Longitudinal ComBat: A Method for Harmonizing Longitudinal Multi-scanner Imaging Data
Kristin Linn, PhD University of Pennsylvania
Neuroimaging is a major underpinning of modern neuroscience research and the study of brain development, abnormality, and disease. Combining neuroimaging datasets from multiple sites and scanners can increase statistical power for detecting biological effects of interest. However, technical variation due to differences in scanner manufacturer, model, and acquisition protocols may bias estimation of these effects. Originally proposed to address batch effects in genomic data sets, ComBat has been shown to be effective at removing unwanted variation due to scanner in cross-sectional neuroimaging data. We propose an extension of the ComBat model for longitudinal data and demonstrate its performance using simulations and longitudinal cortical thickness data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) study. We demonstrate that longitudinal ComBat controls Type I error and has higher power for detecting changes in thickness over time compared to alternative methods such as naively applying cross-sectional ComBat to the longitudinal thickness trajectories.