Exploring bias due to below-limit-of-detection values in influenza vaccine antibody modeling: A case study and instructional guide for the CIVIC study.
Yang Ge, Andreas Handel, Philippe J Giabbanelli, Jennifer Lemacks, Tammy Greer, Pooja Raynee, Justin Bahl, Amanda L Skarlupka, Kevin K Dobbin, Ted M Ross, Ye Shen
Author Information
Yang Ge: College of Public Health, The University of Georgia, Athens 30606, GA, USA.
Andreas Handel: College of Public Health, The University of Georgia, Athens 30606, GA, USA; Center for the Ecology of Infectious Diseases, The University of Georgia, Athens 30606, GA, USA.
Philippe J Giabbanelli: Virginia Modeling, Analysis, and Simulation Center (VMASC), Old Dominion University, Suffolk 23435, VA, USA.
Jennifer Lemacks: School of Health Professions, The University of Southern Mississippi, Hattiesburg 39402, MS, USA.
Tammy Greer: School of Psychology, The University of Southern Mississippi, Hattiesburg 39402, MS, USA.
Pooja Raynee: School of Health Professions, The University of Southern Mississippi, Hattiesburg 39402, MS, USA.
Justin Bahl: College of Public Health, The University of Georgia, Athens 30606, GA, USA; Department of Infectious Diseases and Institute of Bioinformatics, The University of Georgia, Athens 30606, GA, USA.
Amanda L Skarlupka: National Cancer Institute, Bethesda 20814, MD, USA.
Kevin K Dobbin: School of Public Health, Augusta University, Augusta, GA 30912, USA.
Ted M Ross: Cleveland Clinic Florida, Port St. Lucie 34952, FL, USA; Center for Vaccines and Immunology, The University of Georgia, Athens 30606, GA, USA.
Ye Shen: College of Public Health, The University of Georgia, Athens 30606, GA, USA. Electronic address: yeshen@uga.edu.
In many laboratory assay datasets, missing values due to a limit of detection (LOD) are not uncommon. We observed this issue in our CIVIC-UGAFLUVAC hemagglutination inhibition assay (HAI) dataset. The standard imputation method recodes these values as either equal to the LOD or LOD/2. However, ignoring censoring can lead to falsely significant results in research. In this study, we explored the bias in modeling vaccine HAI titer increase. Moreover, we modified the titer increase modeling within the interval censoring framework to adjust for bias in parameter estimates. Our method provided less biased results compared to the standard imputation method. We anticipate that this study will serve as a case study and instructional guide for future vaccine research.
