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

Area of Science:

• Neuroscience
• Psychology
• Data Science

Background:

• Neuroimaging research often assumes equal variance across demographic groups.
• Differences in variance (heteroscedasticity) across racialized populations are understudied.
• Understanding variance is crucial for equitable AI and scientific interpretation.

Purpose of the Study:

• To investigate racialized heteroscedasticity in neuroimaging and behavioral data.
• To assess how variance differences impact predictive modeling performance and reliability.
• To highlight the importance of variance structure in diverse population studies.

Main Methods:

• Analysis of 4,736 participants from the Adolescent Brain Cognitive Development (ABCD) cohort.
• Examination of neuroimaging (e.g., functional imaging) and behavioral data variance.
• Simulation analyses to model the propagation of variance differences into predictions.

Main Results:

• Consistent patterns of racialized heteroscedasticity were observed across modalities and behaviors.
• Black participants showed greater variance in multiple neuroimaging and behavioral measures.
• Predictive models exhibited higher residual dispersion and prediction variance in Black participants.

Conclusions:

• Variance structure, not just central tendency, critically influences model performance and reliability.
• Racialized differences in variance can create disparities in prediction error and reliability.
• Addressing variance heterogeneity is essential for equitable and generalizable neuroscientific models.