Predictors of Abstinence From Heavy Drinking During Follow-Up in COMBINE

Objective: Although the primary focus of clinical trials is on between-group comparisons during treatment, these studies can also yield insights into which patient characteristics predict longer term outcomes. Our goal was to identify predictors of good outcome during the 1-year follow-up in the Combined Pharmacotherapies and Behavioral Interventions for Alcohol Dependence (COMBINE) Study. Method: We constructed classification trees and a deterministic forest to predict no heavy drinking days during the last 8 weeks of the 1-year follow-up in COMBINE, based on more than 100 baseline predictors and drinking outcomes during the treatment phase of the study. The COMBINE sample was randomly split into a training and a validation data set. Logistic regression models were fit to compare the predictive performance of tree-based methods and classical methods. Results: A small tree with only two splits and four nodes based on abstinence and good clinical outcome during treatment had fair classification accuracy in the training and the validation samples: area under the curve (AUC) of 71% and 70%, respectively. Drinking outcomes during treatment were the strongest predictors in the deterministic forest. Logistic regression analyses based on four main effects (good clinical outcome, level of drinking during treatment, age at onset of alcohol dependence, and feeling more energetic) had slightly better classification accuracy (AUC = 74%). Conclusions: End-of-treatment outcomes were the strongest predictors of long-term outcome in all analyses. The results emphasize the importance of optimizing outcomes during treatment and identify potential subgroups of individuals who require additional or alternative interventions to achieve good long-term outcome.