Socioeconomic status (SES) has been linked to brain-based markers, but most studies rely on conventional statistical methods that overlook the complexity and inherent multicollinearity of the brain. We trained elastic net models to predict SES from multimodal neuroimaging data: diffusion tensor imaging (DTI), structural MRI (sMRI), and resting-state functional connectivity (RSFC) data. Neural features independently predicted SES, and demographic information only minimally enhanced performance. The income and multimodal models performed best; accordingly, the best-performing primary model predicted income using multimodal data, achieving AUCs of 0.75 (test) and 0.811 (train) without demographic information and 0.779 (test) and 0.836 (train) with demographic information. The performance of the secondary multimodal models for predicting income had a positive relationship with income disparity; expectedly, the best performing model distinguished between children from the top and bottom ~ 10-20% of income brackets, reaching AUCs of 0.81 (test) and 0.969 (train) without demographic information and 0.863 (test) and 0.986 (train) with demographic information. Among the modalities, DTI was the most discriminative, followed by sMRI. Globally distributed along with executive functioning (EF) and language features were the most discriminative. Multimodal neuroimaging can predict SES, especially income, even without demographic data, and the most discriminative features tended to be measurements of white matter integrity and organization; more globally distributed than isolated to specific regions; and linked to cognitive control and language.
Kavari Hercules (Sat,) studied this question.