The primary objective of randomized clinical trials is to assess average treatment effects. However, due to between-subject heterogeneity, treatments may work only in a subset of the population and may not work or may be even counter-therapeutic in another subset. For such treatments, average effect may be small to non- existent and hence these treatments may be underutilized in a population for whom they might provide significant benefit. Since alcohol dependence is characterized by high patient heterogeneity and treatment effects have been shown to be in the small to medium range, it is crucial to identify specific covariates (moderators) that stratify the population into subgroups fo which treatments have differential effects. The usual approach has been to consider baseline predictors one at a time or to test treatment effects among predefined endophenotypes. This limits the potential for discovery of important combinations of predictor variables that might moderate treatment effects. In this application we propose to apply tree-based and forest-based methods that address the limitations of considering predictors one at a time. Decision trees allow identification of subgroups of subjects for whom there are significant differences in effectiveness of treatments. They can consider a large pool of predictor variables and empirically derive the strongest predictors/moderators that identify subgroups of patients with differential treatment effects. Compared to classical classification methods, tree-based methods may be easier to use in clinical settings because they require evaluation of simple decision rules rather than mathematical equations. We will apply the methods to the U.S.-based COMBINE study that evaluated efficacy of naltrexone, acamprosate and CBI;consider the results in the context of the subject-matter literature;and based on the results for COMBINE formulate and test hypotheses using classical statistical methods on the European PREDICT study that was designed to be comparable to COMBINE. Our goal is to identify robust predictors and moderators of treatment effects during treatment (Specific aims 1 and 2 respectively) and during follow-up (Specific aim 3) and to inform clinicians making treatment decisions in different subgroups of alcohol- dependent patients using simple to interpret and externally validated decision rules.
This project will provide easily interpretable and bust decision rules identifying subgroups of patients with good outcome irrespective of the type of treatment and subgroups for which specific treatments or treatment combinations may be most effective. The rules will be developed using a sophisticated and powerful approach called tree-based methods on data from the COMBINE Study that tested the efficacy of naltrexone, acamprosate and a specialized behavioral intervention for alcohol dependence. The results will be externally validated on the PREDICT study that tested the efficacy of naltrexone, acamprosate and CBI as an augmentation strategy in a different population. The ultimate goal of this approach is to inform clinicians making treatment decisions for different subgroups of alcohol-dependent patients and to present the decision rules in simple form that can be widely applied in clinical practice.
