How should race be best considered in understanding brain-behavior relationships? Implications for understanding onset of engagement in addictive behaviors and subsequent problems

Abstract

Developmental onset has been linked to addictive problems and severity, including for behavioral addictions like gambling disorder. Individual differences associated with race have been linked to addictive behaviors and disorders in complex manners. Race is understood as a multidimensional construct encompassing biological and social influences. This raises questions about how race should be conceptualized and modeled in brain-behavior relationship studies relevant to addictive behaviors and disorders. Here, we consider two recent publications involving early initiation of substance use (EISU). They derive potentially different conclusions, perhaps in part relating to how race and ethnicity were considered in analyses. Implications for behavioral addictions are explored.

Early engagement in potentially addictive behaviors has been linked to subsequent problems, including for gambling (Rahman et al., 2012). Race has also linked to onset of engagement in addictive behaviors and subsequent disorders in complex manners. Currently, there is considerable interest in understanding brain-behavior relationships, and there has been discussion regarding how best to consider race in understanding brain-behavior relationships (Dhamala et al., 2024; Osayande et al., 2024). Here, we use as an example two recent publications regarding early initiation of substance use (EISU) that have considered in analyses race in different manners, seemingly reaching different conclusions, as an opportunity to explore how race might best be considered in understanding brain-behavior relationships related to behavioral addictions.

Using data from the Adolescent Brain Cognitive Development (ABCD) study, Miller and colleagues (Miller et al., 2024) investigated relationships between neuroanatomy and EISU. Their findings revealed that individual variations in global brain structures, as well as subcortical and cortical anatomy during late childhood, were linked to EISU in adolescence. In separate ABCD analyses, Green and colleagues concluded that race is one of the strongest predictors of EISU and highlighted the limited predictive utility of neuroimaging (Green et al., 2024). In this commentary, we consider the role of race in the relationship between brain structures and EISU and discuss implications for understanding addictive behaviors more broadly.

In traditional regression analyses, race has often been treated as a confounding variable that warrants covariation to isolate relationships between brain features and EISU that are independent of race. This approach considers that race may influence both the brain and EISU but is not necessarily part of a causal pathway linking to EISU. However, treating race as a confounder has sparked debate, as race has more recently been considered a social construct that serves as a proxy for systemic inequities, environmental exposures, and sociocultural factors, rather than solely a factor linked to inherent biological measures (Flanagin, Frey, Christiansen, & Committee, 2021).

In two recent studies using longitudinal data from the ABCD study (Green et al., 2024; Miller et al., 2024), one investigated the relationship between brain structure and EISU in youth (Miller et al., 2024). Controlling for potential confounding effects due to covariates including baseline age and age-squared, sex, pubertal status, familial relationship, and MRI scanner model, their study used separate linear mixed effects models with family nested within study site as random intercepts to assess relationships between EISU and 297 imaging-derived phenotypes (IDPs) at baseline. EISU was defined as initiation of use of alcohol (i.e., sipping or full drinks of alcohol), nicotine, or cannabis, or any other illicit substance use prior to age 15 years. Their study found that differences in a total of 39 global and local brain features existed prior to any substance initiation. Using the same ABCD dataset, Green et al. (2024) evaluated the predictive power of 420 variables, including sociodemographic factors, self-reported behaviors, neurocognitive performance, hormones, parenting practices, and structural neuroimaging, through elastic net regression (Green et al., 2024). This study concluded that sociodemographic variables were the strongest statistical predictors of EISU, with the top three predictors identified as religious preference (Mormon or Jewish) and race. Notably, the authors highlighted the limited contribution of neuroimaging to prediction accuracy in this context. These two studies offered complementary but arguably contrasting perspectives on the predictors of substance initiation.

Of note, Miller et al. (2024) did not include race as covariates in analyses, as they communicated that race is likely associated with both neuroanatomical variability and EISU and may moderate associations between them (Miller et al., 2024). To further investigate the role of race in the reported associations between IDPs and EISU, we conducted additional analyses using data from ABCD Release 5.1, the most recent release currently available. We evaluated how incorporating race into the model influenced relationships. Specifically, replicating the linear mixed model (LMM) framework employed by Miller and colleagues (Miller et al., 2024), we (1) conducted the same analyses using the latest release of the ABCD data; (2) added race as a covariate to the model; (3) included the first 10 genetic principal components (PCs) as covariates to account for genetic ancestry and evaluated whether EISU was associated with any IDPs after adjusting for population stratification; (4) evaluated potential interaction effects between race and EISU on IDPs; and (5) assessed potential interaction effects between race and EISU on IDPs while controlling for genetic ancestry. The first 10 genetic PCs were extracted from the gen_y_pihat.csv file with variable names genetic_pc_1, genetic_pc_2, …, genetic_pc_10.

These analyses were motivated by considering that race-related confounding may manifest in different ways depending on how race may impact brain structure, EISU, or both. For instance, race may reflect systemic inequities that shape both brain development and EISU through discriminatory mechanisms. Alternatively, race may capture sociocultural influences, such as community norms or practices, that may influence both exposure and outcome pathways. Additionally, race may overlap with biological factors (Swilley-Martinez et al., 2023). To address this complexity, we included the top 10 genetic principal components (PCs) as proxy measures of biological determinants of race. This approach allows us to consider potential biological contributions of race related to genetics as opposed to social factors, enabling assessment of whether neuroanatomical differences related to EISU persist and whether race as a social construct moderates the brain-EISU relationship. Greater attention is warranted to the rationale for these proposed analyses, particularly in light of the multifaceted ways in which race intersects with biological, sociocultural, and systemic factors. These considerations are important for refining our understanding of the brain's role in substance use and addiction vulnerability.

Our data showed that race is independently associated with brain measures and EISU. Of 297 brain features assessed, 264 showed significant associations with race after false-discovery rate (FDR) adjustment, suggesting independent association between race and brain features. Furthermore, we also note that race was significantly associated with EISU (p-value < 0.001) with and without controlling for the covariates. In this sample, percentages of participant with EISU were 37.3% (n = 2,146), 23.3% (n = 333), 30.7% (n = 659), 26.67% (n = 60), and 35.38% (n = 398) for White, Black, Hispanic, Asian, and others respectively. These findings suggest potential confounding effects of race on the brain-EISU relationships. Table 1 summarizes additional results. First, without controlling for race as a covariate in the model, significant associations were identified between 28 IDPs and EISU (Table 1), similar to the findings reported by Miller and colleagues (Miller et al., 2024). Second, controlling for race as a covariate in the model, there were no significant associations between IDPs and EISU surviving multiple-comparison adjustment with an FDR of 0.05. Third, controlling for the top 10 genetic PCs as covariates in the model, there were similarly no significant associations between IDPs and EISU. Additionally, in this sample, race does not moderate relationships between brain features and EISU with or without adjusting for genetic ancestry (data not shown). To ensure model robustness, we assessed multicollinearity between self-reported race and the top 10 genetic PCs when examining potential interaction effects between race and EISU on IDPs, while controlling for genetic ancestry. We used the generalized variance inflation factor (GVIF) to evaluate multicollinearity. The adjusted GVIF^(1/(2*df)) for the categorical race variable was 1.807 ± 0.031 (range: 1.757–1.865), indicating an acceptable level of multicollinearity (Fox & Monette, 1992). The variance inflation factor (VIF) measures for the first genetic PC ranged from 6.542 to 7.575 across models (mean = 6.996), suggesting moderate collinearity. The maximum VIF for each of the remaining nine genetic PCs was below 3.146, indicating no substantial concerns regarding multicollinearity. The decision by Miller and colleagues to omit race from their analyses invites further consideration regarding the interpretation of neuroanatomical associations (Miller et al., 2024). In the absence of racial adjustment, such associations may be influenced by population stratification, social determinants of health, or structural inequities. When race is conceptualized as a proxy for social constructs and potentially situated within a causal pathway, its role in statistical models becomes complex, warranting careful consideration of whether it should be treated as an exposure, mediator, or otherwise (VanderWeele & Robinson, 2014). In such cases, it might be reasonable to consider race as an exposure. If race is an exposure risk for EISU, then neuroanatomical features might be reasonably treated as mediators. On the other hand, if race functions largely as a biological factor, then it might be most reasonable to consider it as a covariate along with other biological measures such as age, age squared, sex, and pubertal development stages. Not considering population stratification in statistical analyses may result in a phenomenon known as Simpson's paradox or reversal (Kopal, Uddin, & Bzdok, 2023), and this could lead to overestimation of the strengths of the associations between brain structure and other measures like, in this case, EISU. The lack of consensus on the most appropriate treatment of race in such analyses highlights the importance of thoughtful interpretation of analyses. Given that race was associated with both brain features and behavioral outcomes in our study, it is important to account for race-related variability when examining brain–EISU relationships. Drawing on practices from genetic research, where population structure is routinely controlled to avoid spurious findings, a similar level of rigor should be applied in neuroimaging analyses when possible. Sensitivity analyses with and without race as a covariate and moderation models can help assess whether brain–EISU associations differ across racial groups.

Table 1.

FDR-significant neuroanatomical associations from mixed-effect regressions comparing participants reporting any substance use initiation vs. substance-naïve participants

Variable	Primary			Adjust for race			Adjust for genetic PCs
	Beta	P	P FDR	Beta	P	P FDR	Beta	P	P FDR
rh_lateraloccipital_vol	0.073	1.54E−05	0.002	0.034	0.039	0.592	0.029	0.081	0.875
smri_vol_scs_wholeb	0.077	1.84E−05	0.002	0.026	0.126	0.758	0.016	0.367	0.875
smri_vol_scs_subcorticalgv	0.080	2.05E−05	0.002	0.041	0.026	0.516	0.031	0.097	0.875
lh_lateraloccipital_vol	0.073	2.58E−05	0.002	0.039	0.021	0.516	0.040	0.021	0.777
lh_lateraloccipital_thick	0.055	3.69E−05	0.002	0.029	0.028	0.516	0.028	0.038	0.875
smri_vol_cdk_total	0.076	3.75E−05	0.002	0.02	0.256	0.781	0.008	0.643	0.888
smri_area_cdk_total	0.070	1.11E−04	0.005	0.025	0.156	0.758	0.015	0.396	0.875
rh_posteriorcingulate_area	−0.062	1.96E−04	0.007	−0.053	0.001	0.147	−0.050	0.003	0.194
lh_cuneus_thick	0.064	4.00E−04	0.013	0.041	0.025	0.516	0.035	0.059	0.875
lh_lingual_thick	0.061	4.50E−04	0.013	0.029	0.089	0.758	0.025	0.144	0.875
rh_cuneus_thick	0.060	6.94E−04	0.019	0.034	0.054	0.666	0.027	0.127	0.875
lh_middletemporal_sulc	0.071	7.58E−04	0.019	0.043	0.040	0.592	0.035	0.096	0.875
rh_lateraloccipital_thick	0.044	8.47E−04	0.019	0.014	0.285	0.781	0.010	0.436	0.875
lh_superiorfrontal_thick	−0.046	9.21E−04	0.020	−0.016	0.240	0.781	−0.009	0.500	0.875
smri_vol_scs_intracranialv	0.056	1.17E−03	0.023	0.017	0.311	0.781	0.008	0.646	0.888
rh_parahippocampal_thick	0.064	1.43E−03	0.027	0.037	0.063	0.725	0.026	0.190	0.875
rh_lateraloccipital_area	0.050	1.56E−03	0.027	0.037	0.018	0.516	0.038	0.018	0.743
lh_pallidum_vol	0.055	1.64E−03	0.027	0.059	0.001	0.127	0.059	0.001	0.126
rh_caudalmiddlefrontal_thick	−0.052	1.72E−03	0.027	−0.036	0.030	0.516	−0.026	0.120	0.875
lh_parahippocampal_thick	0.063	2.04E−03	0.030	0.039	0.056	0.666	0.024	0.247	0.875
lh_medialorbitofrontal_thick	−0.060	2.13E−03	0.030	−0.038	0.053	0.666	−0.033	0.098	0.875
rh_frontalpole_thick	−0.062	2.31E−03	0.031	−0.044	0.028	0.516	−0.042	0.039	0.875
lh_inferiorparietal_area	−0.049	2.67E−03	0.035	−0.037	0.024	0.516	−0.026	0.115	0.875
rh_pallidum_vol	0.052	3.09E−03	0.038	0.063	3.15E−03	0.093	0.061	0.001	0.126
lh_parsopercularis_thick	−0.051	3.28E−03	0.039	−0.018	0.307	0.781	−0.016	0.362	0.875
rh_medialorbitofrontal_thick	−0.056	3.58E−03	0.041	−0.035	0.070	0.735	−0.031	0.109	0.875
rh_medialorbitofrontal_sulc	0.061	4.06E−03	0.045	0.038	0.076	0.756	0.033	0.123	0.875
rh_cuneus_vol	0.055	4.22E−03	0.045	0.032	0.090	0.758	0.029	0.137	0.875

Although the above example relates to EISU, similar considerations may hold for behavioral addictions like gambling disorder and other proposed conditions, like those related to problematic use of social media (Brand et al., 2022). Developing youth are being introduced to digital technologies at increasingly earlier ages, raising concerns about the potential impact (Christakis & Hale, 2025; Hutton et al., 2024). Relationships between screen media activity (SMA) and brain structure and function in developing youth have been increasingly investigated (Song et al., 2023; Zhao et al., 2022, 2023). In such studies, it is important to consider how best to evaluate the influences of population stratification on brain-SMA associations, assessing the extent to which they may potentially introduce confounding in the observed and reported relationships.

We also note that race has been implicated in complex ways in the use of digital media. For example, in the ABCD baseline data (youth aged 9–10 years), a machine-learning approach identified high-frequency and low-frequency groups based on screen media activity (Song et al., 2023). The high-frequency group differed from the low-frequency group on race, being more likely to be Black and of other race, and these factors were considered in understanding brain-behavior relationships related to resting-state brain activity, SMA, mental health concerns and other factors. Other data suggest that Black, as well as Hispanic, youth may be particularly vulnerable to engaging in high-frequency online activity, with a Pew Research study of 13- to 17-year-old adolescents finding that 54% of Black respondents and 55% of Hispanic respondents reported being online almost constantly, as compared to 38% of White youth (Anderson, Faverio, & Gottfried, 2023). This pattern may in part link to how youth engage with social media. For example, the same Pew Research report found that while 10% of White youth reported being on TikTok almost constantly, the percentages were 20% and 32% for Black and Hispanic youth, respectively (Anderson et al., 2023). Of note, the findings reported here are from the United States, with other jurisdictions possibly having different socio-cultural environments that may influence brain-behavior relationships in both similar and distinct manners. Although substance use and behavioral addictions may differ in social and neurobiological contexts, our prior work identified shared brain structure patterns linked to both early initiation of alcohol use and excessive screen media activity (Zhao et al., 2021, 2022). This suggests some potential common pathways, though distinct mechanisms likely also exist (Brand et al., 2025). Future research should investigate both shared and domain-specific processes, including how race may influence these relationships differently across addiction types.

Funding Statement

Funding sources: This study was supported by grants from the NIH (RF1 MH128614, R01 AA029611). The views presented in this commentary represent those of the authors and not necessarily those of the funding agencies.