A Weighted Quantile Regression Model--Based Multiple Imputation Method for Left-Censored and Partially Observed Biomarker Data

• Activity Number: 498
• Type: Contributed
• Date/Time: Wednesday, August 12, 2015 : 8:30 AM to 10:20 AM
• Sponsor: Section on Nonparametric Statistics
• Author(s): MinJae Lee* and Mohammad Rahbar and John Reveille and Michael Weisman and Michael M. Ward and Lianne Gensler and Matthew Brown
• Companies: The University of Texas Health Science Center and The University of Texas Health Science Center and The University of Texas Health Science Center and Cedars Sinai Medical Center and NIAMS/NIH and UC San Francisco and The University of Queensland Diamantina Institute
• Keywords: Limit of Detection ; Quantile Regression ; Multiple Imputation ; Left Censored Data ; Biomarker ; Genetics and Ankylosing Spondylitis (AS) Pathogenesis study

Abstract

In some patient-based cohort studies, the biomarker data can be censored subject to the detection limits of the assay and it is also possible that the data are not collected in certain time points during follow-up in the cohort. Inappropriate handling of these two issues could lead to biased estimation of regression parameters. We propose the Multiple Imputation (MI) based on censored quantile regression (CQR) that not only accounts for censoring, but also applies the inverse probability weighting technique to adjust for the partially collected data. Through a simulation study we evaluate the performance of our procedure in different scenarios. Our findings indicate that the proposed method performs better than MI based on un-weighted CQR and three different types of complete case analyses given moderate levels of censoring. We demonstrate application of our method to data from Genetics and Ankylosing Spondylitis (AS) Pathogenesis study with an aim to examine the longitudinal association between C-reactive protein (CRP) level and radiographic progression in AS when CRP levels were not collected in the first part of study period and also censored due to the detection limit.