Skip to content
Open navigation
OpenOpen workspace

Statistical validation

A transparent, reproducible comparison of selected browser-engine calculations with SciPy. This is automated parity evidence, not independent human review or certification.

Last updated

Current validation scope

The public manifest contains 48 reference cases: 7 generated with SciPy 1.18.0, plus 41 browser-engine cases independently reproduced with a named package or an embedded transparent formula. Release tests recompute every stored result and compare the browser engine with the same values.

  • 2 Pearson chi-square cases, including expected frequencies, p-values, degrees of freedom, and Cramér's V.
  • 4 two-sided Fisher exact cases covering a small classic table, equal central modes, a zero cell, and skewed margins.
  • 1 directional Somers' d case using SciPy's row-independent, column-dependent convention and its transpose.
  • 10 additional cases covering Yates' correction, the likelihood-ratio G-test, Woolf odds-ratio intervals, signed phi, Pearson's contingency coefficient, pooled lambda, gamma, tau-b and tau-c, pooled Theil's U, and adjusted standardized residuals.
  • 31 CROSSTABS cases covering asymptotic standard errors, approximate T and significance for lambda, Goodman-Kruskal tau, the uncertainty coefficient, gamma, tau-b, tau-c and Somers' d; phi, Cramér's V and the contingency coefficient with p-values; Pearson and Spearman correlations, linear-by-linear association and eta; kappa, McNemar and McNemar-Bowker; odds ratio and relative risks; one-sided and R×C Fisher exact tests; raw and standardized residuals; and Cochran-Mantel-Haenszel, Breslow-Day, Tarone and the Mantel-Haenszel common odds ratio. References come from statsmodels, R (DescTools, fisher.test, mantelhaen.test), SciPy and a 50-digit delta-method calculation.

Open the machine-readable reference manifest. Its numeric tolerance is the larger of 1.0 × 10−14 absolute or 1.0 × 10−10 relative error.

Reference cases

48 cases, each showing its input table and the values the browser engine must reproduce within the published tolerance.

All 48 cases

Pearson chi-square, 2×2 balanced margins

[[20,30],[30,20]]

χ² = 4; df = 1; p = 0.0455002638964; Cramér's V = 0.2. Pearson calculation uses correction=false.

Pearson chi-square, 3×3 table

[[50,10,5],[20,30,10],[5,15,25]]

χ² = 67.869010619; df = 4; p = 6.392728 × 10−14; Cramér's V = 0.446782650866. Pearson calculation uses correction=false.

Fisher's exact test, classic 2×2 table

[[3,1],[1,3]]

Sample odds ratio = 9; two-sided p = 0.485714285714.

Fisher's exact test, equal-probability central modes

[[20000,19999],[19999,20000]]

Sample odds ratio = 1.0001000075; two-sided p = 1.

Fisher's exact test, zero cell

[[0,5],[10,15]]

Sample odds ratio = 0; two-sided p = 0.140036209002.

Fisher's exact test, skewed margins

[[1,99],[8,92]]

Sample odds ratio = 0.116161616162; two-sided p = 0.0348991519721.

Somers' d, SciPy hotel-cleanliness example

[[27,25,14,7,0],[7,14,18,35,12],[1,3,2,7,17]]

d(column|row) = 0.603276611151; d(row|column) = 0.4704595186; symmetric d = 0.528653579219.

Yates continuity-corrected chi-square, 2x2 table

[[12,5],[29,2]]

chiSquare = 2.98601643647; degreesOfFreedom = 1; pValue = 0.083986541715.

Reference: SciPy 1.18.0. Method: scipy.stats.chi2_contingency(table, correction=True)

Likelihood-ratio G-test, 3x3 table

[[50,10,5],[20,30,10],[5,15,25]]

gStatistic = 67.6208218592; degreesOfFreedom = 4; pValue = 7.211644 × 10−14.

Reference: SciPy 1.18.0. Method: scipy.stats.chi2_contingency(table, correction=False, lambda_='log-likelihood')

Sample odds ratio with Woolf 95% confidence interval

[[40,10],[10,40]]

oddsRatio = 16; ciLower = 6.00508571861; ciUpper = 42.6305321848.

Reference: Python standard-library math and SciPy Python 3.13.13; SciPy 1.18.1. Method: OR=(a*d)/(b*c); CI=exp(log(OR) +/- scipy.stats.norm.ppf(0.975)*sqrt(1/a+1/b+1/c+1/d))

Signed phi coefficient, negative association

[[5,50],[50,5]]

phi = -0.818181818182.

Reference: Python standard-library math 3.13.13. Method: phi=(a*d-b*c)/sqrt((a+b)*(c+d)*(a+c)*(b+d))

Pearson contingency coefficient, 3x3 table

[[50,10,5],[20,30,10],[5,15,25]]

contingencyCoefficient = 0.534154419466.

Reference: SciPy 1.18.0 plus Python standard-library math 1.18.0. Method: chi2=scipy.stats.chi2_contingency(table, correction=False).statistic; C=sqrt(chi2/(chi2+N))

Goodman-Kruskal pooled lambda, 3x3 table

[[30,2,3],[4,20,6],[8,5,22]]

symmetric = 0.544715447154; rowDependent = 0.569230769231; colDependent = 0.51724137931.

Reference: NumPy 2.5.1. Method: pooled=(sum(column maxima)+sum(row maxima)-max(row totals)-max(column totals))/(2*N-max(row totals)-max(column totals)); directional lambdas use their corresponding proportional error reduction

Goodman-Kruskal gamma, ordinal 3x3 table

[[30,10,5],[10,20,10],[5,10,30]]

gamma = 0.69696969697.

Reference: Python standard-library arithmetic 3.13.13. Method: gamma=(concordant-discordant)/(concordant+discordant), with pair counts recomputed by direct four-loop enumeration

Kendall tau-b and Stuart tau-c, ordinal 2x4 table

[[20,10,5,2],[5,10,20,30]]

kendallTauB = 0.542832333959; kendallTauC = 0.63629373318.

Reference: Python standard-library arithmetic 3.13.13. Method: tau_b=(C-D)/sqrt((C+D+T_row)*(C+D+T_col)); tau_c=2*m*(C-D)/(N^2*(m-1)); pair and tie counts use direct enumeration

Pooled and directional Theil uncertainty coefficients

[[30,5],[10,5],[5,45]]

symmetric = 0.339807206285; rowGivenCol = 0.286987204028; colGivenRow = 0.416455931073.

Reference: NumPy plus Python standard-library math 2.5.1. Method: I=H(row)+H(column)-H(joint); U(row|column)=I/H(row); U(column|row)=I/H(column); pooled U=2*I/(H(row)+H(column))

Adjusted standardized residuals, 3x3 table

[[50,10,5],[20,30,10],[5,15,25]]

residuals = [[6.777799374399998,-3.720866942175243,-3.830094678535809],[-2.0914753180106986,3.632408916720151,-1.557933303179625],[-5.200344867856302,0.163943497448759,5.906464660549895]].

Reference: NumPy 2.5.1. Method: r_ij=(O_ij-E_ij)/sqrt(E_ij*(1-R_i/N)*(1-C_j/N)), where E_ij=R_i*C_j/N

CROSSTABS nominal measures with ASE and significance, scipy-hotel

[[27,25,14,7,0],[7,14,18,35,12],[1,3,2,7,17]]

phi = 0.699398889904; phiP = 1.477365 × 10−16; cramersV = 0.494549697806; cramersVP = 1.477365 × 10−16; contingencyCoefficient = 0.573131701935; contingencyCoefficientP = 1.477365 × 10−16; lambdaSymmetricValue = 0.271604938272; lambdaSymmetricAse = 0.0554903264662; lambdaSymmetricT = 4.40046568238; lambdaSymmetricP = 1.080188 × 10−5; lambdaRowDependentValue = 0.349514563107; lambdaRowDependentAse = 0.0790827294674; lambdaRowDependentT = 3.69074811138; lambdaRowDependentP = 0.000223595472131; lambdaColumnDependentValue = 0.214285714286; lambdaColumnDependentAse = 0.0482190095947; lambdaColumnDependentT = 4.11159188581; lambdaColumnDependentP = 3.929404 × 10−5; goodmanKruskalTauRowDependentValue = 0.23327249487; goodmanKruskalTauRowDependentAse = 0.0452596683831; goodmanKruskalTauRowDependentP = 1.354755 × 10−15; goodmanKruskalTauColumnDependentValue = 0.113982132265; goodmanKruskalTauColumnDependentAse = 0.0247160789216; goodmanKruskalTauColumnDependentP = 3.435422 × 10−15; uncertaintySymmetricValue = 0.186693049112; uncertaintySymmetricAse = 0.0315718936661; uncertaintySymmetricT = 5.80275510729; uncertaintySymmetricP = 1.729493 × 10−16; uncertaintyRowDependentValue = 0.239411546996; uncertaintyRowDependentAse = 0.0402419196656; uncertaintyRowDependentT = 5.80275510729; uncertaintyRowDependentP = 1.729493 × 10−16; uncertaintyColumnDependentValue = 0.153001967735; uncertaintyColumnDependentAse = 0.0260851572134; uncertaintyColumnDependentT = 5.80275510729; uncertaintyColumnDependentP = 1.729493 × 10−16.

Reference: R DescTools, SciPy and mpmath SciPy 1.18.1; statsmodels 0.15.0; mpmath 1.4.1; R 4.6.1 DescTools 0.99.60. Method: phi/V/C and p: scipy.stats.chi2_contingency + contingency.association; lambda and Goodman-Kruskal tau values/ASE: DescTools Lambda/GoodmanKruskalTau longer CI half-width / qnorm(0.975); uncertainty coefficient values and ASE: 50-digit multinomial delta method; approximate T: value / null-hypothesis SE (first-index modal categories for lambda); tau p: chi2.sf((W-1)(K-1)tau, (R-1)(C-1)); U p: likelihood-ratio chi-square p.

CROSSTABS nominal measures with ASE and significance, job-satisfaction-income

[[1,2,1,0],[3,3,6,1],[10,10,14,9],[6,7,12,11]]

phi = 0.249280518626; phiP = 0.743364725053; cramersV = 0.143922174532; cramersVP = 0.743364725053; contingencyCoefficient = 0.241878517662; contingencyCoefficientP = 0.743364725053; lambdaSymmetricValue = 0.0258620689655; lambdaSymmetricAse = 0.0407052512206; lambdaSymmetricT = 0.626822037187; lambdaSymmetricP = 0.530775896421; lambdaRowDependentValue = 0.0377358490566; lambdaRowDependentAse = 0.0827725398278; lambdaRowDependentT = 0.447680172148; lambdaRowDependentP = 0.654384033452; lambdaColumnDependentValue = 0.015873015873; lambdaColumnDependentAse = 0.0272737997773; lambdaColumnDependentT = 0.578355231254; lambdaColumnDependentP = 0.563024312585; goodmanKruskalTauRowDependentValue = 0.0174033917153; goodmanKruskalTauRowDependentAse = 0.017005844303; goodmanKruskalTauRowDependentP = 0.837784005072; goodmanKruskalTauColumnDependentValue = 0.0197614542848; goodmanKruskalTauColumnDependentAse = 0.0154127977067; goodmanKruskalTauColumnDependentP = 0.776107638613; uncertaintySymmetricValue = 0.0282436024641; uncertaintySymmetricAse = 0.0180662923767; uncertaintySymmetricT = 1.54573859816; uncertaintySymmetricP = 0.661669583155; uncertaintyRowDependentValue = 0.0311565032115; uncertaintyRowDependentAse = 0.0197350851026; uncertaintyRowDependentT = 1.54573859816; uncertaintyRowDependentP = 0.661669583155; uncertaintyColumnDependentValue = 0.0258288020518; uncertaintyColumnDependentAse = 0.0166817408326; uncertaintyColumnDependentT = 1.54573859816; uncertaintyColumnDependentP = 0.661669583155.

Reference: R DescTools, SciPy and mpmath SciPy 1.18.1; statsmodels 0.15.0; mpmath 1.4.1; R 4.6.1 DescTools 0.99.60. Method: phi/V/C and p: scipy.stats.chi2_contingency + contingency.association; lambda and Goodman-Kruskal tau values/ASE: DescTools Lambda/GoodmanKruskalTau longer CI half-width / qnorm(0.975); uncertainty coefficient values and ASE: 50-digit multinomial delta method; approximate T: value / null-hypothesis SE (first-index modal categories for lambda); tau p: chi2.sf((W-1)(K-1)tau, (R-1)(C-1)); U p: likelihood-ratio chi-square p.

CROSSTABS nominal measures with ASE and significance, newsom-education-meals

[[85,16,13],[183,17,27],[161,23,26]]

phi = 0.0840019427573; phiP = 0.421370564413; cramersV = 0.0593983433565; cramersVP = 0.421370564413; contingencyCoefficient = 0.0837071295044; contingencyCoefficientP = 0.421370564413; lambdaSymmetricValue = 0.0134529147982; lambdaSymmetricAse = 0.0140611174055; lambdaSymmetricT = 0.949459035216; lambdaSymmetricP = 0.342387197252; lambdaRowDependentValue = 0.0185185185185; lambdaRowDependentAse = 0.0193386449655; lambdaRowDependentT = 0.949459035216; lambdaRowDependentP = 0.342387197252; lambdaColumnDependentValue = 0; lambdaColumnDependentAse = 0; goodmanKruskalTauRowDependentValue = 0.00342463445999; goodmanKruskalTauRowDependentAse = 0.0034729760427; goodmanKruskalTauRowDependentP = 0.438442264331; goodmanKruskalTauColumnDependentValue = 0.00334385445607; goodmanKruskalTauColumnDependentAse = 0.00370776510874; goodmanKruskalTauColumnDependentP = 0.451298713344; uncertaintySymmetricValue = 0.00403230085739; uncertaintySymmetricAse = 0.00408915701249; uncertaintySymmetricT = 0.984797944158; uncertaintySymmetricP = 0.424337753892; uncertaintyRowDependentValue = 0.00331351776264; uncertaintyRowDependentAse = 0.00336410776308; uncertaintyRowDependentT = 0.984797944158; uncertaintyRowDependentP = 0.424337753892; uncertaintyColumnDependentValue = 0.00514931282981; uncertaintyColumnDependentAse = 0.00521661959698; uncertaintyColumnDependentT = 0.984797944158; uncertaintyColumnDependentP = 0.424337753892.

Reference: R DescTools, SciPy and mpmath SciPy 1.18.1; statsmodels 0.15.0; mpmath 1.4.1; R 4.6.1 DescTools 0.99.60. Method: phi/V/C and p: scipy.stats.chi2_contingency + contingency.association; lambda and Goodman-Kruskal tau values/ASE: DescTools Lambda/GoodmanKruskalTau longer CI half-width / qnorm(0.975); uncertainty coefficient values and ASE: 50-digit multinomial delta method; approximate T: value / null-hypothesis SE (first-index modal categories for lambda); tau p: chi2.sf((W-1)(K-1)tau, (R-1)(C-1)); U p: likelihood-ratio chi-square p.

CROSSTABS nominal measures with ASE and significance, powermutt-lambda

[[424,213,59],[55,188,357]]

phi = 0.617951040009; phiP = 3.426879 × 10−108; cramersV = 0.617951040009; cramersVP = 3.426879 × 10−108; contingencyCoefficient = 0.525680051283; contingencyCoefficientP = 3.426879 × 10−108; lambdaSymmetricValue = 0.423429781228; lambdaSymmetricAse = 0.0209196207685; lambdaSymmetricT = 16.8748352075; lambdaSymmetricP = 6.891702 × 10−64; lambdaRowDependentValue = 0.496666666667; lambdaRowDependentAse = 0.0241169987691; lambdaRowDependentT = 15.9864620342; lambdaRowDependentP = 1.587971 × 10−57; lambdaColumnDependentValue = 0.36964504284; lambdaColumnDependentAse = 0.0197250965535; lambdaColumnDependentT = 16.3392132021; lambdaColumnDependentP = 5.191855 × 10−60; goodmanKruskalTauRowDependentValue = 0.381863487849; goodmanKruskalTauRowDependentAse = 0.0240969141124; goodmanKruskalTauRowDependentP = 4.147815 × 10−108; goodmanKruskalTauColumnDependentValue = 0.197601632609; goodmanKruskalTauColumnDependentAse = 0.0141440188325; goodmanKruskalTauColumnDependentP = 7.355264 × 10−112; uncertaintySymmetricValue = 0.239340219642; uncertaintySymmetricAse = 0.0171895443516; uncertaintySymmetricT = 13.9508477847; uncertaintySymmetricP = 5.043183 × 10−121; uncertaintyRowDependentValue = 0.309574388556; uncertaintyRowDependentAse = 0.0221635348618; uncertaintyRowDependentT = 13.9508477847; uncertaintyRowDependentP = 5.043183 × 10−121; uncertaintyColumnDependentValue = 0.195081444837; uncertaintyColumnDependentAse = 0.0140424980315; uncertaintyColumnDependentT = 13.9508477847; uncertaintyColumnDependentP = 5.043183 × 10−121.

Reference: R DescTools, SciPy and mpmath SciPy 1.18.1; statsmodels 0.15.0; mpmath 1.4.1; R 4.6.1 DescTools 0.99.60. Method: phi/V/C and p: scipy.stats.chi2_contingency + contingency.association; lambda and Goodman-Kruskal tau values/ASE: DescTools Lambda/GoodmanKruskalTau longer CI half-width / qnorm(0.975); uncertainty coefficient values and ASE: 50-digit multinomial delta method; approximate T: value / null-hypothesis SE (first-index modal categories for lambda); tau p: chi2.sf((W-1)(K-1)tau, (R-1)(C-1)); U p: likelihood-ratio chi-square p.

CROSSTABS nominal measures with ASE and significance, yates-2x2

[[12,5],[29,2]]

phi = -0.311128012246; phiP = 0.0311181873293; cramersV = 0.311128012246; cramersVP = 0.0311181873293; contingencyCoefficient = 0.297081266524; contingencyCoefficientP = 0.0311181873293; lambdaSymmetricValue = 0.125; lambdaSymmetricAse = 0.0979942068866; lambdaSymmetricT = 1.14939154227; lambdaSymmetricP = 0.250394565989; lambdaRowDependentValue = 0.176470588235; lambdaRowDependentAse = 0.141234128941; lambdaRowDependentT = 1.14939154227; lambdaRowDependentP = 0.250394565989; lambdaColumnDependentValue = 0; lambdaColumnDependentAse = 0; goodmanKruskalTauRowDependentValue = 0.0968006400042; goodmanKruskalTauRowDependentAse = 0.0890484886712; goodmanKruskalTauRowDependentP = 0.0329258592157; goodmanKruskalTauColumnDependentValue = 0.0968006400042; goodmanKruskalTauColumnDependentAse = 0.0890484886712; goodmanKruskalTauColumnDependentP = 0.0329258592157; uncertaintySymmetricValue = 0.0870382626192; uncertaintySymmetricAse = 0.0797236641439; uncertaintySymmetricT = 1.06537538013; uncertaintySymmetricP = 0.034879371736; uncertaintyRowDependentValue = 0.0713325910033; uncertaintyRowDependentAse = 0.0668263711494; uncertaintyRowDependentT = 1.06537538013; uncertaintyRowDependentP = 0.034879371736; uncertaintyColumnDependentValue = 0.11161259436; uncertaintyColumnDependentAse = 0.100102972464; uncertaintyColumnDependentT = 1.06537538013; uncertaintyColumnDependentP = 0.034879371736.

Reference: R DescTools, SciPy and mpmath SciPy 1.18.1; statsmodels 0.15.0; mpmath 1.4.1; R 4.6.1 DescTools 0.99.60. Method: phi/V/C and p: scipy.stats.chi2_contingency + contingency.association; lambda and Goodman-Kruskal tau values/ASE: DescTools Lambda/GoodmanKruskalTau longer CI half-width / qnorm(0.975); uncertainty coefficient values and ASE: 50-digit multinomial delta method; approximate T: value / null-hypothesis SE (first-index modal categories for lambda); tau p: chi2.sf((W-1)(K-1)tau, (R-1)(C-1)); U p: likelihood-ratio chi-square p.

CROSSTABS nominal measures with ASE and significance, synthetic-agreement-3x3

[[22,2,2],[5,7,14],[0,2,36]]

phi = 0.850737755012; phiP = 2.406673 × 10−13; cramersV = 0.601562435581; cramersVP = 2.406673 × 10−13; contingencyCoefficient = 0.647974584205; contingencyCoefficientP = 2.406673 × 10−13; lambdaSymmetricValue = 0.522222222222; lambdaSymmetricAse = 0.0709933438407; lambdaSymmetricT = 5.44689876565; lambdaSymmetricP = 5.125562 × 10−8; lambdaRowDependentValue = 0.519230769231; lambdaRowDependentAse = 0.0742413457836; lambdaRowDependentT = 5.64216812364; lambdaRowDependentP = 1.679220 × 10−8; lambdaColumnDependentValue = 0.526315789474; lambdaColumnDependentAse = 0.0887291821498; lambdaColumnDependentT = 4.52267016867; lambdaColumnDependentP = 6.106436 × 10−6; goodmanKruskalTauRowDependentValue = 0.368827502312; goodmanKruskalTauRowDependentAse = 0.0692411435911; goodmanKruskalTauRowDependentP = 1.876067 × 10−13; goodmanKruskalTauColumnDependentValue = 0.475723641908; goodmanKruskalTauColumnDependentAse = 0.0776585566111; goodmanKruskalTauColumnDependentP = 1.774653 × 10−17; uncertaintySymmetricValue = 0.39989922848; uncertaintySymmetricAse = 0.0635121733173; uncertaintySymmetricT = 6.29509374145; uncertaintySymmetricP = 6.482686 × 10−15; uncertaintyRowDependentValue = 0.372824291283; uncertaintyRowDependentAse = 0.0598138078068; uncertaintyRowDependentT = 6.29509374145; uncertaintyRowDependentP = 6.482686 × 10−15; uncertaintyColumnDependentValue = 0.431214533538; uncertaintyColumnDependentAse = 0.0696673575259; uncertaintyColumnDependentT = 6.29509374145; uncertaintyColumnDependentP = 6.482686 × 10−15.

Reference: R DescTools, SciPy and mpmath SciPy 1.18.1; statsmodels 0.15.0; mpmath 1.4.1; R 4.6.1 DescTools 0.99.60. Method: phi/V/C and p: scipy.stats.chi2_contingency + contingency.association; lambda and Goodman-Kruskal tau values/ASE: DescTools Lambda/GoodmanKruskalTau longer CI half-width / qnorm(0.975); uncertainty coefficient values and ASE: 50-digit multinomial delta method; approximate T: value / null-hypothesis SE (first-index modal categories for lambda); tau p: chi2.sf((W-1)(K-1)tau, (R-1)(C-1)); U p: likelihood-ratio chi-square p.

CROSSTABS ordinal measures with ASE and significance, scipy-hotel

[[27,25,14,7,0],[7,14,18,35,12],[1,3,2,7,17]]

gammaValue = 0.716743735886; gammaAse = 0.0550156562196; gammaT = 10.9128482845; gammaP = 1.000709 × 10−27; tauBValue = 0.532744989713; tauBAse = 0.0469246486441; tauBT = 10.9128482845; tauBP = 1.000709 × 10−27; tauCValue = 0.559754766104; tauCAse = 0.0512931868483; tauCT = 10.9128482845; tauCP = 1.000709 × 10−27; somersSymmetricValue = 0.528653579219; somersSymmetricAse = 0.0465642735986; somersSymmetricT = 10.9128482845; somersSymmetricP = 1.000709 × 10−27; somersRowDependentValue = 0.4704595186; somersRowDependentAse = 0.0424745150248; somersRowDependentT = 10.9128482845; somersRowDependentP = 1.000709 × 10−27; somersColumnDependentValue = 0.603276611151; somersColumnDependentAse = 0.0527090192662; somersColumnDependentT = 10.9128482845; somersColumnDependentP = 1.000709 × 10−27.

Reference: R DescTools and mpmath SciPy 1.18.1; statsmodels 0.15.0; mpmath 1.4.1; R 4.6.1 DescTools 0.99.60. Method: gamma, tau-c and Somers' d (column dependent): DescTools values and ASE (longer CI half-width / qnorm(0.975)); tau-b and Somers' d (row dependent): 50-digit multinomial delta method; symmetric Somers' d ASE: 2*ASE(tau-b)*sqrt(Dr*Dc)/(Dr+Dc); approximate T: value / null SE with pair counts from direct four-loop enumeration; p: two-sided normal.

CROSSTABS ordinal measures with ASE and significance, job-satisfaction-income

[[1,2,1,0],[3,3,6,1],[10,10,14,9],[6,7,12,11]]

gammaValue = 0.221100917431; gammaAse = 0.117162829395; gammaT = 1.85146715927; gammaP = 0.0641023750891; tauBValue = 0.152352151347; tauBAse = 0.0818360657143; tauBT = 1.85146715927; tauBP = 0.0641023750891; tauCValue = 0.139467592593; tauCAse = 0.075328148217; tauCT = 1.85146715927; tauCP = 0.0641023750891; somersSymmetricValue = 0.151954602774; somersSymmetricAse = 0.0816225222179; somersSymmetricT = 1.85146715927; somersSymmetricP = 0.0641023750891; somersRowDependentValue = 0.14172302264; somersRowDependentAse = 0.0763951129361; somersRowDependentT = 1.85146715927; somersRowDependentP = 0.0641023750891; somersColumnDependentValue = 0.163778457356; somersColumnDependentAse = 0.0878027941356; somersColumnDependentT = 1.85146715927; somersColumnDependentP = 0.0641023750891.

Reference: R DescTools and mpmath SciPy 1.18.1; statsmodels 0.15.0; mpmath 1.4.1; R 4.6.1 DescTools 0.99.60. Method: gamma, tau-c and Somers' d (column dependent): DescTools values and ASE (longer CI half-width / qnorm(0.975)); tau-b and Somers' d (row dependent): 50-digit multinomial delta method; symmetric Somers' d ASE: 2*ASE(tau-b)*sqrt(Dr*Dc)/(Dr+Dc); approximate T: value / null SE with pair counts from direct four-loop enumeration; p: two-sided normal.

CROSSTABS ordinal measures with ASE and significance, newsom-education-meals

[[85,16,13],[183,17,27],[161,23,26]]

gammaValue = -0.00107251876908; gammaAse = 0.0813088103347; gammaT = -0.0131897730967; gammaP = 0.989476388819; tauBValue = -0.0005276709843; tauBAse = 0.0400059819472; tauBT = -0.0131897730967; tauBP = 0.989476388819; tauCValue = -0.000385374224723; tauCAse = 0.0292176538517; tauCT = -0.0131897730967; tauCP = 0.989476388819; somersSymmetricValue = -0.000508070504553; somersSymmetricAse = 0.038519949055; somersSymmetricT = -0.0131897730967; somersSymmetricP = 0.989476388819; somersRowDependentValue = -0.000696005996359; somersRowDependentAse = 0.0527688031751; somersRowDependentT = -0.0131897730967; somersRowDependentP = 0.989476388819; somersColumnDependentValue = -0.000400049236829; somersColumnDependentAse = 0.0303300204442; somersColumnDependentT = -0.0131897730967; somersColumnDependentP = 0.989476388819.

Reference: R DescTools and mpmath SciPy 1.18.1; statsmodels 0.15.0; mpmath 1.4.1; R 4.6.1 DescTools 0.99.60. Method: gamma, tau-c and Somers' d (column dependent): DescTools values and ASE (longer CI half-width / qnorm(0.975)); tau-b and Somers' d (row dependent): 50-digit multinomial delta method; symmetric Somers' d ASE: 2*ASE(tau-b)*sqrt(Dr*Dc)/(Dr+Dc); approximate T: value / null SE with pair counts from direct four-loop enumeration; p: two-sided normal.

CROSSTABS ordinal measures with ASE and significance, yates-2x2

[[12,5],[29,2]]

gammaValue = -0.715976331361; gammaAse = 0.220375675381; gammaT = -1.90229078278; gammaP = 0.0571331500183; tauBValue = -0.311128012246; tauBAse = 0.14310586827; tauBT = -1.90229078278; tauBP = 0.0571331500183; tauCValue = -0.210069444444; tauCAse = 0.110429723125; tauCT = -1.90229078278; tauCP = 0.0571331500183; somersSymmetricValue = -0.297297297297; somersSymmetricAse = 0.13674431806; somersSymmetricT = -1.90229078278; somersSymmetricP = 0.0571331500183; somersRowDependentValue = -0.421602787456; somersRowDependentAse = 0.184942635018; somersRowDependentT = -1.90229078278; somersRowDependentP = 0.0571331500183; somersColumnDependentValue = -0.229601518027; somersColumnDependentAse = 0.118993252622; somersColumnDependentT = -1.90229078278; somersColumnDependentP = 0.0571331500183.

Reference: R DescTools and mpmath SciPy 1.18.1; statsmodels 0.15.0; mpmath 1.4.1; R 4.6.1 DescTools 0.99.60. Method: gamma, tau-c and Somers' d (column dependent): DescTools values and ASE (longer CI half-width / qnorm(0.975)); tau-b and Somers' d (row dependent): 50-digit multinomial delta method; symmetric Somers' d ASE: 2*ASE(tau-b)*sqrt(Dr*Dc)/(Dr+Dc); approximate T: value / null SE with pair counts from direct four-loop enumeration; p: two-sided normal.

CROSSTABS ordinal measures with ASE and significance, synthetic-agreement-3x3

[[22,2,2],[5,7,14],[0,2,36]]

gammaValue = 0.93074501574; gammaAse = 0.0332282670759; gammaT = 11.8109408797; gammaP = 3.427046 × 10−32; tauBValue = 0.722548722914; tauBAse = 0.0512208080398; tauBT = 11.8109408797; tauBP = 3.427046 × 10−32; tauCValue = 0.657037037037; tauCAse = 0.0556295255159; tauCT = 11.8109408797; tauCP = 3.427046 × 10−32; somersSymmetricValue = 0.720406091371; somersSymmetricAse = 0.0510689188792; somersSymmetricT = 11.8109408797; somersSymmetricP = 3.427046 × 10−32; somersRowDependentValue = 0.780466344039; somersRowDependentAse = 0.0543630041496; somersRowDependentT = 11.8109408797; somersRowDependentP = 3.427046 × 10−32; somersColumnDependentValue = 0.668929110106; somersColumnDependentAse = 0.0547526079046; somersColumnDependentT = 11.8109408797; somersColumnDependentP = 3.427046 × 10−32.

Reference: R DescTools and mpmath SciPy 1.18.1; statsmodels 0.15.0; mpmath 1.4.1; R 4.6.1 DescTools 0.99.60. Method: gamma, tau-c and Somers' d (column dependent): DescTools values and ASE (longer CI half-width / qnorm(0.975)); tau-b and Somers' d (row dependent): 50-digit multinomial delta method; symmetric Somers' d ASE: 2*ASE(tau-b)*sqrt(Dr*Dc)/(Dr+Dc); approximate T: value / null SE with pair counts from direct four-loop enumeration; p: two-sided normal.

CROSSTABS Pearson r, Spearman, linear-by-linear and eta with integer scores, scipy-hotel

[[27,25,14,7,0],[7,14,18,35,12],[1,3,2,7,17]]

pearsonValue = 0.594286294584; pearsonAse = 0.0508445898985; pearsonT = 10.1047022798; pearsonP = 1.976155 × 10−19; spearmanValue = 0.601233252065; spearmanAse = 0.0500752564941; spearmanT = 10.2890914365; spearmanP = 5.836495 × 10−20; linearByLinearStatistic = 66.3971255869; linearByLinearP = 3.686393 × 10−16; etaColumnDependent = 0.600920541147; etaRowDependent = 0.614990092133.

Reference: SciPy, NumPy and mpmath SciPy 1.18.1; statsmodels 0.15.0; mpmath 1.4.1; R 4.6.1 DescTools 0.99.60. Method: scores 1..R and 1..C; r and p: scipy.stats.pearsonr / spearmanr on the expanded cases; ASE: 50-digit multinomial delta method with fixed (midrank) scores; T = r*sqrt(W-2)/sqrt(1-r^2); M2 = (W-1)*r^2 with chi2.sf(M2, 1); eta: sqrt(between-group SS / total SS) on the expanded cases.

CROSSTABS Pearson r, Spearman, linear-by-linear and eta with integer scores, job-satisfaction-income

[[1,2,1,0],[3,3,6,1],[10,10,14,9],[6,7,12,11]]

pearsonValue = 0.177200098631; pearsonAse = 0.0906727006167; pearsonT = 1.74564378158; pearsonP = 0.0841403279051; spearmanValue = 0.176891710719; spearmanAse = 0.0949971152405; spearmanT = 1.74250755261; spearmanP = 0.0846901486904; linearByLinearStatistic = 2.98298812069; linearByLinearP = 0.0841437947064; etaColumnDependent = 0.180746013073; etaRowDependent = 0.217094970882.

Reference: SciPy, NumPy and mpmath SciPy 1.18.1; statsmodels 0.15.0; mpmath 1.4.1; R 4.6.1 DescTools 0.99.60. Method: scores 1..R and 1..C; r and p: scipy.stats.pearsonr / spearmanr on the expanded cases; ASE: 50-digit multinomial delta method with fixed (midrank) scores; T = r*sqrt(W-2)/sqrt(1-r^2); M2 = (W-1)*r^2 with chi2.sf(M2, 1); eta: sqrt(between-group SS / total SS) on the expanded cases.

CROSSTABS Pearson r, Spearman, linear-by-linear and eta with integer scores, newsom-education-meals

[[85,16,13],[183,17,27],[161,23,26]]

pearsonValue = 0.000873820917827; pearsonAse = 0.0428031064676; pearsonT = 0.0204742864375; pearsonP = 0.98367246312; spearmanValue = -0.000500839725554; spearmanAse = 0.0434901971458; spearmanT = -0.0117350513844; spearmanP = 0.990641261659; linearByLinearStatistic = 0.000419959648037; linearByLinearP = 0.98365016229; etaColumnDependent = 0.0354347274013; etaRowDependent = 0.0237637635382.

Reference: SciPy, NumPy and mpmath SciPy 1.18.1; statsmodels 0.15.0; mpmath 1.4.1; R 4.6.1 DescTools 0.99.60. Method: scores 1..R and 1..C; r and p: scipy.stats.pearsonr / spearmanr on the expanded cases; ASE: 50-digit multinomial delta method with fixed (midrank) scores; T = r*sqrt(W-2)/sqrt(1-r^2); M2 = (W-1)*r^2 with chi2.sf(M2, 1); eta: sqrt(between-group SS / total SS) on the expanded cases.

CROSSTABS Pearson r, Spearman, linear-by-linear and eta with integer scores, synthetic-agreement-3x3

[[22,2,2],[5,7,14],[0,2,36]]

pearsonValue = 0.785434189441; pearsonAse = 0.0480807699443; pearsonT = 11.9041654353; pearsonP = 4.963632 × 10−20; spearmanValue = 0.772638106331; spearmanAse = 0.0495901601942; spearmanT = 11.416910073; spearmanP = 4.700563 × 10−19; linearByLinearStatistic = 54.9047110689; linearByLinearP = 1.265174 × 10−13; etaColumnDependent = 0.796045903239; etaRowDependent = 0.786063455273.

Reference: SciPy, NumPy and mpmath SciPy 1.18.1; statsmodels 0.15.0; mpmath 1.4.1; R 4.6.1 DescTools 0.99.60. Method: scores 1..R and 1..C; r and p: scipy.stats.pearsonr / spearmanr on the expanded cases; ASE: 50-digit multinomial delta method with fixed (midrank) scores; T = r*sqrt(W-2)/sqrt(1-r^2); M2 = (W-1)*r^2 with chi2.sf(M2, 1); eta: sqrt(between-group SS / total SS) on the expanded cases.

Cohen's kappa and McNemar/McNemar-Bowker, presidential-approval

[[794,86],[150,570]]

mcnemarP = 3.715936 × 10−5; kappaValue = 0.699592668024; kappaAse = 0.0179791650736; kappaT = 28.0770322762; kappaP = 1.868886 × 10−173.

Reference: statsmodels and SciPy SciPy 1.18.1; statsmodels 0.15.0; mpmath 1.4.1; R 4.6.1 DescTools 0.99.60. Method: statsmodels.stats.inter_rater.cohens_kappa (kappa, std_kappa, std_kappa0); 2x2 McNemar: scipy.stats.binomtest(min(b,c), b+c, 0.5); larger tables: statsmodels SquareTable(...).symmetry().

Cohen's kappa and McNemar/McNemar-Bowker, synthetic-agreement-3x3

[[22,2,2],[5,7,14],[0,2,36]]

bowkerStatistic = 12.2857142857; bowkerDegreesOfFreedom = 3; bowkerP = 0.00646584136108; kappaValue = 0.56191588785; kappaAse = 0.0684302655244; kappaT = 7.74792601568; kappaP = 9.340553 × 10−15.

Reference: statsmodels and SciPy SciPy 1.18.1; statsmodels 0.15.0; mpmath 1.4.1; R 4.6.1 DescTools 0.99.60. Method: statsmodels.stats.inter_rater.cohens_kappa (kappa, std_kappa, std_kappa0); 2x2 McNemar: scipy.stats.binomtest(min(b,c), b+c, 0.5); larger tables: statsmodels SquareTable(...).symmetry().

Cohen's kappa and McNemar/McNemar-Bowker, synthetic-bowker-4x4-empty-pair

[[10,3,0,1],[6,12,2,0],[0,4,9,3],[2,0,1,14]]

bowkerStatistic = 3; bowkerDegreesOfFreedom = 6; bowkerP = 0.808846830538; kappaValue = 0.561178922298; kappaAse = 0.0763541933452; kappaT = 7.97996938305; kappaP = 1.463696 × 10−15.

Reference: statsmodels and SciPy SciPy 1.18.1; statsmodels 0.15.0; mpmath 1.4.1; R 4.6.1 DescTools 0.99.60. Method: statsmodels.stats.inter_rater.cohens_kappa (kappa, std_kappa, std_kappa0); 2x2 McNemar: scipy.stats.binomtest(min(b,c), b+c, 0.5); larger tables: statsmodels SquareTable(...).symmetry().

2x2 odds ratio and cohort relative risks with 95% intervals, yates-2x2

[[12,5],[29,2]]

oddsRatio = 0.165517241379; oddsRatioLower = 0.0281238371546; oddsRatioUpper = 0.97411875354; firstColumnRisk = 0.754563894523; firstColumnRiskLower = 0.54766970321; firstColumnRiskUpper = 1.03961688511; secondColumnRisk = 4.55882352941; secondColumnRiskLower = 0.987717556896; secondColumnRiskUpper = 21.0413106735.

Reference: statsmodels SciPy 1.18.1; statsmodels 0.15.0; mpmath 1.4.1; R 4.6.1 DescTools 0.99.60. Method: Table2x2(table).oddsratio / oddsratio_confint; riskratio / riskratio_confint for column 1 and for the column-swapped table (column 2); log-scale normal intervals.

2x2 odds ratio and cohort relative risks with 95% intervals, presidential-approval

[[794,86],[150,570]]

oddsRatio = 35.0837209302; oddsRatioLower = 26.3543532663; oddsRatioUpper = 46.7045220906; firstColumnRisk = 4.33090909091; firstColumnRiskLower = 3.74993555015; firstColumnRiskUpper = 5.00189224665; secondColumnRisk = 0.123444976077; secondColumnRiskLower = 0.100642326608; secondColumnRiskUpper = 0.151414048464.

Reference: statsmodels SciPy 1.18.1; statsmodels 0.15.0; mpmath 1.4.1; R 4.6.1 DescTools 0.99.60. Method: Table2x2(table).oddsratio / oddsratio_confint; riskratio / riskratio_confint for column 1 and for the column-swapped table (column 2); log-scale normal intervals.

Fisher exact test, one-sided in the observed direction, fisher-tea-tasting

[[3,1],[1,3]]

pValue = 0.242857142857; greater = 1.

Reference: SciPy 1.18.1. Method: scipy.stats.fisher_exact(table, alternative='greater' if f11 > E11 else 'less')

Fisher exact test, one-sided in the observed direction, yates-2x2

[[12,5],[29,2]]

pValue = 0.0445547378351; greater = 0.

Reference: SciPy 1.18.1. Method: scipy.stats.fisher_exact(table, alternative='greater' if f11 > E11 else 'less')

Fisher exact test, one-sided in the observed direction, synthetic-lower-tail

[[1,9],[8,3]]

pValue = 0.00580070084714; greater = 0.

Reference: SciPy 1.18.1. Method: scipy.stats.fisher_exact(table, alternative='greater' if f11 > E11 else 'less')

Fisher-Freeman-Halton exact test, job-satisfaction-income

[[1,2,1,0],[3,3,6,1],[10,10,14,9],[6,7,12,11]]

pValue = 0.782684938967.

Reference: R stats R 4.6.1. Method: stats::fisher.test(table)$p.value (network algorithm)

Fisher-Freeman-Halton exact test, synthetic-sparse-3x3

[[3,0,1],[1,4,0],[0,1,5]]

pValue = 0.00245785960072.

Reference: R stats R 4.6.1. Method: stats::fisher.test(table)$p.value (network algorithm)

Fisher-Freeman-Halton exact test, synthetic-sparse-2x4

[[2,0,3,1],[0,4,1,2]]

pValue = 0.0792540792541.

Reference: R stats R 4.6.1. Method: stats::fisher.test(table)$p.value (network algorithm)

Fisher-Freeman-Halton exact test, synthetic-sparse-3x2

[[4,0],[1,3],[0,5]]

pValue = 0.00699300699301.

Reference: R stats R 4.6.1. Method: stats::fisher.test(table)$p.value (network algorithm)

Raw and standardized (Pearson) residuals, scipy-hotel

[[27,25,14,7,0],[7,14,18,35,12],[1,3,2,7,17]]

raw = [[13.481481481481481,8.777777777777779,0.867724867724867,-11.925925925925927,-11.201058201058201],[-8.925925925925926,-5.111111111111111,2.5291005291005284,12.703703703703702,-1.1957671957671963],[-4.555555555555555,-3.666666666666667,-3.396825396825397,-0.7777777777777777,12.396825396825397]]; standardized = [[3.666680503640696,2.179362173319569,0.2394484650359853,-2.741344017414877,-3.3467982014244897],[-2.2366649529986464,-1.1691557058833024,0.6429960579535411,2.690382270498483,-0.3291768708260697],[-1.9327585352432297,-1.4200938936093863,-1.462190776681905,-0.2788866755113585,5.778056180400486]].

Reference: SciPy and NumPy SciPy 1.18.1; statsmodels 0.15.0; mpmath 1.4.1; R 4.6.1 DescTools 0.99.60. Method: E = scipy.stats.contingency.expected_freq(table); raw = O - E; standardized = (O - E)/sqrt(E).

Raw and standardized (Pearson) residuals, synthetic-agreement-3x3

[[22,2,2],[5,7,14],[0,2,36]]

raw = [[14.2,-1.1777777777777776,-13.022222222222222],[-2.8,3.8222222222222224,-1.0222222222222221],[-11.4,-2.644444444444445,14.044444444444444]]; standardized = [[5.084415605679974,-0.6606958729393096,-3.3598354738842997],[-1.002560823655206,2.144145097086062,-0.26374135119228287],[-3.376388603226826,-1.2270649793646629,2.9973147376488867]].

Reference: SciPy and NumPy SciPy 1.18.1; statsmodels 0.15.0; mpmath 1.4.1; R 4.6.1 DescTools 0.99.60. Method: E = scipy.stats.contingency.expected_freq(table); raw = O - E; standardized = (O - E)/sqrt(E).

Stratified 2x2 tests and Mantel-Haenszel common odds ratio, synthetic-three-strata

[[10,5],[3,12],[6,4],[5,9],[8,2],[4,6]]

cochranStatistic = 10.6551524135; cochranP = 0.00109764698332; mantelHaenszelStatistic = 8.79871679017; mantelHaenszelP = 0.0030144249138; mantelHaenszelUncorrectedStatistic = 10.2233698714; mantelHaenszelUncorrectedP = 0.00138672249701; breslowDayStatistic = 0.867879099435; breslowDayP = 0.647951423148; taroneStatistic = 0.866575938212; taroneP = 0.648373753309; commonOddsRatio = 4.99038461538; commonOddsRatioLower = 1.83329801071; commonOddsRatioUpper = 13.5842282401; logCommonOddsRatioStandardError = 0.510925943245; commonOddsRatioP = 0.00165365036816.

Reference: statsmodels and R stats/DescTools SciPy 1.18.1; statsmodels 0.15.0; mpmath 1.4.1; R 4.6.1 DescTools 0.99.60. Method: statsmodels StratifiedTable: test_null_odds(correction=True/False), test_equal_odds(adjust=False/True), oddsratio_pooled(_confint), logodds_pooled_se; cross-checked against R mantelhaen.test and DescTools::BreslowDayTest(correct=FALSE/TRUE) to 1e-9; Cochran's statistic from its definition.

Stratified 2x2 tests and Mantel-Haenszel common odds ratio, ucb-admissions

[[512,313],[89,19],[353,207],[17,8],[120,205],[202,391],[138,279],[131,244],[53,138],[94,299],[22,351],[24,317]]

cochranStatistic = 1.5265578428; cochranP = 0.216629787168; mantelHaenszelStatistic = 1.42694622859; mantelHaenszelP = 0.232263462817; mantelHaenszelUncorrectedStatistic = 1.52460666044; mantelHaenszelUncorrectedP = 0.216923697056; breslowDayStatistic = 18.8255137052; breslowDayP = 0.00207139034992; taroneStatistic = 18.8255012521; taroneP = 0.00207140139788; commonOddsRatio = 0.904696828259; commonOddsRatioLower = 0.771907361759; commonOddsRatioUpper = 1.06032976444; logCommonOddsRatioStandardError = 0.0809889096562; commonOddsRatioP = 0.216214972781.

Reference: statsmodels and R stats/DescTools SciPy 1.18.1; statsmodels 0.15.0; mpmath 1.4.1; R 4.6.1 DescTools 0.99.60. Method: statsmodels StratifiedTable: test_null_odds(correction=True/False), test_equal_odds(adjust=False/True), oddsratio_pooled(_confint), logodds_pooled_se; cross-checked against R mantelhaen.test and DescTools::BreslowDayTest(correct=FALSE/TRUE) to 1e-9; Cochran's statistic from its definition. Strata: R datasets::UCBAdmissions (Gender x Admit by Department A-F).

Market research tools

A second manifest checks the market research tools: rim weighting, MaxDiff, CBC conjoint, latent class analysis, key driver analysis, TURF, NPS, Van Westendorp and Gabor-Granger. It holds 14 cases with 284 reference values. Each case embeds a small simulated dataset; the reference values come from established packages (ipfn, statsmodels, stepmix, dominance-analysis) run in a pinned Python environment, or, where no package implements the convention, from an independent NumPy and SciPy recomputation that is labelled as such. The table below is computed when this page is built, by running the browser engines on the same inputs. Automated parity evidence for the market research tools against established independent implementations; not certification or independent human review.

Market research tools compared with independent reference implementations
ProcedureReferenceCasesValuesMax |Δ|Status
Rim weighting (raking)ipfn 1.4.42473.2 × 10−12Within tolerance
MaxDiffstatsmodels 0.15.01291.4 × 10−8Within tolerance
CBC conjointstatsmodels 0.15.02474.4 × 10−9Within tolerance
Latent class analysisstepmix 3.0.02886.7 × 10−5Within tolerance
Key driver analysisdominance-analysis 1.1.91261.5 × 10−13Within tolerance
TURFNumPy 2.5.4 (independent recomputation)188.9 × 10−16Within tolerance
Net Promoter ScoreSciPy 1.18.1 (independent recomputation)3187.5 × 10−15Within tolerance
Van Westendorp PSMSciPy 1.18.1 (independent recomputation)150Within tolerance
Gabor-GrangerNumPy 2.5.4 (independent recomputation)1160Within tolerance

Open the market research reference manifest. Every value carries its own absolute tolerance. Closed-form procedures are held to about 10−9 or tighter; the choice models allow 10−7 for utilities and 10−6 for standard errors and percentages, because statsmodels uses a numerical Hessian; latent class models allow 10−5 for the log-likelihood, 10−4 for class shares and 10−3 for item probabilities, because both sides stop iterative EM at a convergence threshold.

  • Each procedure is checked on one or two small simulated datasets; other designs, sizes and edge cases are not covered by these cases.
  • Hierarchical Bayes estimation (MaxDiff and CBC), market simulators, NMS purchase-probability curves and bootstrap intervals are not part of this comparison.
  • Latent class models are compared at the best of 50 random starts on each side; a different start count or a harder dataset can reach different local maxima.
  • TURF, NPS, Van Westendorp and Gabor-Granger are compared with an independent recomputation of the stated formulas, not with a third-party package.

Market research tool cases

Every compared value of the 14 cases (the Crosstabs result, the reference result, their difference and the tolerance) is listed on its own page, so this one stays light.

See all 14 market research tool cases

What this does not prove

  • The selected cases do not cover every statistic, input shape, numerical edge case, export, or interface.
  • The cases assume nonnegative frequency-count tables with positive row and column margins.
  • Ordinal measures depend on the supplied row and column category order.
  • The checks do not provide complex-survey variance estimation or establish that a statistic is appropriate for a particular study.
  • Agreement with a reference implementation is not a substitute for a named statistical owner, independent reviewer, or domain-specific review.

Named statistical review is still pending. Read the broader methods and assumptions, and report a discrepancy to support@crosstabs.com.

Independent review packet

The public Phase-1 review request binds the complete statistical scope, Python tool catalog, release artifacts, and current deployment identity in machine-readable JSON. It is preparation material for a qualified reviewer and deliberately does not claim that a review has occurred.

Open the revision-bound independent review request. A genuine approval record must be published by the named reviewer on an independent evidence origin after they inspect the bound revision.

Reference documentation

Formula-backed cases publish the exact equation and generating runtime in each manifest entry; their Python recomputation deliberately uses a separate implementation from the browser engine.

Reference manifest generated .