Description
Find pairs of weighted combinations, one from each of two sets of variables, that correlate as highly as possible, to describe how two blocks of measures relate.
Canonical correlation analysis relates a set of dependent variables to a set of independent variables; multiple regression is its special case with one dependent variable. It forms a linear combination (canonical variate) of each set so that their correlation is maximal, then further pairs uncorrelated with earlier ones, up to the size of the smaller set. Output includes canonical correlations, standardised coefficients, loadings and cross-loadings, canonical scores per case, and redundancy indices showing how much of one set's variance the other set's variates explain. It resembles principal components analysis but maximises between-set correlation rather than within-set variance. A key discipline is checking redundancy: Mazzocchi's example finds three significant functions between attitudes and lifestyle items, yet the first variate explains under 2% of the other set's variance.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Unmigrated
- Stage
- Discovery, Post-launch
- Data
- Quantitative
- Effort
- Days
Steps
- Define two sets of variables measuring different concepts, for example attitude items and behavioural usage measures.
- Check sample size (at least 10 cases per variable is a common guide), linearity and outliers.
- Run CCA (SAS PROC CANCORR, the SPSS CANCORR macro, R cancor) and test how many canonical functions are significant with sequential tests.
- Inspect canonical correlations and standardised coefficients for each significant function.
- Interpret each dimension from loadings (variables with their own variate) and cross-loadings (with the other set's variate), not raw weights.
- Check redundancy indices to see how much variance each set shares with the other's variates, and drop functions that explain trivially little.
- Save canonical scores for segmentation or further modelling if useful.
Sources
- Statistics for Marketing and Consumer Research (Mazzocchi, 2008)
- Methods of Multivariate Analysis (Rencher, 2002)
- Quantitative Social Research Methods (Singh, 2007)
- A User's Guide to Principal Components (Jackson, 1991)
- Linear Statistical Inference and Its Applications (Rao, 1973)
- Multivariate Statistical Inference and Applications (Rencher, 1998)
Also known as
- CCA
- Canonical correlation
- Canonical variates analysis of two variable sets
Categories
- Analytics & Behaviour
- Statistics & Modelling