|
Activity Number:
|
312
|
|
Type:
|
Topic Contributed
|
|
Date/Time:
|
Tuesday, August 4, 2009 : 10:30 AM to 12:20 PM
|
|
Sponsor:
|
Section on Statistics in Epidemiology
|
| Abstract - #305686 |
|
Title:
|
Estimating Diagnostic Accuracy of Linear Combination of Multiple Biomarkers While Accounting for Limits of Detection
|
|
Author(s):
|
Neil J. Perkins*+ and Enrique Schisterman
|
|
Companies:
|
National Institute of Child Health and Human Development and National Institutes of Health
|
|
Address:
|
6100 Executive blvd, 7b03, Rockville, MD, 20852,
|
|
Keywords:
|
linear combination ; limit of detection ; area under the curve ; ROC
|
|
Abstract:
|
The best linear combination (BLC) of multiple biomarkers is estimated parametrically using Su and Liu 1993 which maximizes the area under the receiver operating characteristic curve (AUC). BLC have also been found numerically via non and semiparametric devices. Emerging biomarkers are often subject to limits of detection (LOD), essentially left censoring the data. Common solutions, omitting or naïvely replacement, lead to negatively biased estimates of the AUC and maximum likelihood is heavily dependent on assumptions. We propose a comprehensive approach to estimating AUC of BLC taking into account bias, efficiency and robustness. Non and semiparametric estimates are compared to parametrically estimated AUC for BLC of two biomarkers under correctly and incorrectly specified distributional assumption and exemplified by polychlorinated biphenyls in women with and without endometriosis.
|
