Longitudinal Craving Profiles in Cannabis Use Disorder: A Latent Class Growth Analysis of the Achieving Cannabis Cessation: Evaluating N-Acetylcysteine Treatment (ACCENT) (CTN-0053) Trial

Abstract

Background:

Craving is a core clinical feature of cannabis use disorder (CUD) and a predictor of treatment outcomes, yet its temporal course during treatment is not well characterized. This study aimed to identify latent classes of cannabis craving trajectories among adults with CUD and examine associated baseline predictors and cannabis use outcomes.

Methods:

This was a secondary analysis of the National Drug Abuse Treatment Clinical Trials Network’s CTN-0053 trial, a 12-week, multisite randomized controlled trial of N-acetylcysteine versus placebo for adults with CUD (N=302). Cannabis craving was measured using the Marijuana Craving Questionnaire–Short Form at six treatment timepoints (weeks 1-5, 9 and 12) and one 5-week post-treatment follow-up (week 17). Urine cannabinoid tests were conducted twice weekly throughout treatment and follow-up. Latent class growth analysis identified craving trajectories. The present study aimed to identify latent classes of cannabis craving over 12 weeks of treatment and examine baseline predictors of class membership.

Results:

A four-class solution provided the best fit: low craving (41%), moderate-decreasing craving (38%), moderate-stable craving (11%), and high craving (10%). Participants in higher craving classes exhibited greater baseline anxiety, depression, and obsessive-compulsive symptoms related to cannabis use. The high craving class had the greatest proportion of cannabis positive urine tests (96%) and the lowest urine test completion rate.

Conclusions:

Craving follows heterogeneous trajectories during CUD treatment and is associated with co-occurring mental health symptoms and poorer outcomes. Dynamic craving assessment may support personalized treatment and strategies to prevent return to use.

1. Introduction

Cannabis remains the most commonly used federally prohibited substance in the United States (U.S.), representing a significant challenge for public health monitoring and intervention, particularly in the context of expanding legalization, which has increased availability and may influence patterns of use and treatment needs. In 2022, approximately 61.9 million people in the U.S. — 22% of the population — used cannabis (Substance Abuse and Mental Health Services Administration, 2023). About 1 in 10 people who use cannabis develop cannabis use disorder (CUD) with risk increasing to 1 in 6 if use begins before the age of 18 (The Effects, Risks and Side Effects of Marijuana (Weed, Cannabis, THC) | SAMHSA, 2024). CUD is characterized by continued use of cannabis despite the negative consequences of use causing clinically significant impairment or distress (American Psychiatric Association, 2013) such as difficulty concentrating, short-term memory loss, and loss of motivation.

CUD is a growing public health concern, as its prevalence has more than quadrupled over the past decade (Substance Abuse and Mental Health Services Administration, 2025, 2016). The most recent available data suggest that among people aged 12 or older, 16.7 million (6.%) have been diagnosed with a CUD in the past year (Substance Abuse and Mental Health Services Administration, 2025). Based on an analysis of the 2022 National Survey on Drug Use and Health (NSDUH), (Substance Abuse and Mental Health Services Administration, 2023) reported that, for the first time, the number of people in the U.S. who report daily or near-daily cannabis use (17.7 million) exceeded those reporting daily or near-daily alcohol use (14.7 million), underscoring the substantial public health relevance of cannabis-related harms and the need for early identification and access to appropriate targeted treatments and support services.

Craving—defined as an intense desire to use a substance—is a core feature of substance use disorders (SUDs) that plays a significant role in lower treatment engagement and lower retention, predicts poorer treatment outcomes, and strongly predicts return to cannabis use (Sayette, 2016); however, its heterogeneity over the course of CUD treatment is poorly understood (Cavicchioli et al., 2020; Cless et al., 2023).

Craving can be triggered by environmental cues, stress, or withdrawal symptoms, reinforcing habitual use and making cessation efforts more challenging (Rosenberg, 2013). However, craving may not be experienced in the same way by all individuals. Some studies suggest that craving trajectories over time vary between individuals in intensity, persistence, and cue-reactivity (Fatseas et al., 2015; Filbey et al., 2009; Moore et al., 2014). But how craving severity and its persistence during treatment differ across individuals and how it might predict outcomes remains poorly understood, potentially resulting in missed opportunities for more personalized and effective care.

Previous studies demonstrate that cannabis craving strongly predict return to use (e.g., Enkema et al., 2020), making it a critical target for intervention and monitoring in treatment settings. These studies highlight the importance of craving, but they have relied on mean-level or aggregate change scores, which can obscure clinically meaningful heterogeneity in craving trajectories that could inform personalized treatment strategies. Latent class growth analysis (LCGA) offers a person-centered approach to identify subgroups characterized by different levels and patterns of craving severity across repeated assessments, even when average rates of change are similar (Jung and Wickrama, 2008; Nagin, 2005; Ram and Grimm, 2009). In clinical trial settings, this approach can complement baseline craving assessments by highlighting individuals with persistently elevated craving and distinct clinical correlates that may benefit from targeted intervention strategies.

We conducted a secondary exploratory analysis of a multisite National Drug Abuse Treatment Clinical Trials Network (CTN-0053) randomized controlled trial evaluating N-acetylcysteine (NAC) for adults with CUD (Gray et al., 2017). Although the parent trial did not find significant group differences in cannabis abstinence (Gray et al., 2017) or craving (Roydhouse et al., 2021), craving may serve as a transdiagnostic feature and may vary in meaningful severity-based subgroups of individuals over time. The present study aimed to identify latent classes of cannabis craving trajectories over 12 weeks of treatment and examine baseline predictors of class membership.

2. Methods

2.1. Study design and participants

Participants were enrolled (n=302) in the CTN-0053 trial (Achieving Cannabis Cessation: Evaluating N-Acetylcysteine Treatment [ACCENT]) (Gray et al., 2017).

This US-based six-site clinical trial aimed to evaluate the effect of NAC 1200 mg versus placebo on cannabis use disorder (CUD), with cannabis abstinence as the primary outcome, assessed via biochemically verified urine cannabinoid testing (UCT) conducted throughout treatment and follow-up. Participants were 18-50 years old and met Diagnostic and Statistical Manual of Mental Disorders, 4th edition (American Psychiatric Association (APA), 2000) criteria for cannabis dependence. To be included in the study, participants had to have a positive urine cannabinoid test for cannabis at baseline.

Participants were randomized 1:1 to receive 1200 mg of NAC twice daily or placebo over the course of 12 weeks. All participants received contingency management (CM) throughout the study — an evidence-based intervention in which participants earn monetary incentives (e.g. cash or gift card) in exchange for target behaviors. In this trial, CM reinforced both cannabis abstinence, confirmed by negative urine cannabinoid tests, and attendance at study visits, using an escalating schedule with resets. Participants were seen twice per week throughout the study to receive urine cannabinoid testing and contingency management. During one weekly visit, participants completed self-assessments, such as measures related to mental health and substance use. This clinical trial took place at six sites in South Carolina, Oregon, California, Connecticut, Texas, and Kentucky. At the time of recruitment (January 2014–April 2015), recreational cannabis use was legal only in Oregon, while medical cannabis was legal in Oregon, California, Connecticut, and Texas.

2.2. Measures

Demographic variables (age, sex, race, ethnicity, and education) were measured by self-report at the baseline visit. Baseline psychological characteristics were measured using validated self-report scales: anxiety and depression levels with the Hospital Anxiety and Depression Scale (HADS)(Zigmond and Snaith, 1983) and compulsive thoughts and patterns related to substance use with the Obsessive-Compulsive Drug Use Scale (OCDUS) (Franken et al., 2002).

Our main latent class indicator variable of cannabis craving was measured with the Marijuana Craving Questionnaire (MCQ-SF), a reliable and valid measure of craving in people who use cannabis (Heishman et al., 2009). The MCQ-SF is a 12-item self-assessment measure designed to capture different dimensions of craving, including: 1) compulsivity, an inability to control marijuana use; 2) emotionality, use of marijuana in anticipation of relief from withdrawal or negative mood; 3) expectancy, anticipation of positive outcomes from smoking marijuana; and 4) purposefulness, intention and planning to use marijuana for positive outcomes. Each of these domains represents a specific aspect of the overarching construct of craving. Each item in the MCQ-SF is scored on a Likert scale from 1 (strongly disagree) to 7 (strongly agree). For this analysis, we computed the total summed score, noting that this assumes equal weighting of each item. While the MCQ-SF is multidimensional, prior research suggests that the total score is a valid and interpretable measure of overall craving severity (Roten et al., 2013), allowing examination of patterns of craving severity across repeated assessments. Total summed score ranged from 12 to 84, where a higher score indicates a higher level of marijuana craving.

2.3. Statistical analysis

We conducted a LCGA of MCQ-SF cannabis craving scores collected at baseline, across 12 weeks of treatment (specifically at weeks 1 through 5, week 9, and week 12), and again at a 5-week post-treatment follow-up. LCGA is a specific type of growth mixture modeling (GMM) in which within-class variances of growth parameters (e.g., intercepts and slopes) are fixed to zero, resulting in trajectory classes that are assumed to be internally homogeneous. In contrast, GMM allows for within-class variability in growth factors, capturing heterogeneity both between and within latent classes. LCGA was selected for the present analyses to prioritize parsimony, model stability, and interpretability, and because the primary objective was to identify distinct prototypical outcome trajectories rather than to model individual-level variation within classes. We ran models with 1-6 latent classes, and for each model, we included intercept, slope, quadratic, and cubic terms. When higher-order terms for a newly added class did not contribute meaningfully to the trajectory (i.e., near-zero and non-significant mean estimates), those terms were fixed to zero and the model was re-estimated to improve parsimony and interpretability. This process resulted in “revised” models that retained only growth terms supported by the data while holding within-class variances and covariances fixed across time, consistent with LCGA assumptions. Model selection was based on information criteria (AIC, BIC), the Vuong–Lo–Mendell–Rubin likelihood ratio test, entropy, class size, and substantive interpretability of the resulting trajectories (Nagin, 2005; Nagin and Odgers, 2010). Individuals were assigned to latent classes using modal (most likely) class membership based on posterior probabilities, consistent with standard LCGA practice. We used chi square tests to compare distributions of categorical variables including sex, ethnicity, race, education, and employment status, and ANOVA for continuous variables, including age, baseline MCQ-SF total score, HADS-Anxiety, HADS-Depression, and OCDUS score, and t-tests for continuous variables. Finally, to compare cannabis use outcomes across latent classes, we used the Bolck–Croon–Hagenaars (BCH) method, a three-step approach that incorporates classification uncertainty when estimating differences in distal outcomes across classes. Unlike analyses based on modal class assignment alone, BCH preserves the latent class solution while providing unbiased class-specific estimates of outcome means and tests of between-class differences(Asparouhov and Muthén, 2021). LCGA and BCH analyses were conducted in Mplus (Muthén and Muthén, 2017) and all other analysis were computed in R version 4.4.1. (R Core Team, 2023). We implemented Full Information Maximum Likelihood , robust to data missing at random.

3. Results

Fit statistics for LCGA models examining the marijuana craving score are presented in Table 1. Models labeled with an “R” suffix (e.g., 4R) indicate revised specifications of the corresponding k-class solution in which higher-order growth terms (e.g., quadratic or cubic) that did not meaningfully contribute to class-specific trajectories were fixed to zero and the model was rerun to determine if these modifications improved model fit. These revised models retained the same number of classes as the unconstrained solution but imposed additional parameter constraints to improve parsimony, stability, and interpretability. Thus, the 4R model represents a constrained four-class solution evaluated during model selection, rather than a distinct alternative class structure (Nagin, 2005; Nagin and Odgers, 2010; Ram and Grimm, 2009). After examining all potential models and fit indices, we determined that the 4-class solution had optimal model fit. This included lower AIC (AIC = 17,730) and BIC (BIC = 17,823) than other models, a VLMR p value below alpha (VLMR p = 0.0046), and a high entropy value (entropy = 0.868). The class sizes were all above 10% of the sample size and each class trajectory was clinically relevant (provided a individual group that differed in expression of craving across time). Trajectories of mean cannabis craving scores by most likely class membership are provided in Figure 1. The final 4-class solution included a low craving class (n = 123), a moderate and decreasing craving class (n = 116), a moderate and stable craving class (n = 33), and a high craving class (n = 30).

Table 1.

Model fit statistics for 1-6 class solutions in longitudinal latent class analysis of marijuana craving questionnaire total score in CTN 0053 (N = 302)

Classes	AIC	BIC	VLMR p	Entropy	Class sizes
1	19280.037	19328.272			1 = 302 (100%)
2	18325.822	18392.609	0.0030	0.893	1 = 96 (32%) 2 = 206 (68%)
3	17824.356	17909.696	0.0004	0.893	1 = 38 (13%) 2 = 132 (44%) 3 = 132 (44%)
3R	18100.740	18174.949	0.0003	0.879	1 = 32 (11%) 2 = 192 (64%) 3 = 78 (26%)
4	17730.462	17823.222	0.0046	0.868	1 = 123 (41%) 2 = 116 (38%) 3 = 33 (11%) 4 = 30 (10%)
4R	17732.952	17822.002	0.1586	0.867	1 = 30 (10%) 2 = 121 (40%) 3 = 32 (11%) 4 = 119 (39%)
5	17598.322	17705.925	0.9229	0.847	1 = 25 (8%) 2 = 104 (34%) 3 = 35 (12%) 4 = 35 (12%) 5 = 103 (34%)
6	17537.993	17664.148	0.7228	0.841	1 = 95 (31%) 2 = 91 (30%) 3 = 32 (11%) 4 = 23 (8%) 5 = 24 (8%) 6 = 37 (12%)
6R	17554.166	17672.899	0.7131	0.824	1 = 68 (23%) 2 = 90 (30%) 3 = 31 (10%) 4 = 27 (9%) 5 = 34 (11%) 6 = 52 (17%)

AIC = Akaike information criterion; BIC = Bayesian information criterion; VLMR p = p for Vuong-Lo-Mendell-Rubin test; -R = revised model; Bolded indicates chosen model.

Figure 1.

Mean marijuana craving questionnaire score (scale range = 12-84) in the 4-class solution by most likely latent class membership in CTN 0053 (N = 302)

Class 1 (low craving class): n = 123 (41%); Class 2 (moderate and decreasing craving class): n = 116 (38%); Class 3 (moderate and stable craving class): n = 33 (11%); Class 4 (high craving class): n = 30 (10%). Each circle represents a measurement (weeks 1-5, 9, 12 and 17).

The low craving class represented the largest percentage of the sample (41%). This class had the lowest mean baseline craving (M = 38.6). Craving then declined through three weeks after randomization before leveling off at week 4 (M = 16.0). The moderate and decreasing class was the next largest proportion of the sample (38%). At baseline, the MCQ-SF score was moderate (M = 55.2). This group then saw large declines in the first four weeks, with a mean craving score of 40.3 after one week of treatment, 35.0 after two weeks, and 29.7 after three weeks. The moderate and stable craving class (11% of the sample) began with a similar mean baseline craving to the moderate and decreasing class (M = 55.6). In contrast, however, this group saw much smaller decreases in craving, with a final mean craving score of approximately 40. Finally, the high craving class represented the smallest proportion of the sample (10%). This group began with very high baseline craving (M=72.5) and made only modest improvements throughout the study.

Differences in key demographic and baseline mental health variables by most likely latent class are presented in Table 2. Classes differed by ethnicity where the high craving class included a larger proportion of people identifying as Hispanic (43%) than the low craving class (22%) and the moderate and decreasing class (22%). Other demographic variables, including age, sex, race, education, and employment did not differ in distribution by class membership. Several baseline mental health-related variables varied by most likely class membership. Baseline anxiety score differed by class, such that the low craving class had a lower mean anxiety score (M = 5.0, SD = 3.2) than the moderate and decreasing class (M = 7.3, SD = 3.7), the moderate and stable class (M = 7.4, SD = 4.8), and the high craving class (M = 7.4, SD = 4.8). Similarly, baseline depression differed by class, such that the low craving class had a lower mean depression score (M = 2.8, SD = 2.8) than the moderate and decreasing class (M = 4.5, SD = 3.2), the moderate and stable class (M = 5.1, SD = 3.3), and the high craving class (M = 5.5, SD = 3.9). Finally, there were differences in the OCDUS scale, where those in the low craving class had a lower OCDUS score (M = 17.6, SD = 8.4) than those in the moderate and decreasing class (M = 23.6, SD = 8.0), those in the moderate and stable class (M = 23.0, SD = 7.9), and those in the high craving class (M = 28.4, SD = 9.9).

Table 2.

Differences in key demographic and mental health variables by most likely class membership in 4-class solution for LCGA of marijuana craving score in CTN 0053 (N = 302)

	Overall (N = 302)	1: low craving (n = 123, 41%)	2: moderate, decreasing craving (n = 116, 38%)	3: moderate, stable craving (n = 33, 11%)	4: high craving (n = 30, 10%)
Treatment group
Placebo	149 (49%) [43.7%, 54.9%]	61 (50%) [40.9%, 58.3%]	62 (53%) [44.4%, 62.3%]	14 (42%) [27.2%, 59.2%]	12 (40%) [24.6%, 57.7%]
Treatment	153 (51%) [45.1%, 56.3%]	62 (50%) [41.7%, 59.1%]	54 (47%) [37.7%, 55.6%]	19 (58%) [40.8%, 72.8%]	18 (60%) [42.3%, 75.4%]
Age at randomization
Mean (SD)	30.3 (9.0)	29.8 (9.0)	30.5 (8.9)	31.3 (9.6)	30.8 (9.2)
95% CI	(29.3, 31.3)	(28.2, 31.4)	(28.9, 32.1)	(28.0, 34.5)	(27.5, 34.0)
Sex
Male	216 (72%) [66.2%, 76.3%]	89 (72%) [63.9%, 79.5%]	75 (65%) [55.6%, 72.8%]	29 (88%) [72.7%, 95.2%]	23 (77%) [59.1%, 88.2%]
Female	86 (28%) [23.7%, 33.8%]	34 (28%) [20.5%, 36.1%]	41 (35%) [27.2%, 44.4%]	4 (12%) [4.8%, 27.3%]	7 (23%) [11.8%, 40.9%]
Ethnicity *
Hispanic	65 (22%) [17.3%, 26.5%]	19 (15%) [10.1%, 22.9%]	26 (22%) [15.8%, 30.8%]	7 (21%) [10.7%, 37.8%]	13 (43%) [27.4%, 60.8%]
Non-Hispanic	237 (78%) [73.5%, 82.7%]	104 (85%) [77.1%, 89.9%]	90 (78%) [69.2%, 84.2%]	26 (79%) [62.2%, 89.3%]	17 (57%) [39.2%, 72.6%]
Race
White	176 (58%) [52.6%, 63.7%]	69 (56%) [47.3%, 64.6%]	68 (59%) [49.5%, 67.2%]	22 (67%) [49.6%, 80.2%]	17 (57%) [39.2%, 72.6%]
Other	126 (42%) [36.3%, 47.4%]	54 (44%) [35.4%, 52.7%]	48 (41%) [32.8%, 50.5%]	11 (33%) [19.8%, 50.4%]	13 (43%) [27.4%, 60.8%]
Education
High school or less	117 (39%) [33.4%, 44.3%]	43 (35%) [27.1%, 43.7%]	45 (39%) [30.4%, 47.9%]	14 (42%) [27.2%, 59.2%]	15 (50%) [33.2%, 66.8%]
Some college or more	185 (61%) [55.7%, 66.6%]	80 (65%) [56.3%, 72.9%]	71 (61%) [52.1%, 69.6%]	19 (58%) [40.8%, 72.8%]	15 (50%) [33.2%, 66.8%]
Employment
Working now	155 (51%) [45.7%, 56.9%]	64 (52%) [43.3%, 60.7%]	60 (52%) [42.7%, 60.6%]	15 (45%) [29.8%, 62.0%]	16 (53%) [36.1%, 69.8%]
Other	147 (49%) [43.1%, 54.3%]	59 (48%) [39.3%, 56.7%]	56 (48%) [39.4%, 57.3%]	18 (55%) [38.0%, 70.2%]	14 (47%) [30.2%, 63.9%]
HADS-Anxiety ***
Mean (SD)	6.4 (3.9)	5.0 (3.2)	7.3 (3.7)	7.4 (4.8)	7.4 (4.8)
95% CI	(5.9, 6.8)	(4.4, 5.5)	(6.7, 8.0)	(5.8, 9.1)	(5.7, 9.1)
HADS-Depression ***
Mean (SD)	4.0 (3.3)	2.8 (2.8)	4.5 (3.2)	5.1 (3.3)	5.5 (3.9)
95% CI	(3.6, 4.3)	(2.3, 3.3)	(3.9, 5.1)	(4.0, 6.2)	(4.1, 6.9)
Age of first THC use
Mean (SD)	15.2 (3.4)	15.2 (3.1)	15.4 (3.8)	15.5 (3.7)	14.2 (2.5)
95% CI	(14.8, 15.6)	(14.6, 15.7)	(14.7, 16.1)	(14.2, 16.8)	(13.3, 15.1)
OCDUS Score ***
Mean (SD)	21.6 (9.1)	17.6 (8.4)	23.6 (8.0)	23.0 (7.9)	28.4 (9.9)
95% CI	(20.6, 22.6)	(16.1, 19.1)	(22.1, 25.0)	(20.3, 25.7)	(24.9, 31.9)

^*p < 0.05;

^***p < 0.001;

Categorical variables are presented as n (%) [95% CI]. Continuous variables are presented as mean (SD) with 95% CI; HAD = hospital anxiety and depression scale; OCDUS = obsessive-compulsive drug use-marijuana scale; THC = Δ⁹-tetrahydrocannabinol.

Results of BCH testing examining urine cannabinoid testing by most likely class membership are presented in Supplemental Table 1. Those in the low craving class had the lowest number of positive screens (M = 12.41, SE = 0.89) and proportion of positive tests (M = 0.71, SE = 0.04). Notably, those in the high craving class had both the highest proportion of positive tests (M = 0.96, SE = 0.04) and the lowest number of total tests (M = 14.58, SE = 1.94), indicating that these individuals missed the urine tests during that week. Proportion of positive tests and number of tests collected were similar for number of positive and total tests, and proportion of positive tests, between the moderate/decreasing and moderate/stable craving groups.

4. Discussion

In this secondary analysis of a multisite randomized controlled CUD treatment trial, we identified four distinct severity-based craving trajectories, primarily distinguished by baseline craving levels and persistence, over the course of a 12-week NAC versus placebo in combination with contingency management intervention, for both groups, and subsequent follow-up. These trajectories—low, moderate-decreasing, moderate-stable, and high craving— reflect differences in participants’ subjective experiences of craving during treatment, despite a shared clinical protocol. Importantly, individuals in the higher craving classes exhibited greater baseline symptoms of anxiety, depression, and OCDUS, underscoring the relevance of co-occurring mental health symptoms in shaping treatment experiences. Although the differences in baseline anxiety and depression scores across craving classes were statistically significant, all depression scores fell within the non-clinical range on the HADS, indicating that observed between-class differences reflect statistical variation rather than clinically meaningful depressive symptomatology.

In contrast, anxiety scores ranged from normal in the low craving class to borderline or clinically significant levels in the higher craving classes, indicating that anxiety may be a more meaningful differentiator between groups. The OCDUS scores are interpreted continuously, with higher scores reflecting greater obsessive-compulsive features of craving. Lastly, individuals in the high craving class exhibited the highest proportion of positive cannabis urinalysis results, suggesting that this may be a prime patient group for intensified interventions.

These findings suggest that craving is not only an important clinical target, but also a potential marker for identifying individuals at elevated risk for poor treatment response and return to use. Prior studies have established the associations between craving, mood symptoms, and treatment outcomes, but few have examined individual craving trajectories during treatment. Our use of LCGA provides a way to summarize changes in craving over time, complementing information from baseline measures and offering insight into variability across participants, although these observed trajectories should be interpreted cautiously, as they may not reflect definitive subpopulations (Mésidor et al., 2022).

Our findings concur with previous studies showing that individuals with low distress tolerance who experience more intense craving tends to exhibit worse treatment outcomes, as they are more likely to return to cannabis use (Buckner et al., 2019). Similarly, cannabis use has consistently been associated with greater severity of depressive symptoms (Lev-Ran et al., 2014; Manrique-Garcia et al., 2012), and some individuals report using cannabis to manage negative affect and withdrawal symptoms (Kosiba et al., 2019). Because cannabis use has been linked to more severe depressive symptoms, higher craving levels—often a sign of more severe use—could also be related to worse depression, although this hypothesis needs to be tested in future research. When cannabis is used as a maladaptive coping mechanism to manage these symptoms, this perceived relief may reinforce continued use and contribute to a cycle of problematic use (Hyman and Sinha, 2008). These patterns may help explain why participants in the high-craving trajectory also showed elevated baseline symptoms of depression and anxiety, as well as higher rates of continued cannabis use. Comparable studies in other SUDs (e.g., alcohol, opioids), have also found persistent high craving trajectories associated with worse outcomes (Bowen et al., 2009; Burgess-Hull et al., 2022; Witkiewitz et al., 2012), yet longitudinal investigations of craving in cannabis use remain limited. This study shows that craving patterns in CUD resemble those observed in other SUDs, reinforcing the idea that craving is a core feature across substance use disorders and highlighting the need for targeted strategies to manage craving (Sayette, 2016).

Although our study analyzed baseline craving, the observed differences across classes suggest that distinguishing and monitoring craving trajectories throughout treatment, rather than relying solely on baseline assessments, may not only enhance intervention targeting but also help identify individuals at greater risk for poor treatment response and return to use. Notably, the moderate-decreasing and moderate-stable trajectories had similar baseline craving levels, yet they differed over time, with the moderate-decreasing group showing reductions in craving while the moderate-stable group remained largely unchanged. Anxiety levels were higher in both groups compared to the low-craving trajectory, emphasizing the clinical importance of monitoring craving dynamically throughout treatment. These findings underscore that craving trajectories offer additional clinically relevant information beyond baseline scores. Individuals with persistently high craving frequently present with elevated symptoms of depression and anxiety, suggesting that co-occurring mental health conditions play a significant role in treatment outcomes (Hyman and Sinha, 2008; Lev-Ran et al., 2014). Therefore, integrated interventions that simultaneously address craving and co-occurring psychological symptoms may offer enhanced benefits. Moreover, interventions focused on craving regulation have been shown to be effective in reducing both craving and cannabis use, improving treatment outcomes (Cousijn and van Duijvenvoorde, 2018).

Continuous craving assessment throughout treatment may allow clinicians to dynamically adjust treatment based on individuals needs (Frohe et al., 2025; Panlilio et al., 2019). Individuals with persistent craving could benefit from adjunctive therapies such as mindfulness-based prevention of return of use, cognitive-behavioral techniques focused on craving management (Bowen et al., 2009; Witkiewitz and Bowen, 2010), or pharmacological treatments when available. This monitoring of craving could improve treatment personalization and subsequently, improve outcomes.

Several limitations should be considered when interpreting these findings. First, the use of self-reported measures of craving and psychological distress may be subject to reporting biases, such as social desirability or inaccurate recall. Yet, all measures of craving are necessarily subjective. Additionally, the sample may not represent all people with CUD, particularly those not seeking treatment, which may limit generalizability (Moberg and Humphreys, 2017). Furthermore, participants received a contingency management intervention, so findings may not generalize to treatment settings without such intervention. Given the secondary and observational nature of this analysis, finding should be interpreted as associations rather than evidence of causal relationships.

Also, this method of analysis is largely empirically-driven, necessitating follow-up studies to confirm class membership meaning and interpretation. Finally, other relevant factors such as trauma history, personality traits, or the use of cannabis as a ineffective coping mechanism for negative affect were not examined in this study but may play an important role in craving patterns and should be considered in future research. Future studies should evaluate the influence of tobacco and other substance use on craving trajectories and treatment outcomes. Future research should explore these factors and replicate these findings in larger, more diverse samples, including participants with a broader range of demographics and clinical profiles. In particular, samples could include individuals with co-occurring CUD and psychiatric conditions to examine whether craving trajectories and treatment responses differ in these populations. Pooling data across multiple clinical trials may facilitate such efforts.

This study identified distinct craving trajectories among individuals receiving treatment for CUD, highlighting meaningful heterogeneity in subjective craving experiences over time. Individuals with higher craving levels at baseline exhibited more severe baseline symptoms of depression, anxiety, and OCDUS, and were more likely to show poorer treatment outcomes, as evidenced by a higher proportion of positive cannabis urinalysis results. These findings underscore the clinical importance of routinely assessing craving, not only as a treatment target, but also as a potential marker for identifying individuals at greater risk of poor response and return of use. Tailored interventions that address both craving and co-occurring mental health symptoms may be especially beneficial for those with persistent high craving (Buckner et al., 2013; Wallis et al., 2022).

By identifying distinct craving trajectories, this study supports efforts toward more personalized and risk-informed approaches in the treatment of CUD, where dynamic monitoring of craving—easily collected via phone or ecological momentary assessment (EMA)—could help identify individuals at risk of return of use and enable early intervention.

Highlights.

• Four distinct cannabis craving trajectories were identified during treatment: low, moderate-decreasing, moderate-stable, and high craving.

• Higher craving trajectories were associated with greater baseline anxiety, depression, and obsessive-compulsive symptoms, underscoring the role of co-occurring mental health in treatment response.

• Participants in the high craving class had the poorest treatment outcomes, with the highest rates of cannabis-positive urine tests and lowest test completion.

• Latent class growth analysis revealed craving heterogeneity, emphasizing the need for dynamic craving monitoring to inform personalized interventions.

• Findings support craving as a clinical marker for treatment tailoring, with potential to guide targeted interventions and improve outcomes for individuals with cannabis use disorder.