Objective: This study aims to 1) examine the predictors of drug use at high school 2) build a predictive model for drug use using artificial neural network and compare its performance to logistic regression model. Methods: Youth Risk Behavior Surveillance System (YRBSS) 2015 data were used for this study. The YRBSS was developed in 1990 to monitor priority health risk behaviors that contribute markedly to the leading causes of death, disability, and social problems among youth and adults in the

more »

... United States. All the participants who were eligible were randomly assigned into 2 groups: training sample and testing sample. Two models were built using training sample: artificial neural network and logistic regression. We used these two models to predict the risk of Drug Usein the testing sample. Receiver operating characteristic (ROC) were calculated and compared for these two models for their discrimination capability and a curve using predicted probability versus observed probability were plotted to demonstrate the calibration measure for these two models. Results: About 18.1% of 8711 students were drug users, about 19.1% among the female and 17.1% among the male. According to the logistic regression, students who had rides in a car driven by someone who is drinking were more likely to have drug use. Students who never tried cigarette smoking were less likely to use drug. Students who drank often were more likely to use drug. Student who used marijuana often were more likely to use drug. Heterosexual students were less likely to use drug. Students who slept 4 hours or less daily were more likely to use drug. Students who did not speak English well were less likely to be a drug user. According to this neural network, the top 5 most important predictors were 'being black', Q99 (How well do you speak English), 'being Asian', Q68 (sexual orientation), Q88 (On an average school night, how many hours of sleep do you get?). For training sample, the ROC was 0.84 for the Logistic regression and 0.88 for the artificial neural network. Artificial neural network performed better clearly. In testing sample, the ROC was 0.83 for the Logistic regression and 0.80 for the artificial neural network. Artificial neural network had worse performance. As to calibration measure, predictions made by the neural network are (in general) less concentrated around the 45-degree line (a perfect alignment with the line would indicate an ideal perfect calibration) than those made by the Logistic model. Conclusions: In this study, we identified several important predictors for drug use e.g., cigarette smoking,drinking, sexual orientation. This provided important information for educators as well as parents provide timely intervention. We built a predictive model using artificial neural network as well as logistic regression to provide a tool for early detection. As to performance of these two models, logistic regression and neural network had a similar discriminating capability.