R/compute.threshold.YI.AROC.sp.R

In ROCnReg: ROC Curve Inference with and without Covariates

compute.threshold.YI.AROC.sp <-
function(object, newdata, ci.level = 0.95, parallel = c("no", "multicore", "snow"), ncpus = 1, cl = NULL) {
  if(class(object)[1] != "AROC.sp") {
    stop(paste0("This function can not be used for this object class: ", class(object)[1]))
  }
  names.cov <- all.vars(object$formula)[-1]
  if(!missing(newdata) && !inherits(newdata, "data.frame"))
    stop("Newdata must be a data frame")
  if(!missing(newdata) && length(names.cov) != 0 && sum(is.na(match(names.cov, names(newdata)))))
    stop("Not all needed variables are supplied in newdata") 

  if(missing(newdata)) {
    newdata <- cROCData(object$data, names.cov, object$group)
  } else {
    newdata <- as.data.frame(newdata)
    newdata <- na.omit(newdata[,names.cov,drop = FALSE])
  }  
  p <- seq(0, 1, length = 500)
  np <- length(p)
  npred <- nrow(newdata)

  # Compute the AROC
  sigma0 <- summary(object$fit)$sigma
  data.d <- (object$data[object$data[,object$group] != object$tag.h,])[!object$missing.ind$d,]
  n1 <- nrow(data.d)
  pre.placement.values <- (data.d[,object$marker] - predict(object$fit, newdata = data.d))/sigma0

  if(object$est.cdf.h == "normal") {
    u1 <- 1 - pnorm(pre.placement.values)
  } else {
    res0p <- object$fit$residuals/sigma0
    F0res <- ecdf(res0p)
    u1 <- 1 - F0res(pre.placement.values)
  }
  AROC <- numeric(np)
  for(i in 1:np){
    AROC[i] <- sum(u1 <= p[i])/n1
  }

  # Compute YI and associated threshold values
  difbb <- AROC -  p
  FPF <- mean(p[which(difbb == max(difbb))])
  YI <- max(difbb)

  pred0 <- predict(object$fit, newdata = newdata)
  if(object$est.cdf == "normal") {
    csf0_inv <- qnorm(1-FPF)
  } else {
    csf0_inv <- quantile(res0p, 1-FPF, type = 1)
  }
  
  thresholds <- pred0 + sigma0*csf0_inv
  #thresholds <- qnorm(1 - FPF, mean = pred0, sd = sigma0)

  res <- list()
  res$call <- match.call()
  res$newdata <- newdata
  res$thresholds <- thresholds
  res$YI  <- YI
  res$FPF <- FPF
  res
  
}