Predictors of Abstinence, No Heavy Drinking Days, and a 2-Level Reduction in World Health Organization Drinking Levels During Treatment for Alcohol Use Disorder in the COMBINE Study

Abstract

Background:

Data from trials of medications for alcohol use disorder (AUD) can be used to identify predictors of drinking outcomes regardless of treatment, which can inform the design of future trials with heterogeneous populations. Here, we identified predictors of abstinence, no heavy drinking days, and a 2-level reduction in World Health Organization (WHO) drinking levels during treatment for AUD in the Combined Pharmacotherapies and Behavioral Interventions (COMBINE) Study.

Methods:

We utilized data from the COMBINE Study, a randomized placebo- controlled trial evaluating the efficacy of naltrexone and acamprosate, both alone and in combination, for AUD (n = 1168). A tree-based machine learning algorithm was used to construct classification trees predicting abstinence, no heavy drinking days, and a 2-level reduction in WHO drinking levels in the last 4 weeks of treatment, based on 89 baseline variables.

Results:

The final tree for predicting abstinence had one split based on consecutive days abstinent prior to randomization, with a higher proportion of subjects achieving abstinence among those classified as abstinent for >2 versus ≤2 consecutive weeks prior to randomization (66% vs. 29%). The final tree for predicting no heavy drinking days in the last 4 weeks of treatment had three splits based on consecutive days abstinent, age, and total Alcohol Dependence Scale score at baseline. Seventy-three percent of the subjects classified as abstinent for >2 consecutive weeks prior to randomization had no heavy drinking days in the last 4 weeks of treatment. Among those classified as abstinent ≤2 consecutive weeks prior, three additional splits showed that younger subjects (age ≤44 years; 37%), and older subjects (age >44) with a total Alcohol Dependence Scale score >13 and complete abstinence (56%) or other drinking goals (35%), were less likely to have no heavy drinking days than older subjects with a total Alcohol Dependence Scale score ≤13 (67%). The final tree for predicting a 2-level reduction in WHO levels had no splits.

Conclusions:

Consecutive days abstinent prior to randomization may predict absti- nence and no heavy drinking days and total Alcohol Dependence Scale score and age may predict no heavy drinking days. The 2-level reduction in WHO levels outcome may be less likely to discriminate based on multiple patient characteristics.

Introduction

Clinical trials of medications for alcohol use disorder (AUD) typically contain heterogeneous groups of individuals,(Litten et al., 2015) which can make it difficult to identify active medication treatment effects versus placebo.(Falk et al., 2019b, Gueorguieva et al., 2015b) Although moderator (i.e., subgroup) analyses can be conducted to determine whether there are any potential between-group treatment differences in medication effects by baseline variables,(Gueorguieva et al., 2014) analyses focused on identifying predictors of drinking outcomes regardless of treatment can inform the design of future trials. In particular, these analyses can help identify patient characteristics on which to stratify randomization, to ensure balanced treatment groups, and whether future trials should focus on the recruitment of specific participants (e.g., those who are less likely to have good outcomes regardless of treatment).(Gueorguieva et al., 2014) While previous studies have identified a number of consistent predictors of alcohol treatment outcomes, including motivation, alcohol-related self-efficacy, psychopathology, and alcohol dependence severity,(Adamson et al., 2009) these findings are based on traditional subgroup analyses, which consider one variable at a time, and fail to account for the fact that patient subgroups are more often defined by combinations of different characteristics.(Gueorguieva et al., 2015b)

Unlike traditional subgroup analyses, tree-based machine learning methods allow for the identification of combinations of patient characteristics that may be associated with beneficial outcomes.(Zhang H, 2010) These methods, which can overcome the assumptions required for standard linear model methods (e.g., normality and linearity), recursively partition trial populations into similar groups based on an outcome of interest. By allowing a large pool of predictors variables to be considered, tree-based analyses can be particularly informative for clinical settings, where decision rules can be made based on multiple patient variables. According to a previous tree-based analysis of COMBINE Study, the largest United States trial of pharmacotherapy (naltrexone, acamprosate) for AUD to date, consecutive days abstinence prior to randomization and drinking goals were found to be reliable predictors of no heavy drinking days.(Gueorguieva et al., 2015a, Gueorguieva et al., 2014) However, that evaluation used two older approaches to construct trees, and it is unclear which of the same predictors would be identified for the same outcomes using newer machine learning approaches with potentially more favorable characteristics (Supplementary Table 1). Moreover, to the best of our knowledge, no previous evaluations have assessed the relationship between predictors and two additional drinking reduction outcomes, abstinence and the 2-level reduction in World Health Organization (WHO) drinking levels (e.g. WHO 2-level reduction from high to low risk based on risk drinking level).

To guide AUD medication approval, the FDA emphasizes “responder” outcomes, which include total abstinence (i.e., patients who do not drink at all) or no heavy drinking days (i.e., patients who do not have any heavy drinking days) during a given period.(FDA, February 2015) Although abstinence has historically been considered the most desirable AUD goal,(Witkiewitz et al., 2017) drinking reductions can lead to improvements in psychosocial and physical function, mortality, and morbidity,(Kline-Simon et al., 2013, Kline-Simon et al., 2017, Laramée et al., 2015, Witkiewitz, 2013) and may be the preferred treatment target for many individuals.(Aubin and Daeppen, 2013, Gastfriend et al., 2007) Recently, there have been proposals to include a 2-level reduction in WHO drinking levels as the primary endpoint for AUD clinical trials.(Falk et al., 2019a, Witkiewitz et al., 2017) Studies have evaluated the sensitivity and validity of the endpoint,(Hasin et al., 2017, Witkiewitz et al., 2017) and the 2-level reduction has already been endorsed by the European Medicine Agency.(Witkiewitz et al., 2017) If the 2-level reduction in WHO drinking levels outcome will guide future drug approvals,(Falk et al., 2019a) it is important to assess whether combinations of patient characteristics may be associated with this outcome of AUD trials.

In order to inform the design of future AUD trials, we sought to identify potential predictors of abstinence, no heavy drinking days, and a 2-level reduction in WHO drinking levels using data from the COMBINE Study. According, we utilized a machine learning algorithm to construct classification trees predicting these responder and drinking reduction outcomes, based on 89 baseline predictors.

Materials and Methods

Study Sample

In the COMBINE Study, participants were randomized in a 2 × 2 × 2 design to one of eight conditions that received medical management with 16 weeks of naltrexone (100 mg/day) or acamprosate (3 g/day), both, and/or both placebos, with or without a combined behavioral intervention.(Anton et al., 2006) The COMBINE Study included subjects with AUD with more than 14 (women) or 21 (men) drinks per week, at least 2 heavy drinking days during a consecutive 30-day period within the 90 days prior to the baseline evaluation, and attained at least 4 days of abstinence.(Anton et al., 2006) For the current analyses, we excluded participants whose drinking at baseline was classified as low-risk according to WHO drinking criteria (i.e., ≤ 20 grams/day women and ≤ 40 grams/day for men) (52 of 1220, 4.3% of the analytic sample). These participants are generally not enrolled in AUD trials and would not be able to attain a 2-level reduction in WHO drinking levels (e.g., they can only reduce one level to abstinence).

Drinking Outcomes

Abstinence

Abstinence from alcohol was defined as no drinking during the last 4 weeks of double-blind treatment period. There were 70 (6.0%) subjects coded as not abstinent due to missing data.

No heavy drinking days

No heavy drinking days was defined as no heavy drinking during the last 4 weeks of double-blind treatment. As a secondary analysis, we also evaluated no heavy drinking during the last 8 weeks of double-blind treatment. This will allow our findings to be compared to a previous evaluation of predictors using data from the COMBINE Study and different classification tree methodology, where no heavy drinking was measured during the last 8 weeks of double-blind treatment.(Gueorguieva et al., 2014) There were 67 (5.7%) subjects coded as heavy drinking due to missing values.

WHO 2-level reduction

Using previously described methods,(Witkiewitz et al., 2017) WHO risk drinking levels were calculated using patient reports (Form-90 and Timeline Follow-Back interview) of the number of standard drinks, defined as 0.6 ounces of absolute alcohol consumed and converted to grams of pure alcohol (0.6 ounces = 14 g). WHO risk levels were then calculated based on the average grams of alcohol consumed per day (i.e., drinks per day). Although abstinence was not included as a separate category in the original WHO risk levels, we included it as the lowest risk level. For the baseline period, we used data from the 4 weeks prior to the screening appointment. For the within-treatment WHO risk level, WHO risk was defined as the average grams of alcohol consumed over the last 4 weeks of the study. A reduction in risk level was computed by subtracting the end-of-treatment WHO risk levels from the baseline WHO risk level. Missing data was coded as no-change (i.e., non-response).

Predictors

All potential baseline predictors were identified in the COMBINE Study protocol and publication, as previously described.(Gueorguieva et al., 2015b) For continuous variables, missing values were imputed using multiple imputation (PROC MI, SAS). For categorical variables, an additional missing category was created. For original and continuous variables with >5 levels were re-classified into 4–5 categories to avoid over-representation as splitters and to improve interpretability.(Gueorguieva et al., 2015b) All 89 predictors are described in Supplementary Table 2.

Tree-based analyses

After evaluating recent reviews focused on subgroup identification methods for precision medicine,(Loh et al., 2019) we selected the Generalized, Unbiased, Interaction Detection and Estimation (GUIDE)(Loh, 2021) tree-based machine learning method to generate classification trees. GUIDE is a machine learning algorithm that recursively partitions data to form binary trees. GUIDE was selected based on the favorable properties and features, including the ability to consider many categorical and continuous predictor variables and minimize bias in split variable selection.(Loh, 2021, Loh et al., 2019) In each step of tree building, chi-squared tests are used to evaluate associations between predictor variables and the outcome of interest. GUIDE selects the most important predictors, which are the tree splitters, and the data are subdivided into groups that are as homogeneous on the outcome as possible. Successive partitioning is used to identify the unique combinations of predictor variables and predictor cut-off points that are the best at capturing subgroups that differ the most on the outcome.(Kim et al., 2019) After the GUIDE approach grows a large tree, a 10-fold cross-validation pruning method is used to reduce the size of the tree. For our analyses, we utilized the default parameters of GUIDE, including unit misclassification costs and estimated priors.(Loh, 2021) With estimated priors, the assumption is that our sample is a simple random sample of the target population.(Loh, 2021) Each terminal node is assigned to the outcome with the posterior probability that is higher than its counterpart in the root node.(Loh and Zhou, 2021) Lastly, to provide additional information about the potentially important variables considered by the GUIDE approach, we reported the 10 most important predictors, in order, at the root node, which are identified by using a weighted sum of chi-squared statistics obtained from a short (four-level) unpruned tree with conditional tests for pairwise interactions and a permutation-based step for bias correction.(Loh and Zhou, 2021) The GUIDE approach does not provide information about the direction of prediction for these variables unless they are included in the final classification trees.

Results

After excluding participants whose drinking at baseline was classified as low risk according to WHO drinking criteria, there were 1168 eligible participants. Of these, 398 (34.1%) achieved abstinence, 483 (41.4%) had no heavy drinking days, and 850 (72.8%) achieved a 2-level reduction in WHO drinking levels in the last 4 weeks of treatment.

Abstinence

For abstinence in the last 4 weeks of treatment, the top 10 predictors at root node were consecutive days abstinent prior to randomization, drinking goal, creatinine, total Alcohol Dependence Scale score, prior detoxification treatment, Alcoholics Anonymous attendance, University of Rhode Island Change Assessment Scale overall readiness, total bilirubin, having received naltrexone versus placebo naltrexone, and total protein at baseline (Table 1).

Table 1.

Top ranked variables at the root node

Predictors	Abstinence	Abstinence from heavy drinking	2-level reduction in WHO drinking levels
Treatment
Received naltrexone vs. placebo naltrexone	9
Received CBI and MM vs. MM only			5
Demographic
Age		2	8
Marital status		10	2
Race/ethnicity			6
Alcohol consumption, severity, and goals
Consecutive days abstinent prior to randomization	1	1
Alcohol Dependence Scale	4	3
Age of onset		5	10
Drinking goal	2	4
WHO drinking risk levels 3 months prior to baseline			1
Percent days abstinent			3
Percent no heavy drinking			4
Peak blood alcohol concentration		6
Prior alcohol treatment
Detoxification	5
Alcoholics anonymous attendance	6
Laboratory analyses
Creatinine	3	5
Total protein	10	8	9
Bilirubin	8
Glucose			7
Smoking and drug use
Current smoker		6
Cannabis use
Profile of Mood States (POMS)
POMS Confusions Subscale		7
POMS Fatigue Subscale
University of Rhode Island Change Assessment Scale	7
WHO Quality of Life
WHO Environmental Domain		9

The final tree for predicting abstinence in the last 4 weeks of treatment had one root-node split based on consecutive days abstinent prior to randomization (Figure 1). According to the tree, a higher proportion of subjects (66%) achieved abstinence among those classified as abstinent for >2 consecutive weeks compared to those who were classified as abstinent for ≤2 consecutive weeks before treatment (29%).

Figure 1. Final classification tree for predicting abstinence.

At each split, an observation goes to the left branch if and only if the predictor condition/splitting value listed to the left of the node is satisfied. The total sample sizes are printed below the final (terminal) nodes. The predicted proportions of individuals achieving abstinence are printed to the right or left of the terminal nodes, and are colored yellow if they are below 50% and orange if they are above 50%. The second best split variable at root node is drinking goal.

No heavy drinking days

For no heavy drinking days in the last 4 weeks of treatment, the top 10 predictors were consecutive days abstinent prior to randomization, age, total Alcohol Dependence Scale score, drinking goal, age at onset, peak blood alcohol concentration, Profile of Mood States Confusion Subscale, total protein, WHO Quality of Life Environment Domain, and marital status at baseline (Table 1).

The final classification tree for predicting no heavy drinking during the last 4 weeks had 4 splits based on consecutive days abstinent prior to randomization (root node), age, Alcohol Dependence Scale score, and drinking goal (Figure 2). However, there was no significant split variable for the drinking goal. Nearly three-quarters (73%) of the subjects classified as abstinent for >2 consecutive weeks prior to randomization had no heavy drinking days in the last 4 weeks of treatment. Among those classified as abstinent ≤2 consecutive weeks prior to randomization, three additional splits showed that younger subjects (age ≤44 years; 37%), and older subjects (age >44) with a total Alcohol Dependence Scale score >13 and complete abstinence (56%) or other drinking goals (35%), were less likely to have no heavy drinking days in the last 4 weeks of treatment than older subjects with a total Alcohol Dependence Scale score ≤13 (67%).

Figure 2. Final classification tree for predicting no heavy drinking days.

At each split, an observation goes to the left branch if and only if the predictor condition/splitting value listed to the left of the node is satisfied. The total sample sizes are printed below the final (terminal) nodes. The predicted proportions of individuals achieving abstinence are printed to the right or left of the terminal nodes, and are colored yellow if they are below 50% and orange if they are above 50%. The circle with dashed lines represents a node with no significant split variables. Even when no statistically significant split variables are found, GUIDE continues to split the node using the most significant of the non-significant variables. It then relies on pruning to find out if the split should be retained. This approach accounts for the fact that good splits may result in later nodes. In this example, individuals with complete abstinence (1) or other drinking goals (0) go to the right branch. The second best split variable at root node, which is not included in the final classification tree, is age.

2-level reduction in WHO levels

For a 2-level reduction in WHO levels in the last 4 weeks of treatment, the top predictors, in order from 1 to 10, were WHO drinking risk levels 3 months prior to baseline, marital status, percent days abstinent, percent no heavy drinking days, having received Combined Behavioral Intervention and medical management vs. medical management only, race (white vs. other), glucose, age, total protein, and age of onset of alcohol consumption.

The final classification tree for predicting a 2-level reduction in WHO levels using estimated priors had no splits (Figure 3).

Figure 3. Final classification tree for predicting a 2-level reduction in WHO levels.

No splits were identified using the GUIDE approach.

Sensitivity analyses

For no heavy drinking days during the last 8 weeks of treatment, the top 10 predictors were consecutive days abstinent prior to randomization, age, drinking goal, total Alcohol Dependence Scale score, creatinine levels, smoking status, Profile of Mood States Confusion Subscale, POMS Fatigue Subscale, cannabis use, and WHO Quality of Life Environment Domain. The final classification tree for predicting no heavy drinking had three splits based on consecutive days abstinent prior to randomization (root node), age, and Alcohol Dependence Scale scores at baseline (Figure S1). Over two- thirds (67%) of the subjects classified as abstinent for >2 consecutive weeks at baseline had no heavy drinking days. Among those classified as abstinent ≤2 consecutive weeks at baseline, two additional splits showed that younger subjects (age ≤44), and older subjects (age >44 years) with a total Alcohol Dependence Scale score >13, were less likely to have no heavy drinking days (28% and 36%, respectively) than older subjects with a total Alcohol Dependence Scale score ≤13 (57%).

Discussion

In this secondary evaluation of shared individual participant data from the COMBINE Study, the largest U.S. trial of pharmacotherapy for AUD to date, we used a tree-based machine learning algorithm to identify potential predictors of abstinence, no heavy drinking days, and a 2-level reduction in WHO drinking levels in the last 4 weeks of treatment. We found that there were at least 2 outcomes where consecutive days abstinent, total Alcohol Dependence Scale score, drinking goal, creatinine, total protein, and age were among the top 10 predictors. According to the final classification trees, consecutive days abstinent prior to randomization may be a strong predictor of abstinence and no heavy drinking days in the last 4 weeks of treatment. In addition, Alcohol Dependence Scale scores and age may predict no heavy drinking days. However, the final classification tree for predicting a 2-level reduction in WHO levels had no splits. These findings highlight opportunities to use the results from classification trees to identify and recruit specific subjects into future trials that measure abstinence and no heavy drinking days, and suggest that the 2-level reduction in WHO risk drinking level may be a useful endpoint for all subjects.

Beyond consecutive days abstinent prior to randomization, which was identified in the final classification trees for abstinence and no heavy drinking days in the last 4 weeks of treatment, there were several baseline predictors – Alcohol Dependence Scale scores, drinking goal, creatinine, total protein, age, and age of onset - that came out among the top 10 ranked variables at the root node for at least 2 outcomes. Dependence severity, which can be measured using several different questionnaires, is one of the most widely studied predictors of AUD treatment outcomes.(Adamson et al., 2009) In fact, according to a previous review, across all dependence severity measures, nearly half of the individual studies found that lower severity was associated with a better outcome.(Adamson et al., 2009) Pretreatment drinking goal, especially complete abstinence, has also been found to be associated with various positive treatment outcomes.(Bujarski et al., 2013) In tree-based and logistic regression analyses of the COMBINE Study, considering the same predictors, and PREDICT, a smaller trial among AUD patients from inpatient facilities randomly assigned to naltrexone, acamprosate, or placebo, drinking goal and Alcohol Dependence Scale score emerged as important predictors.(Gueorguieva et al., 2014) Since we also identified total Alcohol Dependence Scale score as a potentially important splitting variable for no heavy drinking days, assessing total Alcohol Dependence Scale score and drinking goals of trial participants at the beginning of trials could potentially help identify the most appropriate interventions to achieve the outcomes desired by patients (e.g. abstinence vs. drinking reduction outcomes). (Yoshimura et al., 2021) (O’Malley et al., 2015) Although these characteristics may have value in matching individuals to treatment in research and clinical practice, this would need to be formally evaluated.

Creatinine was identified as a potentially important predictor for abstinence and no heavy drinking in the last 4 weeks of treatment. This finding may not be surprising because previous studies have suggested that drinkers in higher alcohol intake categories have reduced creatinine levels compared to nondrinkers.(Schaeffner et al., 2005) Another laboratory measure, total protein, was identified as a potentially important predictor for all outcomes. Patients with AUD often have protein deficiencies, which may contribute to liver disease and other alcohol-related disorders.(Lieber, 2004) Unlike a previous tree-based analyses of the COMBINE Study, gamma-glutamyl transferase level was not identified as a potentially important laboratory measure.(Gueorguieva et al., 2014) However, similar to that evaluation, we found that age at baseline was a potentially important predictors for no heavy drinking. Typically, demographic variables have been found to be poor predictors of treatment outcomes in AUD clinical trials,(Adamson et al., 2009) but older age tends to be associated with better outcomes across treatment groups.(Gueorguieva et al., 2014) Although not all of the laboratory predictors were included in the final classification trees, these measures are routinely obtained at baseline for AUD pharmacotherapy studies. Therefore, their predictive abilities can be assessed in future evaluations, and pharmacotherapy studies could look at improvements in creatinine and total protein at the end of treatment.

We found that consecutive days abstinent prior to randomization predicted abstinence and no heavy drinking days in the last 4 weeks of treatment. For both outcomes, approximately two-thirds of the subjects classified as abstinent for >2 consecutive weeks achieved abstinence or had no heavy drinking days. Studies have consistently suggested that the ability to sustain abstinence before treatment can decrease the likelihood of heavy drinking.(Stout, 2000) For instance, in our previous evaluation using different classification tree, deterministic forest, and logistic regression-based analyses, two consecutive weeks of abstinence prior to randomization was found to be the optimal split across multiple classification tree approaches for abstinence from heavy drinking during the last 8 weeks of treatment.(Gueorguieva et al., 2014) These findings were confirmed in an external validation using data from PREDICT.(Gueorguieva et al., 2014) However, unlike in the COMBINE Study, consecutive days abstinent was defined as >3 weeks of abstinence in PREDICT, which may indicate that even longer pre-treatment periods of abstinence also predict a good response.(Gueorguieva et al., 2014) Altogether, these findings suggest that individuals with lower levels of abstinence prior to starting treatment may benefit from additional support to decrease their likelihood of heavy drinking and increase their likelihood of achieving abstinence. Furthermore, given that subjects who were able to maintain abstinence for >2 weeks are more likely to have good outcomes, these individuals may not be as appropriate for evaluations of novel treatments for AUD. Therefore, future trials could save resources by focusing their recruitment on subjects who do not achieve this level of pre-treatment abstinence. At a minimum, pretreatment abstinence should be considered as a variable for balancing or stratifying randomization to treatment groups, which has been successful in trials of extended-release naltrexone and varenicline.(Garbutt et al., 2005, O’Malley et al., 2018)

The final classification tree for predicting a 2-level reduction in WHO levels in the last 4 weeks of treatment had no splits. Our findings may suggest that unlike abstinence and no heavy drinking days, this outcome is less likely to discriminate based on multiple patient characteristics. This feature may further justify the use of the 2-level reduction in WHO levels as a primary efficacy endpoint, along with total abstinence and no heavy drinking days, for phase-3 AUD pharmacotherapy clinical trials supporting regulatory approvals in the United States.(Witkiewitz et al., 2017) However, it is also possible that our findings reflect that fact that AUD is a multifaceted disorder. Moreover, with nearly three-quarters of the COMBINE Study population achieving a 2-level reduction in WHO drinking levels in the last month of treatment, it may be more difficult to discriminate who will be successful on this endpoint. Both of these characteristics can make it difficult to identify strong predictors.(Gueorguieva et al., 2015a, Gueorguieva et al., 2014, Gueorguieva et al., 2015b) Future evaluations, using the same or different classification tree approaches, should evaluate whether predictors are identified for a 2-level reduction in WHO drinking levels in AUD trials.

This study has several limitations. First, we focused our evaluation on one machine learning method - the GUIDE approach. However, according to our own searches and a previous systematic evaluation of statistical methods in subgroup identification, this approach likely outperforms other methods.(Liu et al., 2019) Second, GUIDE is a purely statistical approach, and does not consider the relative clinical importance of various predictor variables. Furthermore, although we report the 10 most important predictors identified for each classification tree, the GUIDE approach does not provide information about the direction of prediction for variables that are not included in the final trees. Third, the external validity of the identified splitting variables in unclear, and our conclusions are limited to the COMBINE Study sample. That is, these findings may not be representative of other patient populations treated for AUD and trials with different inclusion/exclusion criteria. Fourth, we coded subjects with missing drinking data as heavy drinkers with no change in WHO risk level, which can produce biased estimates when the proportion of missing data is large or when the assumption is unrealistic.(Hallgren and Witkiewitz, 2013) However, drinking data in the COMBINE Study are fairly complete. Lastly, we utilized previously generated categorical predictors, with clinically informed groupings. However, it is unclear where the same or different cut-off values would have been identified if all predictors were considered as continuous variables.

Conclusion

In this study, which used a tree-based machine learning algorithm to identify predictors of treatment outcomes in the COMBINE Study, consecutive days abstinent prior to randomization was identified as a strong predictor of abstinence and no heavy drinking days in the last 4 weeks of treatment. In addition, Alcohol Dependence Scale scores and age may predict no heavy drinking days. Therefore, measuring consecutive days abstinent and Alcohol Dependence Scale scores of trial participants at the beginning of trials can help identify the most appropriate interventions and outcomes. Although potentially important predictors were identified at the root node for a 2-level reduction in WHO levels in the last 4 weeks of treatment, the final classification tree generated by the GUIDE approach had no splits. While this does not suggest that patient characteristics are not associated with this endpoint, it does indicate that there was no classification tree that was the best at capturing subgroups that differed the most on this endpoint. Future evaluations, using the same or different classification tree approaches, should evaluate whether predictors are identified for a 2-level reduction in WHO drinking levels in AUD trials.