Modeling tobacco dependence using penalized regression and machine learning: The predictive roles of emotions, habits, and intention to quit

This study aimed to predict nicotine dependence levels among university students in Türkiye by modeling the Fagerström Test for Nicotine Dependence (FTND) score as a continuous outcome using penalized regression and machine learning (ML) approaches. A cross-sectional study was conducted with a stratified sample of 1,120 university students, examining 24 psychological, behavioral, motivational, and sociodemographic variables. Six predictive models were compared: linear regression, Least Absolute Shrinkage and Selection Operator (LASSO), Ridge regression, Elastic Net, Support Vector Regression (SVR), and tuned extreme gradient boosting (XGBoost). Penalized regression models and the tuned XGBoost algorithm demonstrated largely comparable predictive performance. XGBoost achieved the highest predictive accuracy (<i>R</i>² = 0.621; adjusted <i>R</i>² = 0.575) and the lowest RMSE (1.678). However, the LASSO model yielded nearly identical performance (<i>R</i>² = 0.620; RMSE = 1.680) while offering greater interpretability through embedded variable selection. Given the negligible difference in predictive accuracy, LASSO was selected as the primary model due to its clinical utility and ability to identify modifiable risk factors. The LASSO model explained approximately 59% of the variance in FTND scores and identified 11 significant predictors, including emotional symptoms (e.g., irritability and depressive mood), habitual smoking, intention to quit, daily cigarette consumption, and family smoking. This study highlights the value of interpretable modeling approaches and underscores that nicotine dependence is shaped by a complex interplay of psychosocial, behavioral, and demographic determinants in addition to physiological factors.