Comparing Predictive Models of Pain Reliever Misuse and Abuse
The misuse and abuse of prescription opioids (MUPO) has become a major health crisis in the U.S. Predictive modeling provides a useful approach for analyzing pain reliever misuse and abuse and identifying features that contribute to MUPO. This study compared ten classification models using four performance metrics: accuracy, sensitivity (i.e. recall), precision, and f1-score. Data from three years of the National Survey on Drug Use and Health (2015-2017) was combined into a sample of N = 170317
... ample of N = 170317 observations. Twenty-six percent of respondents reported using prescription opioid medications in the past year, 11% reported misusing or abusing pain relievers, and 2% reported heroin use. The classifier models were fit to a training set using sixteen features that included demographic variables, medications, and illicit drugs. The binary target variable was pain reliever misuse and abuse. Model performance was evaluated on the testing set. The f1-score was used as the performance metric due to unbalanced classes in the data. Logistic regression, random forests, and decision trees had the highest f1-scores compared to other models. All three models identified cocaine use as the most informative feature for predicting pain reliever misuse and abuse. Amphetamine use was selected as the second most important variable by logistic regression and random forests. The models differed in the importance they assigned to heroin use, tranquilizer use, age category, and mental health as predictors of pain reliever misuse and abuse. Advantages and limitations of the classifier models are considered and the tradeoff between model complexity and interpretability is discussed.