MerQuR+Statistics Glossary

Statistics Glossary → Tukey's test for nonadditivity

Tukey's test for nonadditivityconcept

A one-degree-of-freedom test for a two-way layout with a single observation per cell (e.g. persons × items) that checks whether the row and column effects are additive, i.e. whether there is a multiplicative interaction between them.

In a scale where each person answers each item once, person and item effects are assumed to be additive: a person's score is the sum of the person's general level, the item's general easiness and random error. With one observation per cell, the full interaction term cannot be estimated. Tukey (1949) modelled the interaction in a particular form, proportional to the product of the row and column effects, which splits off a one-degree-of-freedom 'nonadditivity' component from the residual sum of squares and tests it with an F test.

A significant result means the additivity assumption is violated, in which case results built on Cronbach's alpha and the persons × items ANOVA may be misleading. Tukey's method also provides an estimate of the power to which the data could be raised (e.g. log, square root) to make them additive. The test captures only one specific form of interaction and can miss others.

An everyday example

Twenty people answer 10 items on an attention test. Additivity expects a strong participant to be ahead of a weak one by roughly the same margin on every item. If, instead, strong participants barely differ from weak ones on easy items but pull far ahead on hard ones, person and item effects interact multiplicatively, and Tukey's test picks up that pattern.

In MerQur

It is given, together with a power-transformation estimate, in the 'Scale ANOVA + Tukey additivity + Friedman/Cochran + Hotelling T²' extra output of Cronbach's Alpha and Reliability Analysis.

Analyses where this concept appears

Cronbach's Alpha and Reliability Analysis

Related concepts

Analysis of variance (ANOVA) table