(Converted as ePoster, See Poster Gallery) Predicting School Dropout Among South Korean Adolescents: A Machine Learning Approach

Background/Purpose:

Dropping out of school is associated with negative consequences in various domains in adolescence and across the life-course. However, a large body of literature on school dropout has focused on the western context, limiting our ability to apply information about risk factors to other socio-cultural settings. Also, research on school dropout has primarily focused on estimating parameters to identify whether independent variables are associated with the dependent variable with statistical significance. For intervention purposes however, accurate prediction of adolescents with the greatest risk of school dropout is imperative. The current study applies an ecological systems framework and machine learning techniques to identify a set of variables that are most important in accurately predicting the experience of school dropout among South Korean adolescents.

Methods:

Data from two independent panel studies collected by National Youth Policy Institute in South Korea were used in this study: Dropout Youth Panel Study (DYPS; N=609, age=16.84, girls=56.16%) and Korean Children and Youth Panel Study (KCYPS; N=1,646, age=15.90, girls=50.73%). We applied machine learning algorithms to predict school dropout using three analytic methods: logistic regression, random forest, and decision tree. School dropout was made based on the information to which the dataset adolescents belong (0=KCYPS, 1=DYPS). Total of 36 predictors including personal variables (e.g., gender, depression, delinquent behaviors), micro variables (e.g., family structure, peer attachment, teacher relationship, academic involvement), and exo variables (e.g., collective efficacy, media time use on weekend/weekdays) were used in the analyses. The rate of training data was set at 70 and 10-fold cross validation was used to train the dataset. The remaining data set of 30 was used as test data. The model performances were evaluated using AUC and accuracy.

Results:

Among three analyses, there were featured predictors of school dropout. Results from random forests showed that relative importance of predictors were truancy, smoking, media time use on weekdays, drinking, family structure, collective efficacy, teacher relationship, and group bullying victimization in order. Decision tree analysis revealed that truancy has the largest effect on the classification of school dropout. Logistic regression showed that truancy (AOR=27.54, p<0.001), group bullying victimization (AOR=4.93, p<0.001), spending time in media on weekdays (AOR=4.45, p<0.001), smoking (AOR=2.39, p<0.001), drinking alcohol (AOR=1.86, p<0.05), low teacher relationships (AOR=0.52, p<0.01), living without both parents (AOR=0.29, p<0.001), and low collective efficacy (AOR=0.27, p<0.001) were related to school dropout. The performance of machine learning models ranged 0.83-0.93 of AUC and 0.88-0.89 of accuracy in the classification of school dropout.

Conclusions/Implications:

In using machine learning techniques, we were able to extract the salient predictors for identifying school dropout among a host of factors across adolescent’s ecological systems. All ecological systems--including personal characteristics (truancy, smoking, drinking), micro (family structure, teacher relationship, group bullying victimization), exo (media time use on weekdays, collective efficacy)--were consistently identified as significant predictors of school dropout across various models. This information may be used to guide school-level prevention efforts by allowing social workers and educators to develop early warning systems against school dropout and accurately screen adolescents with high risk.