Skip to main content
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2026 Jun 1.
Published in final edited form as: Addict Behav. 2025 Feb 25;165:108273. doi: 10.1016/j.addbeh.2025.108273

Finding Purpose: Integrated Latent Profile and Machine Learning Analyses Identify Purpose in Life as an Important Predictor of High-Functioning Recovery after Alcohol Treatment

Frank J Schwebel 1, Adam D Wilson 1, Matthew R Pearson 1, Matison W McCool 1, Katie Witkiewitz 1
PMCID: PMC12184658  NIHMSID: NIHMS2061791  PMID: 40020459

Abstract

Background:

Recent investigations of recovery from alcohol use disorder (AUD) have distinguished subgroups of high and low functioning recovery in data from randomized controlled trials of behavioral treatments for AUD. Analyses considered various indicators of alcohol use, life satisfaction, and psychosocial functioning, and identified four recovery profiles from AUD three years following treatment.

Objectives:

The present study integrates these profiles into a two-part machine learning framework, using recursive partitioning and random forests to distinguish a) clinical cut-points across 28 end-of-treatment biopsychosocial measurements that are predictive of high or low functioning recovery three years after treatment; and b) a rank-ordered list of the most salient variables for predicting individual membership in the high-functioning recovery sub-groups. Methods: This secondary data analysis includes individuals (n=809; 29.7% female) in the outpatient arm of Project MATCH who completed the end-of-treatment assessment and three-year follow-up batteries.

Results:

Recursive partitioning found individuals with low depressive symptoms and less than 25% drinking days were more likely to be in a high functioning recovery profile (68%), whereas those with at least mild depressive symptoms and low purpose in life were more likely to be in a low functioning recovery profile (70%). Random forests identified purpose in life, social functioning, and depressive symptoms as the best predictors of recovery profiles.

Conclusions:

Recovery profiles are best predicted by variables often considered of secondary interest. We demonstrate the utility of two machine learning approaches, highlighting how random forests can overcome recursive partitioning limitations.

Keywords: alcohol use disorder, recovery, purpose in life, machine learning, random forests


An inherent tension between evidence-based medicine and precision/personalized medicine is increasingly recognized by researchers and clinicians (Fava, 2020; Horwitz & Singer, 2017; Lobitz et al., 2019). The originators and initial champions of the evidence-based medicine movement (Evidence-Based Medicine Working Group, 1992; Sackett et al., 1996) were aligned with Cochrane’s (1973) arguments and advocated for a data-driven clinical science approach to treatment, while also being careful to carve out a role for clinical expertise and judgment (Sackett et al., 1996). However, evidence-based medicine has shifted towards a reliance on randomized clinical trials (RCTs) and meta-analyses, which are based on average effects across heterogenous samples, to dictate patient care. This approach improved upon the prior reliance on what some have dubbed ‘eminence-based medicine,’ based in opinion, popularity, personality, and prestige, yet evidence-based medicine often overlooks subgroup and individual differences, and has thus failed as a comprehensive model for informing patient care (Fava, 2020; Horwitz & Singer, 2017).

Personalized or precision medicine was advanced as a concept to improve prevention and treatment strategies by accounting for individual differences (Collins & Varmus, 2015). The initiative seeks to leverage innovation and “big data” coming from areas like genomics, bioinformatics, and wearable health technologies, but is agnostic as to how those data might best be organized and analyzed for maximum clinical utility. A concurrent call for a shift away from evidence-based medicine and towards ‘medicine-based evidence’ (Concato, 2012; Druss, 2005; Horwitz et al., 2017) begins to address this data organization and analysis concern by imagining large searchable databases of individual patient profiles and matching patients with archived profiles to inform response probabilities to a given medication or therapy. This call for increased integration of the idiographic with the nomothetic has been echoed by leading researchers in psychiatry and psychology (cf. Fava, 2020; Hayes et al., 2019; Hofmann & Hayes, 2019).

An additional concern with the current evidence-based medicine paradigm is that it relies primarily on theory-driven confirmatory hypothesis testing, often using null-hypothesis significance testing (NHST) as its benchmark for evidence. Resistance to exploratory data analytic techniques and strict adherence to a ‘research grounded in theory’ dogma has often been supported by the idea that data-mining and other exploratory techniques run the risk of ‘capitalizing on chance.’ This is perhaps true when the data are being explored with NHST (p<0.05) approaches. A goal, therefore, might be to develop flexible, exploratory data analytic techniques that do not rely on NHST, and can also integrate or inform a more idiographic way forward to develop a medicine-based evidence foundation and a viable precision medicine approach to treatment, although care must still be taken to avoid capitalizing on chance – ideally through replication of findings in independent datasets.

We argue a currently feasible first step toward precision medicine could include the application of latent profile analyses to identify subgroups of people in large RCT datasets, in combination with machine learning algorithms to determine the best predictors of outcomes for each subgroup. Witkiewitz and colleagues (2019) employed latent profile analysis (LPA) to distinguish subgroups of recovery, in the data (n=806) from the outpatient arm of Project MATCH, a multi-site RCT of behavioral treatments for alcohol use disorder (AUD) (Project MATCH Research Group, 1997), a prevalent and growing global health concern (Casswell & Rehm, 2020; GBD 2016 Alcohol Collaborators, 2018). The analyses considered indicators of alcohol use, life satisfaction, and psychosocial functioning, and identified four latent profiles of recovery from AUD three years following treatment. The profiles were identified as (1) low functioning frequent heavy drinkers, (2) low functioning infrequent heavy drinkers, (3) high functioning occasional heavy drinkers, and (4) high functioning infrequent drinkers. With the latent profile model as a base for parsing heterogeneity in patient outcomes, machine learning algorithms, such as recursive partitioning, can be applied to provide clinical cut-points to predict treatment outcomes.

Current Study

The present investigation integrates the previously identified latent profiles into a two-part algorithmic machine learning framework. We used recursive partitioning to identify clinical cut-points across 28 end-of-treatment biopsychosocial measurements that are predictive of 1) high versus low functioning recovery three years after treatment, and 2) high functioning occasional heavy drinking versus high functioning infrequent drinking three years after treatment. We used random forests to create a rank-ordered list of variables the algorithm found most salient when predicting individual membership in one of the two high-functioning recovery subgroups and examined variable minimal depth.

Materials and Methods

Participants and Procedure

The present study is a secondary analysis of data from the outpatient arm of a multi-site randomized clinical trial, Project MATCH (Project MATCH Research Group, 1997). Participants (N=952) sought outpatient treatment for AUD, were recruited at nine research centers in the United States, and were randomized to one of three treatments (Twelve-Step Facilitation, Motivational Enhancement Therapy, or Cognitive Behavioral Therapy). Of the 952 participants in the outpatient sample, 806 participants (84.7%) had alcohol consumption data in the 3-year follow-up assessment and were included in the analyses. DSM-III-R criteria were used to assess alcohol dependence (n=732, 90.8%) and alcohol abuse (n=74, 9.2%).

Measures were administered at baseline, during the 12 weeks of treatment, immediately after treatment (3-months post-baseline), six-months post-treatment (9-months post-baseline), 12-months post-treatment (15-months post-baseline), and three years after treatment (39-months post-baseline). All procedures were approved by the IRB of record.

Measures

Three-Year Follow-Up Latent Profile Indicators

Indicators of latent profiles at three-years following treatment were derived from the Form-90 (Miller, 1996), a calendar-based method to obtain reports of alcohol/drug use in the previous 90-day period, the Drinker Inventory of Consequences (DrInC; Miller et al., 1995), the Psychosocial Functioning Inventory (PFI; Feragne et al., 1983), and the Addiction Severity Index (ASI; McLellan et al., 1992). Indicators from the Form-90 included percent drinking days (PDD), percent heavy drinking days (PHDD, i.e., 4 or more drinks in a day for women, 5 or more drinks in a day for men), and drinks per drinking day (DDD). Marijuana and other illicit drug use were coded as binary (0=no use, 1=any use). The DrInC was used to measure alcohol-related negative consequences (Cronbach’s α=0.97). These variables were used to identify the drinking patterns for each profile. The PFI was used to measure social behavior (Cronbach’s α=0.83), and four binary items measuring satisfaction with life, living situation, relationship, and social/recreational functioning over the past 30 days (Cronbach’s α=0.79). The ASI was used to measure employment and experiences of “serious depression,” cognitive difficulty (“trouble understanding, concentrating, or remembering”), and “serious anxiety or tension” in the past 30 days. All items with binary indicators were coded 0=employed or symptom not present, and 1=unemployed or symptom present (Cronbach’s α=0.63). These variables were used to identify the functioning level for each profile.

End of Treatment Features Included in Machine Learning Prediction

To determine the final selection of features measured at the end of treatment to include in the analyses as predictors, we excluded any immutable demographic variables, and selected all available biopsychosocial variables that were indicative of potential intervention targets, processes, or mechanisms of change. The variables selected (n=28) covered a range of domains, including mental health, physical health, quality of life, purpose in life, alcohol consumption, other substance use, social support, social roles and behavior, stress, employment, living situation, AA involvement, abstinence self-efficacy, and temptation to drink. Table 1 provides a summary of each of the variables, the measures, the scale details, and whether the variables were also used in the three-year follow-up latent profile analyses. Supplementary Table 1 reports the variance or variation ratio for each predictor variable. Supplementary Figure 1 reports the correlations between each predictor variable.

Table 1.

Brief Summary of End of Treatment Variables Entered as Predictors of 3-Year Recovery Profiles.

Measure Scale Details End of Treatment Variables Used in LPA at 3-year Minimal Depth

Form-90 (Miller, 1996) Calendar method, captures drinking over past 90 days Percent drinking days (PDD)Percent heavy drinking days (PHDD) Yes 3.65
Captures a wide range of substance use and other functioningCalculated World Health Organization (WHO) Risk Drinking Levels in 3 ways Yes 3.91
Yes 2.77
Drinks per Drinking Day (DDD)Average Drinks per Day (AVSD) No 3.00
(see Witkiewitz et al., 2019): Yes 8.50
0 = no change or increase Yes 7.54
1 = 1 WHO Risk level reductions Any marijuana useAny other drug use (not marijuana) No 8.12
2 = 2 WHO Risk level reductions No 6.62
3 = 3 WHO Risk level reductions Illicit drug use No 7.41
4 = 4 WHO Risk level reductionsBinarized to those with and without at least 1-level reduction in WHO risk levelBinarized to those with and without at least a 2-level reduction in WHO risk level Received alcohol, drug, or mental health treatment No 8.14
No 7.68
Attended at least 1 AA or 12-step meeting NoNoNo 4.559.309.25
Any days spent at work
Under medical care
WHO Risk Level ChangeWHO Risk 1-Level ChangeWHO Risk 2-Level Change
Drinker Inventory of Consequences (DrInC, Miller et al., 1995) 45 items, 4-point response scale DrInC Total Score Yes 3.19
Composite Outcome Variable (Cisler & Zweben, 1999) Calculated from Form-90 and DrInC 1 = abstinent; 2 = non-heavy without problems; 3 = heavy without problems; 4 = heavy with problems No 5.59
Purpose in Life (PIL; Crumbaugh & Maholick, 1976) 20 items, 7-point response scale PIL Total Score No 2.61
Psychosocial Functioning Inventory (PFI; Feragne et al., 1983) 19 items, 4- to 5-point response scales Social Behavior Role ScaleRoommate-Housemate ScaleDissatisfied with leisure, social, or recreational activities YesYesYes 3.094.688.57
Addiction Severity Index (ASI, McLellan et al., 1992) 27-items, various response scales Psychiatric Severity Composite Score No 3.83
“Experienced serious depression”“Experienced serious anxiety or tension”“Experienced trouble understanding, concentrating, or remembering” DepressionAnxietyCognition YesYesYes 9.338.178.89
Beck Depression Inventory (BDI; Beck et al., 1988) 21 items, 4-point response scale BDI Total Score No 2.66
Alcohol Abstinence Self-Efficacy Scale (AASE; DiClemente et al., 1994) 40 items, 5-point response scale TemptationConfidence NoNo 2.503.39
Alcoholics Anonymous Involvement (AAI) Scale (Tonigan et al., 1996) 13 items, various response scales AAI Composite Score No 4.52

Analytic Approach

Latent Profiles of Three Year Outcomes

Latent profile analyses were conducted in Mplus version 8 (Muthén & Muthén, 2017) using a weighted maximum likelihood function, which provides the estimated variance-covariance matrix for all available data, thus all data were included in the models (n=806). See Witkiewitz et al (2019) for full details on the latent profile analyses. Individual participant membership in a particular profile is by definition probabilistic in latent profile analysis. Relative entropy is used to evaluate how well the latent profile solution classifies individuals into latent profiles and is strongly correlated with correct profile assignment (entropy>0.80 is considered good classification precision) (Wang et al., 2017). The classification precision of the four-profile model was excellent (entropy=0.92), suggesting on average participants were expected to be correctly classified 92% of the time. As such, most likely latent profile membership was derived for each participant and the highest likelihood of being classified profile 1 (low functioning frequent heavy drinkers; n=127), 2 (low functioning infrequent heavy drinkers; n=130), 3 (high functioning occasional heavy drinkers; n=136), or 4 (high functioning infrequent drinkers; n=413). Profiles 1 and 2 were characterized by having below average psychosocial functioning and more alcohol-related negative consequences whereas profiles 3 and 4 had above average psychosocial functioning and fewer alcohol-related negative consequences. Profiles 2 and 4 had many people who abstained from alcohol, whereas nobody in profiles 1 and 3 abstained. Distinctions in drinking frequency and quantity were derived from percent drinking days, percent heavy drinking days, and drinks per drinking day values (see Witkiewitz et al., 2019 for more details). Profile membership was extracted and used in the machine learning analyses as outcome variables. Specifically, we examined expected membership in a) any high-functioning profile among the entire sample (profiles 1 and 2 [n=257] vs. profiles 3 and 4 [n=549]), and b) the occasional heavy drinking versus infrequent profiles among the high-functioning profiles (profile 3 [n=136] vs. profile 4 [n=413]).

Machine Learning Algorithms

Recursive partitioning analyses were completed using the ‘rpart’ package for R (Therneau et al., 2015). Decision trees were grown to determine which end of treatment indicators, and levels of those indicators, best predicted membership in high-functioning profiles at three years post-treatment. Single trees were grown, and to guard against overfitting, 10-fold cross-validation and the 1 standard error rule (Hastie et al., 2009) were used to create parsimonious pruned models predicting membership in the high-functioning profiles. With recursive partitioning, cut-points are empirically derived for continuous measures based on maximizing the homogeneity of the resulting nodes (i.e., a node with low probability of membership in the high-functioning profile vs. a node with high probability of membership in the high-functioning profile). Given the unbalanced sample size in each profile, we applied weights to each sample to maximize predictive ability. We set ‘minbucket’ to 5 to balance model parsimony with risk of overfitting.

The ‘caret’ package in R (Kuhn, 2008) was used to tune a meta-parameter ‘mtry’ to determine the optimal number of random variables to be selected on each run of the random forest analyses. A random forest ensemble of trees (n=1000) were grown using the ‘randomForest’ package in R (Liaw & Wiener, 2002), with mtry=6 as the optimized meta-parameter to predict high- or low-functioning profile membership. ‘Out of bag’ (OOB) estimates of prediction accuracy and variable importance measures were computed for the random forest. Median imputation was used for any missing cases. Variable importance, or relative predictive utility of variables used across all 1000 decision trees grown, was assessed using mean decrease in Gini impurity. This metric is the average of any single variable’s ability to improve classification accuracy, and decrease node impurity, weighted depending on how many bootstrapped samples reached that node (split on that variable) in each tree grown in the random forest. We used the ‘randomForestExplainer’ package in R (Paluszynska et al., 2020) to examine the distribution of minimal depth among predictor variables. Minimal depth measures the average position of each variable across all trees in the forest, with a smaller depth value indicating that a variable appears higher in the trees more frequently. This approach provides an additional method of cross-validating the variable importance findings.

Sensitivity Analyses to Test Additional Analytic Approaches

Given the novelty of our analytic approach, it is useful to consider other potential analytic methods to address our research question. One alternative approach would be to use packages such as ‘party’ or ‘partykit’ (Hothorn & Zeileis, 2024) instead of ‘rpart’ and ‘caret’. These packages rely on conditional rather than marginal variable importance, thereby accounting for correlations between variables. These packages can also produce more parsimonious models, potentially reducing two trees (comparing Profiles 1+2 vs. 3+4 and Profile 3 vs. 4) into a single tree to compare likely outcomes across all four profiles. Finally, we considered using cforest from partykit as an alternative to the random forests conducted in caret.

Results

Recursive Partitioning Tree

The variable at the top of each tree best splits the data to maximize the homogeneity of the nodes below. The variables splitting the lower nodes show the hierarchical interaction of all measured predictors that led to an individual being classified in a ‘terminal’ node. Figure 1 shows the most salient end of treatment features in determining 3-year high functioning recovery. The uppermost node shows participants in the full sample (n=806) had a 68% probability of being members of one of the two high functioning recovery profiles 3 years after treatment, consistent with our prior findings (Witkiewitz et al., 2019).

Figure 1.

Figure 1.

Likelihood of High Functioning Profile Membership (Profiles 3+4)

Beck Depression Inventory (BDI) score post-treatment (three months post-baseline), best predicted high-functioning recovery three years later. This was split by people’s purpose in life and post-treatment percent heavy drinking days. People with BDI scores greater than or equal to 6.5 (on the left of the figure; n=310) and Purpose in Life scores below 109.5 had a 70% probability of being classified in a low functioning profile (n=261), whereas people with Purpose in Life scores greater than or equal to 109.5 had a 56% probability of being classified in a high functioning profile (n=49). People with BDI scores less than 6.5 (on the right of the figure; n=496) and greater than or equal to 25% heavy drinking days had a 70% probability of being classified in a low functioning profile (n=58), whereas people with less than 25% heavy drinking days had a 68% probability of being classified in a high functioning group (n=438). The pruned model explained 37% of the variability in the sample and, on average, it explained 29% of the variability in the bootstrap cross-validation procedure (relative error=0.63, cross-validation error=0.71, SE=0.02). However, the overall model was only 69.7% accurate in classifying cases, only slightly more accurate than classifying all cases as being high functioning (68.1%). See Table 2 for complete model fit results.

Table 2.

Model fit metrics

Sensitivity Specificity PPV NPV Accuracy AUC F1

Predicting high functioning profiles 64.6% 72.1% 52.0% 81.3% 69.7% 0.69 0.61
Predicting high functioning, infrequent heavy drinking profile (tree) 60.5% 71.3% 86.5% 37.3% 63.2% 0.69 0.64
Predicting high functioning, infrequent heavy drinking profile (forest) 55.6% 77.0% 52.0% 78.5% 69.5 0.64 0.55

Note: PPV = Positive predictive value. NPV = Negative predictive value. AUC = Area under the curve.

Figure 2 shows the pruned tree with the most salient end-of-treatment features in the high functioning sample (n=549; profiles 3 and 4) that determined membership in the high functioning occasional heavy drinking recovery profile three years post-treatment. Post-treatment percent drinking days was the best predictor of high functioning recovery. The social behavior role subscale of the PFI, the temptation subscale of the Alcohol Abstinence Self-Efficacy Scale, and the composite outcome variable (a categorical classification of treatment outcomes derived from Form-90 and DrInC data ranging from abstinence to heavy drinking with problems) also were salient predictors. People who reported greater than or equal to 28.3% drinking days had a 72% probability of being classified in the high functioning occasional heavy drinker profile (n=101) whereas people who reported less than 28.3% drinking days had a 57% probability of being classified in the high functioning infrequent drinker profile (n=448). The left side of the tree in Figure 2 depicts a hierarchical interaction of predictors such that the 159 people who had fewer drinking days, higher psychosocial functioning, higher temptation to drink, and a non-abstinent outcome post-treatment had a 59% probability of being classified in the high functioning occasional heavy drinker profile. Similarly, the 102 people who had fewer drinking days and lower psychosocial functioning had a 73% probability of being classified in the high functioning infrequent drinker profile. The pruned model explained 32% of variability in the sample and on average it explained 13% in the bootstrap cross-validation procedure (relative error=0.68, cross-validation error=0.87, SE=0.03). However, the overall model was only 63.2% accurate in classifying cases, which was less accurate than classifying all cases as being high functioning infrequent drinkers (75.2%). See Table 2 for complete model fit results and Supplementary Table 2 for confusion matrices.

Figure 2.

Figure 2.

Likelihood of High Functioning Occasional Heavy Drinker Profile Membership (Profile 3)

Random Forest

Random forest training and testing was performed, growing 1000 trees using 6 randomly selected variables in each tree in each of the 1000 bootstrapped samples. Each bootstrapped run held out roughly 1/3 of the cases to test all trees on new data for prediction accuracy. The Out-of-bag (OOB) estimated error rate of 23.53% indicated that the ensemble of trees built with random samples of the 28 predictors used in this study were able to correctly classify individuals approximately 76% of the time.

A rank ordered variable importance plot of the 28 predictors ranked by mean decrease in Gini impurity is included in Figure 3 (confusion matrix in Supplementary Table 2). Predictive importance in random forest models can be conceptualized as the average percent reduction in impurity (heterogeneity of outcomes) that a predictor was able to accomplish whenever it was used to split a group into two more homogenous groups. The temptation subscale of the Alcohol Abstinence Self-Efficacy Scale was the most important predictor across all trees grown in the random forest (19.2%). Purpose in Life (16.1%) was the second most important predictor followed by Beck Depression score (15.1%). Other variables with average decreases in Gini impurity above 10% include drinks per drinking day, average drinks per drinking day, the social behavior subscale of the Psychosocial Functioning Inventory, the confidence subscale of the Alcohol Abstinence Self-Efficacy Scale, scores on the Drinkers Inventory of Consequences.

Figure 3. Random forest variable importance rankings of all 28 predictor variables.

Figure 3.

Note. ASE Tempt = Temptation Scale of the Alcohol Abstinence Self-Efficacy Scale, Purpose in Life = Purpose in Life total score, BDI = Beck Depression Inventory total score, DDD = Drinkers per Drinking Day from Form-90, Avg. DPD = Average Drinks per Day from Form-90, Soc. Behav. = Social Behavior Role Scale of the Psychosocial Functioning Inventory, ASE Conf. = Confidence Scale of the Alcohol Abstinence Self-Efficacy Scale, DrInC = Drinker Inventory of Consequences total score, PDD = Percent drinking days from Form-90, Psych. Severity = Psychiatric Severity Composite Score from the Addiction Severity Index, PHDD = Percent Heavy Drinking Days from Form-90, AA Involvement = Composite Score of Alcoholics Anonymous Involvement scale, WHO Change = WHO Risk Level Change from Form-90, Roommate = Roommate-Housemate Scale of the Psychosocial Functioning Inventory, Comp. Outcome = Composite Outcome Variable from Form-90 and DrInC, Treatment = received alcohol, drug, and/or mental health treatment, 12-Step = attended at least 1 AA or 12-step meeting from Form-90, Medical Care = under medical care from Form-90, Other Drug Use = any other drug use (not marijuana) from Form-90, Work = any days spent at work from Form-90, Anxiety = anxiety item from Addiction Severity Index, Illicit Drug use = any illicit drug use from the Form-90, Marijuana Use = any marijuana use from Form-90, Dissatisfaction = Dissatisfaction with leisure, social, or recreational activities from Psychosocial Functioning Inventory, 2-level WHO Red. = whether they achieved at least a 2-level reduction in WHO Drinking Risk levels from the Form-90, Cognition = Cognition item from the Addiction Severity Index, 1-level WHO Red. = whether they achieved at least a 1-level reduction in WHO Drinking Risk levels from the Form-90, Depression = depression item from Addiction Severity Index.

Average minimal depth is displayed in Table 1 and the distribution of minimal depth is displayed in Supplementary Figure 2. The temptation subscale of the Alcohol Abstinence Self-Efficacy Scale was the most important predictor (minimal depth: 2.50), followed by Purpose in Life scores (2.61), and Beck Depression (2.66). The other significant predictors of minimal depth were drinks per drinking day (2.77), average drinks per drinking day (3.00), the social behavior subscale of the Psychosocial Functioning Inventory (3.09), Drinkers Inventory of Consequences total score (3.19), Alcohol Abstinence Self-Efficacy Confidence subscale (3.39), percent drinking days (3.65), the psychiatric severity subscale of the Addiction Severity Index (3.83), percent heavy drinking days (3.91), Alcoholic Anonymous Involvement scale score (4.52), WHO risk level change (4.55), and the Roommate-Housemate scale of the Psychosocial Functioning Inventory (4.68).

Post-Hoc Sensitivity Analyses

First, we conducted a post-hoc sensitivity analysis, using partykit and its ctrees function to build a single predictor tree (see Supplementary Figure 3). We also replicated the comparisons presented in this manuscript with ctrees, yielding a nearly identical tree and fit indices for Profiles 1+2 vs. 3+4 (see Supplementary Figures 4 and 5). Interestingly, we found better model fit when replicating our second tree comparing Profiles 3 vs. 4 (see Supplementary Table 3).

Second, we applied cforest from partykit as an alternative to the random forests conducted in caret, but this approach also yielded inferior model fit indices and similar variables among the top 10 most important variables (see Supplementary Materials for a full comparison, fit indices, and confusion matrices).

Discussion

The present investigation sought to integrate latent profile analyses and machine learning approaches to determine which of 28 potential treatment targets, processes and mechanisms measured at the end of three months of outpatient treatment for AUD were most predictive of high functioning recovery measured three years after treatment. Two exploratory machine learning approaches, recursive partitioning and random forests, were applied. The outcomes in the models were membership in a) either high-functioning profile (profiles 3 and 4), and b) membership in the high functioning occasional heavy drinking profile (profile 3) versus membership in the high functioning infrequent drinking profile (profile 4).

The best end-of-treatment predictors of eventual membership in a high-functioning recovery profile were depressive symptoms, purpose in life, and percent heavy drinking days. Individuals with low depressive symptoms (<6.5 on a scale wherein 10–18 indicates mild to moderate depression; Beck et al., 1988) and less than 25% heavy drinking days had the highest likelihood of being in one of the two high-functioning recovery classes Those with low purpose in life and higher depressive symptoms (>6.5) had the lowest likelihood of being in one of the two recovery classes. The model explained 37% of variability in the sample.

Subdividing the high functioning profiles, expected membership in the profile 3 was best predicted by percent drinking days, psychosocial functioning, temptation to drink, and a composite alcohol treatment outcome variable score. Individuals reporting more drinking days were most likely to be in the high functioning occasional heavy drinker profile. Individuals reporting fewer drinking days and lower levels of psychosocial functioning were most likely to be classified as being in profile 4 (73.2%). Although this tree had some predictive utility, it was less likely to accurately classify people than chance, indicating it may be unlikely to replicate.

Recursive partitioning yields trees that are visually intuitive and present predictors and cut-points that best distinguish between individuals. Although creating parsimonious trees necessitates some degree of subgroup identification to avoid overfitting, our approach could be used in a precision medicine framework by extending it to many AUD treatments across diverse samples. With a sufficient sample, it would be theoretically possible to prospectively assign patients to the treatment(s) that are most likely to be effective for them with less risk of overfitting.

Unfortunately, recursive partitioning is a “greedy” algorithm that selects whichever predictor best separates the groups. This approach means a single tree may not provide full information on the relative importance of all predictors available. We addressed this by using an ensemble machine learning approach, random forest. The temptation subscale of the Alcohol Abstinence Self-Efficacy scale was the most important predictor among the trees grown in the random forest followed by purpose in life and depression. Additional variables of importance were drinks per drinking days and average drinks per drinking day, psychosocial functioning, the confidence subscale of the Alcohol Abstinence Self-Efficacy scale, and negative drinking consequences. The top 6 predictors from the random forest trees were identical to the top 6 variables identified by the minimal depth analysis.

Overall, we found evidence of concordance across our analytic approaches. Beck Depression Inventory (BDI) scores and Purpose in Life scores emerged as top variables in predicting high functioning recovery (Profiles 1 and 2 vs. 3 and 4), in the random forest model, and in the minimal depth analysis. Purpose in Life is particularly noteworthy as an end-of-treatment predictor. Although it has no corollary in the derived latent recovery profiles at 3-years post-treatment, it ranked as the second most important predictor in the random forest model and minimal depth analysis and was the second splitting variable in predicting high functioning recovery. Despite typically receiving less clinical focus compared to alcohol-related outcomes (e.g., consumption, self-efficacy, consequences), both traditional (Alcoholics Anonymous; Grapevine Inc., 1947) and modern approaches to resolving AUD acknowledge its significance. Depression fits within the broader context of psychosocial functioning, yet BDI scores were not used to derive the initial latent classes. However, the link between alcohol and depression is well-established (McHugh & Weiss, 2019) making it an expected, salient predictor. The relative importance of BDI and purpose in life may be driven by their intercorrelation. However, conditional variable importance results also identified BDI and purpose in life as highly important predictors, ranking them as 5th and 7th most important variables, respectively.

Drinking metrics (e.g., percent drinking days, percent heavy drinking days), were consistently important across all our models. Psychosocial functioning distinguished between higher functioning profiles (3 vs. 4) and was the sixth most important predictor in the random forest model and minimal depth analysis. These findings are expected given these variables were included in the initial latent profile analysis and fit well within the expanded recovery definition.

Lastly, the temptation subscale of the Alcohol Abstinence Self-Efficacy scale was a salient predictor in distinguishing between the high functioning profiles and was the top predictor in the random forest model and minimal depth analysis. This subscale assesses ones’ temptation to drink in high stress situations involving strong urges, cravings, or withdrawal symptoms. Research has demonstrated a strong positive relationship between temptation self-efficacy and treatment outcomes in AUD (Kadden & Litt, 2011), despite its limited focus in treatment. The temptation subscale is also noteworthy due to its potential link with the addiction cycle domains of negative emotionality and incentive salience (Koob & Volkow, 2016; Kwako et al., 2016) and its use in subtyping the domains of reward (incentive salience) and relief (negative emotionality) drinking in prior studies (Hebden et al., 2024; Roos et al., 2017). The addiction cycle proposes that most people begin as “reward drinkers,” in which drinking is largely maintained through positive reinforcement. However, with sustained heavy drinking, withdrawal symptoms increase, and drinking shifts from positive to negative reinforcement—people begin drinking to relieve unpleasant sensations (e.g., withdrawal symptoms) and avoid negative consequences (e.g., relief drinkers) (Koob & Volkow, 2016; Verheul et al., 1999). The findings of the current study suggest that the temptation subscale may also be important for differentiating between high functioning recovery groups.

Alternative Approaches

We tested several different approaches for machine learning, including several different packages. The alternative approaches did not yield substantively different conclusions and/or better classification accuracy. One alternative approach that was not tested in the current study would have been to use Structural Equation Modeling (SEM) Trees (Brandmaier et al., 2013, 2016; Serang et al., 2021) instead of a latent profile analysis (LPA). Because the original LPA that identified these recovery profiles used covariates, we expect the SEM Tree results would be similar (Jacobucci et al., 2017). Additionally, prior work by our team found that the outcomes assessed in the LPA were quite stable over time, suggesting reliability of the profiles over time (Witkiewitz et al., 2019).

Limitations and Future Directions

Machine learning methods, and indeed the present analyses, also have drawbacks. Although recursive partitioning and random forest algorithms handle binary, categorical, and continuous variables equally well, the decisions are driven by a singular focus to find the next optimal split and may not be clinically significant. Random forests help mitigate this risk by randomly selecting subsets of predictors for each run, giving all predictors a chance to show their predictive value while growing the forest. That said, machine learning approaches are still susceptible to capitalizing on chance and efforts must be taken to prevent over-fitting and cross-validate results. Additionally, the integration of the latent profiles into the machine learning analyses was accomplished by turning a probabilistic estimate of profile membership into a discrete value. This is defensible insofar as the entropy (an index strongly correlated with probability of correct profile membership) was very high in the latent profile analysis, but is not always ideal. Attempts have been made to employ recursive partitioning and random forest algorithms within latent variable frameworks, such as SEM Trees and show promise, but have not been extended to include latent profile analysis. Given the exploratory nature of this approach, a replication in a new sample would increase confidence in the findings, as would comparing groups identified by SEM Trees compared to a latent profile analysis.

We are intrigued by the limited ability to predict membership in the high-functioning occasional heavy drinking profile better than chance. This indicates exploratory research methods consciously chosen to avoid the problems associated with p-values produce null or findings that may be unlikely to generalize. This improves confidence in the methods, but also raises an interesting substantive question. If the 28 potential treatment targets available for analyses had limited value predicting recovery status three years later, what might have greater predictive value? Future research investigating additional biopsychosocial variables, as well as environmental and contextual characteristics is warranted (Acuff et al., 2023; Witkiewitz & Tucker, 2024).

Conclusion

The current investigation presents a novel way to integrate latent profile analysis and machine learning to delineate treatment targets that may best support long-term recovery from AUD. Latent profile analysis can be used to identify profiles of similar patients based on underlying attributes. The applicability and suitability of these models to discovering groups of similar patients–before, during, or after treatment–make latent profile analysis and associated analyses valuable in the shift towards precision medicine and personalized mental health care. Similarly, machine learning methods, have the capacity to improve clinical decision making and improve predictive accuracy. Tree-based modeling approaches handle high dimensional data well and are robust to non-linear data and outliers compared to traditional linear modeling. Many machine learning algorithms can analyze data quickly and do not rely on NHST, making them especially appropriate for “big data” and exploratory analyses, although they still can capitalize on chance.

Further, we make no claims that these newer approaches constitute a methodological panacea for the challenges ahead on the path toward precision or personalized medicine. We only suggest the pursuit may benefit by allowing exploration and prediction into a space long dominated by explanation and causal hypotheses (Shmueli, 2010). More generally, we hope the present investigation promotes renewed attempts at methodological innovation with an eye toward substantive shifts in client functioning, outcomes, and care.

Supplementary Material

Supplementary materials

Highlights.

  • Machine learning reveals predictors of recovery from alcohol use disorder (AUD)

  • Purpose in life, social functioning, and depression top predictors of AUD recovery

  • Low depressive symptoms and <25% drinking days predict high-functioning recovery

  • Recovery from AUD linked to psychosocial factors often viewed as secondary

Acknowledgements:

The authors would like to thank Jacqueline Luchini for her contributions to the manuscript.

Preparation of this manuscript was supported in part by grants from the National Institute on Alcohol Abuse and Alcoholism (NIAAA; T32 AA018108, F31 AA026773, and R01 AA022328) and National Institute on Drug Abuse (NIDA; K23 DA058015). NIAAA and NIDA had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Footnotes

Conflict of Interest

FJS, ADW, MRP, MM declare no conflicts of interest. KW is a member of the Alcohol Clinical Trials Initiative (ACTIVE) Workgroup, which has been supported previously, but not in the past 36 months, by Abbott/Abbvie, Amygdala Neurosciences, Arbor Pharmaceuticals, GSK, Indivior, Janssen, Lilly, Pfizer, and Schering Plough, but in the past 36 months its activities were supported by Alkermes, Dicerna, Ethypharm, and Otsuka.

Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

References

  1. Acuff SF, MacKillop J, & James G Murphy. (2023). A contextualized reinforcer pathology approach to addiction | Nature Reviews Psychology. Nature Reviews Psychology, 2, 309–323. [DOI] [PMC free article] [PubMed] [Google Scholar]
  2. Beck AT, Steer RA, & Carbin MG (1988). Psychometric properties of the Beck Depression Inventory: Twenty-five years of evaluation. Clinical Psychology Review, 8(1), 77–100. 10.1016/0272-7358(88)90050-5 [DOI] [Google Scholar]
  3. Brandmaier AM, Prindle JJ, McArdle JJ, & Lindenberger U (2016). Theory-guided exploration with structural equation model forests. Psychological Methods, 21(4), 566–582. 10.1037/met0000090 [DOI] [PubMed] [Google Scholar]
  4. Brandmaier AM, von Oertzen T, McArdle JJ, & Lindenberger U (2013). Structural equation model trees. Psychological Methods, 18(1), 71–86. 10.1037/a0030001 [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Casswell S, & Rehm J (2020). Reduction in global alcohol-attributable harm unlikely after setback at WHO Executive Board. Lancet (London, England), 395(10229), 1020–1021. 10.1016/S0140-6736(20)30476-1 [DOI] [PubMed] [Google Scholar]
  6. Cisler RA, & Zweben A (1999). Development of a composite measure for assessing alcohol treatment outcome: Operationalization and validation. Alcoholism, Clinical and Experimental Research, 23(2), 263–271. 10.1111/J.1530-0277.1999.TB04109.X [DOI] [PubMed] [Google Scholar]
  7. Cochrane AL (1973). Effectiveness and efficiency: Random reflections of health services. Nuffield Provincial Hospitals Trust. [Google Scholar]
  8. Collins FS, & Varmus H (2015). A new initiative on precision medicine. The New England Journal of Medicine, 372(9), 793–795. 10.1056/NEJMp1500523 [DOI] [PMC free article] [PubMed] [Google Scholar]
  9. Concato J (2012). Is it time for medicine-based evidence? JAMA, 307(15), 1641–1643. 10.1001/jama.2012.482 [DOI] [PubMed] [Google Scholar]
  10. Crumbaugh JC, & Maholick LT (1976). Purpose in Life Scale. [Google Scholar]
  11. DiClemente CC, Carbonari JP, Montgomery RP, & Hughes SO (1994). The Alcohol Abstinence Self-Efficacy scale. Journal of Studies on Alcohol, 55(2), 141–148. 10.15288/jsa.1994.55.141 [DOI] [PubMed] [Google Scholar]
  12. Druss BG (2005). Medicine-based evidence in mental health. Psychiatric Services, 56(5), 543. 10.1176/appi.ps.56.5.543 [DOI] [PubMed] [Google Scholar]
  13. Evidence-Based Medicine Working Group. (1992). Evidence-based medicine. A new approach to teaching the practice of medicine. JAMA, 268(17), 2420–2425. 10.1001/jama.1992.03490170092032 [DOI] [PubMed] [Google Scholar]
  14. Fava GA (2020). The Decline of Pluralism in Medicine: Dissent Is Welcome. Psychotherapy and Psychosomatics, 89(1), 1–5. 10.1159/000505085 [DOI] [PubMed] [Google Scholar]
  15. Feragne MA, Longabaugh R, & Stevenson JF (1983). The psychosocial functioning inventory. Evaluation & the Health Professions, 6(1), 25–48. 10.1177/016327878300600102 [DOI] [PubMed] [Google Scholar]
  16. GBD 2016 Alcohol Collaborators. (2018). Alcohol use and burden for 195 countries and territories, 1990–2016: A systematic analysis for the Global Burden of Disease Study 2016. Lancet (London, England), 392(10152), 1015–1035. 10.1016/S0140-6736(18)31310-2 [DOI] [PMC free article] [PubMed] [Google Scholar]
  17. Hastie T, Tibshirani R, & Friedman J (2009). The Elements of Statistical Learning: Data Mining, Inference, and Prediction (Vol. 2). Springer. [Google Scholar]
  18. Hayes SC, Hofmann SG, Stanton CE, Carpenter JK, Sanford BT, Curtiss JE, & Ciarrochi J (2019). The role of the individual in the coming era of process-based therapy. Behaviour Research and Therapy, 117, 40–53. 10.1016/j.brat.2018.10.005 [DOI] [PubMed] [Google Scholar]
  19. Hebden HM, Swan JE, Claus ED, Pentkowski NS, & Witkiewitz K (2024). Longitudinal stability of reward and relief drinking phenotypes in community and treatment-seeking individuals who engage in heavy drinking. Alcohol, Clinical & Experimental Research, 48(3), 567–579. 10.1111/acer.15274 [DOI] [PMC free article] [PubMed] [Google Scholar]
  20. Hofmann SG, & Hayes SC (2019). The Future of Intervention Science: Process-Based Therapy. Clinical Psychological Science, 7(1), 37–50. 10.1177/2167702618772296 [DOI] [PMC free article] [PubMed] [Google Scholar]
  21. Horwitz RI, Hayes-Conroy A, Caricchio R, & Singer BH (2017). From Evidence Based Medicine to Medicine Based Evidence. In American Journal of Medicine (Vol. 130, Issue 11, pp. 1246–1250). 10.1016/j.amjmed.2017.06.012 [DOI] [PubMed] [Google Scholar]
  22. Horwitz RI, & Singer BH (2017). Why evidence-based medicine failed in patient care and medicine-based evidence will succeed. Journal of Clinical Epidemiology, 84, 14–17. 10.1016/j.jclinepi.2017.02.003 [DOI] [PubMed] [Google Scholar]
  23. Hothorn T, & Zeileis A (2024). A Toolkit for Recursive Partytioning [R package partykit version 1.2–22]. CRAN: Contributed Packages. 10.32614/CRAN.PACKAGE.PARTYKIT [DOI] [Google Scholar]
  24. Jacobucci R, Grimm KJ, & McArdle JJ (2017). A Comparison of Methods for Uncovering Sample Heterogeneity: Structural Equation Model Trees and Finite Mixture Models. Structural Equation Modeling: A Multidisciplinary Journal, 24(2), 270–282. 10.1080/10705511.2016.1250637 [DOI] [PMC free article] [PubMed] [Google Scholar]
  25. Kadden RM, & Litt MD (2011). The role of self-efficacy in the treatment of substance use disorders. Addictive Behaviors, 36(12), 1120–1126. 10.1016/j.addbeh.2011.07.032 [DOI] [PMC free article] [PubMed] [Google Scholar]
  26. Koob GF, & Volkow ND (2016). Neurobiology of addiction: A neurocircuitry analysis. The Lancet Psychiatry, 3(8), 760–773. 10.1016/S2215-0366(16)00104-8 [DOI] [PMC free article] [PubMed] [Google Scholar]
  27. Kwako LE, Momenan R, Litten RZ, Koob GF, & Goldman D (2016). Addictions Neuroclinical Assessment: A neuroscience-based framework for addictive disorders. Biological Psychiatry, 80(3), 179–189. 10.1016/j.biopsych.2015.10.024 [DOI] [PMC free article] [PubMed] [Google Scholar]
  28. Lobitz G, Armstrong K, Concato J, Singer BH, & Horwitz RI (2019). The Biological and Biographical Basis of Precision Medicine. Psychotherapy and Psychosomatics, 88(6), 333–340. 10.1159/000502486 [DOI] [PubMed] [Google Scholar]
  29. McHugh RK, & Weiss RD (2019). Alcohol Use Disorder and Depressive Disorders. Alcohol Research : Current Reviews, 40(1), arcr.v40.1.01. 10.35946/ARCR.V40.1.01 [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. McLellan AT, Kushner H, Metzger D, Peters R, Smith I, Grissom G, Pettinati H, & Argeriou M (1992). The Fifth Edition of the Addiction Severity Index. Journal of Substance Abuse Treatment, 9(3), 199–213. 10.1016/0740-5472(92)90062-s [DOI] [PubMed] [Google Scholar]
  31. Miller WR (1996). Form 90: A structured assessment interview for drinking and related behaviors (Project MATCH).
  32. Miller WR, Tonigan JS, & Longabaugh R (1995). The Drinker Inventory of Consequences (DrInC). An Instrument for Assessing Adverse Consequences of Alcohol Abuse. Test Manual. … MATCH Monograph Series, 4. [Google Scholar]
  33. Muthén LK, & Muthén BO (2017). Mplus User’s Guide (Eighth Edition). Muthén and Muthén. [Google Scholar]
  34. Paluszynska A, Biecek P, & Jiang Y (2020). randomForestExplainer: Explaining and Visualizing Random Forests in Terms of Variable Importance. In CRAN: Contributed Packages. 10.32614/CRAN.package.randomForestExplainer [DOI] [Google Scholar]
  35. Project MATCH Research Group. (1997). Title: Matching alcoholism treatments to client heterogeneity: Project MATCH posttreatment drinking outcomes. Journal of Studies on Alcohol, 58, 7–29. [PubMed] [Google Scholar]
  36. Roos CR, Mann K, & Witkiewitz K (2017). Reward and relief dimensions of temptation to drink: Construct validity and role in predicting differential benefit from acamprosate and naltrexone. Addiction Biology, 22, 1528–1539. 10.1111/adb.12427 [DOI] [PMC free article] [PubMed] [Google Scholar]
  37. Sackett DL, Rosenberg WMC, Gray JAM, Haynes RB, & Richardson WS (1996). Evidence based medicine: What it is and what it isn’t. BMJ (Clinical Research Ed.), 312(7023), 71–72. 10.1136/bmj.312.7023.71 [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. Serang S, Jacobucci R, Stegmann G, Brandmaier AM, Culianos D, & Grimm KJ (2021). M plus Trees: Structural Equation Model Trees Using M plus. Structural Equation Modeling: A Multidisciplinary Journal, 28(1), 127–137. 10.1080/10705511.2020.1726179 [DOI] [Google Scholar]
  39. Shmueli G (2010). To explain or to predict? Statistical Science, 25(3), 289–310. 10.1214/10-STS330 [DOI] [Google Scholar]
  40. Tonigan JS, Connors GJ, & Miller WR (1996). Alcoholics Anonymous Involvement (AAI) scale: Reliability and norms. Psychology of Addictive Behaviors, 10(2), 75–80. 10.1037/0893-164X.10.2.75 [DOI] [Google Scholar]
  41. Verheul R, Van Den Brink W, & Geerlings P (1999). A three-pathway psychobiological model of craving for alcohol. Alcohol and Alcoholism, 34(2), 197–222. 10.1093/ALCALC/34.2.197 [DOI] [PubMed] [Google Scholar]
  42. Wang M-C, Deng Q, Bi X, Ye H, & Yang W (2017). Performance of the entropy as an index of classification accuracy in latent profile analysis: A Monte Carlo simulation study. Acta Psychologica Sinica, 49(11), 1473. 10.3724/sp.j.1041.2017.01473 [DOI] [Google Scholar]
  43. Witkiewitz K, & Tucker JA (2024). Whole person recovery from substance use disorder: A call for research examining a dynamic behavioral ecological model of contexts supportive of recovery. Addiction Research & Theory. https://www.tandfonline.com/doi/abs/10.1080/16066359.2024.2329580 [DOI] [PMC free article] [PubMed]
  44. Witkiewitz K, Wilson AD, Pearson MR, Montes KS, Kirouac M, Roos CR, Hallgren KA, & Maisto SA (2019). Profiles of recovery from alcohol use disorder at three years following treatment: Can the definition of recovery be extended to include high functioning heavy drinkers? Addiction, 114, 69–80. 10.1111/add.14403 [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplementary materials

RESOURCES