Racialized Heteroscedasticity in Connectome-Based Predictions of Brain-Behavior Relationships

Background: Racialized heteroscedasticity, characterized by unequal variance in neuroimaging data across racial groups, may challenge the assumptions underlying brain-behavior predictive models. These models often assume homoscedasticity, potentially leading to biased predictions, particularly when applied across diverse populations. This study investigates the presence of heteroscedasticity in functional connectivity data from the Adolescent Brain Cognitive Development (ABCD) study and examines its implications for predictive modeling using resting-state fMRI (rs-fMRI) data.Methods: We analyzed resting-state functional connectivity from 5,084 participants in the ABCD study, comparing variance in rs-fMRI connectivity between Black and White participants across 35,912 non-redundant connectivity edges. Levene’s test, adjusted for demographic and socioeconomic covariates, was used to assess differences in variance, with False Discovery Rate (FDR) correction applied. Connectome-Based Predictive Modeling (CPM) was employed to predict performance across 15 behavioral and cognitive tasks, and 95% confidence intervals were generated for prediction intervals via bootstrapping over 1,000 iterations.Results: Significant differences in variance were observed in 10.5% (n = 3,768) of connectivity edges, with 98.5% (n = 3,712) showing greater variance in Black participants. CPM analysis revealed significantly wider prediction intervals for Black participants across 10 of 15 tasks, particularly in measures of working memory, cognitive flexibility, and inhibitory control. These results suggest that increased variance in functional connectivity in Black participants contributes to greater prediction uncertainty in brain-behavior models.Conclusion: Our findings underscore the importance of accounting for racialized heteroscedasticity in neuroimaging-based predictive models. Failing to account for differences in variance may exacerbate health disparities and lead to less reliable predictions for Black participants. Addressing these disparities is essential for the development of more equitable predictive models in neuroscience and psychiatry.