Department of Biostatistics Seminar/Workshop Series

Detection limit methods and their performance in regression on biomarker measurement

Huiyun (William) Wu, PhD

Associate in Biostatistics, Department of Biostatistics, Vanderbilt-Ingram Cancer Center/Biostatistics Shared Resource

Wednesday, September 14, 1:30-2:30pm, MRBIII Conference Room 1220

A joint work with Cindy Chen and Tatsuki Koyama

BACKGROUND: Biomarkers have the potential to improve understanding of disease diagnosis and prognosis. Biomarker levels that fall below the assay detection limits (DLs), however, compromise the appli¬cation of biomarkers in research and practice.

METHODS: Most existing methods to handle non-detections focus on a scenario in which the response variable is subject to detection limit; only a few methods consider explanatory variables when dealing with detection limits. We propose a Bayesian approach for generalized linear model with explanatory variables subject to lower detection limit and compare our method to other four commonly used methods, i.e., single replacement, deletion, multiple regression, and regression in order statistics.

RESULTS: In the simulation studies, we compared the proposed Bayesian approach to the four methods by evaluating bias, standard error, root mean of square error, 95%CI coverage percentage, and type I error rate. We also applied the Bayesian and other four methods in a real study, in which a panel of cytokine biomarkers was studied for their association with acute lung injury (ALI).

CONCLUSIONS: The proposed Bayesian approach outperformed all other methods at most criteria. We also found that IL8 was associated with a moderate increase in risk for ALI in the model based on the Bayesian approach