lars.cross.validation: Runs cross-validation of LARS regression

In knights-lab/MWAS: Biomarker Discovery in Microbiome-Wide Association Studies (MWAS)

Description

Runs cross-validation of LARS regression

Usage

lars.cross.validation(x, y, nfolds = 10, folds = NULL, verbose = FALSE,
  ...)

Arguments

x	A matrix of numeric predictors
y	A factor or logical of responses
nfolds	number of cross-validation folds (nfolds==-1 means leave-one-out)
folds	optional, if not NULL then uses these fold assignments (one fold index for each row in x)
verbose	Use verbose output
...	additional parameters for lars

Value

:

y	Observed values
predicted	CV predicted values
fractions	Vector of LASSO fractions used across folds
best.fraction	Best LASSO fraction used
MSE	Mean squared error
RMSE	Root mean squared error
nfolds	Number of folds used (or -1 for leave-one-out)
params	List of additional parameters, if any
coefficients	Feature-by-fold matrix of coefficients of features across folds
final.model	Final lars model trained on whole data set

Examples

# generate fake data: 50 samples, 20 variables
x <- matrix(rnorm(2000),100)
rownames(x) <- sprintf('Sample%02d',1:100)
colnames(x) <- sprintf('Feature%02d',1:20)
# generate fake response variable, function of columns 1-5 plus noise
y <- rowSums(sweep(x[,1:5],2,runif(5),'*')) + rnorm(100,sd=.5)
names(y) <- rownames(x)

# 10-fold cross validation with predictions
res.lars <- lars.cross.validation(x,y)
# plot predictions
plot(res.lars$y, res.lars$predicted, xlab="Observed", ylab="Predicted (hold-out)")

# correlation
cat('Pearson correlation of predictions with truth?\n')
print(cor.test(res.lars$y, res.lars$predicted))

# plot fraction of times each variable was included in a fold
sorted.importances <- sort(rowMeans(res.lars$coefficients > 0), decreasing=TRUE)
# bar plot of top ten variables
par(mar=c(4.5,6,2,1),las=2) # larger margin on left, perpendicular axis labels
barplot(rev(sorted.importances[1:10]),horiz=TRUE, xlab='Frequency of selection')

knights-lab/MWAS documentation built on May 20, 2019, 12:52 p.m.