547 – Contributed Oral Poster Presentations: Biopharmaceutical Section
Flipped-data Survival Analysis for Metabolomics Data with Non-detects
Eric Siegel
University of Arkansas for Medical Sciences
Metabolomics data sets typically have a large number of "non-detects": features that are left-censored at the limit of detection (LOD). To analyze data with non-detects, modern methods such as maximum-likelihood estimation and multiple imputation have been deployed in fields as diverse as AIDS research and environmental monitoring. However, such modern methods have yet to be deployed in the metabolomics research arena, where currently prevailing practices for analyzing data with non-detects instead seem to be (A) exclude the non-detects from the analysis, or (B) assign to the non-detects a value such as 0.5 or 1.0 times the LOD. Tekwe et al. applied survival-analysis methods to left-censored proteomics data, but their methods were limited to accelerated failure-time models in which parametric distributions were assumed. Here, I apply to left-censored metabolomics data the non-parametric survival-analysis method of Helsel, in which data are 'flipped' by subtracting them from a suitably large number, then analyzed using Kaplan-Meier methods and the log-rank test. I use simulation to compare Helsel's method both to currently prevailing practices and to parametric survival regression.