Multi-site Normative Modeling of Diffusion Tensor Imaging Metrics Using Hierarchical Bayesian Regression

Abstract

Multi-site imaging studies can increase statistical power and improve the reproducibility and generalizability of findings, yet data often need to be harmonized. One alternative to data harmonization in the normative modeling setting is Hierarchical Bayesian Regression (HBR), which overcomes some of the weaknesses of data harmonization. Here, we test the utility of three model types, i.e., linear, polynomial and b-spline - within the normative modeling HBR framework - for multi-site normative modeling of diffusion tensor imaging (DTI) metrics of the brain’s white matter microstructure, across the lifespan. These models of age dependencies were fitted to cross-sectional data from over 1,300 healthy subjects (age range: 2–80 years), scanned at eight sites in diverse geographic locations. We found that the polynomial and b-spline fits were better suited for modeling relationships of DTI metrics to age, compared to the linear fit. To illustrate the method, we also apply it to detect microstructural brain differences in carriers of rare genetic copy number variants, noting how model complexity can impact findings.