Using conventional and machine learning propensity score methods to examine the effectiveness of 12-step group involvement following inpatient addiction treatment. (1st October 2021)

• Record Type: Journal Article
• Authors: Costello, Mary Jean; Li, Yao; Zhu, Yeying; Walji, Alyna; Sousa, Sarah; Remers, Shannon; Chorny, Yelena; Rush, Brian; MacKillop, James
• Is Part Of: Drug and alcohol dependence. Volume 227(2021)
• Publication Date: 2021-10-01
• Subjects: Substance use disorders -- 12-Step -- Causal inference -- Propensity score -- Treatment outcomes -- Abstinence; Drug abuse -- Periodicals; Alcoholism -- Periodicals; 616.86
• DOI: 10.1016/j.drugalcdep.2021.108943
• Languages: English
• ISSNs: 0376-8716

Abstract

Highlights: Used propensity score-based methods to reduce selection bias (or confounding). Applied conventional and machine learning techniques (RF and BART). High 12-step involvement demonstrated a positive effect on 12-month abstinence. Each method effectively reduced potential confounding influences on the outcome. Convergence in findings across methods strengthens confidence in the effect. Abstract: Background: Continuing care following inpatient addiction treatment is an important component in the continuum of clinical services. Mutual help, including 12-step groups like Alcoholics Anonymous, is often recommended as a form of continuing care. However, the effectiveness of 12-step groups is difficult to establish using observational studies due to the risks of selection bias (or confounding). Objective: To address this limitation, we used both conventional and machine learning-based propensity score (PS) methods to examine the effectiveness of 12-step group involvement following inpatient treatment on substance use over a 12-month period. Methods: Using data from the Recovery Journey Project – a longitudinal, observational study – we followed an inpatient sample over 12-months post-treatment to assess the effect of 12-step involvement on substance use at 12-months (n = 254). Specifically, PS models were constructed based on 34 unbalanced confounders and four PS-based methods were applied: matching, inverse probability weighting (IPW), doubly robust (DR) with matching, and DR with IPW. Results: Each PS-based method minimized the potential of confounding from unbalanced variables and demonstrated a significant effect ( p < 0.001) between high 12-step involvement (i.e., defined as having a home group; having a sponsor; attending at least one meeting per week; and, being involved in service work) and a reduced likelihood of using substances over the 12-month period (odds ratios 0.11 to 0.32). Conclusions: PS-based methods effectively reduced potential confounding influences and provided robust evidence of a significant effect. Nonetheless, results should be considered in light of the relatively high attrition rate, potentially limiting their generalizability.