Ordinal Reliability Analysis

Cronbach's alpha is computed on the raw (Pearson) covariance matrix, which is attenuated when responses are clustered into a small number of ordered categories. Ordinal alpha instead uses the polychoric correlation matrix, which estimates correlations between the latent continuous variables underlying the categories. Removing the categorisation artifact raises the correlation estimates and therefore the reliability coefficient.

Ordinal omega is generally preferred because it does not assume that all items have equal relationships to the latent construct (tau-equivalence), an assumption Cronbach's alpha — and ordinal alpha — both require. In most real scales items differ in how strongly they load on the factor, so omega is the more defensible coefficient. Report both if reviewers expect alpha for comparability with prior literature.

A common minimum is around 200 respondents. Polychoric correlation estimation can be unstable with smaller samples, especially when items have extreme marginal distributions (e.g., almost all respondents choose the same category). Simulation studies suggest that below 100–150 cases, polychoric-based estimates may be noisier than Pearson-based ones.

Dichotomous items require tetrachoric rather than polychoric correlations (polychoric correlation reduces to tetrachoric for two categories, so the method is technically valid). However, for purely binary data, KR-20 or a tetrachoric-based omega is more commonly reported. Ordinal reliability analysis is most beneficial when items have three or more ordered categories.

No. Reliability coefficients — ordinal or otherwise — measure the proportion of score variance attributable to common sources, but they do not distinguish between one common source and several. A multidimensional scale can produce a high reliability coefficient if all subscales correlate. Always accompany reliability estimation with an exploratory or confirmatory factor analysis to assess dimensionality.