variable.selection: Tune the number of variables on each component

In bootsPLS: Bootstrap Subsamplings of Sparse Partial Least Squares - Discriminant Analysis for Classification and Signature Identification

Description

Performs a testing procedure to choose the number of variable on each component

Usage

variable.selection(object,ncomp,alpha,limit)

Arguments

object	a ‘bootsPLS’ object', as obtained from bootsPLS
ncomp	How many component?
alpha	Level of the test
limit	Vector of maximal number of genes to include on each component

Details

The testing procedure ranks the variables by decreasing frequency in object$frequency, for each component. Random subsamplings are constructed, a spls.hybrid is fitted on the internal learning set and a prediction is made on the internal test set. The testing procedure evaluates the gain in classification accuracy over the random subsamplings when a new variable is added from a decreasing frequency. This is done by on-sided t-test of level alpha. See the reference below for more details on the multiple testing procedure.

Value

A 'variable.selection' object is returned for which plot is available.

pval	pvalue obtained from the testing procedure.
opt	Number of components chosen by the procedure
keepX.constraint	Variables chosen on each of the ncomp components
object	input ‘bootsPLS’ object
alpha	input level of the test

References

Rohart et al. (2015). A Molecular Classification of Human Mesenchymal Stromal Cells

See Also

spls.hybrid, plot, fit.model, prediction

Examples

## Not run: 
data(MSC)
X=MSC$X
Y=MSC$Y
dim(X)
table(Y)

boot=bootsPLS(X=X,Y=Y,ncomp=3,many=5,kCV=5)
out=variable.selection(boot,ncomp=2)

out2=variable.selection(boot,ncomp=2,limit=c(40,40))


## End(Not run)

bootsPLS documentation built on May 30, 2017, 2:12 a.m.

spls.hybrid, midway between PLS and sPLS

bootsPLS index