Statistics Glossary → Fisher classification functions and canonical discriminant functions
Fisher classification functions and canonical discriminant functionsconcept
In discriminant analysis, the linear score functions computed separately for each class to assign observations, and the linear components that best separate the groups.
With Fisher's classification functions, a separate linear equation is built for each class: the coefficients are the inverse of the pooled within-group covariance matrix times the class mean, and the log of the prior class probability is added to the constant. A new observation is plugged into every equation and assigned to the class with the highest score. This rule gives the same result as the linear discriminant (LDA) rule under the assumption that the classes are multivariate normal with a common covariance.
Canonical discriminant functions are the linear components that maximize the ratio of between-group to within-group variance; there are at most min(k - 1, p) of them (k classes, p variables). The first function gives the direction that best separates the groups; the importance of each function is judged by its eigenvalue and canonical correlation, and its significance by Wilks' Λ. Canonical functions are used to display the groups in a low-dimensional plot and to interpret which variables drive the separation, whereas classification functions are for direct assignment.
An everyday example
A botanist wants to tell three iris species apart from sepal and petal measurements. A 'score formula' is computed for each species; the measurements of a new flower are put into all three formulas, and it is assigned to the species with the highest score. The canonical functions, in turn, show the species as separate clouds on a two-axis plot.
In MerQur
'Additional output: Fisher classification function coefficients' and 'Canonical score plot (LD1 × LD2 + centroids)' in Discriminant Analysis; the predicted class, D scores and membership probabilities can be saved to the dataset.
Analyses where this concept appears
Discriminant Analysis (LDA / QDA)
Related concepts
Prior class probabilities (discriminant analysis)Multivariate test statistics (Wilks' Λ, Pillai's trace, Hotelling-Lawley, Roy)