RVAideMemoire (version 0.9-45-2)

DA.valid: Cross-validation in Discriminant Analysis

Description

Performs a cross-validation to assess the prediction ability of a Discriminant Analysis. Linear Discriminant Analysis (from lda), Partial Least Squares - Discriminant Analysis (from plsda) and Correspondence Discriminant Analysis (from discrimin.coa) are handled. Two methods are implemented for cross-validation: leave-one-out and M-fold.

Usage

DA.valid(model, method = c("Mfold", "loo"), crit.lda = c("plug-in", "predictive",
  "debiased"), crit.plsda = c("mahalanobis.dist", "centroids.dist", "max.dist"),
  crit.cda = c("mahalanobis", "euclidian"), M = 10, nrep = 20)

Arguments

model

object of class "lda" (from lda), "plsda" (from plsda) or "coadisc" (from

method

what kind of validation to use, matching one of "Mfold" or "loo" (see below). Default is "Mfold".

crit.lda

this determines how the parameter estimation is handled in LDA cross-validation. With "plug-in" (the default) the usual unbiased parameter estimates are used and assumed to be correct. With "debiased" an unbiased estimator of the

crit.plsda

prediction method to be applied for PLS-DA cross-validation. Should be a subset of "mahalanobis.dist" (default), "centroids.dist" or "max.dist".

crit.cda

prediction method to be applied for CDA cross-validation. Should be a subset of "mahalanobis" (default) or "euclidian".

M

the number of folds in the M-fold cross-validation.

nrep

the number of repetitions of the whole procedure in the M-fold cross-validation.

Value

• model
  type of discriminant model used.
• method
  method used for cross-validation.
• crit.lda
  method for parameter estimation in LDA cross-validation.
• crit.plsda
  distance used in PLS-DA cross-validation.
• crit.cda
  distance used in CDA cross-validation.
• M
  number of folds in the M-fold cross-validation.
• nrep
  number of repetitions in the M-fold cross-validation.
• tab
  results of cross-validation.

Details

Leave-one-out cross-validation is not recommended with PLS-DA. When working on a LDA, the prior probabilities for the models fitted on the training datasets are extracted from model (they are not evaluated from the training dataset itself).

See Also

lda, predict.lda, plsda, perf, discrimin.coa, predict.coadisc

Examples

# With a LDA
require(MASS)
data(iris)

model.LDA <- lda(iris[,1:4],iris$Species)
DA.valid(model.LDA)