🧬 biology

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

This study of 4,736 ABCD cohort participants reveals that racialized heteroscedasticity, characterized by consistently higher variance in neuroimaging and behavioral data for Black participants, systematically induces disparities in prediction error and model reliability even when mean differences are absent, highlighting variance structure as a critical determinant of model generalizability across diverse populations.

Original authors: Christopher Fields, Matthew Rosenblatt, Joseph Aina, Jannat Thind, Annie Harper, Chyrell Bellamy, Xin Zhou, Alexandra Potter, Hugh Garavan, Nicholas Allgaier, Micah Johnson, Raimundo Rodriguez, Fahmi

Published 2026-06-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Christopher Fields, Matthew Rosenblatt, Joseph Aina, Jannat Thind, Annie Harper, Chyrell Bellamy, Xin Zhou, Alexandra Potter, Hugh Garavan, Nicholas Allgaier, Micah Johnson, Raimundo Rodriguez, Fahmi Khalifa, Deanna Barch, Dustin Scheinost

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The Big Idea: It's Not Just About the Average, It's About the Spread

Imagine you are trying to predict how fast a car will go based on the size of its engine. Usually, scientists look at the average speed of different groups of cars to see if there are differences.

This paper argues that scientists have been ignoring a crucial part of the story: how much the speeds vary within each group.

The researchers looked at brain scans and behavior tests from nearly 5,000 teenagers (mostly White and Black participants) in a massive study called the ABCD Study. They discovered a phenomenon they call "Racialized Heteroscedasticity."

In plain English, this means: The "spread" or "variability" of data is different for Black participants compared to White participants.

The Analogy: The Two Archery Teams

Imagine two archery teams, Team A and Team B, shooting arrows at a target.

Team A (White participants): Their arrows land in a tight cluster right around the bullseye. They are very consistent.
Team B (Black participants): Their arrows are also centered around the bullseye (the average is the same!), but they are scattered all over the target. Some hit the bullseye, some hit the edge, and some miss the target entirely.

The Problem: If you build a computer program (a predictive model) to guess where the next arrow will land based on the team's training, the program will work very well for Team A because their arrows are predictable. But for Team B, the program will be much more "confused." Even if the program guesses the average spot correctly, it will be wrong about where the individual arrow actually lands much more often for Team B.

The paper found that in brain imaging and behavior tests, Black participants often showed this "scattered" pattern (higher variance), while White participants showed a "tight cluster" pattern (lower variance).

What They Found

The Brain Scans: When looking at how different parts of the brain talk to each other (functional imaging), the "noise" or variability was significantly higher for Black participants. It was like the brain signals were more "jittery" or diverse in this group.
The Behavior Tests: The same thing happened with tests of personality, attention, and mood. Black participants showed a wider range of scores.
The Prediction Models: When the researchers built computer models to predict behavior based on brain scans, the models worked differently for the two groups.
- For the group with the "tight cluster" (lower variance), the predictions were stable and reliable.
- For the group with the "scattered arrows" (higher variance), the predictions were all over the place. The model's error rate was higher, and its confidence was lower, even if the overall math looked okay on paper.

Why Does This Happen?

The paper is careful to say this is not because Black brains are "different" or "broken" in a biological sense.

Instead, they suggest this variability comes from life experiences. Think of it like this:

If you live in a very stable, predictable neighborhood with lots of resources, your daily life might follow a very similar pattern to your neighbors (low variance).
If you live in a neighborhood where you face more stress, discrimination, or unpredictable challenges, your daily experiences might vary much more wildly from day to day or from person to person (high variance).

The paper calls this "Racialized Heteroscedasticity" to emphasize that these differences in "spread" are likely caused by how society treats different groups, not by their race itself.

The Takeaway

The main lesson of this paper is that variability matters just as much as the average.

If scientists only look at the "average" brain or behavior, they might miss the fact that their computer models are actually much less reliable for certain groups. It's like a weather app that predicts "sunny" for everyone. It might be right for a city with stable weather, but if you live in a city with wild, unpredictable storms, that "sunny" prediction is useless because the variability is so high.

The authors conclude that to make fair and accurate brain-science tools for everyone, researchers need to stop assuming that all groups have the same amount of "spread" in their data. They need to account for this extra variability to ensure their models work well for everyone, not just the groups with the most consistent data.

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

Problem Statement
Neuroimaging research frequently assumes that the variance structure of brain and behavioral data is equivalent across population subgroups, a condition known as homoscedasticity. However, this assumption is rarely tested. The authors posit that "racialized heteroscedasticity"—systematic differences in variance between racialized groups—may exist due to unequal exposure to stress, environmental adversity, and social resources. If unaddressed, violations of homoscedasticity can lead to biased predictions, reduced model reliability, and obscured meaningful variability in higher-variance populations. This study investigates whether variance structures differ between Black and White participants in the Adolescent Brain Cognitive Development (ABCD) cohort and how these differences propagate into brain-behavior predictive modeling.

Methodology
The study utilized baseline data from 4,736 participants (3,892 White, 844 Black) from the ABCD Study. The analysis proceeded in four stages:

Variance Quantification in Neuroimaging: The authors tested for variance differences across six neuroimaging modalities: resting-state fMRI (rsfMRI), task-based fMRI (Monetary Incentive Delay, Stop-Signal Task, and EN-Back), diffusion-weighted imaging (DWI), and T1-weighted structural MRI (SMRI). Using Levene's test, they compared the variance of 35,778 functional connectivity edges and 151 structural features between racial groups. Analyses were conducted both uncorrected and with conservative corrections (covariate regression for age, sex, income, and motion; site correction via ComBat; and balanced sample sizes).
Variance Quantification in Behavioral Measures: Variance differences were assessed across 36 behavioral measures spanning cognition, personality, emotional regulation, and behavioral health, again using Levene's test with FDR correction.
Predictive Modeling Analysis: Connectome-based Predictive Modeling (CPM) was employed to predict the 36 behavioral measures from whole-brain rsfMRI connectivity using five-fold cross-validation (stratified by family to prevent data leakage). The authors evaluated prediction performance (correlation and Mean Squared Error) and, crucially, the variance of predicted values ( $\Delta$ Variance) between racial groups. Complementary models (Kernel Ridge Regression, Random Forest, and Artificial Neural Networks) were also tested in supplementary analyses.
Simulation: To establish a mechanistic link, the authors simulated linear regression scenarios where two groups shared identical mean relationships between predictors ( $X$ ) and outcomes ( $Y$ ) but differed in variance. This isolated the effect of variance structure on prediction accuracy (MSE and correlation) without confounding variables.

Key Results

Neuroimaging Variance: Across all modalities, 12.44% to 46.36% of features showed significant variance differences between Black and White participants. Functional imaging modalities (rsfMRI and task-based fMRI) consistently exhibited a higher proportion of features with greater variance in Black participants (ranging from 62.37% to 91.69% of significant features). In contrast, DWI showed greater variance in White participants for most features, while SMRI showed no significant difference in the number of high-variance features between groups. These patterns persisted even after applying rigorous corrections for covariates, site effects, and sample imbalance, though the total number of significant features decreased.
Behavioral Variance: Of the 36 behavioral measures, 30 showed significant variance differences after FDR correction. Notably, 27 of these measures exhibited greater variance in Black participants, including Child Behavior Checklist scores (e.g., rule-breaking, aggression) and personality domains (e.g., sensation-seeking).
Predictive Model Performance: In CPM, 20 behavioral measures yielded modest but significant predictions ( $r > 0.1$ ). While overall performance metrics were comparable, 15 of these 20 models showed higher correlation between observed and predicted values for Black participants, yet 18 showed higher Mean Squared Error (MSE). This conflict suggests that higher correlation does not necessarily equate to lower error in the presence of heteroscedasticity.
Prediction Dispersion: The variance of predicted values was significantly higher for Black participants in 16 of the 20 significant models. Two models (rule-breaking and sensation-seeking) showed FDR-corrected significant differences in prediction variance, with Black participants exhibiting greater dispersion.
Simulation Findings: The simulations demonstrated that increasing the variance of the outcome variable ( $Y$ ) in one group directly increased the MSE for that group, even when the underlying relationship between $X$ and $Y$ was identical. Increasing the variance of the predictor ( $X$ ) increased the correlation between observed and predicted values. These results confirm that variance structure alone can generate subgroup disparities in prediction accuracy and create conflicting performance metrics.

Significance and Claims
The authors claim that their findings expose a critical, untested assumption in population neuroimaging: that variance structure is sufficiently similar across subgroups to allow for standard inferential and predictive workflows. The study demonstrates that:

Variance Structure as a Determinant: Variance structure, rather than just central tendency (mean differences), is a stable and reproducible property that significantly influences model performance, generalizability, and reliability.
Mechanism of Bias: The paper proposes that racialized heteroscedasticity provides a mechanistic explanation for why neuroimaging-based predictive models often perform worse for minoritized populations, even when training data is balanced. The increased variance in Black participants leads to higher prediction error (MSE) and greater uncertainty, which standard metrics may obscure.
Equity and Clinical Translation: The authors argue that ignoring these variance differences can exacerbate health disparities by deploying unreliable models in clinical settings. They emphasize that "racialized heteroscedasticity" reflects socially structured exposures shaping variability, not inherent biological differences.
Methodological Imperative: The study concludes that variance structure should be regularly examined across subgroups as a primary analytical step, rather than a secondary assumption. They suggest that future work must move beyond balancing means to addressing variance asymmetry to develop more equitable and accurate predictive models.

The authors remain modest regarding the generalizability of the magnitude of these differences to all populations or developmental stages, noting that the study is limited to a single cohort and binary racial categorization. However, they assert that the existence of such variance differences is a robust finding that necessitates a shift in how neuroimaging data is interpreted and modeled.