Correcting AUC's for Measurement Error

• Activity Number: 511
• Type: Contributed
• Date/Time: Wednesday, August 1, 2012 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): Bernard Rosner*+ and Weiliang Qiu and Robert Glynn
• Companies: Harvard Medical School and Harvard Medical School and Harvard Medical School
• Keywords: Risk prediction ; measurement error ; probit transformation ; c-statistic

Abstract

Risk prediction models are used frequently in epidemiologic and clinical work. Risk scores are often derived from regression models based on risk factors, at least some of which are measured with error. It has been reported that ignoring measurement error can cause biased under-estimation of AUC, in many cases, resulting in misleading interpretation of the efficacy of a risk factor. Several methods have been proposed in the literature to correct AUC estimates for measurement error. Limitations of most existing methods include (a) assuming a normal distribution of risk scores; (b) assuming there is only one predictor and no other covariates in the model; and (c) not considering comparing two corrected AUCs derived from the same set of subjects. We propose an AUC correction method using the probit transformation to address the above three potential limitations. Both real data and simulation studies show good performance of the proposed AUC correction method in terms of bias and coverage probability. An example is given based on the Framingham Heart Study Risk Score.