Regression tree applications to studying alcohol-related problems among college students

Abstract

Machine learning algorithms hold promise for developing precision medicine approaches to addiction treatment yet have been used sparingly to identify predictors of alcohol-related problems. Recursive partitioning, a machine learning algorithm, can identify salient predictors and clinical cut points that can guide treatment. This study aimed to identify predictors and cut points of alcohol-related problems and to examine result stability in two separate, large data sets of college student drinkers (n = 5,090 and 2,808). Four regression trees were grown using the "rpart" package in R. Seventy-one predictors were classified as demographics (e.g., age), alcohol use indicators (e.g., typical quantity/frequency), or psychosocial indicators (e.g., anxiety). Predictors and cut points were extracted and used to manually recreate the tree in the other data set to test result stability. Outcome variables were alcohol-related problems as measured by the Alcohol Use Disorder Identification Test and Brief Young Adult Alcohol Consequences Questionnaire. Coping with depression, conformity motives, binge drinking frequency, typical/heaviest quantity, drunk frequency, serious harm reduction protective behavioral strategies, substance use, and psychosis symptoms best predicted alcohol-related problems across the four trees; coping with depression (cut point range: 1.83-2.17) and binge drinking frequency (cut point range: 1.5-2.5) were the most common splitting variables. Model fit indices suggest relatively stable results accounting for 17%-30% of the variance. Results suggest the nine salient predictors, particularly coping with depression motives scores around 2 and binge drinking frequency around two to three times per month, are important targets to consider when treating alcohol-related problems for college students. (PsycInfo Database Record (c) 2024 APA, all rights reserved).

Figure 1.

Decision tree 1 predicting BYAACQ score in ART (Training sample, top panel) and PSST (validation sample, bottom panel) The top row of each tree depicts the most important variable for splitting the data (i.e., drinking to cope with depression). Each successive variable below contributes to accurately splitting the data. Variables on different levels (e.g., drinking to cope with depression, binge drinking episodes, serious harm reduction PBS use) suggest an implicit interaction between these variables at the cut points. Note: BYAACQ=Brief-Young Adult Alcohol Consequences Questionnaire; ART=Addictions Research Team; PSST=Protective Strategies Study Team; PBS=Protective behavioral strategies

Figure 2.

Decision tree 2 predicting AUDIT Problems score in ART (Training sample, top panel) and PSST (validation sample, bottom panel) The top row of each tree depicts the most important variable for splitting the data (i.e., drinking to cope with depression). Each successive variable below contributes to accurately splitting the data. Variables on different levels (e.g., drinking to cope with depression, substance use symptoms, heaviest quantity) suggest an implicit interaction between these variables at the cut points. Note: AUDIT=Alcohol Use Disorder Identification Test; ART=Addictions Research Team; PSST=Protective Strategies Study Team; PBS=Protective behavioral strategies

Figure 3.

Decision tree 1 predicting BYAACQ score in PSST (Training sample, top panel) and ART (validation sample, bottom panel) The top row of each tree depicts the most important variable for splitting the data (i.e., binge drinking episodes). Each variable below contributes to accurately splitting the data. Variables on different levels (e.g., binge drinking episodes, drinking to cope with depression, typical quantity consumed) suggest an implicit interaction between these variables at the cut points. Note: BYAACQ=Brief-Young Adult Alcohol Consequences Questionnaire; ART=Addictions Research Team; PSST=Protective Strategies Study Team

Figure 4.

Decision tree 2 predicting AUDIT Problems score in PSST (Training sample, top panel) and ART (validation sample, bottom panel) The top row of each tree depicts the most important variable for splitting the data (i.e., substance use symptoms). Each successive variable below contributes to accurately splitting the data. Variables on different levels (e.g., substance use symptoms, conformity motives, typical quantity) suggest an implicit interaction between these variables at the cut points. Note: AUDIT=Alcohol Use Disorder Identification Test; ART=Addictions Research Team; PSST=Protective Strategies Study Team; PBS=Protective behavioral strategies