loocv.discrimin: Leave-one-out cross-validation for a 'discrimin' analysis

In sdray/ade4: Analysis of Ecological Data: Exploratory and Euclidean Methods in Environmental Sciences

Leave-one-out cross-validation for a discrimin analysis

Description

Leave-one-out cross-validation to test the existence of groups in a discrimin analysis.

Usage

## S3 method for class 'discrimin'
loocv(x, nax = 0, progress = FALSE, ...)
## S3 method for class 'discloocv'
print(x, ...)
## S3 method for class 'discloocv'
plot(x, xax = 1, yax = 2, ...)

Arguments

x	the discrimin analysis on which cross-validation should be done
nax	list of axes for mean overlap index computation (0 = all axes)
progress	logical to display a progress bar during computations (see the progress package)
xax, yax	the numbers of the x-axis and the y-axis
...	further arguments passed to or from other methods

Details

This function returns a list containing the cross-validated coordinates of the rows. The analysis on which the discrimin was computed is redone after removing each row of the data table, one at a time. A discrimin analysis is done on this new analysis and the coordinates of the missing row are computed by projection as supplementary element in the new discrimin analysis. This can be useful to check that the groups evidenced by the discrimin analysis are supported.

Value

A list with:

- XValCoord: the cross-validated row coordinates

- PRESS: the Predicted Residual Error Sum for each row

- PRESSTot: the sum of PRESS for each bca axis

- Oij_disc: the mean overlap index for the discriminant analysis

- Oij_XVal: the mean overlap index for cross-validation

- DeltaOij: the spuriousness index

Author(s)

Jean Thioulouse

See Also

loocv.dudi loocv.between

Examples

## Not run: 
# Data = skulls
data(skulls)
pcaskul <- dudi.pca(skulls, scan = FALSE)
facskul <- gl(5,30)
diskul <- discrimin(pcaskul, facskul, scan = FALSE)
xdiskul <- loocv(diskul, progress = TRUE)
oijdisc <- xdiskul$Oij_disc
oijxval <- xdiskul$Oij_XVal
Doij <- (oijxval - oijdisc)/0.5*100
pst1 <- paste0("Skulls discrimin randtest: p=", round(randtest(diskul)$pvalue, 4), 
", Oij = ", round(oijdisc,2))
pst2 <- paste0("Skulls cross-validation: Oij = ", round(oijxval,2), ", dOij = ",
round(Doij), "%")
if (adegraphicsLoaded()) {
	sc1 <- s.class(diskul$li, facskul, col = TRUE, psub.text = pst1, ellipseSize=0,
	chullSize=1, plot = FALSE)
	sc2 <- s.class(xdiskul$XValCoord, facskul, col = TRUE, psub.text = pst2,
	ellipseSize=0, chullSize=1, plot = FALSE)
	ADEgS(list(sc1, sc2), layout=c(2,2))
} else {
	par(mfrow=c(2,2))
	s.class(diskul$li, facskul, sub = pst1)
	s.class(xdiskul$XValCoord, facskul, sub = pst2)
}
data(chazeb)
pcacz <- dudi.pca(chazeb$tab, scan = FALSE)
discz <- discrimin(pcacz, chazeb$cla, scan = FALSE)
xdiscz <- loocv(discz, progress = TRUE)
oijdiscz <- xdiscz$Oij_disc
oijxvalz <- xdiscz$Oij_XVal
Doijz <- (oijxvalz - oijdiscz)/0.5*100
pst1 <- paste0("Chazeb discrimin randtest: p=", round(randtest(discz)$pvalue, 4), 
", Oij = ", round(oijdiscz,2))
pst2 <- paste0("Chazeb cross-validation: Oij = ", round(oijxvalz,2), ", dOij = ", 
round(Doijz), "%")
if (adegraphicsLoaded()) {
	tabi <- cbind(discz$li, pcacz$tab)
	gr1 <- s.class(tabi, xax=1, yax=2:7, chazeb$cla, col = TRUE, plot = FALSE)
	for (i in 1:6) gr1[[i]] <- update(gr1[[i]], psub.text = names(tabi)[i+1],
	plot = FALSE)
	pos1 <- gr1@positions
	pos1[,1] <- c(0, .3333, .6667, 0, .3333, .6667)
	pos1[,2] <- c(.6667, .6667, .6667, .3333, .3333, .3333)
	pos1[,3] <- c(.3333, .6667, 1, .3333, .6667, 1)
	pos1[,4] <- c(1, 1, 1, .6667, .6667, .6667)
	gr1@positions <- pos1
	sc1 <- s1d.gauss(discz$li, chazeb$cla, col = TRUE, psub.text = pst1,
	plot = FALSE)
	sc2 <- s1d.gauss(xdiscz$XValCoord, chazeb$cla, col = TRUE, psub.text = pst2,
	plot = FALSE)
	ADEgS(list(gr1[[1]], gr1[[2]], gr1[[3]], gr1[[4]], gr1[[5]], gr1[[6]], sc1, sc2))
} else {
	dev.new()
	sco.gauss(discz$li[,1], as.data.frame(chazeb$cla), sub = pst1)
	dev.new()
	sco.gauss(xdiscz$XValCoord[,1], as.data.frame(chazeb$cla), sub = pst2)
}

## End(Not run)

sdray/ade4 documentation built on March 30, 2024, 12:33 a.m.