Other CV criteria

By default, cv.ncvreg() returns the cross-validated deviance:

data(Heart)
X <- Heart$X
y <- Heart$y
cvfit <- cv.ncvreg(X, y, family='binomial')
head(cvfit$cve)
# [1] 1.286990 1.266179 1.243593 1.224601 1.208662 1.195316

In addition, summary.cv.ncvreg() returns an estimated $R^2$ , signal-to-noise ratio (SNR), and for logistic regression, a misclassification error (PE, for prediction error):

head(summary(cvfit)$r.squared)
# [1] 0.003282658 0.023810623 0.045612276 0.063567043 0.078374606 0.090592299
head(summary(cvfit)$snr)
# [1] 0.00329347 0.02439140 0.04779219 0.06788211 0.08503955 0.09961682
head(summary(cvfit)$pe)
# [1] 0.3463203 0.3463203 0.3463203 0.3463203 0.3463203 0.3463203

It is very important to note here that these measures are based on out-of-sample CV predictions, and therefore not artificially inflated by overfitting, as would happen if we used the predictions from ncvreg() directly.

In addition, cv.ncvreg() offers the option to return the cross-validated linear predictors (returnY=TRUE), which allows the user to calculate any prediction criteria they wish. For example, here is how one can calculate the cross-validated AUC using the auc() function from the pROC package:

cvfit <- cv.ncvreg(X, y, family='binomial', returnY=TRUE)
auc <- apply(cvfit$Y, 2, pROC::auc, response=y, quiet=TRUE)
head(auc)
# [1] 0.5858133 0.6841163 0.7043978 0.7094681 0.7116204 0.7137314
plot(cvfit$lambda, auc, log='x', las=1, bty='n', xlab=expression(lambda), 
     xlim=rev(range(cvfit$lambda)), type='l')
abline(v=cvfit$lambda[which.max(auc)], lty=2, col='gray')

Patrick Breheny