Biomarker Combination with Partially Observed Gold Standard

• Activity Number: 523
• Type: Topic Contributed
• Date/Time: Wednesday, August 12, 2015 : 10:30 AM to 12:20 PM
• Sponsor: Mental Health Statistics Section
• Author(s): Danping Liu* and Ashok Chaurasia and Zheyu Wang
• Companies: NIH and Eunice Kennedy Shriver National Institute of Child Health and Human Development and The Johns Hopkins University
• Keywords: Biomarker combination ; fetal growth ; imputation ; inverse probability weighting ; missing gold standard ; verification bias

Abstract

In disease prediction, a combination of multiple biomarkers often improves the diagnostic accuracy. Existing methods, such as logistic regression or AUC maximization, both require fully observed disease outcome. However, the true disease condition may be missing in practice because it is expensive or harmful to ascertain. In estimating the ROC curve, it is well-known that the complete-case analysis often leads to biased estimator, known as "verification bias". It is unclear how verification bias affects the estimation of the biomarker combination. This paper is motivated from the Scandinavian Infant Growth Project, in which a cohort of pregnant women underwent multiple ultrasound examinations during pregnancy, but only a proportion of the infants received further follow-up after birth. Our focus is to predict overweight infants at the one-year follow-up using the ultrasound measurements. In this paper, we propose several approaches for biomarker combination that can handle missing disease status, based on reweighting and imputation techniques. These estimation procedures are compared through empirical bias calculation, simulation studies, and analysis of the fetal growth data.