auc function

ROCR

I appreciate this information, which is new for me. Up to now, i was
using the function

  get.auc <- function(statistic, label, negative, positive)
  {
      xmove <- as.numeric(label == negative)
      ymove <- as.numeric(label == positive)
      stopifnot(xmove + ymove == 1)
      rank.stat <- rank(statistic, ties.method="min")
      steps <- aggregate(cbind(xmove, ymove), by=list(rank.stat), sum)
      n <- nrow(steps)
      x <- c(0, cumsum(steps[n:1, 2]))
      y <- c(0, cumsum(steps[n:1, 3]))
      sum(diff(x) * (y[1:n] + y[2:(n+1)]))/(2*max(x)*max(y))
  }

CRAN package ROCR allows to compute many different measures and
visualisations of classifier performance. In particular, AUC may
be computed as follows

  library(ROCR)

  n <- 50
  label <- ordered(rep(c("c1", "c2"), length=n))
  set.seed(12345)
  statistic <- rnorm(n) + (label == "c2")
  pred <- prediction(statistic, label)
  AUC <- performance(pred, "auc")@y.values[[1]]
  cbind(AUC, diff=AUC - get.auc(statistic, label, "c1", "c2"))
  #         AUC diff
  # [1,] 0.7392    0

The difference is not always exactly zero, but is at the level
of the machine rounding error.

Petr Savicky.

On Thu, Jan 20, 2011 at 3:04 PM, He, Yulei <he at hcp.med.harvard.edu> wrote:

Hi, there.

Suppose I already have sensitivities and specificities. What is the quick
R-function to calculate AUC for the ROC plot? There seem to be many R
functions to calculate AUC.

Thanks.

Yulei