Canonical Correlation Analysis
Also known as: CCA, canonical variate analysis, canonical analysis, multiple canonical correlation
Canonical Correlation Analysis (CCA) is a multivariate statistical method that identifies pairs of linear combinations — one from each of two variable sets — such that the correlation between each pair is maximised. Introduced by Harold Hotelling in his landmark 1936 Biometrika paper, CCA provides the most general linear framework for studying the association between two multivariate batteries of measurements, and many classical procedures (multiple regression, MANOVA, discriminant analysis) are special cases of it.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use CCA when you have two conceptually distinct multivariate batteries measured on the same observations and you want to characterise the overall linear association between them. Typical settings include relating psychological test scores to neuroimaging measures, linking environmental predictors to ecological response variables, or jointly analysing questionnaire subscales. Key assumptions: both variable sets are jointly multivariate normal; observations are independent; sample size is substantially larger than the total number of variables (a common guideline is n ≥ 10(p + q)); and neither covariance matrix is singular (high multicollinearity within a set should be addressed before analysis). When the normality assumption cannot be met or the variables are not continuous, regularised CCA or kernel CCA extensions may be preferable.
Strengths & limitations
- Provides the single most comprehensive linear summary of the association between two variable sets, capturing multiple orthogonal dimensions of relationship in one analysis.
- Encompasses multiple regression, MANOVA, discriminant analysis, and the t-test as special cases, giving it exceptional theoretical generality.
- Canonical loadings (structure coefficients) facilitate substantive interpretation of which original variables drive each canonical dimension.
- The sequential testing procedure via Wilks' lambda allows principled determination of how many significant dimensions of association exist.
- Requires large samples relative to the number of variables; with small n the canonical correlations are upwardly biased and the solution is unstable.
- Sensitive to multicollinearity within each variable set; near-singular within-set covariance matrices can produce degenerate solutions.
- Canonical weights are difficult to interpret directly because they reflect partial relationships conditioned on all other variables in the set; canonical loadings are more interpretable but still require care.
- The method finds linear combinations only; genuinely nonlinear associations between the two sets will not be detected without kernel or deep extensions.
Frequently asked
How is CCA different from multiple regression?
Multiple regression predicts a single outcome variable from a set of predictors, making it a special case of CCA in which the Y set contains only one variable and p* = 1. CCA generalises this to any number of variables in both sets and simultaneously finds all orthogonal dimensions of association, not just the best overall prediction of one outcome.
How many canonical variates should I retain?
Apply the step-down Wilks' lambda test: start by testing all s dimensions together (overall significance), then sequentially remove the first canonical correlation and test the remaining dimensions. Retain those dimensions for which the test is significant at your chosen alpha level. Some researchers also apply a practical threshold such as retaining only dimensions with ρ* > 0.30 or with a meaningful redundancy index.
What is the minimum sample size needed?
A widely cited rule of thumb is n ≥ 10(p + q), where p and q are the numbers of variables in the two sets. With fewer observations the sample canonical correlations are upwardly biased, the weight vectors are unstable, and cross-validation is essential. Regularised or sparse CCA variants are better suited to high-dimensional, smaller-sample situations.
Should I use canonical weights or canonical loadings for interpretation?
Prefer canonical loadings (structure coefficients), which are simple correlations between each original variable and its canonical variate. Canonical weights behave like partial regression coefficients — they suppress the contribution of correlated variables and can be misleading in the presence of multicollinearity. Loadings are more stable and easier to interpret substantively.
Sources
- Hotelling, H. (1936). Relations between two sets of variates. Biometrika, 28(3–4), 321–377. DOI: 10.1093/biomet/28.3-4.321 ↗
- Anderson, T. W. (2003). An Introduction to Multivariate Statistical Analysis (3rd ed.). Wiley. ISBN: 978-0471360919
- Tabachnick, B. G., & Fidell, L. S. (2019). Using Multivariate Statistics (7th ed.). Pearson. ISBN: 978-0134790541
How to cite this page
ScholarGate. (2026, June 3). Canonical Correlation Analysis. ScholarGate. https://scholargate.app/en/statistics/canonical-correlation-analysis
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Discriminant AnalysisStatistics↔ compare
- Factor AnalysisResearch Statistics↔ compare
- Multiple Linear RegressionStatistics↔ compare
- Partial Least SquaresMachine learning↔ compare