Racialized Heteroscedasticity in Neuroimaging Features, Behavior Measures, and Neuroimaging-Based Predictive Models.

Neuroimaging studies rarely test whether the variance structure is equivalent across population subgroups. Here, in 4,736 participants from the Adolescent Brain Cognitive Development (ABCD) cohort, we examine racialized heteroscedasticity (i.e., differences in variance across racialized groups) in neuroimaging and behavioral data and test how these differences in variance propagate into predictive modeling. Across neuroimaging modalities, behaviors, and predictive frameworks, variance differences exhibited consistent patterns, indicating that variance structure is a stable property across domains within the dataset. Simulation analyses demonstrated that such differences directly induce subgroup disparities in prediction error and reliability, even in the absence of mean differences. Across neuroimaging modalities, multiple measures demonstrated greater variance in Black participants, particularly in functional imaging modalities. Similar variance patterns were observed in behavioral measures, and predictive models exhibited greater residual dispersion and prediction variance in Black participants even when overall performance metrics were comparable. These findings position variance structure, rather than central tendency, as a critical determinant of model performance, generalizability, and reliability across diverse populations.