statistics-2.0.0

A major release: the Octave-only hypothesis tests share one layout and report effect sizes with their confidence intervals, and repeated measures models, effect sizes for mean differences, drift detection and three chart objects arrive.

Requires datatypes 1.5.0 or newer, which changes how tables convert; see its release notes.

Requires Octave >= 11.1.0. pkg test statistics passes 20,700 of 20,700.

Standing notice, unchanged since 1.9.0. The nan and the tablicious packages are incompatible with this one.

Incompatible changes. Code calling these functions must be checked:

• chi2test and mcnemar_test return [h, p, stats], and ztest2 returns [h, p, ci, stats]. p = chi2test (x) still runs but now returns h.

• correlation_test, hotelling_t2test and hotelling_t2test2 name their statistics in stats as MATLAB does: tstat, zval, CorrCoef, t2stat.

• regression_ftest (y, X, keep) fits both models itself; its former F statistic and degrees of freedom were wrong.

• bartlett_test and levene_test are removed; vartestn runs both tests.

• The SVM's KernelScale divides the predictors and the polynomial kernel is (1 + u'v)^q, as in MATLAB, so every Gaussian, polynomial and sigmoid fit changes.

• jackknife returns one row per jackknife sample, and lime reads predictors holding levels as MATLAB does.

• Effect sizes. meanEffectSize gives Cohen's d, Glass's delta, Cliff's delta and four more, with exact or bootstrap intervals that match MATLAB's; gardnerAltmanPlot draws them beside the data. The Octave-only tests, and ranksum, signrank, kruskalwallis and friedman, report effect sizes with their intervals in stats.

• Repeated measures: fitrm and RepeatedMeasuresModel, with ranova, mauchly, epsilon, manova, multcompare and the sphericity corrections, over the new studentized range distribution (stdrcdf and its family).

• Two-sample and drift tests: knntest, mmdtest and detectdrift, with nomdist and nomdist2 for nominal data.

• Charts: heatmap, scatterhistogram and parallelplot, each returning an object whose properties redraw it.

• Accurate tails. P-values below 1e-16 are now computed across 35 functions and classes, where they were 0. betainv, tinv, tcdf and finv keep their digits far into the tails: tinv (1e-6, 1) was -44721 where the quantile is -318310.

• Fixes.

  • chi2test's homogeneous and marginal models now agree with R.
  • binotest's one-sided p-values were swapped.
  • fitglme with a normal response gave standard errors off by up to a factor of 53.
  • correlation_test's rank tests failed.
  • linkage corrupted the tree on Inf or NaN distances.
  • nancov ([], []) returns a 2-by-2 NaN matrix, as in MATLAB.

Every change is listed in NEWS.