Diagnostic Likelihood Ratios: Hypothesis Testing and 95% CI possible?

[Jack Tay]

I am evaluating a scenario where a standard test is done (S), followed by an advanced test read in a standard or original way (O) and then re-read by a dedicated reader (R), all evaluated against a gold standard (G). I would like to determine the incremental value of O followed by R and would like to seek some advice regarding the best methodology using Stata.

I have generated the usual sensitivity, specificity and ROC parameters for each step in the series together with LR+ and LR- using roccomp (which compares AUC only) and roctab:

First I generated S using a logistic regression of several factors then using predict Pr(G)_S to store the values. Then I added O to the logistic regression and repredicted to get Pr(G)_S+O. Next I added R to the regression to get Pr(G)_S+O+R. This culminated in 3 models. Any comments on such a method to determine the incremental value of a series of tests?

I am interested in comparing the LR+ and LR- of these 3 models rather than their AUC. How would I do that in Stata and is it possible to get 95% CI?

Appreciate any guidance. Thank you in advance.

[Jack Tay]

Hi Statalist,

Am still hoping for comments but just wanted to post an update on progress that I thought may help those new to this: 95% CI for LR+ and LR- can be obtained using -diagt- which is an unofficial Stata command created by Paul Seed based on -diagtest- by Aurelio Tobias. You may find the details here:

http://aureliotobias.weebly.com/uplo...iagt_stb59.pdf

The program can be found by search through Stata itself to obtain the latest version. LRs can then be compared based on 95% CI width though it is not a direct test.

My own questions regarding the methodology of my approach remain, plus whether hypothesis testing of LRs remains possible or meaningful.