Jannat Kaur Thind

Jannat Thind*, Joseph Aina, Jalen Grayson, Fahmi Khalifa, Raimundo Rodriguez, Matthew Rosenblatt, Dustin Scheinost, Christopher Fields* Sociodemographic heteroscedasticity refers to unequal variance in outcome or prediction error across social groups. In psychiatric prediction models, this, when stratified by race, can reflect exposure to structural adversity that models fail to capture. We examined whether brain-based models for adolescent substance use show evidence of racialized heteroscedasticity, focusing on prediction uncertainty and residual variance.  Using ABCD baseline neuroimaging data (ages 9–10), we trained connectome-based predictive models (CPM) on resting-state fMRI and structural connectivity features to predict lifetime substance use summaries (alcohol, nicotine, cannabis, and overall SUD) from up to seven follow-up sessions (n=4,000–9,000). Models were cross-validated using 10-fold splits, with additional evaluation using leave-three-site-clusters-out. We computed correlations between predicted and observed values and tested for group differences in outcome variance and prediction interval width by race. Outcome measures were log-transformed (log[1+x]) to reduce skew. Covariates included age, gender, parental income, and head motion.    Prediction accuracy fell within expected ranges for psychiatric models (e.g., rs-fMRI: r² = 0.08–0.13), yet prediction intervals were significantly wider for Black participants (Levene's test on 120-iteration bootstrapped CPM prediction variance [Black vs White]; SUD: F = 4.98, p = 0.001; Alcohol: F = 8.97, p < 0.001; Nicotine: F = 6.47, p < 0.001; Cannabis: F = 7.15, p < 0.001). Observed and predicted outcome variance was consistently higher for Black and Native American youth than White youth, even after adjusting for covariates. These patterns are held across imaging modalities and behavioral outcomes.  These disparities in prediction variance may scale when applied across clinical settings, increasing the reproduction of structural bias. Attending to group-level variance is necessary for equitable model development.