Calculate and plot AUC

AUC.R

###################################################
##
## Functions for calculating AUC and plotting ROC
## Corey Chivers, 2013
## [email protected]
##
###################################################


## Descrete integration for AUC calc
## Δx.y1 + 1/2Δx.Δy  <- summation of trapezoids
desc_integrate<-function(x,y)
{
   f<-cbind(x,y)
   ## Sort by x, then by y (assending)
   f<-f[order(f[,1],f[,2]),] 
   dint<-0
   x<-f[,1]
   y<-f[,2]
   dint<-sapply(2:length(x),function(i){
      (x[i]-x[i-1])*y[i-1] + 0.5*(x[i]-x[i-1]) * (y[i]-y[i-1])})
   dint<-sum(dint)
   return(dint)
}

## This is a handy generic.
add_error_bars<-function(data,error,dimensions=1,...)
{
   for(i in 1:length(data[,1]))
   {
      # y axis is 1st dimension
      arrows(data[i,1],data[i,2],data[i,1],error[i,1],angle=90,...)
      arrows(data[i,1],data[i,2],data[i,1],error[i,2],angle=90,...)

      if(dimensions==2)
      {
         arrows(data[i,1],data[i,2],error[i,3],data[i,2],angle=90,...)
         arrows(data[i,1],data[i,2],error[i,4],data[i,2],angle=90,...)
      }
   }
}


####################################################################
## Calculate the AUC and optionally plot the ROC
##  **Usage**
## d: a vector of logicals (0,1)
## pred: a vector of predicted values on range [0,1]
## plot: logical - plot or not
## error_bars: logical - add error bars or not
## ci: atomic vector - confidence interval width for error bars
## res: atomic vector - resolution of the thresholds to test
####################################################################
AUC<-function(d,pred,plot=FALSE,error_bars=FALSE,ci=0.95,res=100,add=FALSE,...)
{
   n<-length(d)
   dt<-seq(0,1,length.out=res)
   tp<-numeric(res)
   fp<-numeric(res)
   fn<-numeric(res)
 
   error<-array(dim=c(res,4)) # <tp upper, tp lower, fp upper, fp lower>
   sapply(1:res,function(i)
   {
      tp[i]<<- sum( d[pred > dt[i] ] )/ sum(d)
      fp[i]<<- sum( d[pred > dt[i] ] == 0 )/ sum(!d)
      fn[i]<<- sum( d[pred < dt[i] ] )/ sum(d) 
   

      #Calculate CI based on the beta distribution
      alpha_tp<-sum( d[pred > dt[i] ] ) + 1
      beta_tp<- sum(d) - sum( d[pred > dt[i] ] ) + 1
      error[i,1]<<-qbeta((1-ci)/2,alpha_tp,beta_tp)   #ci% bounds based on beta dist
      error[i,2]<<-qbeta(1-(1-ci)/2,alpha_tp,beta_tp)

      alpha_fp<- sum( d[pred > dt[i] ] == 0 ) + 1
      beta_fp<- sum(!d) - sum( d[pred > dt[i] ] == 0 ) + 1
      error[i,3]<<-qbeta((1-ci)/2,alpha_fp,beta_fp)   #ci% bounds based on beta dist
      error[i,4]<<-qbeta(1-(1-ci)/2,alpha_fp,beta_fp)
   })

   # Which threshold value minimises
   # the sum of the error rates.
   opt_thresh<-dt[which.min(fp+fn)]

   # Ensure collisions at 0,0 and 1,1
   fp<-c(1,fp,0)
   tp<-c(1,tp,0)

   # Integrate the ROC
   auc<-desc_integrate(fp,tp)

   if(plot)
   {
      if(add)
      {
         lines(fp,tp,type='b',pch=20)
      }else{
         plot(fp,tp,type='b',pch=20,xlim=c(0,1),ylim=c(0,1),...)
         text( 0.8,0.2,paste('AUC =',round(auc,3)) )
         abline(0,1,lty=2)
      }

      if(error_bars)
         add_error_bars(cbind(fp[2:(res+1)],tp[2:(res+1)]),error,dimensions=2,length=0.01)
   }
   return(list(auc=auc,opt_thresh=opt_thresh))
}