Analysis of Longitudinal Metabolite Data with Substantial Missingness and Batch Effects

• Activity Number: 575 - Statistical Methods for Batch Effect Correction and Cell Type Deconvolution
• Type: Contributed
• Date/Time: Wednesday, July 31, 2019 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics in Genomics and Genetics
• Author(s): Evan Sticca* and Audrey E Hendricks and Stephanie P Gilley and K Michael Hambidge and Nancy F Krebs and Sarah J Borengasser
• Companies: University of Colorado Anschutz Medical Campus and University of Colorado Denver and University of Colorado Anschutz Medical Campus and University of Colorado Anschutz Medical Campus and University of Colorado Anschutz Medical Campus and University of Colorado Anschutz Medical Campus
• Keywords: Metabolomics; batch effects; pregnancy; nutrition; missing data; longitudinal

Abstract

High-throughput metabolomics allows for larger sample sizes often needed for longitudinal data analyses. Study constraints can make optimal study design challenging. We present strategies for a longitudinal metabolomics study with missing data and batch effects as part of a secondary analysis of the Women First (WF) trial. In WF, women randomly received 1 of 3 nutritional interventions through delivery: ? 3 months before conception (Arm 1, N=47), 11 weeks of gestation (Arm 2, N=48), or none (Arm 3, N=39). 134 Guatemalan women had 27 metabolites assayed at baseline, 12 weeks (Arms 1 and 2 only), and 34 weeks. Samples for arm and time combinations were assayed in two distinct batches. To address batch effects, data were normalized to baseline within each batch. To control for missing data, we reparametrized arm to supplement status enabling analysis at all timepoints. We detect 1 time and supplement interaction, 21 metabolites that change over time, and 2 metabolites that change by intervention. Here, we show how normalization and reparameterization can be used to correct for batch effects and missingness when optimal study design is not possible.