A Simulation Study to Compare Modeling Methods for Analyzing Biomarker Data Subject to a Limit of Detection (LOD)

• Activity Number: 35
• Type: Contributed
• Date/Time: Sunday, August 4, 2013 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Charles Rose*+ and Ryan E. Wiegand
• Companies: CDC and Centers for Disease Control and Prevention
• Keywords: Limit of Detection (LOD) ; Censoring ; Biomarker Data ; Gaussian pdf-cdf Model ; Mixture Model

Abstract

A potential challenge in analyzing biomarker data arises when the data are subject to a limit of detection (LOD). Here LOD is defined as the limit at which the assay values cannot be reliably detected. Several methods have been commonly used to analyze biomarker data subject to LOD. These methods include imputing a value for LOD (e.g, ½ LOD) and modeling using a Gaussian distribution, using LOD as the cut-point to dichotomize and modeling using logistic regression, and logistic-Gaussian or pdf-cdf Gaussian mixture models. The logistic-Gaussian models the LOD values using logistic regression and the Gaussian pdf models values > LOD. The Gaussian pdf-cdf mixture distribution models values =LOD and >LOD using the cdf and pdf, respectively. We conduct a simulation study, assuming cross-sectional data to compare treatment groups, to compare the methods by assessing bias and statistical power under a range of censoring percent and LOD values. Our simulation illustrates that the pdf-cdf Gaussian mixture model provides the best power and the least bias when =50% of the outcomes are LOD values. Results illustrate that if the percent =LOD reaches 70% all methods perform comparably.