Adolescent delinquency has increased in severity and frequency in South Korea. With the recent surge in assault, fraud, and sexual offense among adolescents, practitioners and researchers have emphasized the need to take preventive measures for identifying adolescents with heightened risk. Theories point to various predictors of adolescent delinquency: for example, individual traits, ties with the conventional society (social bond), association with delinquent peers (social learning), victimization (strain), prior delinquency (labeling), and community characteristics (social disorganization). However, methodologically, considering a comprehensive set of theoretically relevant variables that predict probation in a single model is difficult in regression models. The current study applies machine learning techniques using a vast pool of variables to predict probation experience among adolescents in South Korea.
The current study integrated two independent datasets: (a) Data from the nationally representative Fourth International Self-Report Delinquency Study (ISRD4) of South Korean adolescents were used (n=1422, boy=47.40%, age=15.59); and (b) data using the same ISRD questionnaire were collected from youth on probation in South Korea (n=301, boy=80.07%, age=16.69). We applied machine learning analysis using a random forest algorithm to predict adolescents on probation and examined partial dependence plots with the top five important predictors. We also used 10-fold cross-validation to prevent overfitting and evaluated the model performance using AUC and accuracy. The dependent variable was probation experience (0=ISRD sample, 1=probation sample). A total of 90 predictors were used such as demographic variables (e.g., gender, migration status), micro variables (e.g., impulsivity, delinquency), and exo variables (e.g., community crime, time spent online).
The AUC was 0.96 for the mean of the 10-fold train dataset and 0.95 for the test dataset, meaning excellent discrimination. Accuracy was 0.92 with train data and 0.93 with test data in the classification of youth on probation. Random forest analysis results showed that the most crucial predictor in determining adolescents on probation was participating in part-time jobs during adolescents’ free time. Next, adolescents’ perceptions of computer crime/police, robbery, police contact, and age appeared as important predictors in order. Partial dependence plots indicated that adolescents who engaged in more part-time jobs in their free time, had a positive perception toward police, had robbery experience, contacted police before, and were older, were more likely to experience probation.
Study results supported social bond theory's claim that adolescent’s time spent in structured activities, under the supervision of adults is an important protective factor against delinquency. This is an important finding given the emphasis on education in South Korea, and hence, adolescent leisure and work has received little attention for understanding the lives of adolescents. In South Korea, adolescent recurrence in probation is nearly twofold than that of adults. By identifying prior experience of delinquency and police contact as major predictors of probation as proposed by labeling theory, our results point to the seriousness of this issue, and address the need for allocating resources to prevent repetition.