Predicting juvenile delinquency and criminal behavior in adulthood using machine learning

Transparency and Open Practices Statement

This is a two-stage registered report: For the Stage I submission, we used public-use data ² from Waves I and V of the longitudinal Add Health Study (K. M. Harris & Udry, 2018) to test the functionality of the syntax for estimating the measurement models of the outcomes (i.e., juvenile delinquency and criminal behavior) and the ML models (i.e., elastic net and GBMs). Importantly, the ML prediction models proposed in the Stage 1 submission were run in a single iteration of the outer resampling scheme to check for any syntax errors and ensure that the analysis pipeline executed as intended. The full analyses were subsequently conducted using 100 iterations of the outer resampling scheme to obtain more stable and generalizable performance estimates across different data splits (see below for more information).

We confirmed that the proposed research has not been conducted at the time of the Stage I submission. We did not have access to the full data set. After in-principle acceptance of our initial submission, we requested access to the full data set, analyzed the data as proposed, and completed the manuscript. All syntax to reproduce the preliminary analyses of Stage I and the final analyses of Stage II are available in an online repository (https://osf.io/c8bxp/). However, access to the data set is subject to the regulations of the Add Health Study (we have included a link ³ to the download page in the online repository). The time-stamped Stage I of the registered report is available at https://osf.io/xrd4y. The analysis of the Add Health data was approved by the Ethics Committee of the Faculty of Human Sciences of the University of Kassel (EK-Nr. 202318).

Sample

Data are from Waves I and V of the longitudinal Add Health Study (K. M. Harris & Udry, 2018), collected in the United States in 1994 and 2018. Data were obtained through a multistage, school-based sampling procedure (for a detailed description see K. M. Harris, 2013; K. M. Harris et al., 2019) that randomly selected approximately 200 adolescents per high school were randomly chosen, yielding a nationally representative core sample of adolescents from the United States, further supplemented by some oversampled subpopulations (e.g., with respect to ethnicity, genetics, and disabilities). The public-use data consist of N = 6,504 adolescents for Wave I (M_age = 15.54 years, SD_age = 1.79, range: 11–21; 51.61% women, 48.39% men), in comparison to N = 20,745 adolescents in the full sample. Specifically, it consists of a randomly selected half of the nationally representative core sample and a randomly selected half of an oversample of highly educated African Americans.

Measures

With respect to the categorical outcome variables, we assigned the various delinquent behaviors at Wave I to three categories: drug offenses (e.g., drinking beer, wine, or liquor), property offenses (e.g., painting graffiti or signs on someone else’s property), and violent offenses (e.g., hurting someone badly enough to need bandages or care). This categorization of outcomes is based on preliminary analyses conducted on the publicly available use file (for more information on the statistical modeling and the dimensionality of the outcomes, see Appendix A in the online Supplemental Material). We assigned the criminal behavior at Wave V based on the information about whether individuals were charged with drug offenses (e.g., driving under the influence of drugs), property offenses (e.g., fraud, forgery, or embezzlement), and violent offenses (e.g., aggravated assault, intentional manslaughter). Please note that the items used to measure delinquency are not identical across measurement waves (see Tables S4 and S5 in the online Supplemental Material for the complete list of items). The offenses are tailored to the developmental stage of the offender, which is why we refer to the constructs as juvenile delinquency and criminal behavior, respectively.

With respect to predictors, we used two sets: The first predictor set consists of 291 variables from 14 broad categories, including demographics, family- and friend-related variables, personality, mental and physical health variables, and so on (see Table 1). For the second predictor set with 216 variables, the individual items were aggregated into scales wherever possible to allow for a more reliable assessment (see Supplemental Appendix B for the exact methodological procedure of aggregation and Supplemental Table S3 for the scales). Both sets of predictors are presented in Table S2 in the online Supplemental Material. To predict criminal behavior at Wave V, we also included the items that assessed juvenile delinquency at Wave I.

OPEN IN VIEWER

Table 1.

Predictor Categories, Exemplary Topics, and Number of Predictors.

Category	Exemplary topics	n
1. Demographics (DEMO)	Age, sex, racial origin	13
2. Family (FAM)	Relationship to parents, joint activities	54
3. Neighborhood (NBHD)	Feeling safe, talking to neighbors	14
4. Friends (FRI)	Spending time together	12
5. Romantic Relationships & Sexuality (ROM)	Contraception, pregnancy, partners	18
6. Education (EDU)	Suspension, grades, attitudes toward school	38
7. Employment & Finances (FIN)	Work hours, allowance, financial aid	8
8. Personality (PERS)	Intelligence, decision-making	20
9. Mental Health (MHE)	Suicide attempts, feelings	32
10. Physical Health (PHE)	Physical development, health services	43
11. Risk Behavior (RISK)	Availability of drugs and guns at home	13
12. Life Expectations & Aspirations (EXP)	Perceived chances of getting AIDS	9
13. Religion (REL)	Importance of religion, time spent praying	5
14. Extracurricular Activities (ACT)	Hobbies, media consumption	12
	Total	291

Note. The last column indicates the number of predictors within a category.

ML Analyses

For the prediction of deviant and criminal behavior and the validation of predictive accuracy, we employed a nested cross-validation approach (Pargent et al., 2023). Nested cross-validation combines an inner and an outer validation loop to ensure the strict separation of training and testing samples. Consequently, all data preprocessing and the subsequent model validation were performed independently on the training and testing data to avoid information leakage (see also Kapoor & Narayanan, 2023). Each iteration of the outer validation loop consisted of the following steps: First, the full dataset was split into training data (80%) and testing data (the remaining 20%), taking into account the nested data structure (participants nested in high schools). Then, any preprocessing of the data (imputation of missing data with the k-nearest neighbors algorithm, standardization, etc.) was done separately for the training and testing data. In the inner validation loop, the model was trained using the training data and ten-fold cross-validation. Moreover, we used the same cross-sectional weights (as in the measurement models) for model training. The predictive accuracy of the models was calculated using independent testing data over 100 iterations of the outer validation loop for an unbiased estimate. The following indices were used for model evaluation: Explained variance (R²), the Root Mean Squared Error (RMSE), and the Mean Absolute Error (MAE). To examine the degree of overfitting, all indices were calculated separately for the training and testing samples over the 100 iterations. In addition to the predictive accuracy of the different methods, we examined the variable importances. We focused our discussion of important variables on the best-performing model approach, but we also compared the overlap of the most important variables across the different methods to examine their degree of generalizability.

We compared the predictive accuracy of linear regressions with two ML algorithms, namely elastic net regularized regressions and GBMs. Elastic net regressions are a form of regularized regression designed to reduce model variance and prediction error by increasing the bias of parameter estimates through a penalty term (Helwig, 2017; Kuhn & Johnson, 2013). Elastic net regressions often provide good predictive accuracy because they strike a balance between ridge regression and least absolute shrinkage and selection operator (LASSO) regression by performing both shrinkage and feature selection (Zou & Hastie, 2005). GBMs, on the other hand, are tree-based algorithms that incorporate nonlinear and higher-order interaction effects into the modeling without making any assumptions about the functional relationships linking predictor variables to an outcome (James et al., 2021). GBMs use an ensemble of decision trees, where each tree fits the residual error of the previous one, resulting in improved predictive accuracy (Schroeders et al., 2022). For more information on the specific tuning parameters and the nested resampling approach please see the annotated syntax at (https://osf.io/c8bxp/). See Appendix C in the online Supplemental Material for the statistical software used.

Predicting Juvenile Delinquency and Adult Criminal Behavior

The predictive performances of the ML analyses and the linear regressions as baseline models for adolescence (Wave I) are presented in Figure 1 (averaged values for R², MAE, and RMSE across 100 iterations of the outer resampling loop are provided in Table S8 in the online Supplemental Material). The outcome was a factor score representing the different types of juvenile offenses, constructed using multiple age-specific indicators. The key findings are as follows: The best item-level predictions in the independent test sample were achieved using elastic net regression for drug offenses (e.g., drinking beer, wine, or liquor) with R² = .57 (GBM: .55; linear regression: .56). Predictions for juvenile violent offenses (e.g., number of serious physical fights) achieved R² = .44 using elastic net regression (GBM: .41; linear regression: .43), while property offenses (e.g., stealing something worth less than $50) were predicted using elastic net regression with R² = .39 (GBM: .38; linear regression: .38).

OPEN IN VIEWER

Figure 1.

Explained Variance Across Various Juvenile Offenses and Modeling Approaches (Wave I).

In addition, several methodological observations are worth noting. Overfitting, as indicated by the difference in performance between the training and test datasets, was minimal due to the large sample size (Pargent et al., 2023; Yarkoni & Westfall, 2017). Furthermore, the two predictor sets—the item-level set with 291 predictors and the partially aggregated scale-level set with 216 predictors—yielded essentially identical predictive performance, which is largely due to the substantial overlap between the sets, which shared 209 predictors. Thus, no effect of reliability improvement was observed from using scales instead of items. Finally, the different modeling approaches produced similar prediction results for overall test performance. However, compared to simple linear regression, elastic net regression demonstrated slight performance improvements and reduced overfitting due to its regularization properties (Zou & Hastie, 2005). In summary, the elastic net with predictors at the item-level yielded the best predictive results.

Figure 2 illustrates the predictive performances for adult criminal behavior at Wave V (averaged values for R², MAE, and RMSE across 100 iterations of the outer resampling loop are provided in Supplemental Table S9). When interpreting the figures, it is important to note that the items assessing deviant behavior in adulthood—focusing on whether participants had ever been charged with a criminal offense—are qualitatively different from those measuring juvenile delinquency. Although the results from the most recent measurement timepoint of the Add Health Study shared some similarities with findings from the first wave, the most notable difference was a significant decline in predictive accuracy. Using elastic net regression drug-related offenses (e.g., ever been charged with possession, sale, use, etc. of marijuana) could be predicted in the independent test sample at the item level with R² = .16 (GBM: R² = .15; linear regression: R² = .13). For violent offenses (e.g., ever been charged with simple assault), the corresponding values were R² = .15 for elastic net regressions (GBM: R² = .14; linear regression: R² = .12), whereas for property offenses (e.g., ever been charged with theft), the values were R² = .13 for elastic net regressions (GBM: R² = .13; linear regression: R² = .10). This is most likely because only predictors from Wave I were used in the prediction, and the time lag between the assessment of the predictors and the outcome was nearly 22 years (for similar temporal effects, see Jankowsky, Steger, & Schroeders, 2024).

OPEN IN VIEWER

Figure 2.

Explained Variance Across Adult Criminal Behavior and Modeling Approaches (Wave V).

From a methodological perspective, several findings stand out: The smaller sample size in Wave V compared to Wave I resulted in substantially higher overfit (performance difference between training and test datasets), particularly in the (unregularized) linear regression models. Differences between predictor sets (item vs. scale level) were generally negligible, except for linear regression models, where aggregating items into scales mitigated overfit. When comparing the average performance metrics in independent test samples, the elastic net model at the item level consistently delivered the best predictive results.

Key Predictors of Juvenile Delinquency and Adult Criminal Behavior

Since the elastic net models were the best performing models across all conditions (i.e., offense type and measurement time points), we used the regression coefficients of the elastic net, averaged across the 100 iterations, to display the most important variables (see Figure 3; for the exact values, see Supplemental Table S10). Compared to other variable importance measures (Grömping, 2006), regression coefficients offer the advantage of allowing comparisons of relationship magnitude in a familiar metric, while also preserving the direction of the association. In the following, we present a few of the key findings that will be revisited in the discussion.

OPEN IN VIEWER

Figure 3.

Variable Importance Across Offense Types and Measurement Waves.

The following observations can be made for Wave I: First, many small, individual contributions from different variables collectively contributed to the predictions. This can partly be attributed to the statistical method (i.e., shrinkage due to regularization) but also underscores the interplay of numerous influential factors. Second, regardless of the type of offense, questions about whether close friends used drugs regularly (e.g., marijuana or alcohol; H1TO33 and H1TO29) or whether the respondent had ever received an out-of-school suspension (H1ED7) were the most predictive. Third, the three most frequently occurring categories in the top 20 across outcomes were “Risk Behavior,” “Romantic Relationships and Sexuality,” and “Education” (categories are displayed at the right of each needle plot in Figure 3).

The following observations could be made for Wave V: First, among the most influential predictors, reported acts of juvenile delinquency stand out, making up more than one third of the top variables (these variables describing past offenses are labeled with the abbreviation POF, past offenses, in Figure 3). Second, for violent, property, and drug-related offenses, sex emerged as the most predictive variable. Unlike Wave I, where sex ranked outside the top 20 predictors (except for violent offenses), men were more frequently charged with crimes in Wave V (presented in Figure 3 as a protective effect favoring women). Third, regardless of the type of crime, the availability and use of both legal and illegal drugs played an important role.

Prediction of Juvenile Delinquency

The prediction of juvenile delinquency relied on many small effects, supporting the idea that most complex psychological phenomena are “determined by a multitude of causes and that any individual cause is likely to have only a small effect” (Götz et al., 2021, p. 205). Nonetheless, the models predicting juvenile delinquency performed surprisingly well—especially compared to the models predicting adult criminal behavior—apparently due to the temporal proximity between predictors and outcomes. However, there is also a qualitative difference in the self-reported outcomes at the different measurement points, which may affect model performance: Juvenile delinquency in Wave I was measured using a much lower threshold (delinquent behavior within the dark figure), whereas Wave V applied the stricter criterion of formal charges, which require detection, investigation, and sufficient evidence.

As outlined in the introduction, no single theory of crime has yet succeeded in providing a comprehensive explanatory model. Instead, criminal behavior must be understood as a complex interplay of biological, psychological, and social factors. This complexity is also evident in the present study, which identifies numerous variables associated with criminal behavior in adolescence and adulthood. The most prominent predictors included variables from the areas of “Risk Behavior” (including antisocial behavior) and “Romantic Relationships and Sexuality,” peer group dynamics, and “Education” (i.e., severe problems at school). Regarding the theoretical accounts presented in the introduction, it is worth pointing out which groups of variables were not identified as relevant. Neither the extensive set of family-related variables (n = 54) nor the neighborhood variables (n = 14) had a significant incremental impact, which could be interpreted as evidence against theories that primarily focus on these societal factors. Nevertheless, it is likely that these variables lay the foundation for antisocial behavior (e.g., Horwitz et al., 2001), which later manifests during adolescence and is reflected in the recorded variables.

Effective risk prediction and interventions depend on empirically validated criminogenic risk factors. Bonta and Andrews (2017) synthesized findings from multiple meta-analyses and highlighted eight risk factors that consistently predicted offending. These factors, referred to as the “Central Eight,” are divided into two categories: the “Big Four”—antisocial cognition, antisocial associates, antisocial personality patterns, and a history of antisocial behavior—which are considered strong predictors of offending, and the “Moderate Four”—family/marital issues, school/work difficulties, leisure/recreation deficits, and substance abuse—which are viewed as moderate predictors of criminal behavior (Bonta & Andrews, 2017). Several widely used risk-assessment tools are based upon or overlap conceptually with the Central Eight, including the Level of Service Inventory-Revised (LSI-R; Andrews & Bonta, 2011), the Youth Level of Service/Case Management Inventory (YLS/CMI; Hoge & Andrews, 2011), the Historical-Clinical-Risk Management-20 (HCR-20; Webster et al., 1997), the Violence Risk Appraisal Guide (VRAG; G. T. Harris et al., 2015), and the Structured Assessment of Violence Risk in Youth (SAVRY; Borum et al., 2003).

In the same vein, in our ML models predicting juvenile delinquency, we found strong evidence for the Central Eight risk factors. Antisocial associates (i.e., relationships with individuals who engage in or encourage criminal behavior) were particularly important, especially when combined with any form of drug use. Several variables that also emerged as relevant predictors could be subsumed under antisocial cognition and antisocial personality patterns (e.g., trouble with teachers). Of the Moderate Four, school-related problems (e.g., problems finishing homework) and substance abuse were especially prominent. Somewhat surprisingly, the variables we assigned to the “Romantic Behavior and Sexuality” were also associated with juvenile offenses. On the one hand, this could indicate advanced puberty: Adolescents who mature earlier may experience a mismatch between their biological maturity and societal roles, leading them to engage in delinquency as a means of asserting their independence (Moffitt, 1993). On the other hand, these variables could also be understood as indirect markers of antisocial behavior if the focus is on short-term sexual experiences.

Prediction of Adult Criminal Behavior

It is a common theme in both psychology and predictive modeling that the best predictor of future behavior is past behavior. Consequently, juvenile delinquency played a significant role in predicting adult criminal behavior, which also made the relevant predictor set more homogeneous. Whereas in Wave I, the overlap of predictor sets for the different offenses types (drug-related, property, and violent offenses) included only seven variables, which increased to 16 shared variables in Wave V. This is not because the outcome becomes more unidimensional (the average factor correlation of the three-dimensional model was only slightly higher at .66 vs. .62), but rather because juvenile delinquency is particularly effective at predicting adult criminal behavior. Importantly, this finding does not imply causality, and the explained variance remained relatively low at approximately 15%. Nevertheless, variables assessing drug use, minor theft, and physical fights in adolescence were relevant predictors of later criminal behavior. It therefore seems appropriate that most risk-assessment tools include items assessing the history of antisocial behavior, in particular previous convictions (for an overview see Singh et al., 2018). In addition to prior criminal or transgressive behavior, three key variables—biological sex, the use of legal or illegal drugs, and school-related problems—stood out as important predictors of being charged with crimes in adulthood. These variables will be briefly discussed in the following sections.

Limitations and Future Research

The present analyses and findings come with several limitations, which, however, provide opportunities for future research. In a longitudinal study such as the Add Health Study, which spans a long period of approximately 22 years (1996–2018), several societal shifts and legal changes must be considered when interpreting the results. A clear example is the partial legalization of marijuana: Approximately 2 years after the Wave I assessment, in 1996, California became the first state to legalize medical marijuana through the “Compassionate Use Act.” At Wave V in 2018, marijuana was legalized for recreational use in 11 states and Washington, D.C. (“Legality of Cannabis by U.S. Jurisdiction,” 2025), even though it remains illegal at the federal level under the Controlled Substances Act. Drug-related offenses and their criminal liability, therefore, have a special status compared to other types of crimes considered in the present analyses. In addition to the temporal context, the presented findings are specific to the U.S. population. The extent to which the results, particularly regarding overall predictive accuracy or the most influential variables, generalize to other populations remains a topic for future research.

Another limitation of the present analysis is that we chose the question, “Have you ever been charged with [offense]” as the outcome for Wave V, rather than one of the quantifying questions (“How many times have you been charged with [offense]” or “How many times have you been convicted or pleaded guilty of [offense]”). The answer to the questions concerning the number of charges and convictions, respectively, follows the typical form of count variables—a zero-inflated distribution—which introduces unique challenges related to the modeling process, evaluation metrics, and the interpretation of results (Sidumo et al., 2024). Nonetheless, such count variables can be used to identify extreme subgroups of chronic offenders and then to compare their associations with those observed in this study.

The last limitation to be addressed is the issue of dropout. A potential systematic dropout of men (only 43% participation in Wave V) and individuals likely exhibiting antisocial personality traits, mental health issues, and so on could have decreased the prevalence rates of criminal behavior but not necessarily affect the bivariate relationships. There are two additional arguments against this limitation: First, Brownstein et al. (2018, p. 6) reported in their analysis of Wave IV that “there is little statistical evidence of differences in delinquency levels between non-responders and responders” compared to Wave I. Second, in our assessment, the response rates for such a large-scale study remain sufficiently high to ensure representativeness when using replication weights, as was done in this study.