MerQuR+Statistics Glossary

Statistics Glossary → Unidimensionality

Unidimensionalityconcept

The assumption that responses to the items of a scale are driven essentially by a single latent trait (an ability or attitude).

Most item response theory models (Rasch, 2PL, 3PL) assume that each person is described by a single latent trait level (θ) and that all items measure that same trait. If unidimensionality fails, for instance if a maths test measures both computation and reading comprehension, a single θ estimate mixes the two abilities and the item parameters are biased. A closely related assumption is local independence: once θ is held fixed, responses to the items should be independent of one another.

Unidimensionality is usually assessed from the eigenvalues of the item correlations (tetrachoric or polychoric for binary and ordinal items): a large ratio of the first eigenvalue to the second (3 or more, as a rough rule) suggests a single dominant dimension. The fit of a one-factor confirmatory model can also be tested. Real scales are rarely perfectly unidimensional; the practical question is whether the dominant dimension is strong enough to make the influence of secondary dimensions negligible.

An everyday example

A driving theory test is meant to measure knowledge of traffic rules. But if some questions are written in long, convoluted sentences, success on them depends on reading speed as well as knowledge of the rules. If the first eigenvalue of the polychoric matrix is 8 and the second 4 (a ratio of 2), the test may not be measuring traffic knowledge alone; if simplifying the convoluted questions raises the ratio to 5, unidimensionality has been strengthened.

In MerQur

In Item Response Theory, 'Outputs: Unidimensionality (eigenvalue ratio)' gives the eigenvalues of the tetrachoric/polychoric correlation matrix and the ratio of the first to the second eigenvalue.

Analyses where this concept appears

Item Response Theory (IRT / Rasch)

Related concepts

Eigenvalue and eigenvectorItem difficulty, discrimination and guessing (1PL, 2PL, 3PL)