testdim_mbplsda {packMBPLSDA}

Test of number of components by two-fold cross-validation for a multi-block partial least squares discriminant model

Description

Function to perform a two-fold cross-validation in order to select the optimal number of dimensions of a multi-block partial least squares discriminant model, according to the classification error rate or to the area under ROC curve

Usage

testdim_mbplsda(object, nrepet = 100, algo = c("max", "gravity", "threshold"),
threshold = 0.5, bloY, outputs = c("ER", "ConfMat", "AUC"), cpus = 1)

Arguments

object

an object created by mbplsda_nfX

nrepet

integer indicating the number of repetitions

algo

character vector indicating the method(s) of prediction to use (see details)

threshold

numeric indicating the threshold, between 0 and 1, to consider the categories are predicted with the threshold prediction method.

bloY

integer vector indicating the number of categories per variable of the Y-block.

outputs

character vector indicating the wanted outputs (see details)

cpus

integer indicating the number of cpus to use when running the code in parallel

Details

Three different algorithms are available to predict the categories of observations. In the max, and respectively the threshold algorithms, numeric values are calculated from the matrix of explanatory variables and the regression coefficients. Then, the predicted categorie for each variable of the Y-block is the one which corresponds to the higher predicted value, respectively to the values higher than the indicated threshold. In the gravity algorithm, predicted scores of the observations on the components are calculated. Then, each observation is assigned to the observed category of which it is closest to the barycentre in the component space.

Available outputs are Error Rates (ER), Confusion Matrix (ConfMat), Aera Under Curve (AUC).

Value

TRUEnrepet

number of repetitions

TruePosC.max, .gravity, .threshold

statistical description of percentages of true positive observations per category, evaluated on the calibration dataset, with the different algorithms (TPcM for "max", TPcG for "gravity", TPcT for "threshold"), for a number of components ranging from 1 to its maximum value

TruePosV.max, .gravity, .threshold

statistical description of percentages of true positive observations per category, evaluated on the validation dataset, with the different algorithms (TPvM for "max", TPvG for "gravity", TPvT for "threshold"), for a number of components ranging from 1 to its maximum value

TrueNegC.max, .gravity, .threshold

statistical description of percentages of true negative observations per category, evaluated on the calibration dataset, with the different algorithms (TNcM for "max", TNcG for "gravity", TNcT for "threshold"), for a number of components ranging from 1 to its maximum value

TrueNegV.max, .gravity, .threshold

statistical description of percentages of true negative observations per category, evaluated on the validation dataset, with the different algorithms (TNvM for "max", TNvG for "gravity", TNvT for "threshold"), for a number of components ranging from 1 to its maximum value

FalsePosC.max, .gravity, .threshold

statistical description of percentages of false positive observations per category, evaluated on the calibration dataset, with the different algorithms (FPcM for "max", FPcG for "gravity", FPcT for "threshold"), for a number of components ranging from 1 to its maximum value

FalsePosV.max, .gravity, .threshold

statistical description of percentages of false positive observations per category, evaluated on the validation dataset, with the different algorithms (FPvM for "max", FPvG for "gravity", FPvT for "threshold"), for a number of components ranging from 1 to its maximum value

FalseNegC.max, .gravity, .threshold

statistical description of percentages of false negative observations per category, evaluated on the calibration dataset, with the different algorithms (FNcM for "max", FNcG for "gravity", FNcT for "threshold"), for a number of components ranging from 1 to its maximum value

FalseNegV.max, .gravity, .threshold

statistical description of percentages of false negative observations per category, evaluated on the validation dataset, with the different algorithms (FNvM for "max", FNvG for "gravity", FNvT for "threshold"), for a number of components ranging from 1 to its maximum value

ErrorRateC.max, .gravity, .threshold

statistical description of prediction error rates per category, evaluated on the calibration dataset, with the different algorithms (ERcM for "max", ERcG for "gravity", ERcT for "threshold"), for a number of components ranging from 1 to its maximum value

ErrorRateV.max, .gravity, .threshold

statistical description of prediction error rates per category, evaluated on the validation dataset, with the different algorithms (ERvM for "max", ERvG for "gravity", ERvT for "threshold"), for a number of components ranging from 1 to its maximum value

ErrorRateCglobal.max, .gravity, .threshold

statistical description of global prediction error rates, evaluated on the calibration dataset, with the different algorithms (ERcM.global for "max", ERcG.global for "gravity", ERcT.global for "threshold"), for a number of components ranging from 1 to its maximum value

ErrorRateVglobal.max, .gravity, .threshold

statistical description of global prediction error rates, evaluated on the validation dataset, with the different algorithms (ERvM.global for "max", ERvG.global for "gravity", ERvT.global for "threshold"), for a number of components ranging from 1 to its maximum value

AUCc

statistical description of aera under ROC curve values per category, evaluated on the calibration dataset, if all Y-block variables are binary, for a number of components ranging from 1 to its maximum value

AUCv

statistical description of aera under ROC curve values per category, evaluated on the validation dataset, if all Y-block variables are binary, for a number of components ranging from 1 to its maximum value

AUCc.global

statistical description of global aera under ROC curve values, evaluated on the calibration dataset, if all Y-block variables are binary, for a number of components ranging from 1 to its maximum value

AUCv.global

statistical description of global aera under ROC curve values, evaluated on the validation dataset, if all Y-block variables are binary, for a number of components ranging from 1 to its maximum value

Note

at least 30 cross-validation repetitions may be recommended

Author(s)

Marion Brandolini-Bunlon (<[e-mail]>) and Stephanie Bougeard (<[e-mail]>)

References

Stone, M. (1974). Cross-validatory choice and assessment of statistical predictions. Journal of the Royal Statistical Society B, 36(2), 111-147.

Brandolini-Bunlon, M., Petera, M., Gaudreau, P., Comte, B., Bougeard, S., Pujos-Guillot, E.(2019). A new tool for multi-block PLS discriminant analysis of metabolomic data: application to systems epidemiology. Presented at 12emes Journees Scientifiques RFMF, Clermont-Ferrand, FRA(05-21-2019 - 05-23-2019).

Brandolini-Bunlon, M., Petera, M., Gaudreau, P., Comte, B., Bougeard, S., Pujos-Guillot, E.(2019). Multi-block PLS discriminant analysis for the joint analysis of metabolomic and epidemiological data. Metabolomics, 15(10):134

Brandolini-Bunlon, M., Petera, M., Gaudreau, P., Comte, B., Bougeard, S., Pujos-Guillot, E.(2020). A new tool for multi-block PLS discriminant analysis of metabolomic data: application to systems epidemiology. Presented at Chimiometrie 2020, Liege, BEL(01-27-2020 - 01-29-2020).

See Also

mbplsda plot_testdim_mbplsda packMBPLSDA-package

Examples

data(status)
data(medical)
data(omics)
data(nutrition)
ktabX <- ktab.list.df(list(medical = medical[,1:10], 
nutrition = nutrition[,1:10], omics = omics[,1:20]))
disjonctif <- (disjunctive(status))
dudiY   <- dudi.pca(disjonctif , center = FALSE, scale = FALSE, scannf = FALSE)
bloYobs <- 2
modelembplsQ <- mbplsda(dudiY, ktabX, scale = TRUE, option = "uniform", scannf = FALSE, nf = 3)
resdim <- testdim_mbplsda(object = modelembplsQ, nrepet = 30, threshold = 0.5, 
bloY = bloYobs, cpus = 1, algo = c("max"), outputs = c("ER"))