MerQuR+Statistics Glossary

Statistics Glossary → Redundancy index

Redundancy indexconcept

A measure in canonical correlation analysis of how much of the total variance in one set of variables is explained by a canonical variate of the other set.

The canonical correlation (Rc) gives the relationship between linear combinations of the two sets, but not how well those combinations represent the variables in their own sets. Stewart and Love's redundancy index combines both pieces of information: the redundancy of set Y for a canonical function = (the mean squared loading of the Y variables on their own canonical variate) × Rc². Summed over all canonical functions, the redundancies equal the average R² from regressing each Y variable on all the X variables.

The index is not symmetric: the share of Y explained by X can differ from the share of X explained by Y. A high canonical correlation can arise from a strong link between only small parts of the two sets; the redundancy index then stays low and guards against over-interpretation.

An everyday example

In a study, set X holds three lifestyle variables (sleep, exercise, diet) and set Y four health indicators. The first canonical correlation is 0.70 (Rc² = 0.49), but the health indicators' mean squared loading on their own canonical variate is only 0.30. The redundancy is 0.30 × 0.49 ≈ 0.15: the lifestyle combination explains only about 15% of the total variance of the health indicators.

In MerQur

In Canonical Correlation it is reported as the redundancy index for each canonical function; in PLS-SEM, the 'Q2 (blindfolding)' option gives the Stone-Geisser Q2 by the cross-validated redundancy approach.

Analyses where this concept appears

Canonical CorrelationPLS-SEM (Partial Least Squares Path Modeling)

Related concepts

Cross-loadings