A Semi-Parametric Roc Approach To Assessing Biomarkers Subject To A Measurement Error And Limit Of Detection

• Activity Number: 297
• Type: Contributed
• Date/Time: Tuesday, August 2, 2011 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Weijie Chen*+
• Companies: U.S. Food and Drug Administration
• Keywords: ROC analysis ; biomarker ; measurement error ; limit of detection ; maximum likelihood estimation ; semi-parametric model

Abstract

Quantitative biomarkers are emerging to discriminate between two clinically useful conditions. The measured biomarker levels are often corrupted with a random error, which leads to an ROC curve lower than that of the true levels of biomarkers. A solution for correcting such random errors is to repeat the measurements. Due to the limit of detection (LoD) of the instruments, the biomarkers are deemed to be immeasurable when the measured level is below some threshold. Parametric ROC methods have been proposed to analyze repeated measurements of biomarkers. However, parametric methods rely on a strong assumption that the data follows a normal distribution. We investigated a semi-parametric ROC approach that relies on a much looser assumption: the data are related to normal distributions by an implicit monotonic transformation. Maximum likelihood estimation is used to estimate the error-corrected ROC parameters and quantify the amount of measurement error. Extensive simulations show that our method is robust across a broad spectrum of experimental conditions including large measurement errors, substantial LoD, and deviations from the normal distribution.