BuyseTest (version 3.0.5)

auc: Estimation of the Area Under the ROC Curve (EXPERIMENTAL)

Description

Estimation of the Area Under the ROC curve, possibly after cross validation, to assess the discriminant ability of a biomarker regarding a disease status.

Usage

auc(
  labels,
  predictions,
  fold = NULL,
  observation = NULL,
  direction = ">",
  add.halfNeutral = TRUE,
  null = 0.5,
  conf.level = 0.95,
  transformation = TRUE,
  order.Hprojection = 2,
  pooling = "mean"
)

Value

An S3 object of class BuyseTestAUC that inherits from data.frame. The last line of the object contains the global AUC value with its standard error.

Arguments

labels

[integer/character vector] the disease status (should only take two different values).

predictions

[numeric vector] A vector with the same length as labels containing the biomarker values.

fold

[character/integer vector] If using cross validation, the index of the fold. Should have the same length as labels.

observation

[integer vector] If using cross validation, the index of the corresponding observation in the original dataset. Necessary to compute the standard error when using cross validation.

direction

[character] ">" lead to estimate P[Y>X], "<" to estimate P[Y<X], and "auto" to estimate max(P[Y>X],P[Y<X]).

add.halfNeutral

[logical] should half of the neutral score be added to the favorable and unfavorable scores? Useful to match the usual definition of the AUC in presence of ties.

null

[numeric, 0-1] the value against which the AUC should be compared when computing the p-value.

conf.level

[numeric, 0-1] the confidence level of the confidence intervals.

transformation

[logical] should a log-log transformation be used when computing the confidence intervals and the p-value.

order.Hprojection

[1,2] the order of the H-projection used to linear the statistic when computing the standard error. 2 is involves more calculations but is more accurate in small samples. Only active when the fold argument is NULL.

pooling

[character] method used to compute the global AUC from the fold-specific AUC: either an empirical average "mean" or a weighted average with weights proportional to the number of pairs of observations in each fold "pairs".

Details

The iid decomposition of the AUC is based on a first order decomposition. So its squared value will not exactly match the square of the standard error estimated with a second order H-projection.

References

Erin LeDell, Maya Petersen, and Mark van der Laan (2015). Computationally efficient confidence intervals for cross-validated area under the ROC curve estimates. Electron J Stat. 9(1):1583–1607.

Examples

library(data.table)

n <- 200
set.seed(10)
X <- rnorm(n)
dt <- data.table(Y = as.factor(rbinom(n, size = 1, prob = 1/(1+exp(1/2-X)))),
                 X = X,
                 fold = unlist(lapply(1:10,function(iL){rep(iL,n/10)})))

## compute auc
auc(labels = dt$Y, predictions = dt$X, direction = ">")

## compute auc after 10-fold cross-validation
auc(labels = dt$Y, prediction = dt$X, fold = dt$fold, observation = 1:NROW(dt))