help with ROC curve analysis

[LightsEternity]

Hi there,

I know this is my first post, but tried looking on the forum and couldn't find a specific answer to my question.

I'm using SPSS 17 to do my data analysis.

I'm currently evaluating multiple different scoring systems for their ability to predict mortality using ROC curves.

Higher scores = worse

I've compared the scores using AUC's and determined which one is better.

My Question is I would like to find a specific cutoff/cutpoint for the best score.

Using SPSS, it appears that they automatically generate a set of different cut points allowing me to pick. It displays the cutoff points and their sensitivity and 1-specificity.

This may be a naive question, but is there anyway to determine the 95% CI for each cutpoints' sensitivity / 1-specificity

Thanks!

[biztek]

I dont know, if there is posibility to calculate 95% confidence interval for (sensitivity / 1-specificity). But if you want to find the optimal cut-off point (as I understand you want) you can use Youden index (J = max[SEi + SPi - 1]) or a method, that take in to consideration the cost of a fals-positive decision and the cost of a false negative decision, that was described by Zweig in 1993

[bukharin]

This recent paper proposes an alternative to Youden's index, and it also cites a number of papers describing methods for calculating CIs for Youden's index:
Liu X. Classification accuracy and cut point selection. Stat Med. 2012 Feb. 3.