Statistical methods to address heterogeneity in brain imaging and behavioral studies

Abstract

Heterogeneity in cognitive and behavioral traits poses a major challenge for imaginggenetics. Variation arises from multiple sources, including biological differences at genomic, neural, and environmental levels, as well as technical artifacts such as scanner effects in multi-site studies. These diverse sources of variability obscure associations and can lead to irreproducible findings if not appropriately modeled. These sources are becoming more and more common in the world of imaging genetics as focus is shifting to large and diverse consortia study designs. In the first project, we introduce AdjHE-RE, a method-of-moments SNP heritability estimator that incorporates scanner effects as a random effect. AdjHE-RE reduces nuisance dimensionality, yields unbiased estimates under simulation, and provides up to anorder of magnitude speed up over standard restricted maximum likelihood approaches. The second project leverages AdjHE-RE’s efficiency to perform phenome wide heritability analysis on functional brain phenotypes. SNP heritability estimates were consistently low especially when considering the functional connectivity of high consensus regions of interest (ROIs). The third project addressed biological sources of heterogeneity by modeling latent subsets through PPMx-common. PPMx-common is a Bayesian nonparametric model that jointly detects subgroups and estimates population-level associations. Application to the ABCD study revealed no reproducible global association between temperament and cognition, but identified three reproducible subtypes with distinct cognitive–temperamental profiles that were largely transdiagnostic. These projects introduces and demostrates different methodologies to handle heterogeneity in the ABCD dataset. These methods of dealing with heterogeneity aim to improve the interpretability, replicability, and scalability of imaging genetics and cognitive analyses.