1

I am performing a meta-analysis to compare the diagnostic accuracy of different modalities on the same population. I have constructed SROC curves and have calculated the AUC values in each case. Is there anyway to compare the AUC values to check for significant difference? Thank you.

Giuseppe Biondi-Zoccai
  • 2,244
  • 3
  • 19
  • 48
user86013
  • 11
  • 1
  • 1
    This would be a very low power test. The gold standard test would be to test for added value in a full modeling context using a likelihood ratio $\chi^2$ test. – Frank Harrell Aug 15 '15 at 15:56
  • Thank you for the prompt response. I am a beginner to this. Could you please elaborate on that and suggest some sources I can look at to learn more? – user86013 Aug 15 '15 at 16:12
  • 1
    I go into this in my book _Regression Modeling Strategies_ (2nd edition coming 2015-09-14) but this is covered in many guises throughout the statistical modeling literature. Try to phrase this through model specification in a multivariable regressionion model (such as the binary logistic model if your $Y$ is binary). Fit a model with the smaller set of features, then fit a model containing all the features and do the likelihood ratio test for the added value of the additional features. You can also test whether one set of features adds more to the model than another set of features. – Frank Harrell Aug 15 '15 at 16:19

0 Answers0