Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2023 Nov 1.
Published in final edited form as: J Psychoactive Drugs. 2022 Jan 24;54(5):419–428. doi: 10.1080/02791072.2021.2014081

Using Decision Trees to Identify Salient Predictors of Cannabis-Related Outcomes

Frank J Schwebel 1, Dylan K Richards 2, Rory A Pfund 3, Verlin W Joseph 4, Matthew R Pearson 5; Marijuana Outcomes Study Team*
PMCID: PMC9308832  NIHMSID: NIHMS1766981  PMID: 35067209

Abstract

Cannabis use continues to escalate among emerging adults and college attendance may be a risk factor for use. Severe cases of cannabis use can escalate to a cannabis use disorder which is associated with worse psychosocial functioning. Predictors of cannabis use consequences and cannabis use disorder symptom severity have been identified; however, they typically employ a narrow set of predictors and rely on linear models. Machine learning is well-suited for exploratory data analyses of high-dimensional data. This study applied decision tree learning to identify predictors of cannabis user status, negative cannabis-related consequences, and cannabis use disorder symptoms. Undergraduate college students (N = 7000) were recruited from nine universities in nine states across the U.S. Among the 7 trees, 24 splits created by 15 distinct predictors were identified. Consistent with prior research, one’s beliefs about cannabis were strong predictors of user status. Negative reinforcement cannabis use motives were the most consistent predictors of cannabis use disorder symptoms and past month cannabis use was the most consistent predictor of probable cannabis use disorder. Typical frequency of cannabis use was the only predictor of negative cannabis-related consequences. Our results demonstrate that decision trees are a useful methodological tool for identifying targets for future clinical research.

Keywords: machine learning, decision tree, recursive partitioning, cannabis, college students

Introduction

A national survey of emerging adults (i.e., 18–25-year-olds; Arnett, 2000, 2005) demonstrates continued escalation of the prevalence of cannabis use among this population (Schulenberg et al., 2020). Specifically, 40% and 27% of emerging adults reported using cannabis in the past year and past month, respectively (Schulenberg et al., 2020). College attendance has been associated with a 51% increased probability of using cannabis in the past year for emerging adults who had not used cannabis by the 12th grade (Miech et al., 2017). A large, multi-site survey of cannabis use among college students found that 90.8% of past month users experienced at least one negative consequence related to their use (Pearson et al., 2017b) and cannabis use in college is associated with a variety of negative academic consequences (Arria et al., 2015; Phillips et al., 2015).

In the most severe cases, cannabis use escalates into a clinically significant concern, cannabis use disorder (CUD). The prevalence of past year CUD is estimated to be 8.6% among U.S. college students (Arterberry et al., 2020) and CUD prevalence rates increase with time since the initiation of cannabis use (Han et al., 2019). A CUD diagnosis is related to worse psychosocial functioning among regular cannabis users (Foster et al., 2018). Given the public health burden of cannabis use and CUD among college students, research is needed to identify correlates to target for intervention.

Multiple predictors have been identified across studies, including personality traits (e.g., sensation seeking; Galbraith & Connor, 2015), normative perceptions (Pearson et al., 2017a), cannabis use motives (e.g., coping, Moitra et al., 2015), and use of protective behavioral strategies (Pedersen et al., 2016). However, these studies are limited due to their focus on a narrow set of predictors and reliance on linear models. Focusing on a narrow set of predictors prohibits the functional relation between independent variables, and linear models are limited in their classification ability (Strobl et al., 2009). Instead, analyses considering the combination of multiple independent variables to understand their predictive value in explaining cannabis user status, negative cannabis-related consequences, and CUD (symptoms) may be of benefit.

Machine learning is a branch of quantitative methodology arising from computer science and artificial intelligence (Michalski et al., 2013). A subtype of machine learning, decision tree learning, is a promising tool for explaining the predictive value of multiple independent variables related to cannabis-related outcomes. Decision tree learning involves developing parsimonious predictive models and provides decision rules for predicting both categorical and continuous outcomes. The algorithm for decision tree learning finds the split on a predictor variable that best distinguishes between two distinct groups on an outcome. Following each split, the same algorithm using all possible predictor variables (including variables from the previous split), determines the next split. The algorithm is repeated until each terminal node contains a relatively homogeneous subsample. Decision tree models are well-suited to handle high-dimensional data and can process large number of predictor variables simultaneously (Strobl et al., 2009). Recursive partitioning is one type of decision tree learning model useful for exploratory data analyses.

Machine learning approaches have been used to explore predictors of cannabis use, negative cannabis-related consequences, and CUD. Wilson and colleagues (2018) used decision tree learning to identify predictors of lifetime and past month cannabis use status and negative cannabis-related consequences among 8,141 U.S. college students across 11 universities (Wilson et al., 2018). For both lifetime and past month cannabis use status, the best set of predictors were cannabis norms variables and identification as a cannabis user. For negative cannabis-related consequences, the best set of predictors were frequency of cannabis use, use of cannabis protective behavioral strategies, and positive/negative urgency (i.e., impulsivity-like traits when experiencing positive/negative emotions, respectively). Parekh and Fakim (2021) used machine learning algorithms to identify predictors of daily cannabis use among U.S. adults (N=253,569), and found that current cigarette use, male gender, poor mental health, cognitive decline, and depression were predictors of daily cannabis use. Behavioral factors such as HIV risk behaviors, abnormal sleep patterns, and obesity were also important predictors. Rajapaksha and colleagues (2020) compared machine learning models to predict risk of developing a CUD among individuals who regularly use cannabis (N=94). The authors found that age, enjoyment when initially smoking cigarettes, total Impulsivity Sensation-Seeking Scale score, cognitive instability score of the Barratt Impulsivity Scale, and neuroticism, openness, and conscientiousness scores of the Neuroticism, Extraversion, and Openness inventory were important predictors of regular cannabis use. These findings suggest machine learning algorithms such as decision trees may provide insight into targets for cannabis intervention.

The present study applied decision tree learning to identify predictors of cannabis-related outcomes. Prior literature has examined predictors of cannabis use, negative cannabis-related consequences, and risk of developing CUD. Our goal was to examine factors that might contribute to problematic cannabis use and CUD symptoms. Thus, we sought to replicate and extend prior work by examining all possible predictors of cannabis use, negative cannabis-related consequences, and CUD symptoms in our data using a large sample of college students.

Methods

Participants and Procedure

Participants were 7,000 U.S. undergraduate college students recruited in 2016 and 2017 from universities in two states where recreational cannabis use was legal (CO, WA), three states where medical cannabis was legal (CA, NM, NY), and four states where cannabis use was illegal (FL, TN, TX, VA). Convenience sampling was used to recruit participants enrolled in psychology courses to complete an online survey for partial course credit. Among the overall sample, the average age was 20.59 (SD = 4.12), 66.8% female, and 56.3% White. Further characteristics of this sample have been reported elsewhere (Richards et al., 2021). The IRB at each participating university approved the study procedures1.

Measures

Supplemental Table 1 provides a summary of how each predictor was measured. A planned missingness design (Graham et al., 2006) was used to minimize participant burden so sample sizes varied across measures.

Cannabis User Status

A single item asked whether participants ever used cannabis in their lifetime (lifetime use), and a single item asked if participants used cannabis in the past month (past month use).

Cannabis Use

The Marijuana Use Grid (MUG; Pearson et al., 2020) was used to assess past-month cannabis use. The MUG asks participants to report cannabis use in a week-long calendar which represents typical weekly cannabis use during the past month. Specifically, participants report the times per day they used cannabis in six, 4-hour long blocks of time, and, within those time periods, the estimated quantity of cannabis consumed in grams. We calculated the frequency (e.g., summed nonzero values) and quantity (e.g., summed grams consumed) of cannabis use for a typical week.

Cannabis Consequences

The 21-item Marijuana Consequences Questionnaire Short Form (MACQ; Simons, Dvorak, Merrill, & Read, 2010) was used to capture domains of negative cannabis-related consequences during the past 30 days. Participants were queried on whether they experienced each consequence due to their marijuana use in the past 30 days (0 = no, 1 = yes). A total score was calculated to represent the overall number of unique negative cannabis-related consequences experienced in the past 30 days.

DSM-5 Cannabis Use Disorder Symptoms

Past month cannabis users completed measures of CUD symptoms. All completed the Self-Reported Symptoms of Cannabis Use Disorder (SRSCUD, Richards et al., 2020), a 13-item measure designed to capture symptoms of CUD on a continuum of severity as defined by the DSM-5. Each item was assessed on a 4-point scale (0 = not at all, 1 = very little, 2 = somewhat, 3 = to a great extent). We examined the average of all SRSCUD items as a continuous outcome (Model 4).

To minimize response burden/measure redundancy, participants were randomized to complete one of the other three measures of CUD symptoms: the 8-item Cannabis Use Disorder Identification Test-Revised (CUDIT-R; Adamson et al., 2010, n = 632), the 6-item Cannabis Abuse Screening Test (CAST; Legleye et al., 2007, n = 679), and the 5-item Severity of Dependence Scale (SDS; Gossop et al., 1995, n = 712). We examined the continuous scores for each of these measures (sum for CUDIT-R, Model 5; average for CAST, Model 6; average for SDS, Model 7).

Predictor Variables

A total of 178 predictor variables were entered into the models. Predictors were 23 demographic factors (e.g., age, sex/gender), 40 cannabis use indicators (e.g., age of first use, typical frequency, typical quantity), and 115 psychosocial/cannabis use-related constructs (e.g., Protective Behavioral Strategies-Marijuana, UPPS-P) (for a complete list of predictors see Supplemental Table 1). All predictors were included in each model unless otherwise noted.

Analysis Plan

We tested 7 models with different baseline predictor variables using decision tree learning using the ‘rpart’ package (Therneau & Atkinson, 2019) in R (R Core Team, 2019). Decision tree learning was selected as the analytic approach due to the relative ease and intuitiveness of model interpretation. Measures of CUD were modeled using both continuous scoring and dichotomous coding as being above (1) or below (0) the reported cutoff for probable CUD for each individual measure. Models 1 and 2 were dichotomous measures (0=no, 1=yes) of lifetime cannabis use and past month cannabis use, respectively. For Models 1 and 2, measures that were only assessed among past month cannabis users were excluded as predictors given that they were missing for all non-users. Therefore, Models 1 and 2 had 138 total predictors and Models 3–7 had 178 predictors. Model 3 was a continuous measure of negative cannabis-related consequences measured using the MACQ. Each measure of CUD symptoms was examined as a continuous outcome (SRSCUD Model 4, CUDIT-R Model 5, CAST Model 6, and SDS Model 7). To create parsimonious models and to decrease the risk of overfitting, branches that did not improve prediction accuracy in cross-validation (k = 10) were removed or “pruned”. A complexity parameter of 0.03 was selected (Ture & Omurlu, 2018). Default settings were retained for the minimum N to make a split (minsplit = 20), maximum depth (30), and minimum N in a terminal node (minsplit/3).

Results

In the full sample of 7000 participants, 57.3% (n = 3979) reported using cannabis in their lifetime, and 52.3% (n = 2077) reported using cannabis in the past month. Among past month users, our sample used cannabis on average 5.86 time periods (i.e., 4-hour blocks of time, SD = 7.91) during a typical week, averaging around 6.03 grams per week (SD = 12.04, Winsorized). Our sample reported experiencing an average of 3.64 negative cannabis-related consequences (SD = 3.90) in the past month. Based on established cutoff scores for each of our CUD symptom measures (e.g., Adamson et al., 2010, Gossop et al., 1995, Legleye et al., 2007, Richards et al., 2020), 7.7% (CAST) to 34.4% (SRSCUD) of our sample were determined to have a probable CUD, reflecting the different sensitivity and specificity for each measure.

Decision Tree Learning

Decision trees are presented in Figures 1 and 2. These trees are visually intuitive, and may be viewed while reading the results. Table 1 contains a list of the significant predictors identified by the decision trees. Table 2 contains model fit metrics.

Figure 1.

Figure 1.

Open in a new tab

Decision tress predicting lifetime cannabis use status (Model 1), past month cannabis use status (Model 2), and negative cannabis-related consequences (Model 3)

Figure 2.

Figure 2.

Open in a new tab

Decision tress predicting CUD symptom score as measured by the Self-Reported Symptoms of Cannabis Use Disorder (SRSCUD; Model 4), Cannabis Use Disorder Identification Test-Revised (CUDIT-R; Model 5), Cannabis Absue Screening Test (CAST; Model 6), and Severity of Dependence Scale (SDS; Model 7)

Table 1.

Summary of Predictors Identified Across All Models

M SD
Predictors Model: 1 2 3 4 5 6 7
Ease of Obtaining Cannabis 3.73 1.31 X
Pros of Cannabis 2.77 1.08 X
Cons of Cannabis 3.39 1.06 X
Time Spent-Substance Use Activities 3.85 20.92 X
Close Friend Descriptive Norms Frequency 6.53 10.81 X
Typical Frequency of Cannabis Use 5.86 7.91 X
Conformity Motives 1.42 0.74 X X X
Coping Motives 2.10 1.06 X X
Past Month Cannabis Frequency 5.13 8.87 X X X
Money Spent on Cannabis 57.88 363.67 X
% Time Using Cannabis Alone 15.88 27.01 X
Protective Behavioral Strategies (PBS) 4.34 1.02 X
Lack of Emotional Clarity 2.23 1.00 X
% Black Market Purchases 33.51 41.32 X
GPA 1.82 0.69 X
Number of splits 2 3 1 5 3 6 4
Open in a new tab

Note. Model 1=Lifetime cannabis use. Model 2=Past-month cannabis use. Model 3=Negative cannabis-related consequences (MACQ). Model 4=SRSCUD continuous score. Model 5= CUDIT-R continuous score. Model 6= CAST continuous score. Model 7=SDS continuous score.

Table 2.

Model Fit Metrics

Sensitivity Specificity PPV NPV Accuracy MAE MSE RMSE
Model 1 72.56% 76.75% 69.92% 78.98% 74.96% --- --- ---
Model 2 61.37% 82.81% 76.51% 70.15% 72.58% --- --- ---
Model 3 --- --- --- --- --- 2.92 16.90 4.11
Model 4 --- --- --- --- --- 0.31 0.19 0.44
Model 5 --- --- --- --- --- 3.55 21.38 4.62
Model 6 --- --- --- --- --- 0.44 0.34 0.58
Model 7 --- --- --- --- --- 1.43 4.28 2.07
Open in a new tab

Note. PPV = Positive Predictive Value. NPV = Negative Predictive Value. MAE = Mean absolute error. MSE = Mean squared error. RMSE = Root mean square error.

Lifetime Cannabis Use

Two variables contributed to the final tree predicting lifetime cannabis user status (see Figure 1-Model 1): ease of obtaining cannabis and pros of cannabis. In the overall sample (n=6940), 57.3% of individuals reported lifetime cannabis use. Among individuals who reported it being “probably impossible,” “very difficult,” or “fairly difficult” to obtain cannabis (n = 2142), 27.9% reported lifetime cannabis use (classification: non-cannabis user). Among individuals who reported it being “fairly easy” or “very easy” to obtain cannabis (n = 4798), 70.5% reported lifetime cannabis use. This group was split based on perceived pros of cannabis based on the Marijuana Decision Balance scale. The response scale for the pros of cannabis subscale ranged from 1 = Not Important to 5 = Extremely Important. Among individuals (n = 1151) who reported potential pros of cannabis as not being very important to their decision to use/not use cannabis (< 2.16), 42.6% reported lifetime cannabis use (classification: non-cannabis user). Among individuals (n = 3647) with rated pros of cannabis as more important to their decision to use/not use cannabis (≥ 2.16), 79.3% reported lifetime cannabis use (classification: cannabis user).

Past Month Cannabis Use

Three variables contributed to the final tree predicting past month cannabis user status (Figure 1-Model 2): time spent engaging in substance use activities, close friends’ descriptive norms frequency, and cons of cannabis. Among the overall sample (n = 3972), 52.3% reported past month cannabis use. Among individuals who reported spending 0 hours per week engaging in substance use activities (n = 722), 19.7% reported past month cannabis use (classification: non-past month cannabis user). Among individuals who spent more than 0 hours per week engaging in substance use activities (n = 3250), 59.5% reported past month cannabis use. This group was split based on their perceptions of how frequently their close friends use cannabis during a typical week. Among individuals (n = 351) who perceive their close friends to abstain from cannabis during a typical week (< 0.5), 21.9% reported past month cannabis use (classification: non-past month cannabis user). Among individuals (n = 2899) who perceive their close friends to use at least once during a typical week (≥ 0.5), 64.1% reported past month cannabis use. This group was split based on their cons of cannabis use based on the Marijuana Decisional Balance scale. The response scale for the cons of cannabis subscale ranged from 1 = Not Important to 5 = Extremely Important. Among individuals (n = 1151) who rated cons of cannabis as important to their decision to use/not use cannabis (≥ 3.13), 47.5% reported past month cannabis use (classification: non-past month cannabis user). Among individuals (n = 1748) who rated cons of cannabis as less important to their decision to use/not use cannabis (< 3.13), 75% reported past month cannabis use (classification: past month cannabis user).

Negative Cannabis-Related Consequences (MACQ)

One variable contributed to the final tree predictive negative cannabis-related consequences (Figure 1-Model 3): typical frequency of cannabis use based on the Marijuana Use Grid. The mean number of consequences among the overall sample (n = 2036) was 3.64. Individuals who reported using cannabis on 4 or fewer time periods during a typical week (n = 1320) reported an average of 2.52 negative cannabis-related consequences (classification: low consequences). Individuals who reported using cannabis on 5 or more time periods during a typical week (n = 716) reported an average of 5.72 negative consequences (classification: high consequences).

SRSCUD Continuous Symptom Score

Three variables contributed to the final tree predicting CUD symptoms based on the SRSCUD (Figure 2-Model 4): coping motives, past month cannabis frequency, and conformity motives. The average CUD symptom severity was low (M = 1.48; 1 = Not at all, 2 = Very little, 3 = Somewhat, 4 = To a great extent) in the sample of past month cannabis users (n = 2023). Individuals (n = 1052) with low average coping motive scores (< 1.9, less than “2 = Some of the time”) reported below average CUD symptom severity (1.24). This group was further split between an even lower risk group (1.15) who reported using cannabis on less than 7 or fewer days during the past month (n = 745; classification: low CUD symptoms), and a slightly higher risk group (1.46) who used cannabis on 8 or more days during the past month (n = 307; classification: low CUD symptoms). Individuals (n = 971) with moderate to high average coping motive scores (≥ 1.9) reported an average of 1.75 symptoms. This group was split based on average conformity motive scores. Individuals (n = 773) with low to moderate average conformity motive scores (<2.55) reported an average of 1.61 symptoms. This group was split based on past month cannabis frequency. Individuals who used cannabis on 10 or fewer days during the past month (n = 425), reported an average of 1.42 symptoms (classification: low CUD symptoms), whereas those who used cannabis on 11 or more days during the past month (n = 348) reported an average of 1.85 symptoms (classification: low CUD symptoms). Individuals (n = 198) with moderate to high average conformity motive scores (≥ 2.55) reported an average of 2.28 symptoms. This group was split based on average coping motive scores. Individuals (n = 188) with non-maximum average coping motive scores (< 4.9) reported an average of 2.20 symptoms (classification: moderate CUD symptoms). Individuals (n = 10) with maximum average coping motive scores (≥ 4.9) reported an average of 3.66 symptoms (classification: high CUD symptoms).

CUDIT-R Continuous Symptom Score

Two variables contributed to the final tree predicting CUD symptoms based on the CUDIT-R (Figure 2-Model 5): past month cannabis frequency and coping motives. The average CUDIT-R score reported among the overall sample (n = 632) were 8.15. Individuals who used cannabis on 16 or fewer days during the past month (n = 467) reported an average of 6.15 symptoms. This group was split based on average coping motive score. Individuals (n = 214) with a low average coping motive score (< 1.5) reported an average CUDIT-R score of 4.47 (classification: low symptoms) whereas those (n = 253) with a moderate to high average coping motive score (≥ 1.5) reported an average CUDIT-R score of 7.56 (classification: low symptoms). Individuals who used cannabis 17 or more days during the past month (n = 165) reported an average of 13.81 symptoms. This group was split based on average coping motive score. Individuals (n = 88) with a low to moderate average coping motive score (< 2.7) reported an average CUDIT-R score of 11.57 (classification: moderate symptoms), and individuals (n = 77) with a moderate to high average coping motive score (≥ 2.7) reported an average CUDIT-R score of 16.38 (classification: high symptoms).

CAST Continuous Symptom Score

Five variables contributed to the final tree predicting CUD symptoms based on the CAST (Figure 2-Model 6): money spent on cannabis, conformity motives, percent time using cannabis alone, use of cannabis protective behavioral strategies, and lack of emotional clarity. The average CAST score reported among the overall sample (n = 679) was 1.85. Individuals who spent less than $32.50 on cannabis during the past month (n = 468) reported an average of 1.56 symptoms. This group was split based on average conformity motive score. Individuals (n = 42) with a moderate to high average conformity motive score (≥ 2.9) reported an average CAST score of 2.35 (classification: moderate symptoms). Individuals (n = 426) with a low to moderate average conformity motive score (< 2.9) reported an average CAST score of 1.48. This group was split based on percent time using cannabis alone. Individuals (n = 335) who did not report using cannabis alone (< 0.5%) during the past month reported an average CAST score of 1.37 (classification: low symptoms). Individuals (n = 91) who did reporting using cannabis alone (> 0.5%) during the past month reported an average CAST score of 1.89 (classification: low symptoms). Individuals who spent $32.50 or more on cannabis during the past month (n = 211) reported an average of 2.48 symptoms. This group was split based on average conformity motive score such that individuals (n = 27) with a moderate to high average conformity motive score (≥ 2.5) reported an average of 3.33 symptoms (classification: high symptoms) whereas individuals (n = 184) with a low to moderate average conformity motive score (< 2.5) reported an average of 2.36 symptoms. This group was split based on use of protective behavioral strategies. Individuals (n = 51) who reported higher use of protective behavioral strategies (≥ 4.32) reported an average CAST score of 1.83 (classification: low symptoms) whereas individuals (n = 133) who reported lower use of protective behavioral strategies (< 4.32) reported an average CAST score of 2.56 (classification: moderate symptoms). This group was split based on lack of emotion clarity based on the Difficulty in Emotion Regulation Scale. The response scale ranged from 1 = Almost never to 5 = Almost always. Individuals (n = 102) who reported “almost never” or “sometimes” having a lack of emotional clarity experienced an average of 2.38 symptoms (classification: moderate symptoms) and individuals who reported “about half of the time” or more having a lack of emotional clarity experienced an average of 3.15 symptoms (classification: high symptoms).

SDS Continuous Symptom Score

Four variables contributed to the final tree predicting CUD symptoms based on the SDS (Figure 2-Model 7): conformity motives, past month cannabis frequency, percent of black-market purchases of cannabis (i.e., bought not from a dispensary), and grade point average. The average number SDS score reported among the overall sample (n = 712) was 1.55. Individuals (n = 602) with a low average conformity motive score (< 2.10) reported an average SDS score of 1.06. This group was split based on days of past month cannabis use. Individuals who used cannabis 14 or fewer days during the past month (n = 429) reported an average of 0.64 symptoms (classification: low symptoms). Individuals who used cannabis on 15 or more days during the past month (n = 173) reported an average of 2.09 symptoms (classification: moderate symptoms). Individuals (n = 110) with a moderate to high average conformity motive score (≥ 2.10) reported an average SDS score of 4.25. This group was split based on percentage of cannabis purchased through black market sources during the past month. Individuals (n = 60) who purchased less than 15% of cannabis from the black market during the past month reported an average SDS score of 2.73 (classification: moderate symptoms). Individuals (n = 50) who purchased at least 15% of cannabis from the black market during the past month reported an average SDS score of 6.06 (classification: high symptoms). This group was split based on estimated grade point average. Individuals (n = 13) who had an estimated grade point average of an “A” reported an average of 3.15 consequences (classification: moderate symptoms) and individuals (n = 37) who did not have an estimated grade point average below an “A” reported an average of 7.08 consequences (classification: high symptoms).

Discussion

Across all models, 15 distinct predictors were identified: ease of obtaining cannabis, pros of cannabis use, cons of cannabis use, time spent engaging in substance use activities, close friends descriptive norms frequency, typical frequency of cannabis use, conformity motives, coping motives, past month cannabis frequency, money spent on cannabis, percent time using cannabis alone, use of cannabis protective behavioral strategies, lack of emotional clarity, percent of black-market purchases of cannabis, and grade point average (7 trees, 24 splits).

Consistent with Wilson et al. (2018), predictors of user status were largely related to one’s beliefs about cannabis. Although Wilson et al. (2018) identified largely normative beliefs (e.g., beliefs about others’ behaviors) as predictors of user status, we also found higher pros about the effects of cannabis (for lifetime user status) and lower cons about the effects of cannabis (for past month user status) from a marijuana decisional balance measure (Elliott et al., 2011) as predictors of user status. In terms of normative variables, we found the perceived frequency of cannabis use by close friends (i.e., descriptive norms) measured by the Marijuana Norms Grid (Montes et al., 2020) predicted past month user status. We did not find any of the same predictors of regular cannabis use found by Parekh and Fakim (2021) and Rajapaksha and colleagues (2020). This may be due to having used different measures to assess similar constructs (e.g., mental health, personality traits).

From a behavioral economic measure of time allocation, we found time spent engaging in substance use activities was predictive of past month user status. This predictor was a strong indicator of user status despite the measure broadly assessing engagement in substance use activities (as opposed to substance-specific engagement). Similarly, perceived ease of obtaining cannabis was a strong predictor of user status. Individuals who easily obtained cannabis were more likely to be lifetime users as opposed to those who reported difficulty obtaining cannabis.

Typical frequency of cannabis use as assessed by the Marijuana Use Grid (Pearson et al., 2020) was the only salient predictor of negative cannabis-related consequences. This result may provide further support for the consistent relationship found between frequency of cannabis use and consequences (Looby & Earleywine, 2007; Pearson, 2019).

Across the models predicting CUD symptoms (models 4–7), negative reinforcement cannabis use motives (i.e., coping and conformity motives) were the most consistent predictors across models (8 splits). Individuals with higher coping and conformity motive scores typically experienced more CUD symptoms. Past month cannabis use was the next most consistent predictor across models, with individuals who reported more frequent use experiencing more symptoms (4 splits).

Clinically, these results can be applied by considering salient variables alone or in conjunction with the use of existing measures of CUD. A cognitive behavioral therapist might examine the thoughts and other cognitive/behavioral antecedents that lead to an individual using cannabis to cope or conform with others. By understanding the cognitions influencing these behaviors, a therapist can help an individual develop other coping and interpersonal effectiveness skills thus decreasing their need to engage in potentially problematic cannabis use. Teaching coping skills to students before entering college might also be of preventative benefit and might decrease the number of CUD symptoms an individual experiences.

Another therapeutic target might be to focus on frequency of cannabis use, time spent using cannabis, and money spent on cannabis. These targets are common in brief motivational interventions and normative feedback interventions (Saxton et al., 2021) where discussions are often centered on the benefits and consequences of changing the frequency and time spent using cannabis as well as the financial impact of cannabis use. Future research might compare the impact of different components of personalized normative feedback interventions to examine if certain components (e.g., financial and temporal consequences) are more effective at addressing the various targets.

Given the sociodemographic and geographical diversity of our sample, our results are likely to generalize to the broader U.S. college student population. However, it is unclear whether such findings would be similar among non-college attending adults, or in distinct cultural milieu (i.e., across cultures). Although our analytic methods were similar to Wilson et al. (2018), we could not ascertain the degree of replication given the assessment batteries across these studies differed substantially. Further, although we included demographics as predictor variables, most demographic factors are immutable and therefore cannot be treatment targets (e.g., family socioeconomic status, race/ethnicity).

Although other machine learning approaches can be used to cross-validate findings, the typical approaches (e.g., Random Forest models) are more difficult to interpret than a decision tree. Thus, we decided to not cross-validate results in favor of a more interpretable product (e.g., decision tree). Within our sample, we observed overlap in significant predictor variables between Models 4–7 (each outcome variable was a different measure of CUD), suggesting some level of replicability. However, additional work examining the replicability of these predictors is recommended.

The present study applied decision tree learning to identify clinically relevant predictors of cannabis user status, negative cannabis-related consequences, and CUD symptoms. This study provides further support for beliefs about cannabis predicting user status and found that negative reinforcement cannabis use motives are associated with increased CUD symptoms. Findings support the continued use of exploratory data analytic techniques to identify variables that might predict CUD symptoms and probable-CUD diagnosis. Findings may be helpful clinically to diagnose CUD and inform case conceptualization and prevention and treatment targets.

Supplementary Material

1

Acknowledgments

This work was supported by the National Institute of Alcoholism and Alcohol Abuse and the National Institute on Drug Abuse of the National Institutes of Health, award numbers T32 AA018108, F32 AA028712, and K01 AA023233. The content is the sole responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. The authors have no conflicts of interest.

Footnotes

1

IRBs (listed alphabetically): Colorado State University, Old Dominion University, University of Buffalo, University of California, Los Angeles, University of Central Florida, University of Houston, University of Memphis, University of New Mexico, University of Washington

Contributor Information

Frank J. Schwebel, Center on Alcohol, Substance use, And Addictions, University of New Mexico

Dylan K. Richards, Center on Alcohol, Substance use, And Addictions, University of New Mexico

Rory A. Pfund, Center on Alcohol, Substance use, And Addictions, University of New Mexico

Verlin W. Joseph, Center on Alcohol, Substance use, And Addictions, University of New Mexico

Matthew R. Pearson, Center on Alcohol, Substance use, And Addictions, University of New Mexico

References

  1. Adamson SJ, Kay-Lambkin FJ, Baker AL, Lewin TJ, Thornton L, Kelly BJ, & Sellman JD (2010). An improved brief measure of cannabis misuse: the Cannabis Use Disorders Identification Test-Revised (CUDIT-R). Drug and Alcohol Dependence, 110(1–2), 137–143. [DOI] [PubMed] [Google Scholar]
  2. Arnett JJ (2000). Emerging adulthood. A theory of development from the late teens through the twenties. American Psychologist, 55(5), 469–480. [PubMed] [Google Scholar]
  3. Arnett JJ (2005). The developmental context of substance use in emerging adulthood. Journal of Drug Issues, 35, 235–253. 10.1177/002204260503500202 [DOI] [Google Scholar]
  4. Arterberry BJ, Boyd CJ, West BT, Schepis TS, & McCabe SE (2020). DSM-5 substance use disorders among college-age young adults in the United States: Prevalence, remission and treatment. Journal of American College Health, 68, 650–657. [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Budney AJ, & Hughes JR (2006). The cannabis withdrawal syndrome. Current Opinion in Psychiatry, 19(3), 233–238. [DOI] [PubMed] [Google Scholar]
  6. Budney AJ, Hughes JR, Moore BA, & Vandrey R (2004). Review of the validity and significance of cannabis withdrawal syndrome. American Journal of Psychiatry, 161(11), 1967–1977. [DOI] [PubMed] [Google Scholar]
  7. Collins RL, Parks GA, & Marlatt GA (1985). Social determinants of alcohol consumption: The effects of social interaction and model status on the self-administration of alcohol. Journal of Consulting and Clinical Psychology, 53, 189–200. [DOI] [PubMed] [Google Scholar]
  8. Elliott JC, Carey KB, & Scott-Sheldon LA (2011). Development of a decisional balance scale for young adult marijuana use. Psychology of Addictive Behaviors, 25(1), 90–100. [DOI] [PMC free article] [PubMed] [Google Scholar]
  9. Foster KT, Arterberry BJ, Iacono WG, McGue M, & Hicks BM (2018). Psychosocial functioning among regular cannabis users with and without cannabis use disorder. Psychological Medicine, 48, 1853–1861. [DOI] [PMC free article] [PubMed] [Google Scholar]
  10. Galbraith T, & Conner BT (2015). Religiosity as a moderator of the relation between sensation seeking and substance use for college-aged individuals. Psychology of Addictive Behaviors, 29, 168–175. [DOI] [PubMed] [Google Scholar]
  11. Gossop M, Darke S, Griffiths P, Hando J, Powis B, Hall W, & Strang J (1995). The Severity of Dependence Scale (SDS): psychometric properties of the SDS in English and Australian samples of heroin, cocaine and amphetamine users. Addiction, 90(5), 607–614. [DOI] [PubMed] [Google Scholar]
  12. Graham JW, Taylor BJ, Olchowski AE, & Cumsille PE (2006). Planned missing data designs in psychological research. Psychological Methods, 11(4), 323. [DOI] [PubMed] [Google Scholar]
  13. Han B, Compton WM, Blanco C, & Jones CM (2019). Time since first cannabis use and 12-month prevalence of cannabis use disorder among youth and emerging adults in the United States. Addiction, 114, 698–707. [DOI] [PMC free article] [PubMed] [Google Scholar]
  14. Legleye S, Karila L, Beck F, & Reynaud M (2007). Validation of the CAST, a general population Cannabis Abuse Screening Test. Journal of Substance Use, 12(4), 233–242. [Google Scholar]
  15. Looby A, & Earleywine M (2007). Negative consequences associated with dependence in daily cannabis users. Substance Abuse Treatment, Prevention, and Policy, 2(1), 3. [DOI] [PMC free article] [PubMed] [Google Scholar]
  16. Michalski RS, Carbonell JG, & Mitchell TM (Eds.). (2013). Machine learning: An artificial intelligence approach. New York, NY: Springer Science+Business Media. [Google Scholar]
  17. Miech RA, Patrick ME, O’Malley PM, & Johnston LD (2017). The influence of college attendance on risk for marijuana initiation in the United States: 1977 to 2015. American Journal of Public Health, 107, 996–1002. [DOI] [PMC free article] [PubMed] [Google Scholar]
  18. Moitra E, Christopher PP, Anderson BJ, & Stein MD (2015). Coping-motivated marijuana use correlates with DSM-5 cannabis use disorder and psychological distress among emerging adults. Psychology of Addictive Behaviors, 29, 627–632. [DOI] [PMC free article] [PubMed] [Google Scholar]
  19. Montes KS, Richard DK, Pearson MR, & Marijuana Outcomes Study Team (2020). A novel approach to assess descriptive and injunctive norms for college student marijuana use. Addictive Behaviors. [DOI] [PMC free article] [PubMed] [Google Scholar]
  20. Parekh T, & Fahim F (2021). Building risk prediction models for daily use of marijuana using machine learning techniques. Drug and Alcohol Dependence, 225, 108789. 10.1016/j.drugalcdep.2021.108789 [DOI] [PubMed] [Google Scholar]
  21. Pearson MR (2019). A meta-analytic investigation of the associations between cannabis use and cannabis-related negative consequences. Psychology of Addictive Behaviors, 33(3), 190. [DOI] [PMC free article] [PubMed] [Google Scholar]
  22. Pearson MR, Kholodkov T, Gray MJ, & Marijuana Outcomes Study Team. (2017a). Perceived Importance of Marijuana to the College Experience Scale (PIMCES): Initial development and validation. Journal of Studies on Alcohol and Drugs, 78, 319–324. [DOI] [PMC free article] [PubMed] [Google Scholar]
  23. Pearson MR, Liese BS, Dvorak RD, & Marijuana Outcomes Study Team. (2017b). College student marijuana involvement: Perceptions, use, and consequences across 11 college campuses. Addictive Behaviors, 66, 83–89. [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Pearson MR, Marijuana Outcomes Study Team, & Protective Strategies Study Team. (2020). Marijuana Use Grid: A brief, comprehensive measure of cannabis use. Unpublished manuscript.
  25. Pedersen ER, Hummer JF, Rinker DV, Traylor ZK, & Neighbors C (2016). Measuring protective behavioral strategies for marijuana use among young adults. Journal of Studies on Alcohol and Drugs, 77, 441–450. [DOI] [PMC free article] [PubMed] [Google Scholar]
  26. R Core Team (2019). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. URL https://www.R-project.org/. [Google Scholar]
  27. Rajapaksha RMDS, Hammonds R, Filbey F, Choudhary PK, & Biswas S (2020). A preliminary risk prediction model for cannabis use disorder. Preventive Medicine Reports, 20, 101228. 10.1016/j.pmedr.2020.101228 [DOI] [PMC free article] [PubMed] [Google Scholar]
  28. Richards DK, Schwebel FJ, Sotelo M, Pearson MR, & Marijuana Outcomes Study Team (2021). Self-Reported Symptoms of Cannabis Use Disorder (SRSCUD): Psychometric Testing and Validation. Experimental and Clinical Psychoharmacology. [DOI] [PMC free article] [PubMed] [Google Scholar]
  29. Saxton J, Rodda SN, Booth N, Merkouris SS, & Dowling NA (2021). The efficacy of Personalized Normative Feedback interventions across addictions: A systematic review and meta-analysis. PloS one, 16(4), e0248262. [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. Schulenberg JE, Johnston LD, O’Malley PM, Bachman JG, Miech RA & Patrick ME (2020). Monitoring the Future national survey results on drug use, 1975–2019: Volume II, College students and adults ages 19–60. Ann Arbor: Institute for Social Research, The University of Michigan. Available at http://monitoringthefuture.org/pubs.html#monographs [Google Scholar]
  31. Simons JS, Dvorak RD, Merrill JE, & Read JP (2012). Dimensions and severity of marijuana consequences: Development and validation of the Marijuana Consequences Questionnaire (MACQ). Addictive Behaviors, 37(5), 613–621. [DOI] [PMC free article] [PubMed] [Google Scholar]
  32. Strobl C, Malley J, & Tutz G (2009). An introduction to recursive partitioning: Rationale, application, and characteristics of classification and regression trees, bagging, and random forests. Psychological Methods, 14, 323–348. [DOI] [PMC free article] [PubMed] [Google Scholar]
  33. Therneau T & Atkinson B (2019). rpart: Recursive Partitioning and Regression Trees. R package version 4.1–15. https://CRAN.R-project.org/package=rpart [Google Scholar]
  34. Ture M, & Kurt Omurlu I (2018). Determining of complexity parameter for recursive partitioning trees by simulation of survival data and an application on breast cancer data. Journal of Statistics and Management Systems, 21(1), 125–138. [Google Scholar]
  35. Wilson AD, Montes KS, Bravo AJ, Conner BT, Pearson MR, & Marijuana Outcomes Study Team. (2018). Making decisions with trees: Examining marijuana outcomes among college students using recursive partitioning. Clinical Psychological Science, 6, 744–754. [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

1
NIHMS1766981-supplement-1.pdf (196.6KB, pdf)

RESOURCES