Canonical discriminant analysis is a dimension-reduction technique related to principal component analysis and canonical correlation. The methodology used in deriving the canonical coefficients parallels that of a one-way MANOVA. MANOVA tests for equality of the mean vector across class levels. Canonical discriminant analysis finds linear combinations of the quantitative variables that provide maximal separation between classes or groups. Given a classification variable and several quantitative variables, the CANDISC procedure derives canonical variables, linear combinations of the quantitative variables that summarize between-class variation in much the same way that principal components summarize total variation.
The CANDISC procedure performs a canonical discriminant analysis, computes squared Mahalanobis distances between class means, and performs both univariate and multivariate one-way analyses of variance. Two output data sets can be produced: one containing the canonical coefficients and another containing, among other things, scored canonical variables. The canonical coefficients output data set can be rotated by the FACTOR procedure. It is customary to standardize the canonical coefficients so that the canonical variables have means that are equal to zero and pooled within-class variances that are equal to one. PROC CANDISC displays both standardized and unstandardized canonical coefficients. Correlations between the canonical variables and the original variables as well as the class means for the canonical variables are also displayed; these correlations, sometimes known as loadings, are called canonical structures. To aid the visual interpretation of group differences, you can use ODS Graphics to display graphs of pairs of canonical variables from the scored canonical variables output data set.
Given two or more groups of observations with measurements on several quantitative variables, canonical discriminant analysis derives a linear combination of the variables that has the highest possible multiple correlation with the groups. This maximal multiple correlation is called the first canonical correlation. The coefficients of the linear combination are the canonical coefficients or canonical weights. The variable defined by the linear combination is the first canonical variable or canonical component. The second canonical correlation is obtained by finding the linear combination uncorrelated with the first canonical variable that has the highest possible multiple correlation with the groups. The process of extracting canonical variables can be repeated until the number of canonical variables equals the number of original variables or the number of classes minus one, whichever is smaller.
The first canonical correlation is at least as large as the multiple correlation between the groups and any of the original variables. If the original variables have high within-group correlations, the first canonical correlation can be large even if all the multiple correlations are small. In other words, the first canonical variable can show substantial differences between the classes, even if none of the original variables do. Canonical variables are sometimes called discriminant functions, but this usage is ambiguous because the DISCRIM procedure produces very different functions for classification that are also called discriminant functions.
For each canonical correlation, PROC CANDISC tests the hypothesis that it and all smaller canonical correlations are zero in the population. An F approximation (Rao, 1973; Kshirsagar, 1972) is used that gives better small-sample results than the usual chi-square approximation. The variables should have an approximate multivariate normal distribution within each class, with a common covariance matrix in order for the probability levels to be valid.
Canonical discriminant analysis is equivalent to canonical correlation analysis between the quantitative variables and a set of dummy variables coded from the class variable. Canonical discriminant analysis is also equivalent to performing the following steps:
Transform the variables so that the pooled within-class covariance matrix is an identity matrix.
Compute class means on the transformed variables.
Perform a principal component analysis on the means, weighting each mean by the number of observations in the class. The eigenvalues are equal to the ratio of between-class variation to within-class variation in the direction of each principal component.
Back-transform the principal components into the space of the original variables, obtaining the canonical variables.
An interesting property of the canonical variables is that they are uncorrelated whether the correlation is calculated from the total sample or from the pooled within-class correlations. The canonical coefficients are not orthogonal, however, so the canonical variables do not represent perpendicular directions through the space of the original variables.