Predictors of treatment attrition among individuals in substance use disorder treatment: A machine learning approach

Background: Early treatment discontinuation in substance use disorder treatment settings is common and often difficult to predict. We leveraged a machine learning approach (i.e., random forest) to identify individuals at risk for treatment attrition, and specific factors associated with treatment discontinuation. Method: Participants (N = 29,809) were individuals ≥ 18 years who attended substance use disorder treatment facilities in the United States. Using random forest, we aimed to predict three outcomes (1) leaving against medical advice (AMA), (2) discharging involuntarily, and (3) discharging early for any reason. Predictors included participant demographics, substance use the month before and at intake, indices of mental and physical health, as well as treatment center and program type. Findings: We observed low to moderate area under the curve (range = 0.631–0.671), high negative predictive values (range = 0.853–0.965), and low positive predictive values (0.088–0.336) across the three treatment attrition outcomes. The most robust predictors of the three outcomes included treatment center, treatment type, and participant age. Additional predictors of the three outcomes included employment status; reason for treatment; primary drug at intake and frequency of use; prescription opioid, benzodiazepine, or heroin use at intake; living status at intake; and driving under the influence prior to treatment. Conclusions: Our models were able to accurately identify individuals who remained in treatment, but not those who left treatment prematurely. The most robust predictors of treatment discontinuation were treatment center and program type, suggesting that targeting treatment facility features may have a significant impact on reducing treatment attrition and improving long-term recovery.