Using Machine Learning to Identify Key Predictors for Alcohol, Tobacco, and Drug Use in Adolescents

• In Event: Poster Session 05
  In Poster Session: PS 05 Topic - Methods, History, Theory

Abstract

Adolescence is a critical period for the initiation of alcohol, tobacco, and drug use (ATOD). According to recent study, 19.8% of American adolescents used marijuana during the last 30 days, 29% drank alcohol, and 22% smoked (e-)cigarettes (Kann et al., 2018). Early initiation of ATOD use/abuse predicts future problems socially, physically, and psychologically (Friedman, Terras, & Zhu, 2008; McGue & Iacono, 2005). Due to limited sample sizes and heterogeneous American samples, predictors of youth ATOD use vary with multiple regression analyses. Furthermore, the (generalized) linear regression mechanism may fail to account for non-linear relationships. The proposed study will leverage a big data set, the Youth Risk Behavior Survey (YRBS), to identify the key predictors for adolescent ATOD usage with machine-learning approaches (Hindman, 2015; Molina & Garip, 2019). The functional dependence is not constrained by restrictive assumptions but learns from the model training process, opening more possibilities to discover intrinsic relations among variables. We thus expect the sensitivity and precision of the prediction model to be improved when generated by machine-learning techniques.
Methods. The YRBS is a nationwide survey related to youth risk behaviors such as violence, ATOD use, sex behaviors, etc. The 2017 database includes data collected from 1991 to 2017 from 203,363 American 9th-12th graders), with 51% girls, and 42% White, 22% Black, and 27% Hispanic/Latino. For the outcome, seven items related to whether youth used ATODs in the past 30 days were recoded into a dummy variable distinguishing any ATOD use (50.2%) or no use (49.78%).
Analytic Plan. Figure 1 outlines the analytic plan following Sheetal, Feng, and Savani (2020). First, data pre-processing will be performed to create the outcome variable, identify each independent variable’s class (e.g., binary, ordinary), delete observations missing the outcome variable, and exclude variables constructed by the researchers. Second, the data will be randomly split into a model-creation part (n%) and a model-testing part (100-n)%. The optimal percentage will be determined by looping through possible subsets of data as a cross-validation procedure. Third, missing data will be handled for the model-creation and model-testing data separately. Fourth, the model-creation data (n%) will be further randomly divided into k parts using a k-fold cross-validation technique (Anguita et al., 2012) to optimize the parameters for model creation. A total of m iterations will be run to train a prediction model with the model-creation data. Fifth, the prediction model will be tested on the model-testing sample (100-n)%, to evaluate success in predicting ATOD use when generalizing to a new dataset. Cross-entropy loss and other stastical metrics will be used to estimate the accuracy rates for the trained model. Steps four and five will be repeated several times to generate a few trained models applying different machine learning techniques and a traditional logistic regression model. These techniques include neural network modeling, random-forest methods, etc. Finally, certain metrics for predicting ATOD use will be compared among the trained models to see how much they out-perform the traditional logistic regression model. Various packages in Python will be used for the analysis.

Authors

• Hua Lin, Oklahoma State University - Stillwater
  Presenting Author

• Robert E Larzelere, Oklahoma State University - Stillwater
  Non-Presenting Author