Integrative analysis of psychosocial, chronotype, and environmental predictors of aggressive behavior in Adolescents: Insights from machine learning.

Integrative Analysis of Psychosocial, Chronotype, and Environmental Predictors of Aggressive Behavior in Adolescents: Insights from Machine Learning - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice J Psychiatr Res . Author manuscript; available in PMC: 2026 Jun 4. Published in final edited form as: J Psychiatr Res. 2025 Jun 4;189:91–103. doi: 10.1016/j.jpsychires.2025.06.005 Integrative Analysis of Psychosocial, Chronotype, and Environmental Predictors of Aggressive Behavior in Adolescents: Insights from Machine Learning Tam Nguyen-Louie 1 , Michael J McCarthy 2 , Emil F Coccaro 3 , Alejandro D Meruelo 4, * Author information Article notes Copyright and License information 1 University of California San Diego, 9500 Gilman Drive, La Jolla, CA 92093, USA. 2 University of California, San Diego, VA San Diego Healthcare System, 3350 La Jolla Village Dr, San Diego, CA 92161 USA 3 Clinical Neuroscience and Psychotherapeutics Research Unit, Department of Psychiatry and Behavioral Health, The Ohio State University Wexner Medical Center, Columbus, OH, USA. 4 University of California San Diego, 9500 Gilman Drive, La Jolla, CA 92093, USA. * Corresponding Author ameruelo@health.ucsd.edu Author contributions (CRediT Statement) Tam Nguyen-Louie: Conceptualization, Methodology, Writing – Review & Editing Michael J. McCarthy : Methodology, Writing – Review & Editing Emil F. Coccaro: Writing – Review & Editing Alejandro D. Meruelo : Conceptualization, Formal Analysis, Visualization, Writing – Original Draft, Writing – Review & Editing, Funding Acquisition Issue date 2025 Sep. PMC Copyright notice PMCID: PMC12258394  NIHMSID: NIHMS2088810  PMID: 40483849 The publisher's version of this article is available at J Psychiatr Res Abstract Aggressive behavior in adolescents and young adults is a significant public health concern associated with adverse educational, social, and mental health outcomes. This study aimed to identify key predictors of aggression using a cross-sectional dataset from a large, longitudinal U.S. cohort. The outcome was self-reported aggressive behavior, and predictors spanned demographic, psychosocial, behavioral, and contextual domains, including adverse life events, impulsivity, family conflict, peer and school environments, and chronotype. Multiple models were evaluated, including linear regression, a hypertuned random forest, a tuned gradient boosting machine (GBM), XGBoost, and an ensemble model combining random forest and GBM predictions. All models were trained using five-fold cross-validation across five multiply imputed datasets. Linear regression achieved the highest predictive accuracy (r = 0.313; MSE = 40.76), followed closely by the random forest (r = 0.311; MSE = 40.71). The ensemble and GBM models showed slightly lower performance. Across models, key predictors included adverse life events, delayed chronotype, peer network health, family cohesion, and normalized household income. These findings underscore the contribution of environmental and psychological stressors to adolescent aggression, particularly the buffering role of cohesive peer and family relationships. Despite similar predictive accuracy across models, machine learning methods offered advantages for variable importance ranking and interaction discovery. Results highlight the utility of integrating diverse psychosocial, behavioral, and contextual measures to better understand complex behavioral outcomes and inform targeted prevention strategies. Keywords: Aggressive behavior, Adolescents, Adverse life events, Chronotype, Peer environment, Family environment Graphical Abstract Created in BioRender. Meruelo, D.(2025) https://BioRender.com Introduction Aggressive behavior in adolescents and young adults—manifesting through actions such as arguing, fighting, or threatening others—is a pressing public health concern. It is associated with a host of adverse outcomes, including academic underachievement, social difficulties, mental health problems, and increased risk of justice system involvement ( Fauzi et al., 2023 ; Liu et al., 2013 ). Understanding the key predictors of aggression is essential to inform prevention and early intervention strategies. However, aggression does not develop in isolation; it reflects a multifaceted interplay among individual traits, family dynamics, social relationships, and environmental exposures. Research has highlighted a wide array of factors that may influence aggression, including exposure to adverse life events (e.g., trauma, instability) ( Martin-Herz et al., 2022 ; Thompson et al., 2023 ; Tiet et al., 1998 ; Tzouvara et al., 2023 ; Valdes et al., 2024 ; Vergunst et al., 2023 ; Wasserman et al., 2021 ), lower socioeconomic status (SES) ( Fauzi et al., 2023 ), poor sleep patterns (such as later sleep timing or delayed chronotype)( Schlarb et al., 2014 ), high impulsivity ( Miller et al., 2012 ), and low-quality peer and family relationships ( Andreas and Watson, 2009 ; Benson and Buehler, 2012 ; Liu et al., 2013 ; Padilla-Walker et al., 2020 ). These influences are often interrelated. For instance, adverse events may disrupt emotional regulation, while low SES may constrain access to protective resources. Additionally, emerging evidence suggests that dietary quality and eating patterns may also influence emotional and behavioral regulation through neurobiological pathways ( Golomb et al., 2012 ; Meruelo et al., 2024 ). Despite growing attention to these factors, much of the existing literature has examined them in isolation, often using narrowly focused samples. This has limited our understanding of how these predictors interact to influence aggressive behavior—particularly in diverse, population-based cohorts. In addition, while constructs such as impulsivity and emotional dysregulation are often referenced alongside aggression, these terms are not interchangeable. In this study, aggression is conceptualized as a behavioral outcome characterized by externalizing acts of hostility or violence. In contrast, impulsivity reflects a tendency toward rapid, unplanned reactions to stimuli without consideration of consequences, and emotional dysregulation describes difficulties managing emotional responses. Impulsivity and dysregulation are treated as risk factors for aggression, though their pathways of influence may be direct or indirect. To address these conceptual and methodological challenges, we adopt a hypothesis-generating, data-driven approach using machine learning (ML) methods ( Aria et al., 2021 ; Breiman, 2001 ) to examine predictors of aggressive behavior. Our analysis leverages data from the Adolescent Brain Cognitive Development (ABCD) Study ( Volkow et al., 2018 ), a large, diverse, longitudinal cohort of U.S. youth supported by the National Institutes of Health. The ABCD dataset offers a rare opportunity to explore a wide range of behavioral, psychological, and social variables at scale, with the statistical power to detect complex relationships. We employ ensemble-based ML methods, including random forest and gradient boosting, which are particularly well-suited to exploratory analysis in high-dimensional datasets. These algorithms can detect non-linear associations, handle collinearity among predictors, and model interactions and threshold effects without strong parametric assumptions ( Aria et al., 2021 ; Breiman, 2001 ). These advantages position ML as an ideal tool for identifying the most important predictors of aggression and uncovering patterns that may be missed by traditional regression models. In this study, we explore how adverse life events, chronotype, socioeconomic status, peer and family dynamics, impulsivity traits, and dietary patterns collectively contribute to aggressive behavior. Rather than testing predefined hypotheses, our goal is to generate novel insights into the relative importance and interactive effects of these factors. For example, we explore whether late chronotype and poor sleep timing compound the effects of stress, or whether strong family cohesion may buffer against risk associated with impulsivity. Although prior literature informs the selection of variables, we remain agnostic to specific causal relationships and instead aim to uncover the structure of associations using a comprehensive, exploratory framework. Findings from this study may serve as a foundation for future confirmatory work and intervention development. Methods Study Design and Participants This study utilized a cross-sectional design to investigate the predictors of aggressive behavior in adolescents and young adults. Data were drawn from a larger longitudinal project, the Adolescent Brain Cognitive Development (ABCD) study, examining behavioral and mental health outcomes ( “NIH’s Adolescent Brain Cognitive Development (ABCD) Study,” 2018 ). Participants included individuals aged 10–12 years ( Table 1 ), selected based on the availability of complete data for the variables of interest. Table 1. Participant Demographic Characteristics (N = 3,602). This table summarizes demographic characteristics for the full analytic sample. Variable Category Count Proportion Mean ± SD Normalized Household Income – – – 1.39 ± 1.12 Age (in Years) – – – 11.13 ± 1.03 Race Non-Hispanic White 2,728 76.16% – Other/Mixed 524 14.63% – Black 258 7.20% – Asian 72 2.01% – Sex Female 1,681 46.67% – Male 1,921 53.33% – Open in a new tab Outcome Variable: Aggressive Behavior The primary outcome (density plot shown in Figure 1 ) was self-reported aggressive behavior, assessed using a validated questionnaire designed to capture the frequency and intensity of aggressive actions. The Aggressive Behavior Syndrome Scale (T-Score) from the Child Behavior Checklist provides a standardized measure of aggressive behavior tendencies in adolescents. This T-score assesses the frequency and severity of behaviors such as verbal or physical aggression, defiance, and confrontational tendencies, relative to a normative population. Higher scores on this scale indicate greater levels of aggressive behavior. The CBCL aggressive behavior scale has demonstrated excellent internal consistency in normative and clinical samples, with Cronbach’s alpha of 0.86-0.94 ( Achenbach and Rescorla, 2003 ; Bordin et al., 2013 ; Kariuki et al., 2016 ). The measure is widely used in developmental and clinical research to examine the influence of environmental, familial, and neurobiological factors on aggression. Figure 1: Density Plot of T-scores for Aggressive Behavior. Open in a new tab This density plot illustrates the distribution of T-scores for aggressive behavior in the dataset. The x-axis represents T-scores, while the y-axis indicates the density, with the highest density observed at lower aggression scores, reflecting the overall distribution of the population. Predictor Variables Predictors included a range of demographic, psychological, social, and behavioral variables ( Table 2 ; see Supplementary Figure 1 for years of data collection). Table 2. Key Variable Machine Learning Descriptive Statistics (N = 3,602). This table summarizes descriptive statistics for all variables included in the machine learning analyses, derived from the full analytic dataset. Variables are grouped by domain, including demographics, dietary patterns, psychosocial context, chronotype and stress, and behavioral traits. Variable Domain Mean ± SD Min Max Age (in Years) Demographics 11.13 ± 1.03 8.92 13.58 Normalized Household Income Demographics 1.39 ± 1.12 0.07 10.00 Sex (Female) Demographics 0.47 ± 0.50 0.00 1.00 Aggressive Behavior T-score Aggressive Behavior 54.99 ± 6.69 50.00 98.00 Diet Question 1 Dietary Patterns 0.66 ± 0.47 0.00 1.00 Diet Question 2 Dietary Patterns 0.43 ± 0.50 0.00 1.00 Diet Question 3 Dietary Patterns 0.85 ± 0.36 0.00 1.00 Diet Question 4 Dietary Patterns 0.67 ± 0.47 0.00 1.00 Diet Question 5 Dietary Patterns 0.76 ± 0.43 0.00 1.00 Diet Question 6 Dietary Patterns 0.34 ± 0.47 0.00 1.00 Diet Question 7 Dietary Patterns 0.87 ± 0.33 0.00 1.00 Diet Question 8 Dietary Patterns 0.23 ± 0.42 0.00 1.00 Diet Question 9 Dietary Patterns 0.22 ± 0.42 0.00 1.00 Diet Question 10 Dietary Patterns 0.69 ± 0.46 0.00 1.00 Diet Question 11 Dietary Patterns 0.73 ± 0.44 0.00 1.00 Diet Question 12 Dietary Patterns 0.81 ± 0.39 0.00 1.00 Diet Question 13 Dietary Patterns 0.25 ± 0.43 0.00 1.00 Diet Question 14 Dietary Patterns 0.59 ± 0.49 0.00 1.00 Diet Question 15 Dietary Patterns 0.53 ± 0.50 0.00 1.00 Diet Question 16 Dietary Patterns 0.95 ± 0.22 0.00 1.00 Parental Monitoring Psychosocial Factors 4.83 ± 0.43 2.00 5.00 Peer Network Health Psychosocial Factors 11.50 ± 7.94 0.00 27.00 Family Conflict Psychosocial Factors 0.53 ± 0.50 0.00 1.00 School Environment Subscale Psychosocial Factors 20.52 ± 2.59 6.00 24.00 Neighborhood Safety Scale Psychosocial Factors 4.03 ± 0.84 1.00 5.00 Community Cohesion Psychosocial Factors 3.80 ± 0.66 1.00 5.00 Community Control Psychosocial Factors 4.06 ± 0.72 1.00 5.00 Chronotype Chronotype and Stress 27.54 ± 1.59 18.62 41.71 Adverse Life Events Chronotype and Stress 0.93 ± 1.45 0.00 12.00 UPPS Negative Urgency Behavioral Traits 7.63 ± 2.31 4.00 16.00 Past Year Grades Behavioral Traits 3.01 ± 1.86 1.00 12.00 UPPS Sensation Seeking Behavioral Traits 9.58 ± 2.70 4.00 16.00 Open in a new tab Adverse Life Events Adverse life events (density plot shown in Figure 2 ) were assessed using the Life Events Scale ( Wethington, 2016 ), a structured instrument that captures both lifetime and past-year exposure to a range of stressful and potentially traumatic experiences. This measure, which includes content overlapping with Adverse Childhood Experiences (ACEs), was administered through a brief youth survey (approximately 4 minutes) and a complementary parent-report survey (approximately 2 minutes). For the present study, we utilized a cumulative count of endorsed stressful life events for each participant as the primary variable. This total score reflected the number of events experienced across diverse domains, including death or injury of a loved one, personal illness, exposure to crime or violence, peer-related challenges, financial hardship, substance use in the family, legal difficulties, parental separation, school problems, relocation, incarceration, and changes in household composition such as the birth of a sibling or a parent’s job change. Higher scores indicate a greater cumulative burden of stressful life experiences. Figure 2: Density Plot of Adverse Life Events. Open in a new tab This density plot shows the distribution of adverse life events within the dataset. The x-axis represents the number of adverse life events, and the y-axis represents the density. The plot reveals that lower numbers of adverse life events are more common, as indicated by the peak density in this range, reflecting the overall population's distribution. Chronotype Chronotype was obtained from the ABCD Youth Munich Chronotype Questionnaire ( Roenneberg et al., 2003 ). It was operationalized as the midpoint of sleep on free (non-school) days, adjusted for sleep debt during the school week. Specifically, the corrected sleep midpoint was calculated as: Midpoint of Sleep on Free Days – (Sleep Duration on Free Days – Average Weekly Sleep Duration)/2 ( Suh et al., 2017 ) This adjustment accounts for oversleep on free days due to weekday sleep restriction. Chronotype scores were recorded on a 16–40 hour scale, with higher values indicating later sleep timing. For example, a score of 16 corresponds to 4:00 p.m. on the same day, 25.3 corresponds to approximately 1:20 a.m. the next day, and 40 represents 4:00 p.m. the following day ( Burns et al., 2025 ). Socioeconomic Status Household income, serving as a proxy for SES, was collected at baseline from the ABCD Longitudinal Parent Demographics Survey. Income was categorized into ten brackets, ranging from less than $5,000 to $200,000 or more. Income was then normalized by the number of household members. Peer and Family Environments Peer and family environments were evaluated through validated scales assessing social cohesion, parental monitoring, and peer network health ( Hoffman et al., 2019 ). Community Cohesion and Control The predictor variable community cohesion, parent-reported, captures a measure of social cohesion within the participant’s community. It is calculated as the mean score of responses to five items assessing close-knit relationships, willingness to help others, getting along with others, shared values, and trust within the community. Responses of 999 (“Don’t Know”) were recoded to a neutral value of 3 in the calculation of the summary score, and the variable required at least one item to be validly answered. Similarly, community control, parent-reported, measures perceived control within the community environment. It is calculated as the mean score of five items addressing community conditions, including skipping school, graffiti, disrespect, fighting, and budget constraints. Like community cohesion, responses of 999 (“Don’t Know”) were recoded to 3, and the summary score required a minimum of one valid response. Parental Monitoring Scale Parental monitoring was assessed using the Parental Monitoring Scale ( Stattin and Kerr, 2000 ), which evaluates the degree to which parents or guardians are aware of and involved in their child's daily activities, social interactions, and whereabouts. Higher scores indicate greater parental supervision and engagement. Peer Network Health The Peer Network Health scale ( Mason et al., 2004 ) was used to assess the protective and risky qualities of participants’ peer relationships. This composite measure is derived from participant-reported information on each nominated peer’s substance use, behavioral influence, and activity engagement. Items capture whether peers use substances, are daily users, or influence the participant to use or avoid substances. Participants also report on peers' involvement in illegal, violent, or dangerous behaviors, as well as on prosocial aspects, such as offering help with schoolwork or transportation and providing emotional support. Each peer receives a total score ranging from −14 to 14, calculated using a weighted system. Higher scores indicate healthier, more protective peer networks, whereas lower scores reflect greater behavioral risk. The scale has demonstrated strong internal reliability (Cronbach’s alpha = 0.84) and is significantly associated with participants’ self-reported substance use, including alcohol and marijuana ( Mason et al., 2004 ). School Environment The School Environment Subscale from the School Risk and Protective Factors (SRPF) Inventory was used to assess participants’ perceptions of their classroom and school climate. This subscale consists of six self-report items selected from validated instruments including the Academic Competence Scale ( Harter, 2006 ) and the Communities That Care Youth Survey ( Arthur et al., 2007 ). Items were derived from the PhenX Toolkit’s School Social Environment domain but were shortened for feasibility in the ABCD protocol. Items capture experiences of classroom support, student involvement in school rules and activities, and perceived teacher responsiveness (e.g., “In my school, students have lots of chances to help decide things like class activities and rules,” and “My teacher notices when I am doing a good job and lets me know about it”). Responses were summed to create a total score reflecting perceived support and opportunity within the school context. A valid score required responses to at least five of the six items. The subscale demonstrated adequate internal consistency in the ABCD sample (Cronbach’s α = 0.60), and prior analyses showed that scores on this measure significantly distinguished between youth at lower versus higher behavioral risk ( Zucker et al., 2018 ). Higher scores indicate a more positive school climate and were hypothesized to be protective against aggressive behavior in this study. Academic Performance Past year grades were assessed using a detailed scale ranging from A+ (97-100 or 4.0 GPA and higher) to F (0-65 or 0.00 GPA), capturing academic performance across multiple levels of proficiency. This variable was based on parent-reported grades from the prior school year, with descriptors for each grade level that aligned with academic standards, including terms such as "Exceeds Excellent," "Very Good," and "Approaching Proficiency." Responses of "Not Applicable" or "Refuse to Answer" were excluded from analyses. Neighborhood Safety The neighborhood safety variable assessed parental perceptions of neighborhood safety based on three items: The first item measured agreement with the statement, "I feel safe walking in my neighborhood, day or night," with responses ranging from 1 (Strongly Disagree) to 5 (Strongly Agree). The second item assessed agreement with the statement, "Violence is not a problem in my neighborhood," using the same response scale. The third item evaluated agreement with the statement, "My neighborhood is safe from crime," also rated on a scale from 1 (Strongly Disagree) to 5 (Strongly Agree). These three items were averaged to generate a continuous score reflecting overall perceptions of neighborhood safety. The variable required no minimum responses to compute and provided a reliable measure based on established neighborhood assessment protocols ( Echeverria et al., 2004 ; Mujahid et al., 2007 ). Higher scores indicate more positive perceptions of neighborhood safety, a critical environmental factor influencing adolescent and family well-being. Family Conflict Family conflict was assessed using a parent-reported binary variable reflecting perceptions of household communication and emotional regulation. The item, “In our family, we believe you don't ever get anywhere by raising your voice,” was coded as 0 = True and 1 = False. While this single-item, binary measure does not fully capture the multidimensional nature of family conflict, it was selected for its conceptual relevance to emotional reactivity and communication style—factors strongly linked to aggression and externalizing symptoms ( Andreas and Watson, 2009 ; Benson and Buehler, 2012 ). Additionally, this approach minimized missingness compared to longer scales, ensuring broader inclusion in model training. Impulsivity Impulsivity was assessed using selected subscales from the UPPS-P Abbreviated Youth Version a validated self-report instrument designed to measure multidimensional aspects of impulsive behavior. Participants rated items on a 4-point Likert scale ranging from “Not at all true” to “Very true.” In the present study, we focused on two key dimensions: negative urgency and sensation seeking. Negative urgency reflects the tendency to act impulsively in response to negative emotions or distress, capturing the affective component of impulsivity most closely associated with emotional dysregulation and maladaptive coping. Sensation seeking reflects the propensity to pursue novel or exciting experiences, even when such behavior involves risk, and is more closely linked to exploratory tendencies and reward sensitivity. Higher scores on each subscale indicate greater endorsement of the corresponding impulsive trait. These dimensions were selected due to their established associations with externalizing behavior ( Bakhshani, 2014 ; Carvalho et al., 2023 ; Elliott et al., 2023 ; Gröndal et al., 2023 ; Johnson et al., 2017 ; Meruelo et al., 2024 ; Miller et al., 2012 ; Schreiber et al., 2012 ; Velotti and Garofalo, 2015 ; Wasserman et al., 2021 ) and their theoretical relevance to aggressive outcomes. Internal consistency was acceptable for both subscales (Cronbach’s α = 0.75 for negative urgency; α = 0.62 for sensation seeking) ( Lee et al., 2024 ). Dietary Predictors The dietary predictor variables (Diet Question 1 to Diet Question 16) included a range of parent-reported measures capturing typical dietary patterns and nutritional habits of participants over the past year. These included the frequency of consuming whole grains (Diet Question 1), green leafy vegetables (Diet Question 2), other vegetables (Diet Question 3), berries (Diet Question 4), red meats (Diet Question 5), fish (Diet Question 6), poultry or eggs (Diet Question 7), beans (Diet Question 8), nuts (Diet Question 9), and fast or fried food (Diet Question 10). Additional variables assessed whether olive oil was the primary cooking oil (Diet Question 11), butter or margarine usage was limited (Diet Question 12), cheese consumption was infrequent (Diet Question 13), and pastries or sweets were consumed less than five times per week (Diet Question 14). Variables Diet Question 15 and Diet Question 16 addressed whether prenatal vitamins or folic acid supplements were taken before and during pregnancy. Responses were binary (1 = Yes, 0 = No) with 999 indicating “Don’t Know.” Data Preprocessing Initial preprocessing involved identifying and addressing missing data, followed by standardization of continuous predictors. Of the original 3,602 participants, 1,079 (approximately 30%) were excluded in the initial complete-case analysis due to missingness in one or more predictors or the outcome variable, yielding a working dataset of 2,523 participants. To ensure model generalizability and prevent information leakage, we modified our preprocessing pipeline so that standardization was performed after splitting the data into training and testing subsets. Standardization was applied independently within each subset, preserving the integrity of the test data and avoiding inadvertent exposure of test-set information during model fitting. Missing Data Handling and Sensitivity Analyses To evaluate the potential for selection bias resulting from complete-case analysis, we conducted a sensitivity analysis comparing participants included versus excluded from the complete-case dataset on key demographic and psychosocial variables. This analysis revealed several statistically significant differences between groups ( Supplementary Table 1 ). Specifically, excluded participants had slightly later chronotypes, lower perceived neighborhood safety, and marginally lower school environment scores. Standardized mean differences (SMDs) for most variables were below the conventional threshold of 0.1, but some exceeded this threshold, suggesting moderate imbalance and possible selection bias. To address these concerns, we implemented multiple imputation using the mice package in R ( Buuren and Groothuis-Oudshoorn, 2011 ). Five imputed datasets were generated using predictive mean matching, incorporating all analytic variables, including the outcome. Each imputed dataset was processed and analyzed independently, and final results were pooled to generate robust estimates of model performance and predictor importance. This approach allowed us to retain the full analytic sample and mitigate bias introduced by missing data. Machine Learning Models We evaluated multiple predictive modeling approaches to identify key predictors of aggressive behavior, including a standard linear regression model, a tuned random forest, a tuned gradient boosting machine (GBM), an XGBoost model, and an ensemble model that combined predictions from the random forest and GBM. Linear regression was included as a benchmark to assess the added value of machine learning models relative to a traditional parametric approach. Machine learning algorithms were selected for their capacity to model complex nonlinear relationships, accommodate high-dimensional data, and operate without strong parametric assumptions ( Aria et al., 2021 ; Breiman, 2001 ; Chen and Guestrin, 2016 ; Ridgeway, n.d. ; Schonlau and Zou, 2020 ). Each model was trained using five-fold cross-validation within each of the five imputed datasets. For machine learning models, hyperparameters were optimized using grid search procedures. Linear regression was fit directly to each training set without regularization or tuning. Model performance was evaluated on a held-out test set comprising 30% of the sample within each imputed dataset. Evaluation metrics included root mean squared error (RMSE), mean absolute error (MAE), R-squared (R 2 ), and Pearson correlation between actual and predicted values. The ensemble model was constructed by averaging the predicted values from the tuned random forest and GBM models within each test set. This unweighted ensemble was implemented post hoc and did not undergo a separate cross-validation procedure. Instead, each base learner was cross-validated independently, and ensemble performance was evaluated on the same held-out data. Across imputations, the random forest and linear regression models demonstrated comparable predictive accuracy ( Supplementary Table 2 ). Notably, the linear regression model achieved the highest pooled correlation (r = 0.313) and the lowest MSE (40.76), slightly outperforming the random forest model (r = 0.311; MSE = 40.71). Despite this, the random forest was selected as the primary analytic model due to its enhanced ability to capture nonlinear associations and interactions, which were further explored in variable importance and interaction analyses. To evaluate the impact of missing data handling, we also compared the imputed random forest model to a complete-case version. As detailed in Supplementary Table 3 , the complete-case model slightly outperformed the imputed model in correlation (r = 0.331 vs. 0.310) and MSE (36.31 vs. 40.71), but retained fewer participants (N = 1,449 vs. N = 2,523). The imputed dataset thus offered greater statistical power and generalizability, supporting its use for final modeling despite a modest reduction in predictive accuracy. Variable Importance and Statistical Software Variable importance was evaluated using the mean decrease in accuracy metric derived from the random forest model, averaged across imputations. This metric provided insight into which predictors were most influential in forecasting aggressive behavior. All statistical analyses and data visualizations were conducted in R, using the caret ( Kuhn, 2008 ) , randomForest ( Breiman, 2001 ), gbm ( Ridgeway, n.d. ), xgboost ( Chen and Guestrin, 2016 ), and mice ( Buuren and Groothuis-Oudshoorn, 2011 ) packages. Interaction Detection and Visualization To further explore potential nonlinear and interactive effects captured by the random forest model, we conducted a series of post hoc interpretation analyses using the final tuned model from the first imputed dataset. Partial dependence plots (PDPs) ( Molnar et al., 2023 ) were generated to visualize average marginal effects of pairs of predictors, allowing us to examine how the predicted aggression scores varied across combinations of key variables such as adverse life events and community cohesion. Individual conditional expectation (ICE) plots ( Goldstein et al., 2014 ) were also constructed to show subject-level response curves, highlighting heterogeneity in how adverse life events related to aggression across participants. Additionally, we applied a fast variable interaction detection method based on minimal depth, implemented via the randomForestExplainer package ( “Multi-way importance plot,” n.d. ). This approach ranks variable pairs based on how frequently they appear near the top of decision trees in the forest, with smaller average minimal depth values indicating stronger interactions. The results were summarized in tabular form ( Supplementary Table 4 ) and visualized in an interaction plot, providing a global view of the most influential two-way interactions captured by the model. Results A series of machine learning models were developed to identify predictors of aggressive behavior scores using a multimodal dataset. The primary model—a tuned random forest trained on multiply imputed datasets—demonstrated strong predictive performance, with pooled results showing a correlation of r = 0.311 between predicted and actual aggression scores and a mean squared error (MSE) of 40.71 across imputations ( Table 3 and Figure 3 ). Gradient boosting (GBM), XGBoost, and an ensemble model combining GBM and random forest predictions showed slightly lower performance, with the ensemble model achieving r = 0.310 and MSE = 40.74 ( Supplementary Table 2 ). Table 3. Random Forest and Ensemble Model Performance Metrics (Pooled Across Imputations). This table summarizes the performance of the final tuned Random Forest model and an ensemble model combining Random Forest and Gradient Boosting Machine (GBM) for predicting aggressive behavior T-scores. Both models were trained on the multiply imputed dataset (N = 2,523; 5 imputations) using 32 predictors, including age. The Random Forest model was tuned using 5-fold cross-validation with an optimal mtry of 4. The ensemble model, constructed as an unweighted average of RF and GBM predictions, achieved similar performance across all metrics. Pooled correlation was nearly identical (0.3105 vs. 0.3096), as was MSE (40.71 vs. 40.74). These findings suggest that the ensemble model did not meaningfully outperform the tuned Random Forest model and that RF alone is preferred for interpretability and parsimony. Metric Random Forest (Pooled, Tuned) Ensemble (RF + GBM) Number of Samples 2,523 2,523 Number of Predictors 32 32 Cross-Validation Method 5-Fold CV 5-Fold CV (Base models only) Number of Imputations 5 5 Optimal mtry 4 N/A Pooled Correlation (Actual vs. Predicted) 0.3105 0.3096 Pooled Mean Squared Error (MSE) 40.71 40.74 Pooled Root Mean Squared Error (RMSE) 6.39 6.39 Pooled Mean Absolute Error (MAE) 4.80 4.82 Pooled R 2 0.097 0.096 Open in a new tab Figure 3: Random Forest Actual v. Predicted Values Binned Comparison. Open in a new tab This binned heatmap displays the relationship between actual and predicted aggression scores from the Random Forest model, using square bins to address overplotting. Color intensity reflects the number of observations within each bin (log-scaled). The red line represents a linear regression fit, illustrating the overall predictive trend. The binning approach improves clarity in regions of high data density and better visualizes model performance across the outcome range. For comparison, a linear regression model using the same predictors and imputations achieved the highest pooled correlation ( r = 0.313) and a comparable MSE (40.76), suggesting that much of the predictive signal may be captured through linear associations. However, machine learning models, particularly the random forest, provided added value in detecting nonlinear patterns and interaction effects, justifying their use for interpretation and feature interaction analyses. Key predictors of aggression included the cumulative number of adverse life events, chronotype, neighborhood cohesion and safety, impulsivity under negative emotions, and indicators of peer and family support. These variables were consistently ranked high in the random forest variable importance metrics ( Figure 4a and 4b ), which included both mean decrease in accuracy and node purity criteria. Figure 4a: Random Forest Variable Importance (Imputation Model). Open in a new tab This plot displays the average variable importance for predicting aggressive behavior across five imputed datasets using the random forest model. Variable importance was calculated using the percent increase in mean squared error (%IncMSE) when each predictor was permuted. Adverse Life Events emerged as the most influential predictor, followed by contextual and psychosocial variables such as Community Cohesion, Family Conflict, Neighborhood Safety Scale, and Chronotype. While several dietary and demographic variables were included, their predictive contributions were comparatively modest. Figure 4b. Random Forest Model Variable Importance by Node Purity (Imputation Model). Open in a new tab This figure presents variable importance scores based on node purity from the tuned random forest model trained on the first imputed dataset (N = 2,523). Node purity reflects the total decrease in residual sum of squares attributable to splits on each variable across all trees. Higher values indicate that a variable contributed more substantially to reducing impurity and improving model fit. Adverse life events, chronotype, household income, and age were among the most important predictors, consistent with the %IncMSE ranking. The convergence of findings across variable importance metrics strengthens the interpretability and reliability of the model. To address potential bias from listwise deletion, all models were trained on multiply imputed datasets (m = 5), and results were pooled across imputations. Predictive findings and feature rankings remained consistent across datasets, underscoring the robustness of the selected predictors. Nonlinear relationships, interactions and associations were studied. Nonlinear relationships were explored using partial dependence plots ( Figure 5 ) and individual conditional expectation (ICE) plots ( Figure 6 ). These visualizations revealed a dose-response relationship between adverse life events and predicted aggression scores, with steeper predicted increases observed among individuals reporting low peer protection. ICE plots also revealed considerable individual-level variability despite consistent group-level trends. Figure 5: Partial Dependence Heatmap of Predicted Aggression by Adverse Life Events and Family Cohesion. Open in a new tab This heatmap depicts the partial dependence of predicted aggression scores on the interaction between adverse life events (x-axis) and community cohesion (y-axis), based on the random forest model. Warmer colors (yellow) indicate higher predicted levels of aggression. The estimates show that greater exposure to adverse life events is associated with higher predicted aggression, particularly in contexts of low community cohesion (lower values on the y-axis). The interaction pattern suggests that greater community cohesion may mitigate the adverse behavioral effects of life stressors. Figure 6. Individual Conditional Expectation (ICE) Plot of Predicted Aggression by Adverse Life Events. Open in a new tab Individual Conditional Expectation (ICE) plot with partial dependence (blue line) showing the relationship between standardized adverse life events and predicted aggressive behavior scores. Each gray line represents a single participant's predicted outcome trajectory as the number of adverse life events increases, while holding all other variables constant. The blue line reflects the average partial dependence across all individuals. The pattern reveals a consistent positive association, indicating that higher exposure to adverse life events is associated with elevated predicted aggression scores. Interaction effects were examined using both two-way partial dependence surfaces and minimal depth-based interaction detection. The interaction surface between adverse life events and family cohesion ( Figure 5 ) showed that greater cohesion attenuated the effect of adversity on aggression. Friedman’s H-statistic ( Figure 7 ) further confirmed that interactions between Adverse Life Events × Chronotype and Adverse Life Events × Family Cohesion were among the most influential. Figure 7. Top Feature Interactions by Minimal Depth in Random Forest Model. Open in a new tab This bar plot displays the top 30 predictor interactions ranked by mean minimal depth, a metric of interaction strength within the random forest model trained to predict aggression scores. Each bar represents an interaction between two predictors, with shorter depths indicating stronger interactions. The red horizontal line indicates the average minimal depth threshold for comparison. Darker bars reflect more frequently occurring interactions across decision trees. The strongest interactions were observed between adverse life events and chronotype, normalized household income, and community cohesion, suggesting synergistic effects in predicting aggression. The top ten feature interactions, summarized in Supplementary Table 4 , indicated that the influence of adversity was often modulated by contextual and psychosocial factors such as parental monitoring, peer support, and school environment. Finally, binned heatmaps with regression overlays ( Figure 3 ) highlighted approximately linear associations between actual and predicted aggression scores, with particularly strong linear trends observed for several predictors, including chronotype and neighborhood safety. Discussion This study identified key variables associated with aggressive behavior in preadolescents using machine learning techniques, with a focus on random forest modeling and interaction-based interpretation. The findings support and expand upon prior research by highlighting statistically important associations among psychosocial, behavioral, environmental, and biological domains. Our approach, which incorporated multiple imputations and interaction analysis techniques, offers a robust, exploratory framework for understanding aggression as a multifactorial phenotype shaped by cumulative risk exposure and protective contextual factors. Importantly, all results are interpreted in terms of predictive relevance rather than causality, consistent with the cross-sectional design of the study. For comparison, we also evaluated a traditional linear regression model trained on the same imputed datasets. Interestingly, its predictive performance was comparable to the machine learning models, suggesting that many associations were largely linear in nature. Nonetheless, nonlinear models provided unique advantages for interaction detection and interpretation. The cumulative burden of adverse life events showed the strongest predictive association with aggressive behavior in the random forest model. This aligns with extensive literature linking adversity—such as family instability, financial strain, and loss—to aggressive and externalizing behaviors ( Martin-Herz et al., 2022 ; Thompson et al., 2023 ; Tiet et al., 1998 ; Tzouvara et al., 2023 ; Valdes et al., 2024 ; Vergunst et al., 2023 ; Wasserman et al., 2021 ). The interaction analysis revealed that the predictive impact of these life events was significantly modulated by community cohesion and other contextual variables, emphasizing that the effects of stressors are not uniform but shaped by the broader social ecology. These findings underscore the importance of viewing aggression through a systemic lens, where individual risk is embedded within family, peer, school, and neighborhood systems. Chronotype was also among the top predictors, with later sleep timing patterns associated with higher predicted aggression scores. This finding aligns with growing evidence that delayed chronotype is linked to emotional dysregulation and behavioral problems in youth ( McCarthy et al., 2023 ; Schlarb et al., 2014 ). Although the model does not imply causality, the observed association carries important clinical implications. Chronotype may function not only as a marker of behavioral disorganization but also as an early signal of neurobiological vulnerability, particularly in relation to mood and impulse regulation. Delayed chronotype has been linked to dysregulation of circadian rhythms, including shifts in melatonin secretion and altered sleep-wake timing ( Emens and Burgess, 2015 ; Gehrman and Anafi, 2021 ), which may disrupt emotional stability ( Au and Reece, 2017 ) and cognitive control ( Burns et al., 2025 ). Melatonin, a hormone integral to circadian timing, typically begins to rise in the early evening to promote sleep onset ( Crowley et al., 2016 ; Pandi-Perumal et al., 2007 ). In adolescents with a delayed chronotype, this rise is shifted later into the night, contributing to insufficient and irregular sleep schedules. Chronic misalignment between biological and social clocks (e.g., early school start times) can increase sleep debt ( Roenneberg et al., 2019 ), impair prefrontal functioning, and heighten irritability—factors known to contribute to aggression and other externalizing symptoms ( Gollan et al., 2005 ; Rosell and Siever, 2015 ; Schlarb et al., 2014 ). Moreover, delayed sleep phase has been observed in pediatric populations at elevated risk for affective disorders, including bipolar spectrum conditions ( Soehner et al., 2019 ). The predictive signal attributed to chronotype in this study may partially reflect underlying mood instability, irritability, and circadian rhythm disruptions characteristic of early-onset mood disorders ( Meyer and Carlson, 2008 ). Future research should incorporate objective circadian phase markers (e.g., dim light melatonin onset) and examine chronotype alongside clinical screening for internalizing and bipolar symptoms to further clarify these mechanistic links. Lower household income demonstrated a strong association with elevated aggression scores, aligning with prior research linking socioeconomic disadvantage to behavioral risk ( Fauzi et al., 2023 ). Financial stress may exacerbate other risk factors, such as reduced sleep quality, exposure to unsafe neighborhoods, and limited access to supportive social networks. Moreover, the effect of income may differ across demographic groups due to intersecting cultural and structural factors. These findings support calls for intervention strategies that address not only individual behaviors but also the broader social determinants of health. Indicators of peer and family relationships showed strong associations with aggression, reinforcing their importance as contextual predictors. Higher-quality peer networks and greater family cohesion were associated with reduced aggression, consistent with theories of social buffering and emotional scaffolding ( Andreas and Watson, 2009 ; Benson and Buehler, 2012 ; Padilla-Walker et al., 2020 ). Importantly, our interaction analysis revealed that the buffering effect of peer and family support was strongest in the context of high life stress, suggesting a moderating (protective) influence. These results support school-based and family-centered interventions that strengthen relational support as a means to reduce aggressive behavior in high-risk populations. Emotion-driven impulsivity, measured via the UPPS negative urgency scale, was another salient predictor, consistent with its documented association with externalizing tendencies ( Carver and Johnson, 2018 ; Elliott et al., 2023 ; Johnson et al., 2017 ). These findings lend further support to conceptualizing aggression not as a singular trait but as a behavioral outcome shaped by emotional dysregulation, poor coping strategies, and neurobiological sensitivity to reward and punishment. While impulsivity, emotional dysregulation, and aggression are interrelated ( Carvalho et al., 2023 ; Gröndal et al., 2023 ; Schreiber et al., 2012 ; Velotti and Garofalo, 2015 ), it is important to distinguish their respective roles: impulsivity is treated here as a predictor, emotional dysregulation as a potential mechanism, and aggression as the behavioral outcome. Academic performance and perceived school environment demonstrated moderate predictive relevance, contributing additional insight into school-related correlates of aggression. This is consistent with evidence that negative school climates and low academic achievement are associated with externalizing behaviors ( Franco et al., 2022 ; La Salle-Finley et al., 2024 ; Lester and Cross, 2015 ; Lombardi et al., 2019 ; Maxwell et al., 2017 ; O’Brennan et al., 2014 ; Wang et al., 2020 ). Interventions that improve school engagement and connectedness may reduce behavioral issues in vulnerable students. Dietary variables were included as exploratory features and showed modest predictive strength ( Supplementary Table 3 and Figure 6 ), suggesting a limited but potentially relevant role. These results suggest that while diet may play a role in behavioral regulation ( Meruelo et al., 2024 ), it is likely more distal or indirect compared to other psychosocial variables. We caution against overinterpreting the role of diet in isolation. Importantly, aggression—as assessed via the CBCL T-score—is a transdiagnostic symptom that may reflect multiple underlying conditions including ADHD, trauma-related disorders, conduct disorder, and autism spectrum disorder. Our model does not distinguish among these subtypes, which limits its diagnostic specificity. However, this also enhances its generalizability as a screening tool for high-risk behavioral presentations. Future research should integrate structured diagnostic information or neurocognitive phenotyping to better map behavioral subtypes onto specific psychopathological processes. Interaction analysis added further depth to our interpretation. We employed multiple techniques, including partial dependence plots ( Figure 5 ), individual conditional expectation (ICE) plots ( Figure 6 ), and Friedman’s H-statistic via minimal depth analysis ( Figure 7 ), to examine nonlinear and synergistic effects among top associated predictors. These analyses revealed that adverse life events interact most strongly with chronotype, peer health, parental monitoring, and community safety—offering nuanced insights into how cumulative risks cluster and may amplify behavioral outcomes. The modest correlation values (r = 0.310 for random forest, r = 0.259 for the ensemble model) are typical in psychological prediction studies, especially those relying on self-reported behavioral data. These effect sizes are comparable to prior machine learning studies predicting behavioral outcomes ( Hentati Isacsson et al., 2024 ) and reflect the multifactorial, often idiosyncratic nature of aggression. The relatively low R 2 also reflects both measurement noise inherent in self-reported behavioral data and the diffuse, context-dependent nature of aggression in this age group. Importantly, while R 2 values were relatively low (e.g., 0.097), these models still provide actionable insights, especially when combined with interaction-based interpretation tools. These results underscore the importance of combining predictive accuracy with interpretability when designing models for clinical or policy use. This study has several limitations. The cross-sectional nature of the data prohibits causal inference, and it is possible that some predictors (e.g., poor sleep, impulsivity) are consequences, rather than causes, of aggression. Future studies should incorporate more frequent longitudinal assessments to better capture temporal dynamics. Additionally, self-report measures may be biased by social desirability or recall errors. Objective measures of sleep (e.g., actigraphy), impulsivity (e.g., behavioral tasks), and aggression (e.g., teacher or parent ratings) would strengthen future models. The use of a single binary item to assess family conflict may oversimplify a complex construct; future research would benefit from more comprehensive, multi-item assessments of family functioning to improve measurement precision and capture diverse patterns of family dynamics. While the model incorporated a broad array of psychosocial, behavioral, environmental, and biological predictors, it is likely that additional unmeasured factors—such as genetic vulnerability, early neurodevelopmental disruptions, and culturally embedded norms—also influence the emergence of aggressive behavior. These omissions may contribute to unexplained variance and limit the overall predictive power of the model. Furthermore, our findings may not be fully generalizable to populations with differing cultural, geographic, or linguistic characteristics, especially given that many of the contextual variables (e.g., family cohesion, neighborhood safety) may operate differently across sociocultural contexts. Importantly, this study focused on children aged 10 to 12—a developmental period marked by heightened emotional plasticity, rapid cognitive and social change, and emerging identity formation ( Ciccia et al., 2009 ). This preadolescent window is a sensitive phase during which regulatory capacities are still maturing and environmental influences can exert pronounced effects on behavioral trajectories ( Meruelo et al., 2017 ). Studying aggression at this age allows for the identification of early warning signals before the onset of adolescence, when risk behaviors may become more entrenched and difficult to modify. The inclusion of contextual variables such as peer environment, school climate, and parental monitoring supports a systems-level understanding of aggression that is particularly relevant during this transition period. However, greater specificity in future models—such as incorporating detailed measures of family structure (e.g., single-parent households, co-parenting dynamics), cultural orientation (e.g., acculturation, ethnic identity), and intergenerational adversity (e.g., parental trauma exposure)—would further enhance the developmental precision and translational utility of predictive frameworks targeting aggression. Finally, our findings have translational relevance. Sleep hygiene interventions, trauma-informed care, school engagement programs, and family-based therapies all show promise in reducing aggression. Chronotype, in particular, may serve as an early behavioral biomarker for risk and a target for behavioral or pharmacological intervention. By integrating findings across systems—peer, family, school, and biology—this study highlights the potential for coordinated, multi-level approaches to prevention. Supplementary Material Supplementary Material NIHMS2088810-supplement-Supplementary_Material.docx (186.2KB, docx) Applied machine learning models to predict adolescent aggression using a large dataset. Random forest, GBM, XGBoost, and ensemble models were evaluated across imputations. Random forest outperformed other models (r = 0.310; MSE = 40.71; R 2 = 0.097). Linear regression performed similarly (r = 0.313), supporting robustness of findings. Findings highlight the predictive role of adversity, chronotype, and peer context. Acknowledgments: We sincerely thank the participants and their families for their invaluable contributions to the Adolescent Brain Cognitive Development (ABCD) Study. Your time, dedication, and willingness to share your experiences are essential to advancing scientific knowledge in this field. Your participation is the foundation of this research, helping to shape a deeper understanding of adolescent development and informing efforts to improve health and well-being for future generations. We are profoundly grateful for your support—this work would not be possible without you. Funding: This research was supported by the National Institute on Alcohol Abuse and Alcoholism (NIAAA) under grant K23 AA026869, awarded to Alejandro Meruelo, MD, PhD, and by the VA Merit Award BX003431, awarded to Michael McCarthy. Declarations of Interests: The authors declare no known competing financial interests or personal relationships that could have influenced the work reported in this paper. Dr. Coccaro serves as a consultant and a member of the Scientific Advisory Boards for Azevan Pharmaceuticals, Inc., Avanir Pharmaceuticals, Inc., and Boehringer Ingelheim Pharma, Inc. He also receives grant funding from the National Institute of Mental Health (NIMH) and the National Institute on Alcohol Abuse and Alcoholism (NIAAA). Dr. McCarthy receives grant funding from the Veterans Affairs Administration. Dr. Meruelo receives grant funding from the National Institute on Alcohol Abuse and Alcoholism (NIAAA). Data: This study utilized data from the Adolescent Brain Cognitive Development (ABCD) Study ( https://abcdstudy.org ), a longitudinal research initiative funded by the National Institutes of Health (NIH). The ABCD Study is supported by the following grants: U01DA041022, U01DA041028, U01DA041048, U01DA041089, U01DA041106, U01DA041117, U01DA041120, U01DA041134, U01DA041148, U01DA041156, U01DA041174, U24DA041123, and U24DA041147. The data analyzed in this research are publicly available through the National Institute of Mental Health Data Archive (NDA) as part of the ABCD Study. Footnotes Ethics approval and consent: The data used in this study were derived from the Adolescent Brain Cognitive Development (ABCD) Study, which received institutional ethical approval, and all procedures adhered to established research ethics protocols conducted in accordance with the ethical standards outlined in the 1964 Declaration of Helsinki. Informed consent and assent were obtained from all participants and their legal guardians, as applicable. References Achenbach TM, Rescorla L, 2003. Reliability, Internal Consistency, Cross-Informant Agreement, and Stability, and .2003. In Manual for the ASEBA Adult Forms & Profiles: For Ages 18-59: Adult Self-report and Adult Behavior Checklist (pp. 91-125). : ASEBA., in: Manual for the ASEBA Adult Forms & Profiles: For Ages 18-59: Adult Self-Report and Adult Behavior Checklist. ASEBA, Burlington, VY, pp. 91–125. [ Google Scholar ] Andreas JB, Watson MW, 2009. Moderating effects of family environment on the association between children’s aggressive beliefs and their aggression trajectories from childhood to adolescence. Dev Psychopathol 21, 189–205. 10.1017/S0954579409000121 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Aria M, Cuccurullo C, Gnasso A, 2021. A comparison among interpretative proposals for Random Forests. Machine Learning with Applications 6, 100094. 10.1016/j.mlwa.2021.100094 [ DOI ] [ Google Scholar ] Arthur MW, Briney JS, Hawkins JD, Abbott RD, Brooke-Weiss BL, Catalano RF, 2007. Measuring risk and protection in communities using the Communities That Care Youth Survey. Eval Program Plann 30, 197–211. 10.1016/j.evalprogplan.2007.01.009 [ DOI ] [ PubMed ] [ Google Scholar ] Au J, Reece J, 2017. The relationship between chronotype and depressive symptoms: A meta-analysis. J Affect Disord 218, 93–104. 10.1016/j.jad.2017.04.021 [ DOI ] [ PubMed ] [ Google Scholar ] Bakhshani N-M, 2014. Impulsivity: A Predisposition Toward Risky Behaviors. Int J High Risk Behav Addict 3, e20428. 10.5812/ijhrba.20428 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Benson MJ, Buehler C, 2012. Family Process and Peer Deviance Influences on Adolescent Aggression: Longitudinal Effects across Early and Middle Adolescence. Child Dev 83, 1213–1228. 10.1111/j.1467-8624.2012.01763.x [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Bordin IA, Rocha MM, Paula CS, Teixeira MCTV, Achenbach TM, Rescorla LA, Silvares EFM, 2013. Child Behavior Checklist (CBCL),Youth Self-Report (YSR) and Teacher’s Report Form(TRF): an overview of the development of the original and Brazilian versions. Cad. Saúde Pública 29, 13–28. 10.1590/S0102-311X2013000100004 [ DOI ] [ PubMed ] [ Google Scholar ] Breiman L., 2001. Random Forests. Machine Learning 45, 5–32. 10.1023/A:1010933404324 [ DOI ] [ Google Scholar ] Burns J, Li AR, Rohr KE, Thomas ML, McCarthy MJ, Meruelo AD, 2025. The influence of chronotype, socioeconomic status, latitude, longitude, and seasonality on cognitive performance and academic outcomes in adolescents. Sleep Medicine 128, 95–102. 10.1016/j.sleep.2025.01.030 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] van Buuren S, Groothuis-Oudshoorn K, 2011. mice: Multivariate Imputation by Chained Equations in R. Journal of Statistical Software 45, 1–67. 10.18637/jss.v045.i03 [ DOI ] [ Google Scholar ] Carvalho CB, Arroz AM, Martins R, Costa R, Cordeiro F, Cabral JM, 2023. “Help Me Control My Impulses!”: Adolescent Impulsivity and Its Negative Individual, Family, Peer, and Community Explanatory Factors. J. Youth Adolescence 52, 2545–2558. 10.1007/s10964-023-01837-z [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Carver CS, Johnson SL, 2018. Impulsive Reactivity to Emotion and Vulnerability to Psychopathology. Am Psychol 73, 1067–1078. 10.1037/amp0000387 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Chen T, Guestrin C, 2016. XGBoost: A Scalable Tree Boosting System, in: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ‘16. Association for Computing Machinery, New York, NY, USA, pp. 785–794. 10.1145/2939672.2939785 [ DOI ] [ Google Scholar ] Ciccia AH, Meulenbroek P, Turkstra LS, 2009. Adolescent Brain and Cognitive Developments. Top Lang Disord 29, 249–265. 10.1097/TLD.0b013e3181b53211 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Crowley SJ, Suh C, Molina TA, Fogg LF, Sharkey KM, Carskadon MA, 2016. Estimating the dim light melatonin onset of adolescents within a 6-h sampling window: the impact of sampling rate and threshold method. Sleep Med 20, 59–66. 10.1016/j.sleep.2015.11.019 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Echeverria SE, Diez-Roux AV, Link BG, 2004. Reliability of self-reported neighborhood characteristics. J Urban Health 81, 682–701. 10.1093/jurban/jth151 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Elliott MV, Johnson SL, Pearlstein JG, Muñoz Lopez DE, Keren H, 2023. Emotion-related impulsivity and risky decision-making: A systematic review and meta-regression. Clinical Psychology Review 100, 102232. 10.1016/j.cpr.2022.102232 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Emens JS, Burgess HJ, 2015. Effect of Light and Melatonin and other Melatonin Receptor Agonists on Human Circadian Physiology. Sleep Med Clin 10, 435–453. 10.1016/j.jsmc.2015.08.001 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Fauzi FA, Zulkefli NAM, Baharom A, 2023. Aggressive behavior in adolescent: The importance of biopsychosocial predictors among secondary school students. Front Public Health 11, 992159. 10.3389/fpubh.2023.992159 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Franco K, Baumler E, Torres ED, Lu Y, Wood L, Temple JR, 2022. The link between school climate and mental health among an ethnically diverse sample of middle school youth. Curr Psychol 1–11. 10.1007/s12144-022-03016-y [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Gehrman PR, Anafi RC, 2021. Treatment of a patient with a circadian sleep-wake disorder using a combination of melatonin and metoprolol. J Clin Sleep Med 17, 2121–2124. 10.5664/jcsm.9410 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Goldstein A, Kapelner A, Bleich J, Pitkin E, 2014. Peeking Inside the Black Box: Visualizing Statistical Learning with Plots of Individual Conditional Expectation. 10.48550/arXiv.1309.6392 [ DOI ] [ Google Scholar ] Gollan JK, Lee R, Coccaro EF, 2005. Developmental psychopathology and neurobiology of aggression. Development and Psychopathology 17, 1151–1171. 10.1017/S0954579405050546 [ DOI ] [ PubMed ] [ Google Scholar ] Golomb BA, Evans MA, White HL, Dimsdale JE, 2012. Trans fat consumption and aggression. PLoS One 7, e32175. 10.1371/journal.pone.0032175 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Gröndal M, Ask K, Winblad S, 2023. The distinction between irritability and anger and their associations with impulsivity and subjective wellbeing. Sci Rep 13, 10398. 10.1038/s41598-023-37557-4 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Harter S., 2006. The Development of Self-Esteem, in: Self-Esteem Issues and Answers. Psychology Press. [ Google Scholar ] Hentati Isacsson N, Ben Abdesslem F, Forsell E, Boman M, Kaldo V, 2024. Methodological choices and clinical usefulness for machine learning predictions of outcome in Internet-based cognitive behavioural therapy. Commun Med 4, 1–11. 10.1038/s43856-024-00626-4 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Hoffman EA, Clark DB, Orendain N, Hudziak J, Squeglia LM, Dowling GJ, 2019. Stress exposures, neurodevelopment and health measures in the ABCD study. Neurobiol Stress 10, 100157. 10.1016/j.ynstr.2019.100157 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Johnson SL, Tharp JA, Peckham AD, Carver CS, Haase CM, 2017. A Path Model of Different Forms of Impulsivity with Externalizing and Internalizing Psychopathology: Toward Greater Specificity. Br J Clin Psychol 56, 235–252. 10.1111/bjc.12135 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Kariuki SM, Abubakar A, Murray E, Stein A, Newton CRJC, 2016. Evaluation of psychometric properties and factorial structure of the pre-school child behaviour checklist at the Kenyan Coast. Child Adolesc Psychiatry Ment Health 10, 1. 10.1186/s13034-015-0089-9 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Kuhn M., 2008. Building Predictive Models in R Using the caret Package. Journal of Statistical Software 28, 1–26. 10.18637/jss.v028.i0527774042 [ DOI ] [ Google Scholar ] La Salle-Finley T, Yang C, Espelage D, Jimerson SR, 2024. Understanding and Promoting School Climate, Bullying, and Social-Emotional Learning: Transdisciplinary and Transnational Science Advancing Positive Youth Outcomes. School Psychology Review 53, 417–424. 10.1080/2372966X.2024.2386235 [ DOI ] [ Google Scholar ] Lee JY, Kang JH, Shin M-J, Park JH, Song DH, Choi JY, 2024. Validation of a Korean Version of the Short UPPS-P Impulsive Behavior Scale for Children. Psychiatry Investig 21, 208–215. 10.30773/pi.2023.0277 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Lester L, Cross D, 2015. The Relationship Between School Climate and Mental and Emotional Wellbeing Over the Transition from Primary to Secondary School. Psychology of Well-Being 5, 9. 10.1186/s13612-015-0037-8 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Liu J, Lewis G, Evans L, 2013. Understanding Aggressive Behavior Across the Life Span. J Psychiatr Ment Health Nurs 20, 156–168. 10.1111/j.1365-2850.2012.01902.x [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Lombardi E, Traficante D, Bettoni R, Offredi I, Giorgetti M, Vernice M, 2019. The Impact of School Climate on Well-Being Experience and School Engagement: A Study With High-School Students. Front. Psychol 10. 10.3389/fpsyg.2019.02482 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Martin-Herz SP, Haggerty KP, Neilands TB, Sterling M, Christakis DA, 2022. Factors Associated With Trajectories of Externalizing Behavior in Preschoolers. Academic Pediatrics 22, 1212–1220. 10.1016/j.acap.2021.12.027 [ DOI ] [ PubMed ] [ Google Scholar ] Mason M, Cheung I, Walker L, 2004. Substance use, social networks, and the geography of urban adolescents. Subst Use Misuse 39, 1751–1777. [ PubMed ] [ Google Scholar ] Maxwell S, Reynolds KJ, Lee E, Subasic E, Bromhead D, 2017. The Impact of School Climate and School Identification on Academic Achievement: Multilevel Modeling with Student and Teacher Data. Front. Psychol 8. 10.3389/fpsyg.2017.02069 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] McCarthy MJ, Brumback T, Thomas ML, Meruelo AD, 2023. The relations between chronotype, stressful life events, and impulsivity in the Adolescent Brain Cognitive Development (ABCD) study. Journal of Psychiatric Research 167, 119–124. 10.1016/j.jpsychires.2023.10.030 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Meruelo AD, Brumback T, Pelham WE, Wade NE, Thomas ML, Coccaro EF, Nooner KB, Brown SA, Tapert SF, Mrug S, 2024. How Do Anger and Impulsivity Impact Fast-Food Consumption in Transitional Age Youth? AJPM Focus 3, 100208. 10.1016/j.focus.2024.100208 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Meruelo AD, Castro N, Cota CI, Tapert SF, 2017. Cannabis and alcohol use, and the developing brain. Behav. Brain Res 325, 44–50. 10.1016/j.bbr.2017.02.025 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Meyer SE, Carlson GA, 2008. Early-Onset Bipolar Disorder. FOC 6, 271–283. 10.1176/foc.6.3.foc271 [ DOI ] [ Google Scholar ] Miller JD, Zeichner A, Wilson LF, 2012. Personality Correlates of Aggression: Evidence From Measures of the Five-Factor Model, UPPS Model of Impulsivity, and BIS/BAS. J Interpers Violence 27, 2903–2919. 10.1177/0886260512438279 [ DOI ] [ PubMed ] [ Google Scholar ] Molnar C, Freiesleben T, König G, Herbinger J, Reisinger T, Casalicchio G, Wright MN, Bischl B, 2023. Relating the Partial Dependence Plot and Permutation Feature Importance to the Data Generating Process, in: Explainable Artificial Intelligence. Presented at the World Conference on Explainable Artificial Intelligence, Springer, Cham, pp. 456–479. 10.1007/978-3-031-44064-9_24 [ DOI ] [ Google Scholar ] Mujahid MS, Diez Roux AV, Morenoff JD, Raghunathan T, 2007. Assessing the measurement properties of neighborhood scales: from psychometrics to ecometrics. Am J Epidemiol 165, 858–867. 10.1093/aje/kwm040 [ DOI ] [ PubMed ] [ Google Scholar ] Multi-way importance plot, n.d. NIH’s Adolescent Brain Cognitive Development (ABCD) Study, 2018. . Alcohol Res 39, 97. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] O’Brennan LM, Bradshaw CP, Furlong MJ, 2014. Influence of Classroom and School Climate on Teacher Perceptions of Student Problem Behavior. School Ment Health 6, 125–136. 10.1007/s12310-014-9118-8 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Padilla-Walker LM, Stockdale LA, Son D, Coyne SM, Stinnett SC, 2020. Associations between parental media monitoring style, information management, and prosocial and aggressive behaviors. Journal of Social and Personal Relationships 37, 180–200. 10.1177/0265407519859653 [ DOI ] [ Google Scholar ] Pandi-Perumal SR, Smits M, Spence W, Srinivasan V, Cardinali DP, Lowe AD, Kayumov L, 2007. Dim light melatonin onset (DLMO): a tool for the analysis of circadian phase in human sleep and chronobiological disorders. Prog Neuropsychopharmacol Biol Psychiatry 31, 1–11. 10.1016/j.pnpbp.2006.06.020 [ DOI ] [ PubMed ] [ Google Scholar ] Ridgeway G., n.d. Generalized Boosted Models: A guide to the gbm package. [ Google Scholar ] Roenneberg T, Pilz LK, Zerbini G, Winnebeck EC, 2019. Chronotype and Social Jetlag: A (Self-) Critical Review. Biology (Basel) 8, 54. 10.3390/biology8030054 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Roenneberg T, Wirz-Justice A, Merrow M, 2003. Life between clocks: daily temporal patterns of human chronotypes. J Biol Rhythms 18, 80–90. 10.1177/0748730402239679 [ DOI ] [ PubMed ] [ Google Scholar ] Rosell DR, Siever LJ, 2015. The neurobiology of aggression and violence. CNS Spectrums 20, 254–279. 10.1017/S109285291500019X [ DOI ] [ PubMed ] [ Google Scholar ] Schlarb AA, Sopp R, Ambiel D, Grünwald J, 2014. Chronotype-related differences in childhood and adolescent aggression and antisocial behavior--a review of the literature. Chronobiol Int 31, 1–16. 10.3109/07420528.2013.829846 [ DOI ] [ PubMed ] [ Google Scholar ] Schonlau M, Zou RY, 2020. The random forest algorithm for statistical learning. The Stata Journal 20, 3–29. 10.1177/1536867X20909688 [ DOI ] [ Google Scholar ] Schreiber LRN, Grant JE, Odlaug BL, 2012. Emotion Regulation and Impulsivity in Young Adults. J Psychiatr Res 46, 651–658. 10.1016/j.jpsychires.2012.02.005 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Soehner AM, Bertocci MA, Levenson JC, Goldstein TR, Rooks B, Merranko J, Hafeman D, Diler R, Axelson D, Goldstein BI, Hickey MB, Monk K, Phillips ML, Birmaher B, 2019. Longitudinal Associations Between Sleep Patterns and Psychiatric Symptom Severity in High-Risk and Community Comparison Youth. Journal of the American Academy of Child & Adolescent Psychiatry 58, 608–617. 10.1016/j.jaac.2018.09.448 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Stattin H, Kerr M, 2000. Parental monitoring: a reinterpretation. Child Dev 71, 1072–1085. 10.1111/1467-8624.00210 [ DOI ] [ PubMed ] [ Google Scholar ] Suh S, Ryu H, Kim S, Choi S, Joo EY, 2017. Using Mid-Sleep Time to Determine Chronotype in Young Adults with Insomnia-Related Symptoms. Sleep Med Res 8, 107–111. 10.17241/smr.2017.00115 [ DOI ] [ Google Scholar ] Thompson EL, O’Connor KE, Farrell AD, 2023. Childhood adversity and co-occurring post-traumatic stress and externalizing symptoms among a predominantly low-income, African American sample of early adolescents. Dev Psychopathol 35, 383–395. 10.1017/S0954579421001383 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Tiet QQ, Bird HR, Davies M, Hoven C, Cohen P, Jensen PS, Goodman S, 1998. Adverse Life Events and Resilience. Journal of the American Academy of Child & Adolescent Psychiatry 37, 1191–1200. 10.1097/00004583-199811000-00020 [ DOI ] [ PubMed ] [ Google Scholar ] Tzouvara V, Kupdere P, Wilson K, Matthews L, Simpson A, Foye U, 2023. Adverse childhood experiences, mental health, and social functioning: A scoping review of the literature. Child Abuse & Neglect 139, 106092. 10.1016/j.chiabu.2023.106092 [ DOI ] [ PubMed ] [ Google Scholar ] Valdes V, Sacks DD, Nelson CA, Enlow MB, 2024. Stress timing, trauma exposure, and family resilience differentially affect internalizing and externalizing symptoms at 3, 5, and 7 years of age. Commun Psychol 2, 1–13. 10.1038/s44271-024-00151-z [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Velotti P, Garofalo C, 2015. Personality styles in a non-clinical sample: The role of emotion dysregulation and impulsivity. Personality and Individual Differences 79, 44–49. 10.1016/j.paid.2015.01.046 [ DOI ] [ Google Scholar ] Vergunst F, Commisso M, Geoffroy M-C, Temcheff C, Poirier M, Park J, Vitaro F, Tremblay R, Côté S, Orri M, 2023. Association of Childhood Externalizing, Internalizing, and Comorbid Symptoms With Long-term Economic and Social Outcomes. JAMA Network Open 6, e2249568. 10.1001/jamanetworkopen.2022.49568 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Volkow ND, Koob GF, Croyle RT, Bianchi DW, Gordon JA, Koroshetz WJ, Pérez-Stable EJ, Riley WT, Bloch MH, Conway K, Deeds BG, Dowling GJ, Grant S, Howlett KD, Matochik JA, Morgan GD, Murray MM, Noronha A, Spong CY, Wargo EM, Warren KR, Weiss SRB, 2018. The conception of the ABCD study: From substance use to a broad NIH collaboration. Dev Cogn Neurosci 32, 4–7. 10.1016/j.dcn.2017.10.002 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Wang M-T, Degol JL, Amemiya J, Parr A, Guo J, 2020. Classroom climate and children’s academic and psychological wellbeing: A systematic review and meta-analysis. Developmental Review 57, 100912. 10.1016/j.dr.2020.100912 [ DOI ] [ Google Scholar ] Wasserman AM, Wimmer J, Hill-Kapturczak N, Karns-Wright TE, Mathias CW, Dougherty DM, 2021. The Development of Externalizing and Internalizing Behaviors among Youth with or without a Family History of Substance Use Disorder: The Indirect Effects of Early-Life Stress and Impulsivity. Child Psychiatry Hum Dev 52, 978–993. 10.1007/s10578-020-01076-4 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Wethington E., 2016. Chapter 12 - Life Events Scale, in: Fink G (Ed.), Stress: Concepts, Cognition, Emotion, and Behavior. Academic Press, San Diego, pp. 103–108. 10.1016/B978-0-12-800951-2.00012-1 [ DOI ] [ Google Scholar ] Zucker RA, Gonzalez R, Feldstein Ewing SW, Paulus MP, Arroyo J, Fuligni A, Morris AS, Sanchez M, Wills T, 2018. Assessment of culture and environment in the Adolescent Brain and Cognitive Development Study: Rationale, description of measures, and early data. Developmental Cognitive Neuroscience, The Adolescent Brain Cognitive Development (ABCD) Consortium: Rationale, Aims, and Assessment Strategy 32, 107–120. 10.1016/j.dcn.2018.03.004 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary Material NIHMS2088810-supplement-Supplementary_Material.docx (186.2KB, docx) ACTIONS View on publisher site PDF (1.3 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Sociability Under Anthropogenic Change: From Individual Behaviour to Population-Level Consequences.#1167713
• How local social norms shape global environmental action: The role of norm misperceptions in environmental attitude-behaviour cycles.#360821
• Predicting Climate Adaptation Intention and Collective Climate Action: A Supervised Machine Learning Approach#457880
• Mapping how LLMs debate societal issues when shadowing human personality traits, sociodemographics and social media behavior#146599
• Machine Learning and Artificial Intelligence in Toxicology and Environmental Health#164185