Statistical Procedures

Computation lives in the pure-Dart package statkit (no Flutter dependency).

Highlights

• Descriptives.compute
• TTest.oneSample / independentSamples / paired
• OnewayAnova.compute with Levene, tukeyHsd
• ancova — analysis of covariance, Type III sums of squares
• Permutation.random / Permutation.exact, twoSampleMeanDifference, pairedDifference
• Correlation.pearson / spearman / matrix
• Regression.simple / multiple (VIF, DW), stepwiseRegression
• logisticRegression / logisticRegressionFull (binary), multinomialLogistic
• factorPca
• ChiSquareTest.goodnessOfFit / independence
• Nonparametric.* (Mann-Whitney, Wilcoxon, Kruskal-Wallis, Jonckheere-Terpstra, Lilliefors, …)
• Reliability.cronbachAlpha
• KMeans.cluster
• Distribution helpers: normCdf, tTwoTail, chiSquareSf, fSf, …

Procedures added in statkit 0.5.0 (Jonckheere-Terpstra trend, permutation tests, Lilliefors, multinomial logistic regression, ANCOVA) are marked @experimental: their signature may change in a minor release.

Conventions fixed and aligned in v1.4.0

• Binary logistic regression: the IRLS iteration used to oscillate between two points (the working response weight was wrong) and the intercept standard error came off the standardized scale. Coefficients, odds ratios, SEs and p-values now come from the actual maximum-likelihood point.
• Levene's test is centred on the group mean (the PSPP / SPSS convention for ONEWAY, T-TEST and GLM); useMean: false still gives Brown-Forsythe.
• Mann-Whitney U gained the tie correction in its asymptotic variance, with p = 2·Φ(z) (PSPP's Asymp. Sig. (2-tailed)); the continuity-corrected Z / p (SPSS, R) stay available as extra['zContinuityCorrected'] / extra['pContinuityCorrected'].
• KS normality test: with parameters estimated from the sample it now reports the Lilliefors-corrected p (the plain asymptotic p almost never rejected).
• Which cases enter a procedure — user-missing values, pairwise versus row-wise deletion — is spelled out in Missing values.

Tests

cd packages/statkit && dart test

91 cases cover descriptives, distributions, t tests, ANOVA / ANCOVA, correlation, regression, logistic, nonparametric and permutation tests, reliability and K-means. The PSPP 2.1.2 parity cases live in test/pspp_parity_test.dart on the app side and run with CI.

See the statkit API for the full reference.