gamlr (version 0.1)

cvselect: Cross Validation for gamlr

Description

Cross Validated gamma lasso penalty selection

Usage

cvselect(fit, nfold=10, npen=20, verb=TRUE)
cvplot(fit, ...)

Arguments

fit

A fitted gamlr object.

nfold

The number of cross validation folds.

npen

The size of the penalty grid to consider.

verb

Whether to print and plot progress.

...

Extra plot arguments.

Value

• The returned gamlr object is the same as fit, but augmented with
• cvpve
  Out-of-sample proportion of variance explained.
• cvindex
  The index of the CV optimal penalty (within fit$penalty).
• cvgrid
  The grid of actual cross validated penalties.

Details

Performs nfold out-of-sample prediction exercises on n/nfold sized test sets, and uses resulting PVE to choose the penalty size (i.e. the size of the expected L1 regularization weight). See gamlr for details. cvplot can be used to boxplot the results.

References

Taddy 2013, "The Gamma-Lasso".

See Also

gamlr

Examples

n <- 100
p <- 100

xvar <- matrix(ncol=p,nrow=p)
for(i in 1:p) for(j in i:p) xvar[i,j] <- 0.5^{abs(i-j)}
x <- matrix(rnorm(p*n), nrow=n)%*%chol(xvar)
beta <- matrix( (-1)^(1:p)*exp(-(1:p)/10) )
mu = x%*%beta
y <- rbinom(n, prob=1/(1+exp(-2*mu)), size=1)

fit <- cvselect(gamlr(x, y, verb=TRUE, family="binomial"), npen=5, nfold=5)

## plot some model selection criterion	

par(mfrow=c(1,3))
## again, the CV plot
cvplot(fit)
## BF
plot(-log(fit$penalty),fit$BF,ylab="log BF", pch=21,bg="darkorange")
## adj BF
points(-log(fit$penalty),fit$BF-fit$zerodf,pch=21,bg="dodgerblue")
## fitted values and all crit
plot(fit, select="CV", col=8)
abline(v=-log(fit$penalty)[which.min(AIC(fit))], col="green", lty=2) # AIC
abline(v=-log(fit$penalty)[which.min(AIC(fit, k=log(fit$n)))], col="red", lty=2) # BIC
abline(v=-log(fit$penalty)[which.max(fit$BF)], col="darkorange", lty=2) # BF
abline(v=-log(fit$penalty)[which.max(fit$BF-fit$zerodf)], col="dodgerblue", lty=2) # adjBF