mann_whitney_u

mann_whitney_u compares two groups by rank ordering the outcome values. It is the built-in escape from a shaky mean/normality story.

Use it for skewed continuous values, ordinal scores, or outlier-heavy groups when a rank-based comparison is easier to defend than a mean comparison.

Rank-test contract

Field	Required	Notes
inputs	yes	Exactly one tidy table.
valueColumn	yes	Numeric or ordinal outcome.
groupColumn	yes	Column containing group labels.
groupA, groupB	yes	The two group values to compare.

What question it answers

default reports group sizes, U, z, p-value, effect size, and a 95% effect-size confidence interval.

The null is rank/stochastic balance: values from group A are not systematically larger or smaller than values from group B. Do not describe it as a guaranteed median test when the distributions have different shapes.

Limits

• The current assumption-warning pass does not emit Mann-Whitney-specific warnings yet.
• Very small groups make the asymptotic p-value fragile. Use an exact, permutation, or bootstrap approach in Python/R when that distinction matters.
• Both requested groups must exist and have numeric values before the node can produce a result.

Example

- id: mw
  kind: mann_whitney_u
  inputs: [data]
  valueColumn: outcome
  groupColumn: arm
  groupA: control
  groupB: treatment

Related

• t_test - mean-based alternative when assumptions are credible