MerQuR+Statistics Glossary

Statistics Glossary → Polychoric and tetrachoric correlation

Polychoric and tetrachoric correlationconcept

An estimate of the correlation between the continuous, normally distributed latent variables assumed to underlie two ordinal (or binary) variables.

Likert items and yes/no answers can be thought of as coarse categorizations of a tendency that is really continuous. Applying Pearson's correlation directly to these category codes understates the true relationship, especially when there are few categories and the distribution is skewed. The polychoric correlation assumes that two latent variables follow a bivariate normal distribution and that the observed categories arise from cutting these variables at certain thresholds; it estimates the thresholds and the latent correlation by maximum likelihood. Its special case for two binary variables is called the tetrachoric correlation.

Polychoric correlations are widely used in factor analysis and reliability studies with ordinal items (for example ordinal alpha). Their limitations are that the latent normality assumption can be checked only partially, if at all; that estimates can become unstable with sparse or empty cells; and that a matrix built pair by pair may not be positive definite. Biserial and point-biserial correlations serve a similar purpose for one binary and one continuous variable.

An everyday example

A survey asks for 1-5 ratings of 'I wake up feeling energetic' and 'I can concentrate during the day'. Because answers pile up at 4 and 5, the Pearson correlation is 0.38. The polychoric correlation estimates the relationship between the continuous levels of 'alertness' behind the answers and finds 0.52, showing that the items are in fact more strongly related.

In MerQur

It appears as 'Additional output: Polychoric / biserial' in Correlation Analysis, 'Correlation type: Polychoric' in Exploratory Factor Analysis and 'Ordinal α (polychoric correlations)' in Cronbach's Alpha and Reliability Analysis. The unidimensionality check in Item Response Theory is based on the tetrachoric/polychoric matrix.

Analyses where this concept appears

Correlation AnalysisCronbach's Alpha and Reliability AnalysisExploratory Factor Analysis (EFA)Item Response Theory (IRT / Rasch)

Related concepts

Pearson correlation coefficientFactor extraction method (ML, PAF, minres)