Statistical Adjustment of Population Heterogeneity and Generalizability in Neuroscience

Abstract

Advancing population neuroscience, recently large scale studies of human neuroimaging data have been designed to reflect the sociodemographic diversity of the target population. For example, the Adolescent Brain Cognitive Development (ABCD) Study has collected human brain, cognitive, behavioral, social, and emotional data of 11,875 children to characterize psychological and neurobiological development from pre-adolescence to young adulthood. However, existing statistical methods in neuroscience often ignore population heterogeneity and sample selection bias. This dissertation fills in the knowledge gap and aims to develop advanced statistical methods to address heterogeneity and generalizability in neuroscience research. We focus on association studies between brain functions and cognitive abilities in fMRI data across diverse populations.

Image-on-scalar regression has been a popular approach to modeling the association between brain activities and scalar characteristics in neuroimaging research. The associations could be heterogeneous across individuals in the population. In Chapter 2, we introduce the Latent Subgroup Identification in Image-on-Scalar Regression (LASIR), a novel method to identify subgroups of individuals from the population such that: 1) within each subgroup the brain activities have homogeneous associations with the clinical measures; 2) across subgroups the associations are heterogeneous; and 3) the group allocation depends on individual characteristics. We demonstrate that LASIR outperforms existing alternatives for subgroup identification of brain activation patterns with functional magnetic resonance imaging data via comprehensive simulations and applications to the ABCD study.

In Chapter 3, we extend the LASIR method in Chapter 2 with a fully Bayesian framework to model cortical surface fMRI data and capture both brain hemispheres. Our proposed Bayesian approach introduces a random effect at the functional network level to integrate data from both hemispheres within the model. We introduce the Relaxed Thresholded Gaussian Process prior to model sparse and spatially dependent neuroimaging data, enhancing computational efficiency through a relaxation parameter and latent variable structure to improve spatial feature selection and capture piecewise smoothness. We applied this method to the ABCD Study, with fMRI data from the stop signal task. This will allow us to examine associations between task-based brain imaging outcomes and behaviors assessed by the Child Behavior Checklist, and offer insights on studies of impulsive behaviors.

In Chapter 4, we mitigate selection bias in complex association studies by incorporating sample weights into image-on-scalar regression models. We improve multivariate image-on-scalar regression models in the adjustment of sample weights for model inference. To validate our approach, we have conducted extensive simulation studies and applied our method to the ABCD study’s working memory task N-Back fMRI data. Our findings show that, with sample weights adjusted, the image-on-scalar regression model generally detects fewer vertices compared to the weighted model, except for visual networks. When adjusting for control variables, the differences in vertex detection between the weighted and unweighted methods within the Parieto-Occipital and Sensorimotor Hand networks become less pronounced, while more substantial differences are observed in the Ventral Attention and Visual networks. We have demonstrated that our proposed weighted image-on-scalar regression improves the population generalizability of association studies in neuroimaging research.