1
$\begingroup$

I would like to calculate the variance of the AUC of readers (for each reader and averaged results) giving a score(1-5) to specific areas (1-5) of cases using two different modalities.

Test data:

df <-data.frame(
              modality=rep(c("A","B"),each=25,2),
              reader=rep(c("reader1","reader2"),each=50),
              case= rep(c("1","2","3","4","5"),each=5,2),
              area=rep(c("a1","a2","a3","a4","a5"),20),
              score=sample(1:5,100, replace=TRUE),
              disease=rep(sample(0:1,25, replace=TRUE),4),
              stringsAsFactors=FALSE)

> df
    modality  reader case  area score disease
1          A reader1    1   a1     5       0
2          A reader1    1   a2     3       0
3          A reader1    1   a3     4       1
4          A reader1    1   a4     2       0
5          A reader1    1   a5     5       1
6          A reader1    2   a1     1       1
7          A reader1    2   a2     3       1

Suggested Method (from 1):

DBM refers to (2)

Davison and Hinkley refers to (3)

BWC refers to (4)

Manual from [1]

What I know

I found a solution how to do the first step (two-way bootstrap): Bootstrapping hierarchical/multilevel data (resampling clusters)

# dplyr
library(dplyr)

replicate(100, {
  cluster_sample <- data.frame(case= sample(df$case, replace = TRUE))
  dat_sample <- df %>% inner_join(cluster_sample, by = "case")
  dat_sample
})

I know how to calculate the AUC for reader 1 and 2 and the average AUC

library(pROC)
roc1 <- roc(df[which(df$reader=="reader1"&df$modality=="A"),]$disease, df[which(df$reader=="reader1"&df$modality=="A"),]$score)
roc1 <- roc(df[which(df$reader=="reader2"&df$modality=="A"),]$disease, df[which(df$reader=="reader2"&df$modality=="A"),]$score)
rocm <- multiclass.roc(df[which(df$modality=="A"),]$disease, df[which(df$modality=="A"),]$score)

What I cannot replicate

I do not know how to apply the second part of the method stated above to get the correct variances (yellow marked part)

Reference

  1. Gallas BD, Bandos A, Samuelson FW, Wagner RF. A Framework for Random-Effects ROC Analysis: Biases with the Bootstrap and Other Variance Estimators. Communications in Statistics - Theory and Methods. 2009 Jul 23;38(15):2586–603. https://www.tandfonline.com/doi/abs/10.1080/03610920802610084

  2. Dorfman DD, Berbaum KS, Metz CE. Receiver operating characteristic rating analysis. Generalization to the population of readers and patients with the jackknife method. Invest Radiol. 1992 Sep;27(9):723–31.

  3. Davidson, A. C., Hinkley, D. V. (1997). Bootstrap methods and their applications. Cambridge University press.

  4. Beiden SV, Wagner RF, Campbell G. Components-of-variance models and multiple-bootstrap experiments: An alternative method for random-effects, receiver operating characteristic analysis. Academic Radiology. 2000 May;7(5):341–9. https://www.academicradiology.org/article/S1076-6332(00)80008-2/pdf

$\endgroup$
  • 1
    $\begingroup$ This problem is being solved with the help of Brandon D. Gallas and his R-package iMRMC, the implementation is discussed at github.com/DIDSR/iMRMC/issues/147 $\endgroup$ – captcoma Feb 3 '19 at 17:49

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.