testclass: Evaluating a classification method based on several learning...

In MAclinical: Class prediction based on microarray data and clinical parameters

Description

This function evaluates classifiers built using microarray data and/or clinical predictors, based on several pairs of learning and test data sets.

Usage

testclass(x=NULL,y,z=NULL,learningsets,classifier,ncomp=0:3,
varsel=NULL,nbgene=NULL,fold=10,...)

Arguments

x	A n x p matrix giving the gene expression levels of p genes (columns) for n patients (rows).
y	A numeric vector of length n giving the class membership of the n patients, coded as 0,...,K-1 (where K is the number of classes).
z	A n x q data frame giving the q clinical predictors for the n patients. Nominal variables should be given as factors, variables with an at least ordinal scale should be given as numeric.
learningsets	A matrix with niter rows giving the indices of the arrays to be included in the learning sets for the niter iterations, as generated by the function generate.learningsets. The i-th row gives the indices of the arrays included in the learning set for the i-th iteration. For instance, in LOOCV, the i-th row of the matrix learningsets contains all the integers from 1 to n except i. Note that an observation may be included twice or more in the same learning set (for instance in bootstrap sampling).
classifier	The function used to construct a classifier. The function must have the same structure as plsrf_xz_pv.
ncomp	The candidate numbers of PLS components (if PLS dimension reduction is used).
varsel	A niter x p matrix giving the indices of the genes ordered by the chosen gene selection criterion. For example, the element in the first row and the first column is the index of the gene that is ranked best using the first learning set.
nbgene	The number of genes to use for classifier construction. Default is nbgene=NULL, corresponding to all genes.
fold	The number of folds for the pre-validation step. See Boulesteix et al (2008) for more details. Default is fold=10.
...	Other arguments to be passed to the function cforest_control from the party package or to the function svm from the package e1071, depending on the specified classifier.

Details

For an overview of different methods used to generate the learning sets defined by generate.learningsets, see Boulesteix et al (2007). These methods include (repeated) cross-validation, subsampling, bootstrap sampling.

Value

error	A numeric vector of length niter giving the misclassification rate for each iteration.
bestncomp	A numeric vector of length niter giving the best number of (pre-validated) PLS components, as obtained using the model selection method based on the out-of-bag error by Boulesteix et al (returned only for the classifiers plsrf_xz_pv, plsrf_xz, plsrf_x_pv, plsrf_x).
OOB	A list of length niter, whose elements are numeric vectors of the same length as ncomp giving the out-of-bag error of the forest constructed with the corresponding number of (pre-validated) PLS components (returned only for the classifiers plsrf_xz_pv, plsrf_xz, plsrf_x_pv, plsrf_x, rf_z. For rf_z, no model selection is performed: OOB is just the out-of-bag error of the constructed forest.)

Author(s)

Anne-Laure Boulesteix (http://www.ibe.med.uni-muenchen.de/organisation/mitarbeiter/020_professuren/boulesteix/eng.html)

References

Boulesteix AL, Porzelius C, Daumer M, 2008. Microarray-based classification and clinical predictors: On combined classifiers and additional predictive value. Bioinformatics 24:1698-1706.

Boulesteix AL, Strobl C, Augustin T, Daumer D, 2008. Evaluating microarray-based classifiers: an overview. Cancer Informatics 6:77-97.

See Also

testclass_simul, simulate, generate.learningsets, plsrf_xz_pv, plsrf_x_pv, plsrf_xz, plsrf_x, rf_z, svm_x, logistic_z.

Examples

# load MAclinical library
# library(MAclinical)

# Generate data
x<-matrix(rnorm(20000),100,200)
z<-matrix(rnorm(500),100,5)
y<-sample(0:1,100,replace=TRUE)

# Generate learningsets (5-fold CV)
my.learningsets<-generate.learningsets(n=100,method="CV",fold=5)

# Evaluate accuracy of the PLS-PV-RF method
my.eval<-testclass(x=x,y=y,z=z,learningsets=my.learningsets,classifier=plsrf_xz_pv,ncomp=5,
varsel=NULL,nbgene=NULL,fold=10)

# With variable selection
my.varsel<-matrix(0,5,200)
for (i in 1:5)
 {
 my.varsel[i,]<-order(abs(studentt.stat(X=x[my.learningsets[i,],],
 L=y[my.learningsets[i,]]+1)),decreasing=TRUE)
 }

my.eval<-testclass(x=x,y=y,z=z,learningsets=my.learningsets,classifier=plsrf_xz_pv,ncomp=5,
varsel=my.varsel,nbgene=15,fold=10)

MAclinical documentation built on May 2, 2019, 9:30 a.m.