Home › Stats › Overview

Hypothesis tests — Lodestar.Stats

Eighteen families of classical hypothesis test, at scipy.stats 1.18.1 parity. Arrays in, a statistic and a p-value out; nothing is fitted, so every entry point is static.

test	what it asks	entry point
Student / Welch t	do two samples have the same mean?	TTest
Mann-Whitney U	the same question, assuming no shape	MannWhitney
Wilcoxon signed-rank	the same, on paired measurements	Wilcoxon
χ²	do counts match an expected distribution, or are two factors independent?	ChiSquared
Fisher exact	the same for a 2×2 table, at any sample size	FisherExact
Kolmogorov-Smirnov	do two samples share a distribution?	KolmogorovSmirnov
one-way ANOVA	do several groups share one mean?	OneWayAnova
Kruskal-Wallis	the same, assuming no shape	KruskalWallis
Shapiro-Wilk	could this sample be normal?	ShapiroWilk
Bonferroni / BH / BY	how many of these results are chance?	MultipleComparisons
Pearson r	are these two variables related, along a line?	Pearson
Spearman rho	the same, along any rising or falling curve	Spearman
point-biserial	how a variable differs between the two sides of a binary one	PointBiserial
Kendall tau	the same, counted pair by pair	KendallTau
Levene	do several groups share one variance?	Levene
Bartlett	the same, assuming each group is normal	Bartlett
Fligner-Killeen	the same, by ranks	Fligner
Friedman	do several treatments differ on the same subjects?	Friedman
binomial	is this proportion what it was said to be?	Binomial
Anderson-Darling	could this sample be normal, or do several share a distribution?	AndersonDarling

Every family but six returns the same two numbers, TestResult; a t-test also carries its degrees of freedom (TTestResult), a contingency table also carries the table independence would have produced (ChiSquaredContingencyResult), Kolmogorov-Smirnov also carries where and in which direction the two samples parted furthest (KsResult), a Pearson correlation and a binomial test each carry what their intervals need (PearsonResult, BinomialResult), and Anderson-Darling carries the table its statistic is read against (AndersonResult). Spearman's rho over every pair of variables returns both matrices (CorrelationMatrix).

result	carries	returned by
TestResult	a statistic, a p-value	fourteen of the twenty families
TTestResult	+ degrees of freedom, a confidence interval	TTest
ChiSquaredContingencyResult	+ degrees of freedom, the expected table	ChiSquared.Contingency
KsResult	+ where the gap was reached, and its sign	KolmogorovSmirnov
PearsonResult	+ a confidence interval for the correlation	Pearson
BinomialResult	+ a confidence interval for the proportion	Binomial
AndersonResult	+ the critical values and their significance levels	AndersonDarling
CorrelationMatrix	a statistic and a p-value for every pair of variables	Spearman.Matrix

Nine small enums choose what a test asks, and how it answers when the exact and the approximate route disagree.

option	chooses
Alternative	which tail the p-value covers
Variance	whether an independent-samples t-test pools the two variances
Continuity	whether a discrete statistic's normal approximation gets the half-unit correction
ExactMethod	the exact null distribution, its normal approximation, or a choice between them
ZeroMethod	what Wilcoxon does with a pair whose difference is exactly zero
KendallVariant	which normalisation Kendall's tau carries; the p-value is the same either way
Center	which centre Levene's and Fligner-Killeen's tests measure each group's spread around
AndersonKSampleVariant	which form of the k-sample Anderson-Darling statistic is computed
ProportionInterval	which interval a binomial proportion is reported with

The hypothesis-testing guide says which test answers which question, and the Python equivalence table maps each scipy call to its counterpart here.