.py (+2) |
xcorr |
stats::ccf, stats::ar; Stata 18 xcorr | R 4.5.2; Stata 18 | cross-correlations 1e-10; AR coefficients 1e-9; Haugh statistic 1e-9 | — / — | test_xcorr_parity.py (+2) |
xtabond |
plm::pgmm | R 4.5.2; plm 2.6.7 | rel_est<=1e-06, rel_se<=1e-06 | 9.0e-16 / 1.4e-15 | 50_xtabond.py (+2) |
xtdpdsys |
Stata 18 xtdpdsys (built-in); xtabond2 iv(x, eq(diff)) h(2) | Stata 18 MP; xtabond2 SSC 03.07.00 | abdata, n on L.n (and w k): coefficients and SEs at 1e-9 (observed <= 6e-12) for one-step robust, two-step Windmeijer and classical one-step, instrument count equal; xtabond2 with iv(w k, eq(diff)) h(2) reproduces the same numbers. | — / — | test_dynpanel_abdata_parity.py |
xthtaylor |
Stata 18 MP xthtaylor: Hausman-Taylor conventional / vce(robust) / vce(cluster) with one and two endogenous time-varying regressors, Amemiya-MaCurdy (amacurdy) conventional / robust / cluster on balanced panels | Stata 18 MP (born 07 Jun 2023) | 1e-9 relative on coefficients, standard errors, sigma_u, sigma_e, rho and the Wald statistic (observed <= 9.6e-15 on 7 fits of a panel with gaps, plus 4 fits of the Hansen textbook panel) | — / — | test_panel_xt_tools_stata_parity.py (+2) |
xtlsdvc |
Stata xtlsdvc V1.0.4 (Bruno 2005), SSC | Stata 18 MP; xtlsdvc 1.0.4 | abdata: bias-corrected coefficients for initial(ab/ah/bb) x bias(1/2/3) and the AR(1)-only model at rtol 1e-7 (observed 4e-14 to 1.6e-9, the looser end through the Anderson-Hsiao initialiser). Standard errors are excluded by design: xtlsdvc reports the uncorrected LSDV ones, which sp.xtlsdvc reproduces and warns about. | — / — | test_lsdvc_parity.py |
xtnbreg |
Stata 18 xtnbreg, fe / re (Hausman-Hall-Griliches); pglm::pglm(family = negbin, model = 'within' / 'random') | Stata 18 MP; R 4.5.2; pglm 0.2.4 | coefficients / SEs rtol 1e-7 (observed <= 2.7e-8: Stata fe's score at its own estimate is 3e-7, pglm's gradtol), log-likelihood rtol 1e-12 | — / — | test_panel_xtnbreg_parity.py |
xtoverid |
SSC xtoverid 2.1.8 (run as a black box) after xtreg, re vce(robust) and vce(cluster), with and without a time-invariant regressor, full and balanced panel | Stata 18 MP (born 07 Jun 2023); xtoverid 2.1.8; ivreg2 4.1.12 | 1e-9 relative on the statistic and p-value (observed <= 2.5e-13); degrees of freedom exact | — / — | test_panel_xt_tools_stata_parity.py (+2) |
xtpcse |
Stata 18 MP xtpcse, 46 option combinations on a balanced and an unbalanced panel with gaps: correlation(independent / ar1 / psar1) x rhotype(regress / freg / tscorr / dw), hetonly, independent, nmk, np1, casewise / pairwise, noconstant. R pcse for the OLS panel-corrected standard errors. | Stata 18 MP (born 07 Jun 2023); R 4.5.2; pcse 1.9.1.1 | 1e-12 relative on coefficients, standard errors and the AR(1) coefficients (observed <= 2.5e-14; Stata stores the panel-specific ones as a 16-decimal macro); R pcse casewise and pairwise 1e-12 (observed 1.1e-15) | — / — | test_xtpcse_stata.py (+2) |
xtregar |
Stata 18 MP xtregar, 35 fits on a panel with gaps: fe and re with every rhotype, iterated and twostep, onestep, rhof(), lbi, the gap-free panels alone, a time-invariant regressor | Stata 18 MP (born 07 Jun 2023) | 1e-10 relative on coefficients, standard errors and every stored scalar (rho_ar, sigma_u, sigma_e, rho_fov, the R-squared, theta percentiles, F tests, the modified Durbin-Watson and LBI statistics); observed <= 5.1e-15 | — / — | test_xtregar_stata.py (+1) |
xtserial |
Stata xtserial 1.1.0 (Stata Journal st0039; run as a black box) on a panel with gaps, its balanced half and with a time-invariant regressor | Stata 18 MP (born 07 Jun 2023); xtserial 1.1.0 | 1e-9 relative on F, p-value and the residual autocorrelation (observed <= 4.6e-14); degrees of freedom exact | — / — | test_panel_xt_tools_stata_parity.py (+2) |
xtsum |
Stata 18 MP xtsum, every stored result (N, n, Tbar, mean, overall / between / within sd, min, max) for an outcome, a regressor, a time-invariant variable and one with missing values, on a panel with gaps | Stata 18 MP (born 07 Jun 2023) | 1e-9 relative (observed <= 1.8e-14); counts exact | — / — | test_panel_xt_tools_stata_parity.py (+2) |
xttobit |
Stata 18 MP xttobit (random effects): ll() and ul(), ll() only, ul() only, the 12-node default and 40 nodes, a collinear regressor | Stata 18 MP (born 07 Jun 2023) | 1e-6 relative on coefficients, standard errors (sigma_u and sigma_e included), rho, the log-likelihood and the LR test against the pooled tobit (observed <= 4.2e-8 on 5 fits) | — / — | test_xttobit_stata.py (+1) |
yu_elwert_decompose |
R cdgd::cdgd0_manual on independently fitted within-cell lm / within-group glm nuisances | DasGuptR 2.2.0; ddecompose 1.0.0; cdgd 1.0.1 | method='efficient': disparity, baseline, prevalence, effect, selection and their EIF standard errors at 1e-9. method='plugin' has no reference implementation and is covered by its exact additivity identity. | — / — | test_decomp_R_parity.py |
zero_first_stage |
Stata regress / ivregress 2sls and reghdfe / ivreghdfe: the component fits, combined in the do-file | Stata 18 MP; reghdfe 6.13.1; ivreghdfe 1.1.4; ivreg2 4.1.12 | 1e-9 relative on the four component coefficients, their SEs, the IV estimate and SE, and the corrected estimate (observed <= 1e-14) | — / — | test_zero_first_stage_stata.py (+1) |
zip_model |
pscl::zeroinfl(dist="poisson") | R 4.5.2; pscl 1.5.9 | rel_est<=1e-06, rel_se<=1e-06 | 1.2e-07 / 2.8e-08 | 63_zip.py (+2) |
zivot_andrews |
urca::ur.za | R 4.5.2; urca 1.3.4 | minimum t statistic, the path over break dates and the regression at the break 1e-7 rel; break date equal | — / — | test_zivot_andrews_parity.py (+1) |
aligned — 77 functions¶
Agreement within a documented, pre-registered looser tolerance.
| function | reference | versions | tolerance | rel err (R / Stata) | test |
|---|---|---|---|---|---|
ackerberg_caves_frazer |
Stata prodest, method(lp) acf valueadded; R prodest::prodestACF | Stata 18 MP; prodest SSC; R prodest 1.0.2 | StatsPAI returns an exact root of the just-identified moment conditions (criterion < 1e-25), the one nearest the stage-1 coefficients; R stops 2.4e-6 (relative) from that root with criterion 9e-16. Stata's Nelder-Mead stops at non-roots (criterion 4e-6 / 7e-6), which the test records. The parity panel has three roots, all reported in diagnostics['acf_roots']. | — / — | test_prodest_parity.py |
aft |
survival::survreg (Weibull AFT) | R 4.5.2; survival 3.8.3 | coefficients & log-scale 5e-5 abs (observed ~1e-5) | — / — | test_aft_parity.py (+1) |
augsynth |
augsynth::augsynth | R 4.5.2; augsynth 0.2.0 | rel_est<=2e-05, rel_se<=1e-06 | 7.9e-06 / 9.4e-13 | 18_augsynth.py (+2) |
avg_comparisons |
marginaleffects::avg_comparisons on lm / glm (logit, probit, poisson) fits with interactions, factors and squared terms | R 4.5.2; marginaleffects 0.32.0 | estimates 1e-8 relative; standard errors 2e-5 relative (slopes: estimates 1e-6, standard errors 2e-4), the precision of the reference's numerical Jacobian | — / — | test_marginaleffects_r_parity.py (+1) |
avg_predictions |
marginaleffects::avg_predictions on lm / glm (logit, probit, poisson) fits with interactions, factors and squared terms | R 4.5.2; marginaleffects 0.32.0 | estimates 1e-8 relative; standard errors 2e-5 relative (slopes: estimates 1e-6, standard errors 2e-4), the precision of the reference's numerical Jacobian | — / — | test_marginaleffects_r_parity.py (+1) |
avg_slopes |
marginaleffects::avg_slopes on lm / glm (logit, probit, poisson) fits with interactions, factors and squared terms | R 4.5.2; marginaleffects 0.32.0 | estimates 1e-8 relative; standard errors 2e-5 relative (slopes: estimates 1e-6, standard errors 2e-4), the precision of the reference's numerical Jacobian | — / — | test_marginaleffects_r_parity.py (+1) |
balance_vs_randomization |
iTOS::evalBal | iTOS 1.0.3 | the p-values of the matched sample, their minimum, truncated product and count below alpha 1e-8 relative under five settings; the shares of better-balanced simulated experiments within Monte Carlo error only | — / — | test_rosenbaum_itos_parity.py (+1) |
blp |
pyblp 1.2.0 (Conlon & Gortmaker), identical Halton nodes via agent_data | pyblp 1.2.0; numpy 2.2.6; python 3.10.20 | beta, sigma, SEs, objective/N, own elasticities 1e-6 rel (observed <= 1.6e-8) | — / — | test_blp_pyblp_parity.py (+1) |
breakdown_m |
HonestDiD::findOptimalFLCI 0.2.8 (Rambachan & Roth), breakdown by uniroot on the bound facing zero | R 4.5.2; HonestDiD 0.2.8; CVXR 1.8.2 | vs HonestDiD with its Monte-Carlo folded-normal quantile replaced by the exact one: 1e-9 (observed 6.1e-11); vs HonestDiD as shipped: 1e-3 (observed 4.3e-4), the simulation error of .qfoldednormal (1e6 draws, seed 0; 1.96224 vs exact 1.95996 at mu = 0) | — / — | test_did_synth_R_parity.py (+1) |
cate_eval |
grf::rank_average_treatment_effect 2.6.1 (AUTOC, QINI, TOC), fed grf's nuisances; shares sp.rate's operator | R 4.5.2; grf 2.6.1 | Point estimate (AUTOC, QINI) and TOC curve on q = 0.1..1 exact at 1e-10 rel (observed 4.6e-15), including a 30-group tied-priority case against rank_average_treatment_effect.fit. SE is T3: the analytic rank-corrected influence-function SE is compared with grf's half-sample bootstrap (R = 2000) at 6% rel (observed 2.6% AUTOC, 1.7% QINI). | — / — | test_ml_causal_R_parity.py (+1) |
causal_forest |
grf::causal_forest | R 4.5.2; grf 2.6.1 | rel_est<=0.01, rel_se<=0.05 | 8.1e-04 / — | 13_causal_forest.py (+1) |
cbps |
CBPS::CBPS 0.24 (Imai & Ratkovic 2014) | — | ATE over/exact and ATT exact: rel <= 5e-3 (R's optimiser slack). ATT over is NOT pinned to R -- CBPS's ATT gradient mis-scales the balance block by n/n_1 and stops off-stationarity; StatsPAI is asserted to attain strictly better covariate balance instead. | — / — | test_matching_r_parity.py (+1) |
centrality |
R igraph degree / betweenness / closeness / page_rank / eigen_centrality | igraph 2.3.3; sna 2.8; ergm 4.12.0; dyadRobust 0.0.1.0001 | degree, betweenness (normalised and raw), closeness and PageRank at 1e-10 on Zachary's karate club. The eigenvector column is L2-normalised (networkx) where igraph max-scales it: the ratio is constant across nodes to 1e-14, so it is the same vector under a documented normalisation. | — / — | test_network_parity.py |
cfm_decompose |
Stata cdeco, method(logit) 1.0.2 with drprocess (Chernozhukov, Fernandez-Val & Melly) | Stata 18.0 MP; cdeco 1.0.2 01mar2023 (bmelly/Stata counterfactual, Distribution-Date 20220803) | CDFs at the thresholds 1e-8 abs (observed 2.4e-10: Stata's logit stops at its default tolerance); quantiles 1e-12 abs (observed 0) | — / — | test_decomp_qte_parity.py (+1) |
cic |
qte::CiC | R 4.5.2; qte 2.0.0 | rel_est<=1e-06 | 2.4e-15 / 6.0e-03 | 74_cic.py (+2) |
cloglog |
stats::glm(binomial('cloglog')) | R 4.5.2 | coefficients 5e-5 abs (observed ~1e-5; IRLS convergence) | — / — | test_glm_ext_parity.py (+1) |
community_detection |
R igraph::cluster_louvain (T3: randomised on both sides) | igraph 2.3.3; sna 2.8; ergm 4.12.0; dyadRobust 0.0.1.0001 | T3, not T2: 200 seeded runs on karate have mean modularity within four combined standard errors of igraph's 200 runs (0.4157 vs 0.4145), and both reach the same maximum, 0.41979, the known optimum for this graph. | — / — | test_network_parity.py |
comparisons |
marginaleffects::comparisons on lm / glm (logit, probit, poisson) fits with interactions, factors and squared terms | R 4.5.2; marginaleffects 0.32.0 | estimates 1e-8 relative; standard errors 2e-5 relative (slopes: estimates 1e-6, standard errors 2e-4), the precision of the reference's numerical Jacobian | — / — | test_marginaleffects_r_parity.py (+1) |
conditional_lr_ci |
R ivmodel::CLR | ivmodel 1.9.1 | Monte Carlo by construction and graded as such: ivmodel integrates Moreira's conditional distribution while StatsPAI simulates it, so the two cannot agree deterministically. The test asserts the error SHRINKS with n_sim rather than pinning a number -- 3.8e-3 at n_sim=5,000 and 1.7e-4 at 200,000 -- which is the only honest statement about a simulated critical value. The endpoints themselves are bisected off-grid, so the residual is the critical value and not the grid. | — / — | test_weakiv_meta_parity.py |
conformal_ite_interval |
cfcausal naive ITE rule (conformalIteNaive: [L1 - U0, U1 - L0] from counterfactual bands at alpha/2) on the unconditional counterfactual bands | R 4.5.2; cfcausal 0.2.1 (GitHub lihualei71/cfcausal e4ca670); conformalInference 1.1 (GitHub ryantibs/conformal 15c51c6) | 1e-12 relative (observed <= 2.6e-15) with the reference run under StatsPAI's target population | — / — | test_conformal_cfcausal_r_parity.py (+1) |
cox_frailty |
R survival::coxph(... + frailty(id, theta=, sparse=FALSE)); Stata stcox, shared() | R 4.5.2; survival 3.8.3; Stata 18 MP | fixed theta: beta and SE 1e-9, integrated log likelihood 1e-10 (observed 4e-14); theta maximiser 1e-6 vs R optimize (observed 5e-8) and 5e-5 vs Stata e(theta) (observed 7.5e-6) | — / — | test_survival_epi_R_parity.py (+2) |
ctmle |
ctmle::ctmleDiscrete: the greedy and the pre-ordered sequence of candidate estimators | R 4.5.2; ctmle 0.1.2 | 1e-7 relative on the estimate at every step of the sequence (observed 5e-9), order of entry exact, 12 cases | — / — | test_ctmle_R_parity.py (+1) |
demeaned_synth |
augsynth::augsynth(progfunc = 'None', fixedeff = TRUE) 0.2.0 (de-meaned SCM) | R 4.5.2; augsynth 0.2.0; osqp 1.0.0 | gap path and ATT 1e-7 rel (observed 1.2e-9); weights 1e-8 abs | — / — | test_did_synth_synthvar_parity.py (+1) |
dml_model_averaging |
ddml::ddml_plm 0.3.1 (shortstack = TRUE, ensemble_type = 'nnls1'), OLS candidates, shared folds | R 4.5.2; ddml 0.3.1; sandwich 3.1.1 | short-stacking weights 1e-10 (observed 4.4e-14); ddml's final lm(y_r ~ d_r) with intercept and HC1 SE rebuilt from StatsPAI's stacked residuals at 1e-10 (observed 4.8e-16) | — / — | test_ml_causal_dml_parity.py (+2) |
doe_optimal |
AlgDesign 1.2.1.2 optFederov (D, A, I) on the same candidate sets | R 4.5.2; AlgDesign 1.2.1.2 | saturated degree-9 polynomial: the same ten runs (1e-12) and D value (1e-9); 14-run quadratic in three factors: D and A values 1e-9 relative, I value no worse than optFederov's design | — / — | test_joseph_doe_parity.py (+1) |
factorial_design |
FrF2 2.3-5 minimum aberration catalogue (resolution and word length pattern) and DoE.base 1.2-5 GWLP of the catalogued L9 / L18 arrays | R 4.5.2; FrF2 2.3.5; DoE.base 1.2.5 | resolution and word length pattern equal as integers; designs compared up to isomorphism, not run by run | — / — | test_joseph_doe_parity.py (+1) |
fmm |
Stata 18 fmm (regress / logit / poisson, lcprob()); R flexmix 2.3-20 as a second reference for Poisson | Stata 18.0 MP; flexmix 2.3-20 | ll 2.2e-16, coefficients and SEs 1.4e-8 on 12 of 13 cases (budget 1e-6); 2-class logit without lcprob 1e-5 / 2e-5 (observed 1.2e-6 / 5.4e-6): Stata's e(ll) exceeds its own ll at e(b) by 3.6e-7 | — / — | test_fmm_stata_parity.py (+3) |
full_match |
MatchIt::matchit(method = 'full') 4.7.2, optmatch 0.10.8; lm + sandwich::vcovCL | R 4.5.2; MatchIt 4.7.2; optmatch 0.10.8; sandwich 3.1.1 | Total matched distance <= optmatch's, and equal to it to 1e-7 at optmatch tol = 1e-9. Given optmatch's matched sets: ATT, ATE and matched-set-clustered standard error 1e-9. The estimate from StatsPAI's own sets is not pinned (5% screen): the optimum is nearly flat and the sets are not unique. | — / — | test_yuksel_aydede_causal_ml_parity.py (+1) |
garch |
Stata 18 arch; rugarch::ugarchfit 1.5.6 | R 4.5.2; rugarch 1.5.6; Stata 18 | log-likelihood at the reference optimum 1e-12 (observed 8.5e-16); vs Stata: b 1e-5, SE 2e-5 (observed 8.2e-6 / 1e-5); vs rugarch: params 5e-4, SE at their parameters 1e-2 | — / — | test_timeseries_R_parity.py (+2) |
genmatch |
Matching::Match 4.10-15 (Weight = 3, Weight.matrix) | — | Deterministic kernel only: given the same diagonal W, the 1-NN assignment agrees with Matching::Match on all 163 uniquely matched treated units on MatchIt::lalonde. | — / — | test_matching_r_parity.py (+1) |
gp_regress |
rkriging 1.0.2 ordinary kriging, Gaussian kernel: predictions and sd at given hyperparameters, and the fitted interpolator (restricted likelihood) | R 4.5.2; rkriging 1.0.2 | at given hyperparameters 1e-10 absolute on mean, predictions and sd; fitted: hyperparameters 1e-4 relative, predictions and expected improvement 2e-4 absolute | — / — | test_joseph_doe_parity.py (+1) |
grapple |
GRAPPLE::grappleRobustEst (GitHub jingshuw/GRAPPLE 317e837) | GRAPPLE 0.2.2 | sandwich at GRAPPLE's estimates 1e-6 (l2 1e-12); fitted beta 1e-3, tau2 / SE 1e-4 | — / — | test_misc_sens_R_parity.py |
hits |
R igraph::hits_scores | igraph 2.3.3; sna 2.8; ergm 4.12.0; dyadRobust 0.0.1.0001 | Hub and authority vectors are igraph's up to normalisation: L1 here (documented), max = 1 in igraph; the ratio is constant across nodes to 1e-11. | — / — | test_network_parity.py |
hypotheses |
marginaleffects::hypotheses on lm / glm (logit, probit, poisson) fits with interactions, factors and squared terms | R 4.5.2; marginaleffects 0.32.0 | estimates 1e-8 relative; standard errors 2e-5 relative (slopes: estimates 1e-6, standard errors 2e-4), the precision of the reference's numerical Jacobian | — / — | test_marginaleffects_r_parity.py (+1) |
iv_wild_bootstrap |
boottest after ivregress 2sls, small (WRE, Rademacher weights enumerated over 2^12 sign vectors): over- and just-identified, with and without an exogenous control, beta0 = 0 and 0.5, levels 95 / 93.75 / 90, two endogenous regressors; the disjoint weak-IV confidence set with boottest's default and a wide grid | Stata 18; boottest 4.5.3 | p-values exact (counts over 4,096 enumerated sign vectors), t-statistics 1e-9 relative (observed 2e-15); confidence bounds within 0.02 standard errors (observed 0.006 with one endogenous regressor, 0.018 with two) | — / — | test_regcmd_stata_r_parity.py (+2) |
lcsf |
sfaR::sfalcmcross 1.0.1 (2 classes, half-normal) | R 4.5.2; sfaR 1.0.1 | estimates 1e-6 rel (observed 1.1e-8); OIM SEs 1e-6 rel (observed 8.6e-8) | — / — | test_frontier_struct_R_parity.py (+1) |
levinsohn_petrin |
Stata prodest, method(lp) valueadded (Rovigatti & Mollisi); R prodest::prodestLP | Stata 18 MP; prodest SSC; R prodest 1.0.2 | As olley_pakes with materials as the proxy: free-input coefficient at 1e-12 against both; state coefficient at the minimiser, R 1.3e-6 and Stata up to 1.8e-3 away where their optimisers stop. | — / — | test_prodest_parity.py |
lincom |
Stata 18 lincom (after regress / ivregress / logit / poisson) | Stata 18 | 1e-6 rel (observed <= 2.3e-15 on linear fits, <= 2.3e-8 overall) | — / — | test_postestimation_stata_parity.py (+1) |
malmquist |
sfaR::sfacross 1.0.1 per period + sfaR::efficiencies (teBC / teJLMS); TC from sfaR betas | R 4.5.2; sfaR 1.0.1; metafrontier 0.3.1 | EC / TC / M 1e-6 rel (observed 2.7e-7); metafrontier::malmquist_meta(method='sfa') EC_group and 1/TC_group 1e-4 (observed 1.8e-5, reference optim BFGS with finite-difference gradients) | — / — | test_r2_frontier_parity.py (+1) |
margins |
Stata 18 margins, dydx(*) | Stata 18 | 1e-6 rel (observed 2.7e-7 AME / 7.5e-8 SE on probit and logit-with-interaction; <= 1e-10 otherwise) | — / — | test_postestimation_stata_parity.py (+1) |
melogit |
lme4::glmer | R 4.5.2; lme4 2.0.1 | rel_est<=0.0002, rel_se<=1e-06 | 1.7e-04 / 1.4e-08 | 26_glmm_logit.py (+2) |
mmqreg |
Stata 18 MP with SSC mmqreg 2.6 (run as a black box): one and two absorbed sets and none, q(10 / 25 / 50 / 75 / 90), vce default (GLS) / robust / cluster | Stata 18 MP (born 07 Jun 2023); mmqreg 2.6 | Location, scale and quantile coefficients and the 12 x 12 covariance of the three equations 1e-10 (observed <= 4e-12) on 18 fits; density e(fden) and e(qval) 1e-10. On 6 fits where n tau is an integer Stata returns the other end of a tied sample quantile (see note): there the quantile coefficients differ by up to 1.7% and agree to 1e-10 once tau is moved 1e-11 to select that end. | — / — | test_mmqreg.py (+1) |
model_averaging_dml |
ddml::ddml_plm 0.3.1 (shortstack = TRUE, ensemble_type = 'nnls1'), OLS candidates, shared folds | R 4.5.2; ddml 0.3.1; sandwich 3.1.1 | short-stacking weights 1e-10 (observed 4.4e-14); ddml's final lm(y_r ~ d_r) with intercept and HC1 SE rebuilt from StatsPAI's stacked residuals at 1e-10 (observed 4.8e-16) | — / — | test_ml_causal_dml_parity.py (+2) |
mr_raps |
R mr.raps 0.4.3 (simple / overdispersed / overdispersed.robust) | MendelianRandomization 0.10.0; TwoSampleMR 0.7.9; RadialMR 1.2.4; MRPRESSO 1.0; mr.raps 0.4.3 | Simple and L2-overdispersed fits (beta, SE, tau2) at 1e-8. Robust Huber / Tukey: the sandwich reproduces R's SEs at 1e-10 when evaluated at R's own (beta, tau2) and integrate() moments; the fitted values agree to 5e-5 (beta), 5e-4 (SE), 1.5e-3 (tau2) because R stops uniroot at its default tolerance and integrate() at 1.2e-4 -- a test shows StatsPAI's root satisfies the estimating equation more tightly than R's. | — / — | test_mr_R_parity.py |
notears |
notears.linear.notears_linear (authors' implementation, lambda1 = 0.1, least-squares loss) | python 3.11.15; notears 3.0 (github.com/xunzheng/notears 4a9ab19); scipy 1.13.1 | thresholded graph exact on 16 designs (44 edges); edge weights 5e-2 absolute (observed <= 1.2e-2) | — / — | test_notears_reference_parity.py (+1) |
olley_pakes |
Stata prodest, method(op) valueadded (Rovigatti & Mollisi); R prodest::prodestOP | Stata 18 MP; prodest SSC; R prodest 1.0.2 | Simulated unbalanced panel with 31 calendar gaps. Free-input coefficient (stage-1 OLS) at 1e-12 against both, polynomial degree 2 and 3. State coefficient: StatsPAI solves the first-order condition and has a sum of squares no larger than at either reference estimate; the references stop their optimisers early (R BFGS 5.9e-6 away, Stata Nelder-Mead at tolerance 1e-5 up to 4.3e-3 away). | — / — | test_prodest_parity.py |
optimal_design |
Stata 18 power twomeans, knownsds (and cluster m1() m2() rho()): sample sizes, cluster counts and the minimum detectable difference | Stata Stata 18 MP (07 Jun 2023) | sample sizes and cluster counts equal (integers); unrounded n and detectable difference 2.5e-6 relative, rebuilt to 1e-9 | — / — | test_optimal_design_stata_parity.py (+1) |
optimal_match |
optmatch::pairmatch 0.10.8 on a logit propensity score | — | Total matched distance <= optmatch's (1 + 1e-6). The matched pairs are not pinned: the assignment problem is degenerate on this data, so equally optimal solutions report different ATTs. | — / — | test_matching_r_parity.py (+1) |
predictions |
marginaleffects::predictions on lm / glm (logit, probit, poisson) fits with interactions, factors and squared terms | R 4.5.2; marginaleffects 0.32.0 | estimates 1e-8 relative; standard errors 2e-5 relative (slopes: estimates 1e-6, standard errors 2e-4), the precision of the reference's numerical Jacobian | — / — | test_marginaleffects_r_parity.py (+1) |
pscore |
Stata 18 pscore (Becker and Ichino, st0026_2) | Stata 18; pscore st0026_2 | blocks, common-support flags and the list of unbalanced (covariate, block) pairs exact; scores and coefficients 1e-8 relative; t-test p-values 1e-10; log likelihood 1e-9; standard errors 5e-5 | — / — | test_pscore_stata_parity.py (+1) |
qdid |
qte::QDiD 1.3.1 | — | max deviation / scale < 0.08, sign agreement on the large effects and correlation > 0.999. R's quantiles come from BMisc::weighted_quantile (stats::optimize on a piecewise-linear check function, which has plateaus where every point is a minimiser) while sp.qdid interpolates the empirical inverse CDF; on this fixture the gap is at most ~152 currency units against effects running to ~8900. | — / — | test_qdid_parity.py (+1) |
qte |
qte::ci.qte / qte::ci.qtet 1.3.1 (Firpo 2007) | — | max relative deviation < 0.01 on lalonde.exp / lalonde.psid. Both sides minimise the same weighted check function; R's BMisc::weighted_quantile uses a golden-section search whose answer on a plateau is an optimiser artifact, so point-value equality is not asserted. | — / — | test_firpo_qte_parity.py (+1) |
rate |
grf::rank_average_treatment_effect and rank_average_treatment_effect.fit 2.6.1 (AUTOC, QINI, TOC), forest outputs held fixed | R 4.5.2; grf 2.6.1 | Point estimate (AUTOC, QINI) and TOC curve on q = 0.1..1 exact at 1e-10 rel (observed 4.6e-15), including a 30-group tied-priority case against rank_average_treatment_effect.fit. SE is T3: the analytic rank-corrected influence-function SE is compared with grf's half-sample bootstrap (R = 2000) at 6% rel (observed 2.6% AUTOC, 1.7% QINI). | — / — | test_ml_causal_R_parity.py (+1) |
rd_bias_aware_fuzzy |
R RDHonest 1.0.1.9000, RDHonest(y | d ~ x) | R 4.5.2; RDHonest 1.0.1.9000 | estimate, std.error, maximum.bias, conf.low/high, M, first stage: rel 1e-9 at fixed h and M (observed 1.1e-14), 1e-6 with selected h (observed 3.5e-8, optimiser) | — / — |
rd_discrete |
R RDHonest::RDHonest / RDHonestBME 1.0.1.9000 (Kolesar) | R; RDHonest 1.0.1.9000 | estimate, std.error, maximum.bias, conf.low/high rel 1e-9 at fixed h and for BME; 1e-6 when RDHonest selects h and M (optimiser) | — / — | test_rd_open_R_parity.py (+1) |
rdrbounds |
R rdlocrand::rdrbounds 2.0 | R 4.5.2; rdlocrand 2.0 | upper and lower bounds within 4 pooled binomial SEs at 4000 draws (Monte Carlo on both sides, not parity) | — / — | test_rd_iv_rd_R_parity.py (+1) |
rdsensitivity |
R rdlocrand::rdsensitivity / rdrandinf 2.0 | R 4.5.2; rdlocrand 2.0 | per-window estimates rel 1e-9 (rdrandinf observed statistic); randomization p-values within 4 pooled binomial SEs at 4000 draws (Monte Carlo on both sides, not parity) | — / — | test_rd_iv_rd_R_parity.py (+1) |
residual_balance |
balanceHD::approx.balance and residualBalance.ate (optimizer = 'quadprog', fit.method = 'none') | balanceHD 1.0 | weights and pure-weighting estimates 2e-5: balanceHD's quadprog path stops short of the optimum of the strictly convex programme (its objective is 1e-10 to 1e-8 above ours in every setting, asserted) | — / — | test_residual_balance_balancehd.py (+1) |
rlassologit |
R hdm::rlassologit 0.3.2 (glmnet 4.1.10 engine) | R 4.5.2; hdm 0.3.2; glmnet 4.1.10 | support exact; post-Lasso atol 1e-5 (observed rel 1.7e-10); post=False atol 1e-4 (observed 1.3e-6: glmnet coordinate-descent convergence) | — / — | test_rlassologit_parity.py (+1) |
sarar_gmm |
spatialreg::gstsls 1.4.3 (Kelejian-Prucha GS2SLS) | R 4.5.2; spatialreg 1.4.3; spdep 1.4.2 | beta, rho and SEs 1e-9 rel (observed 3.8e-11 / 3.0e-11); lambda 1e-7 rel (observed 6.5e-9) | — / — | test_spatial_survey_R_parity.py (+1) |
scest |
R scpi::scest (w.constr simplex / lasso / ridge / ols / L1-L2, V = 'separate') | scpi 4.0.1; CVXR 1.9.2; clarabel 0.11.2; osqp 1.0.0 | ols and lasso weights at 1e-9 abs; ridge Q / lambda and L1-L2 Q2 at 1e-10. simplex / ridge / L1-L2 weights are bounded by the conic solver: R's objective exceeds StatsPAI's exact optimum by <= 1e-7 relative (CLARABEL gap 1e-8), R's point is feasible, and | w_R - w_py | |
scpi |
R scpi::scpi (effect = 'unit-time', u.missp, u.sigma = HC1, u.order = e.order = 1, rho = type-2, e.method = all) | scpi 4.0.1; CVXR 1.9.2; ECOSolveR 0.6.1; Qtools 1.6.0; quantreg 6.1 | On R's weights: rho, Q.star, u.mean, Omega, Sigma, e.mean at 1e-9; out-of-sample e.var and gaussian / ls / qreg bounds at 1e-9 against R with rrq(method = 'br') (exact LP) and 5e-4 abs against the default Frisch-Newton rrq. In-sample simulation fed R's draws: per-draw median <= 1e-6 and max <= 2e-4 vs ECOS at 1e-12, quantile bounds at 1e-5; vs default ECOS (1e-8) bounds within 2e-3 abs. Average-effect CI = scdataMulti(effect = 'unit') within 5e-4. J > T0 (California, 38 donors) covered. | — / — | test_did_synth_scpi_parity.py (+1) |
shapley_inequality |
Stata shapley2 1.5 (Chavez Juarez) over regress + ineqdeco (Jenkins) | Stata 18.0 MP; shapley2 1.5 10jun15 (SSC) | value function v(S) 1e-11 rel (observed 1.4e-13); Shapley values 1e-6 rel (observed 2.7e-7, shapley2 routes v(S) through float variables); totals 1e-12 | — / — | test_decomp_qte_parity.py (+1) |
shapley_r2 |
R domir 1.3.0 and relaimpo 2.2.7 (lmg); Stata shapley2 1.5 | R 4.5.2; domir 1.3.0; relaimpo 2.2.7; Stata 18.0 MP; shapley2 1.5 10jun15 (SSC) | R 1e-10 rel (observed 8e-15); Stata 1e-6 rel (observed 3.2e-7, shapley2 stores subset R2 as float) | — / — | test_shapley_r2_parity.py (+2) |
slopes |
marginaleffects::slopes on lm / glm (logit, probit, poisson) fits with interactions, factors and squared terms | R 4.5.2; marginaleffects 0.32.0 | estimates 1e-8 relative; standard errors 2e-5 relative (slopes: estimates 1e-6, standard errors 2e-4), the precision of the reference's numerical Jacobian | — / — | test_marginaleffects_r_parity.py (+1) |
sobol_indices |
sensitivity 1.31.0 soboljansen on the same two input matrices (estimator='jansen') | R 4.5.2; sensitivity 1.31.0 | 1e-12 absolute after the divisor rescaling | — / — | test_joseph_doe_parity.py (+1) |
spatial_panel |
splm::spml(model = 'within') 1.6.5; Stata xsmle 1.4.5 fe type(ind) vce(oim) | R 4.5.2; splm 1.6.5; plm 2.6.7; Stata 18; xsmle version 1.4.5 5jun2017 | vs splm: estimates and SEs 1e-7 rel (observed 7.2e-8 / 1.3e-8, the splm optimize floor); vs xsmle (tightened ml tolerances): estimates and beta SEs 1e-9 (observed 2.5e-13 / 8.5e-11), spatial-parameter SE 1e-8 (observed 3.0e-9) | — / — | test_spatial_survey_R_parity.py (+2) |
test |
Stata 18 test (after regress / ivregress / logit / poisson) | Stata 18 | 1e-6 rel (observed <= 2.3e-15 on linear fits, <= 8e-10 on ML fits, far-tail p <= 4.7e-8) | — / — | test_postestimation_stata_parity.py (+1) |
tighten_blocks |
tightenBlock::tighten | tightenBlock 0.1.7 | objective exactly equal on truncated costs, the same block penalty and the same number of dropped blocks at three subset prices | — / — | test_rosenbaum_itos_parity.py (+1) |
trimming |
propensity::ps_trim (method = 'adaptive', the Crump et al. rule, and method = 'ps' with bounds 0.1 / 0.9) | propensity 0.1.0; R 4.5.2 | fixed bounds: the kept rows are R's. Crump: cutoff 1e-10 relative and the same kept rows when the sample equation gamma = 2 mean(g | g <= gamma) has one root; with several roots each side returns a root and the kept sets differ by the rows between them | — / — |
two_criteria_match |
iTOS::makematch with addcaliper / addinteger / addMahal / addNearExact / addquantile | iTOS 1.0.3 | cost matrices 1e-10; optimal objective exactly equal on integer costs and on costs truncated to integers as makematch does; on the costs as given ours is lower and equals an independent LP solution to 1e-10 | — / — | test_rosenbaum_itos_parity.py (+1) |
weakrobust |
R ivmodel::CLR 1.9.1; Stata weakiv 2.4.07 (md small) | R 4.5.2; ivmodel 1.9.1; Stata 18 MP; weakiv 2.4.07 | CLR statistic and p-value vs ivmodel rel 1e-9 (observed 1.2e-11); CLR set vs ivmodel 5e-5 (observed 1.2e-5, ivmodel's uniroot default tolerance); CLR / K / AR vs Stata weakiv 1e-6 (observed 1.1e-7 / 1.1e-7 / 2.1e-8, not bisected further) | — / — | test_rd_iv_R_parity.py (+2) |
xtfrontier |
frontier::sfa | R 4.5.2; frontier 1.1.8 | rel_est<=5e-06, rel_se<=1e-06 | 9.7e-07 / 5.9e-12 | 29_panel_sfa.py (+2) |
xtintreg |
Stata 18 MP xtintreg: default (mvaghermite) fits at 12 and 40 nodes, non-adaptive intmethod(ghermite) intpoints(100), left- and right-censored intervals, a collinear regressor | Stata 18 MP (born 07 Jun 2023) | Estimates and log-likelihood 1e-6 against every Stata fit that reached its maximum (observed <= 2e-7); standard errors 1e-6 against the non-adaptive 100-node fits (observed <= 6.8e-8). Against Stata's default adaptive fits the standard errors differ by up to 0.47% (interval data) and 2.3% (right-censored) on se(sigma_e), compared at 5e-2. | — / — | test_xttobit_stata.py (+1) |
xtivreg |
Stata 18 MP xtivreg: re (G2SLS, ec2sls, nosa), fe, be, fd; vce conventional / robust / cluster; small, first, regress; one and two endogenous regressors; unbalanced, balanced and gapped panels, time-invariant regressors and instruments. R plm within and between IV. | Stata 18 MP (born 07 Jun 2023); R 4.5.2; plm 2.6.7 | 1e-12 relative on coefficients, standard errors, sigma_u, sigma_e, rho and the three R-squared for 33 of 41 Stata fits (observed <= 2.3e-14) and for the 6 plm fits (observed <= 3e-14). The other 8 are random-effects fits whose sigma_e divisor Stata counts differently (see note): under Stata's count 7 agree to 1e-12 on every number and the eighth's sigma_e is reproduced from our within sum of squares; by default sigma_e and rho differ by <= 2.5e-3, coefficients by <= 1.2e-4 and standard errors by <= 8.2e-4. Options (stata_xtivreg_opts.json, 20 fits and 8 first stages): coefficients and standard errors 1e-12 (observed <= 1.6e-13), model chi2 / F, small-sample df and the reported scalars 1e-11 except the Stata conventions listed in the note. | — / — | test_xtivreg_stata.py (+5) |
zinb |
pscl::zeroinfl(dist="negbin") | R 4.5.2; pscl 1.5.9 | rel_est<=3e-06, rel_se<=1e-06 | 9.5e-07 / 4.5e-11 | 64_zinb.py (+2) |
zisf |
Stata chks 1.1 (estimation(zsf) eoption(ml)); R sfa::zsfm 1.2.0 (ZISF / ZISF_Z, likelihood at its optimum) | R 4.5.2; sfa 1.2.0; numDeriv 2016.8.1.1; stata 18; chks 1.1 (chks.pkg dated 20190320) | estimates and OIM SEs 1e-6 rel (observed chks 8.6e-8 / 9.4e-8; sfa likelihood at its optimum 1.3e-8 / 5.4e-8); sfa's reported L-BFGS-B point 5e-5 / 5e-4 (observed 1.6e-5 / 1.8e-4) | — / — | test_r2_frontier_parity.py (+2) |
external-replication — 54 functions¶
Reproduces published-paper numbers; sources in tests/external_parity/PUBLISHED_REFERENCE_VALUES.md.
analytical-only — 373 functions¶
Recovers a known DGP truth / closed-form identity within tolerance; no cross-package reference. See tests/reference_parity/REFERENCES.md.
unverified — 654 functions¶
These are registered public functions with no cross-language or published-reference parity evidence attached yet. This is the honest coverage gap, not a claim of incorrectness — many are frontier methods with no Stata/R sibling to align against. Query any of them with sp.parity_status(name); the closing roadmap lives in docs/dev/parity_status_roadmap.md.