0
$\begingroup$

I am performing a meta-analysis of a diagnostic score going from 0 to 10 based on 30 reports. For each report I have data on number of false/true positive/negative patients for one and in some cases two cutoff points (low and high). I plan to do meta analysis on sensitivity and specificity for these results.

For most reports I also have information on the ROC area of the complete score but for most reports I miss the confidence interval.

I can of course make ROC estimates from the cutoff points but would prefer using the ROC area that is provided from the complete score. Can I use the reported ROC areas and the total number of patients for each report as weights to pool the ROC areas?

Improve this question
$\endgroup$
5
  • $\begingroup$ You can compute confidence intervals from the ROC AUC and number of positives/negatives: doi.org/10.1002/… Would that answer (part of) your question? $\endgroup$
    – Calimo
    Commented Sep 20 at 8:17
  • $\begingroup$ Are you referring to equation (6) from that paper? That seems to require more info. $\endgroup$
    – Wolfgang
    Commented Sep 20 at 10:44
  • 1
    $\begingroup$ Ah, but this can be simplified. doi.org/10.1148/radiology.143.1.7063747 gives an equation for the SE of the AUC (underneath Table II) and using eq. (2), we only need the AUC, the number of cases, and the number of controls. $\endgroup$
    – Wolfgang
    Commented Sep 20 at 10:54
  • $\begingroup$ Thanks for the answers. I am not a mathematician so have problems understand the formulas. Can you show how to estimate the SE for this example, the AUC is 0.94, the number of diseased cases is 116 and number of non-diseased is 184. $\endgroup$
    – Roland
    Commented Sep 24 at 20:17
  • $\begingroup$ See: stats.stackexchange.com/q/291681/1934 $\endgroup$
    – Wolfgang
    Commented Oct 1 at 9:09

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.