Examining dynamic patterns of problematic cannabis use: Results from a multilevel network analysis

Abstract

Young adults who engage in problematic cannabis use report lower work and interpersonal functioning yet are less likely to seek treatment, necessitating alternative methods for assessing and intervening on problematic cannabis use (e.g., mobile health applications to self-monitor drivers of cannabis use in daily life). However, previous work examining maintenance models of problematic cannabis use has primarily focused on modeling predictors of cannabis use as measured using static retrospective report, rather than examining cannabis use as a series of interactions that unfold in everyday life. In this study, we analyzed ecological momentary assessment data (T = 3,230 observations) from 65 young adults who reported problematic cannabis use (CUDIT-R: M = 10.38, SD=4.35) and an interest in reducing their use. We used multilevel network analyses to model associations among biopsychosocial factors that aligned with social learning, self-medication, and experiential avoidance theories of substance use. Network models demonstrated consistent associations between socio-environmental triggers and cannabis cravings, use, and intoxication that were nearly all clinically meaningful in size (bs > .10). Results indicated a statistically significant association between negative and positive affect with cannabis use and intoxication, respectively; however, these associations were not clinically meaningful in size. There were no clinically meaningful associations between coping strategies and cannabis use variables. Findings advance our understanding of cannabis use in everyday life, which is critical for refining more dynamic conceptualization of substance use and improving the precision of clinical assessments.

General Scientific Summary

This study aimed to model, visualize, and interpret the dynamic patterns between biopsychosocial factors associated with problematic cannabis use using a network analytic approach. Findings advance our theoretical understanding of cannabis use processes in everyday life and help to improve the precision of clinical assessment.

Cannabis is one of the most used psychotropic drugs in the United States (Richesson & Hoenig, 2021) with over 50% of young adults reporting lifetime use (Han et al., 2019). Over the past decade, cannabis use has increased nearly 20% nationwide as more states legalize recreational cannabis (Zellers et al., 2022); this increase contributes to the prevalence of problematic cannabis use (i.e., frequent or heavy use that is associated with significant functional problems). The prevalence of problematic cannabis use is highest among young adults and problematic cannabis use among adolescents and young adults is associated with deleterious effects on brain development, cognitive functioning, school performance, educational achievement, and mental disorders (Arria et al., 2015; De Aquino et al., 2018; Han et al., 2019), which is particularly notable given that young adults are less likely to seek treatment for cannabis-related problems (Amiet et al., 2020; Hall et al., 2019). Furthermore, effects on functioning are evident even among individuals who do not meet full diagnostic criteria for cannabis use disorder (Foster et al., 2017; Hasin, 2018). Thus, studying the naturalistic and dynamic factors that maintain problematic cannabis use for young adults in everyday life is critical to improve clinical assessment and to advance the design of more tailored treatment and prevention efforts for young adults.

Maintenance models of problematic cannabis use

Researchers have long sought to understand the complex constellation of biopsychosocial factors that maintain substance use and increase risk for subsequent problems for some (but not all) individuals. There are three core theoretical models that have risen to prominence within the literature: 1) social learning models (Smith, 2021), 2) affect-regulation models (Cooper et al., 2015) or self-medication models (Baker et al., 2004) and 3) models of experiential avoidance or paradoxical effects (Hayes et al., 1996; Moss et al., 2015; Wegner & Zanakos, 1994). Social learning theories of addictive behavior posit that problematic substance use is maintained by reinforcing relationships between individual-level factors (e.g., personality, family history of addiction), substance-related variables (e.g., intensity of craving, frequency of use, intensity of intoxication/withdrawal), and social-environmental variables (e.g., availability of substances, peer-based influences) (Smith, 2021; Strickland & Acuff, 2023). For example, some research has demonstrated that young adults are more likely to use cannabis in social settings, particularly when around others using cannabis (Jackson et al., 2021).

Motivational or affect-regulation models hypothesize that substances are used to enhance positive affect (Cooper et al., 2015) and that substances alleviate negative affect or psychological suffering (also referenced in self-medication models; Khantzian, 2003). For example, research has demonstrated that positive and negative affect motivates later use (Buckner et al., 2015; Shrier et al., 2014; Tyler et al., 2015). Accordingly, substance use may be negatively reinforced by relief from negative affect and positively reinforced by enhancement of positive affect (Baker et al., 2004). As such, research has demonstrated that cannabis use increases positive affect (Henquet et al., 2010) and decreases negative affect (Buckner et al., 2015; Trull et al., 2016). Over time, these reinforcing processes increase tolerance and negative substance-related consequences.

Finally, models describing experiential avoidance or paradoxical effects have been largely studied within the clinical and psychotherapy research fields (Hayes et al., 1996). For example, clinical researchers have suggested that efforts to avoid or escape challenging internal states (e.g., anxiety) via thought suppression (Wegner & Zanakos, 1994) and experiential avoidance (Hayes et al., 1996) may contribute to persistence or even escalation of craving, maintaining problematic substance use (Britton, 2004; Moss et al., 2015). Experiential avoidance can refer to different coping or behavioral strategies, but largely refers to disengagement or purposefully distract from one’s experience (Kashdan, 2006). Alternatively, strategies to decrease experiential avoidance, such as acceptance or engaging in valued action are featured prominently in substance use interventions to reduce the distress associated with increased cravings (e.g., mindfulness-based relapse prevention or acceptance and commitment therapy; (Thekiso et al., 2015; Witkiewitz et al., 2013).

Taking theoretical models into everyday life

Existing models of substance use provide minimal insight as to how biopsychosocial factors might interact in everyday life to maintain cannabis use, complicating the advancement of evidence-based clinical assessments for cannabis use. Data collection methods that enable more precise assessments of cannabis use on more granular timescales can help strengthen understanding of cannabis use processes in everyday life. Ecological momentary assessment (EMA; a type of naturalistic data collection method; Shiffman, 2008) allows researchers to measure symptoms, experiences, and behavior in real-time as an individual goes through their everyday life, often using brief self-report assessments sent to a personal smartphone. Accordingly, researchers using EMA can reduce the retrospective recall bias that is more likely to impact traditional methods of self-report.

Additionally, network analytic approaches allow researchers to conceptualize and model variables as a series of dynamic interactions that assist researchers in examining substantive relationships between biopsychosocial factors as maintained in everyday life (Borsboom et al., 2021). Previous research testing maintenance models of substance use typically examine a subset of cannabis use predictors in isolation, which is misaligned with models of psychological experience and behavior that describe factors as interacting components of a system (e.g., interactions between thoughts, feelings, and behaviors as outlined in cognitive behavioral theories; Beck, 1979). Within a network framework, nodes (which represent factors or constructs) are directly associated with each other through a series of pathways that maintain the system of interest (edges; Borsboom et al., 2021; Bringmann, 2021). Thus, a network approach provides a useful framework for conceptualizing linkages between biopsychosocial factors of cannabis use and may yield inferences more closely aligned with clinical conceptualizations of cannabis use and related behavior.

Furthermore, network frameworks are aligned with a set of statistical tools that can be used to model associations between biopsychosocial factors and cannabis-related variables. A statistical network using EMA data can be constructed to examine how nodes relate with each other contemporaneously (i.e., at the same time), as well as temporally (i.e., from one time point to the next) (Borsboom et al., 2021; Borsboom & Cramer, 2013; Bringmann et al., 2013). Furthermore, examining the degree to which specific nodes are connected to other nodes in the networks (i.e., expected influence; Robinaugh et al., 2016) allows researchers to examine which cannabis-related constructs exert the most influence on other nodes within the network. Thus, the use of EMA data, a network framework, and analytic methods for modeling statistical networks offer a way for researchers to advance understanding of cannabis use in everyday life and, ultimately, improve the precision of clinical assessment.

The current study

The primary aim of this study was to model the dynamic associations among biopsychosocial factors theorized to maintain problematic cannabis use. We recruited a young adult sample reporting problematic cannabis use who were interested in reducing their cannabis use and administered a series of brief, self-report surveys assessing cannabis use, cravings, and intoxication, as well as affect, social-environmental factors, and coping strategies. Multilevel network analyses were used to model associations between factors, as reported on in everyday life. We used a multilevel network analytic approach to explore the contemporaneous and temporal associations between cannabis use and related theoretical factors within a network framework (Borsboom & Cramer, 2013; Bringmann et al., 2013; Epskamp et al., 2024). We aimed to model, visualize, and interpret the statistical associations between these variables to facilitate a critical evaluation of cannabis use processes in everyday life.

Method

Participants

This study enrolled young adults between the ages of 18 to 30 from a large public university in the Northwestern United States who reported problematic cannabis use and who were interested in reducing their use (N = 86, n = 65 analyzed). Eligible participants reported using cannabis on average at least two days per week during the previous month, screened positive for problematic cannabis use using the Cannabis Use Disorder Identification Test – Revised (Adamson et al., 2010; Schultz et al., 2019), and screened positive for either contemplation or action stages of change related to cannabis use according to the Readiness to Change Questionnaire (Rollnick et al., 1992).

Measures

All items were assessed using an EMA design consisting of brief, self-report assessments completed four times per day for at least 14 days. Descriptive statistics and reliability for EMA items can be found in Table 1.

Table 1.

Within-person descriptive statistics for EMA items

Construct	Intensity M (SD), range	Variance M (SD), range	Range M (SD), range	Multilevel reliability
Cannabis use	0.39 (0.43), 0.00–2.65	0.61 (0.31), 0.00 – 1.54	1.17 (1.42), 0.00 – 4.00	ICC2 = 0.27 ICC1 = 0.73
Craving	1.80 (1.26), 0.00–5.75	1.92 (0.75), 0.00–3.85	3.46 (3.74), 0.00–10.00	ICC2 = 0.26 ICC1 = 0.74
Intoxication	1.34 (1.12), 0.00–5.18	2.05 (0.79), 0.00–3.90	3.62 (3.87), 0.00–10.00	ICC2 = 0.19 ICC1 = 0.81
Socio-environmental triggers	0.84 (0.35), 0.21 – 1.64	0.66 (0.16), 0.23 – 0.99	1.01 (0.99), 0.00 – 2.00	ICC2 = 0.20 ICC1 = 0.80
Positive affect	1.46 (1.02), 0.09–4.47	1.04 (0.43), 0.18–2.32	2.31 (2.57), 0.00–9.40	${\overline{\mathrm{\omega }}}_{\text{within}}=.6$
Negative affect	3.82 (1.12), 0.95–6.13	1.69 (0.43), 0.48–3.15	3.96 (3.61), 0.00–10.00	${\overline{\mathrm{\omega }}}_{\text{within}}=.83$
Coping: Accept	4.58 (1.88), 0.00–9.45	1.96 (0.80), 0.00–3.77	4.46 (3.91), 0.00–10.00	ICC2 = 0.44 ICC1 = 0.56
Coping: ValAct	5.19 (1.93), 0.00–10.00	1.89 (0.86), 0.00–3.88	4.92 (4.16), 0.00–10.00	ICC2 = 0.47 ICC1 = 0.53
Coping: Giveup	1.95 (1.50), 0.00–7.55	1.67 (0.75), 0.00–3.24	3.42 (3.71), 0.00–10.00	ICC2 = 0.40 ICC1 = 0.60
Coping: Distract	3.52 (1.80), 0.17–7.15	2.16 (0.72), 0.69–3.89	4.13 (4.01), 0.00–10.00	ICC2 = 0.38 ICC1 = 0.62

Note. These values represent the descriptive statistics that are pooled (i.e., averaged) for each person during the EMA period. Coping: Giveup = Coping with challenging experiences (since the last prompt) by disengaging; Coping: Distract = Coping with challenging experiences (since the last prompt) by engaging in distraction; Coping: Accept = Coping with challenging experiences (since the last prompt) using acceptance; Coping: ValAct = Coping with challenging experiences (since the last prompt) by engaging in valued actions; ICC1 = Intraclass correlation, within-person reliability; ICC2 = Intraclass correlation, between-person reliability.

Cannabis use, craving, and intoxication.

Cannabis use was assessed using a single item, “Since the last prompt, approximately how many times have you used marijuana?” with a response scale ranging from 0 (none) to 4 (four or more). Participants were instructed to consider a period of use as an “episode’” of use (i.e., multiple hits from the same source within a continuous episode were not intended to reflect separate times of use). Cannabis intoxication was assessed using the item, “Since the last prompt, how high did you feel when you felt the most high?” with a response scale ranging from 0 (not at all) to 10 (most ever). Craving was assessed using one item adapted from the Penn Alcohol Craving Scale, a widely used measure of craving for alcohol and other drugs, with well-established reliability and validity (Flannery et al., 1999). This item was phrased, “Since the last prompt, when you wanted to use marijuana the most, how strongly did you want to use?” with a response scale ranging from 0 (not at all) to 10 (most ever).

Social learning factors.

Social and environmental triggers were assessed using two items. The first item assessed the presence of peers using cannabis in the immediate environment, “Were you around other people who were using marijuana?” (response options, 0 = no, 1 = yes) and the second item assessed the availability of cannabis in the immediate environment, “Has marijuana been available for you to use?” (response options, 0 = not available, 1 = available). A composite variable was created to reflect the degree to which socio-environmental triggers were present at each prompt. If participants responded to both variables using ‘0’, then ‘triggers’ was coded as ‘0’, if participants responded to either variable with ‘1’, then ‘triggers’ was coded as 1, and if both participants responded to both variables with ‘1’, then ‘triggers’ was coded as 2. The association between these two variables was moderate and statistically significant, r_multilevel = .41, p < .05.

Affect regulation factors.

Factors reflecting positive and negative affect were assessed using items from the Positive and Negative Affect Schedule (Watson & Clark, 1988). The prompt stem for each affect item was, “Since the last prompt, how ____ have you felt?” Negative affect items included: afraid, bored, guilty, irritable, nervous. Positive affect items included: active, attentive, determined, interested, proud. Each item used a response scale that ranged from 0 (not at all) to 10 (most ever). Affect items were combined to create average, composite variables of positive and negative affect, respectively.

Experiential avoidance or paradoxical effect factors.

Items reflecting one of four coping strategies implicated in theories related to paradoxical effects of thought suppression or experiential avoidance were assessed (Kashdan et al., 2006; Moss et al., 2015). Specifically, the prompt stem: “Since the last prompt, when having challenging or negative experiences, to what extent have you been…” was used to measure three items adapted from the Brief COPE (Carver, 1997) assessing behavioral disengagement (…”giving up trying to deal with [challenging or negative experiences]?”), distraction (…”trying to take your mind off [challenging or negative experiences]?”), and acceptance (…”accepting and learning to live with [challenging or negative experiences]?”), as well as one item assessing committed action adapted from the Acceptance and Action Questionnaire (AAQ-2; Bond et al., 2011). All coping items used a response scale that ranged from 0 (not at all) to 10 (most ever).

Procedure

Participants completed a baseline assessment and brief training prior to completing EMA. Participants completed four surveys per day for approximately 14 days (additional surveys were administered to some participants who reported issues starting the study). Surveys were delivered to the participant’s personal smartphone at random times within three two-hour time windows during the day ranging from 10am – 12pm, 3pm – 5pm, and 8pm – 10pm. Before completing the 10am-12pm survey, participants reported on their experience from 10pm – 10am that morning (e.g., cannabis use that occurred after 10pm). Participants were given two hours to complete each survey (Enkema et al., 2021). Implications related to design-related considerations are discussed further in the Discussion.

Data analytic plan

We constructed a multilevel vector autoregressive model (ML-VAR) and interpreted the presence, valence, and magnitude of associations between cannabis use and theoretical variables at a given point in time, as well as associations from one timepoint to the next. We calculated and examined the expected influence statistic for each node in each network to determine the relative prominence of each node in the network (Robinaugh et al., 2016). Data were analyzed using R, version 4.2.2 (R Core Team, 2023) and the packages, tidyverse, version 1.3.1 (Wickham et al., 2019), misty, version 0.6.2 (Yanagida, 2024), mlVAR, version 0.5 (Epskamp et al., 2024), and qgraph, version 1.9 (Epskamp et al., 2012). We discuss assumptions of network analytic statistical approaches below and further in the Discussion.

Interval spacing and missing data.

An assumption of the ML-VAR statistical model is that variables are measured at approximately even intervals (Epskamp et al., 2018). Thus, to ensure that paths were not estimated across night periods spanning one day to the next, study day and time of prompt were specified in the ML-VAR model, which effectively reduced the maximum planned daily observations available for analysis by one. Additionally, multilevel network analyses have limited ability to handle missing data or time series of unequal length (Epskamp et al., 2018). As mentioned previously, participants were allowed to complete additional days of the study to maximize their number of available observations (i.e., in the event they had trouble starting the study). To standardize the length of each participant’s time series, we excluded observations collected after the modal study length of 21 days. Further, we subset data to include participants who, on average, completed multiple surveys per day during the EMA period to optimize statistical power when constructing the temporal network. Our analyzed sample included data from 65 participants who completed a total of 3,230 observations (M = 49.69, minimum = 32 observations, maximum = 62 observations per participant).¹

Multicollinearity and violations to stationarity and normality.

To reduce the potential for multicollinearity among network nodes we created composite variables for items that assessed similar constructs, including socioenvironmental triggers, positive affect, and negative affect as described above. To account for the potential effects of time (which may reflect a violation to stationarity; Bringmann et al., 2017), we used a detrending process to model the linear effects of study day and time of prompt (Jordan et al., 2020) and modeled the residualized data in our ML-VAR network. The residualized variables exhibited an approximately normal distribution.

Stability of multilevel networks.

To assess the stability of the mlVAR networks, we used a case-drop resampling procedure described in previous work (Jongeneel et al., 2020). We resampled and constructed 100 ML-VAR networks using randomly selected subsets of participants (i.e., 80% of the 65 analyzed participants) and calculated the expected influence statistic from the contemporaneous and temporal resampled networks. Correlations were calculated between the expected influence statistic from each resampled network and the expected influence statistic calculated from the original ML-VAR network, respectively. We examined the number of resampled networks (out of 100) that demonstrated correlations of r > .70 between the original – resampled expected influence statistics and concluded that networks were stable if greater than 50% of the correlations from the original – resampled networks demonstrated original-resampled correlations of r > .70.

Interpreting network models: Expected influence, statistical significance, effect sizes.

We calculated the expected influence for each node in the contemporaneous and temporal networks to characterize each node’s prominence in the model (i.e., the extent to which a given node impacts the rest of the network; Robinaugh et al., 2016). Two expected influence statistics were calculated from temporal networks: 1) the inwards expected influence, which reflects how much one node is impacted by other nodes in the network and 2) outwards expected influence, which reflects the extent to which one node impacts other nodes in the network (Robinaugh et al., 2016). ML-VAR statistical models use a thresholding approach to prune paths with sufficiently small estimates during the estimation process (see Epskamp et al., 2018). Thus, pathways estimated from each model are considered statistically significant at p < .05. However, given our exploratory aims and the relatively large number of estimated pathways in a ML-VAR network, we sought to describe and interpret fixed effects estimates with b ≥ 0.10 as we deemed these pathways most likely to be clinically meaningful in size.

Transparency and openness

The EMA items used in this study are described above. The study design and analyses were not pre-registered. Analytic code used to conduct these analyses is included in supplementary material and data are available by emailing the corresponding author.

Results

Descriptive statistics

Participants analyzed in this study (N = 65) were mostly young (M = 19.72, SD=2.41, range = 18–30 years) men (n = 39, 60.0%) who used cannabis on average 3.66 days (SD=1.72, range = 2 – 7 days) per week (assessed prior to beginning EMA). The mean CUDIT-R score was above the threshold for hazardous use (M = 10.38, SD=4.35, range = 5–26; Adamson et al., 2010; Schultz et al., 2019).

How are cannabis-related constructs associated at similar points in time?

The contemporaneous network demonstrating associations between cannabis-related constructs at a given point in time exhibited adequate stability (100% of resampled networks demonstrated r > .70 between the expected influence statistic calculated from the original versus resampled networks). Pathways from this network are displayed in Figure 1, left panel.

Figure 1. Contemporaneous and temporal associations between cannabis-related constructs.

Note. The left panel demonstrates the contemporaneous associations and the right panel demonstrates temporal associations. Solid edges represent significant, positive associations; dashed edges represent significant, negative associations. For the contemporaneous network (left panel), the proportion to which the ring around each node is shaded represents the amount of expected influence that a given node exerts on the network and * denotes the node with the largest expected influence. For the temporal networks (right panel), the proportion to which the ring around each node is shaded dark grey represents the extent to which a node exerts influence on other nodes in the network (i.e., outwards expected influence) and ** reflects the node with largest outwards expected influence. The proportion to which the ring around each node is shaded light grey represents the extent to which nodes are influenced by other nodes in the network (i.e., inwards expected influence) and *** reflects the node with the largest inwards expected influence. Variable labels include used cannabis (#) = Number of instances of cannabis use since the last prompt; Level intox cannabis = Level of intoxication since the last prompt; Soc-env triggers = Degree of exposure to socio-environmental triggers for cannabis use since the last prompt; Coping: Giveup = Coping with challenging experiences (since the last prompt) by disengaging; Coping: Distract = Coping with challenging experiences (since the last prompt) by engaging in distraction; Coping: Accept = Coping with challenging experiences (since the last prompt) using acceptance; Coping: ValAct = Coping with challenging experiences (since the last prompt) by engaging in valued actions.

The number of times a participant reported using cannabis since the last prompt demonstrated the largest expected influence, suggesting strong associations with other cannabis-related nodes in the network. Associations between cannabis use and level of intoxication (b = 0.70), socio-environmental triggers and cannabis use (b = 0.29), cannabis cravings and level of intoxication (b = 0.24), and socio-environmental triggers and intoxication (b = 0.10) were all clinically meaningful in size. Although pathways between affect and cannabis use and intoxication were present and statistically significant in the network, only the association between cannabis cravings and positive affect (b = 0.15) was clinically meaningful in size.

The association between coping by engaging in valued action and cannabis intoxication was present and statistically significant in the model, although the magnitude of this association was not clinically meaningful in size. However, associations between negative affect and coping via acceptance (b = 0.11) or valued action (b = .17) were clinically meaningful in size, as was the association between positive affect and coping via behavioral disengagement (“giving up”; b = 0.12), There was clear individual-level variation in the magnitude and valence of network edges. Taken together, results demonstrate associations between cannabis craving, use, intoxication, and socio-environmental triggers that were large in magnitude, few associations between affect and cannabis-related variables that were relatively small in magnitude, and virtually no statistically significant associations between coping strategies and cannabis-related variables.

How are cannabis-related variables associated across time?

The temporal network demonstrating lagged associations between cannabis-related constructs from one time point to the next. The temporal network demonstrated lower stability (57% of resampled networks demonstrated r > .70 between the outwards expected influence, 31% of resampled networks demonstrated r > .70 between the inwards expected influence), which reduces the strength of inferences made using the findings from this network. Pathways are displayed in Figure 1, right panel.

Positive affect demonstrated the largest outwards expected influence, suggesting that positive affect was most impactful on other nodes in the network; whereas, coping with challenges by giving up demonstrated the largest inwards expected influence, suggesting that other nodes exerted a stronger influence on one’s reports of behavioral disengagement.

Most variables, except for cannabis use, level of intoxication, and degree of exposure to socio-environmental triggers, demonstrated auto-lagged associations, suggesting some association between levels of a given variable from one timepoint to the next. Of the temporal associations that were present and statistically significant, none were clinically meaningful in size (bs > 0.10). Taken together, results illustrated some statistically significant temporal pathways that were small in magnitude.

Discussion

We used a network analytic approach to model, visualize, and interpret the dynamic associations between cannabis-related variables and theoretically relevant variables of affect, socioenvironmental triggers, and coping strategies. Results from contemporaneous networks demonstrated large associations between cannabis craving, use, intoxication, and socio-environmental triggers, some small associations between affect and cannabis-related variables, and only one association between coping strategies and cannabis-related variables (that was not clinically meaningful in size). Additionally, the stability of temporal associations was limited and the associations demonstrated in the temporal networks were not clinically meaningful in size.

Associations between cannabis use, cravings, intoxication, and socio-environmental triggers were frequently observed in the contemporaneous network. Furthermore, the magnitude of these pathways were nearly all clinically meaningful in size and often the largest in comparison to other associations in the network. Findings align with research demonstrating socio-environmental factors as predictors of cannabis cravings, use, and intoxication and suggest these associations are likely concurrent in nature (or interact at timescales shorter than every few hours as measured here). Accordingly, interventions to reduce craving (e.g., mindfulness; (Enkema et al., 2021; Enkema & Bowen, 2017; Witkiewitz et al., 2013), as well as social- and environmental-specific interventions, such as practicing refusal skills, engaging in non-substance focused activities with peers, encouraging social interactions with non-substance-using peers, or restructuring one’s plans to avoid situations where substances are more likely to be available (Coughlin et al., 2023) are most likely to impact one’s cannabis use (and subsequent) intoxication as experienced in everyday life.

Additionally, we observed a large, positive, concurrent association between cannabis cravings and positive affect, suggesting that cravings for cannabis were associated with experiences of heightened positive affect (perhaps reflecting an anticipatory outcome of use). Although there were statistically significant associations between cannabis use and negative affect, as well as cannabis intoxication and positive affect, we considered these associations too small in magnitude to be clinically meaningful. These findings suggest that increased cravings are contemporaneously associated with enhanced positive affect (as measured every few hours); however, other associations with affect may be too small to be clinically meaningful or operate on substantially different timescales than those measured here. Taken together, these findings are largely in keeping with the mixed or null findings from other EMA studies examining affect-regulation models of cannabis use (e.g., Wycoff et al., 2018).

Finally, there were virtually no associations between coping strategies and cannabis use variables, as might be expected by experiential avoidance or paradoxical effects models (Moss et al., 2015). Notably, participants recruited in this sample screened positive for problematic cannabis use but was not a treatment-seeking sample and thus, demonstrated greater variability in motivations to use (or practice) coping strategies to reduce cannabis use. However, some associations between affect and coping strategies were statistically significant or clinically meaningful in size, suggesting that further research is needed with larger sample sizes to characterize the temporal associations between affect, coping strategies, and subsequent edges with cannabis use variables with greater stability.

Our work helps to characterize the fluctuations of cannabis cravings, use, and intoxication in everyday life and linkages to variables highlighted in existing theories, including socio-environmental triggers, affect, and coping strategies. Results demonstrated some concordance with pathways implicated in existing substance use theories (e.g., large, contemporaneous associations between proximity to socioenvironmental triggers and cannabis use). However, pathways between affect and cannabis use, as well as pathways between coping strategies and use were sparse. We believe this pattern of findings (specifically, the lack of alignment with existing theory) reflects a relatively common outcome that stems from attempts to use existing theoretical models to design research studies aimed to test hypotheses around substance use processes in everyday life. ² Indeed, the lack of timescale specificity around how biopsychosocial factors might fluctuate or be associated with cannabis use (and related variables) makes it significantly challenging to generate and test hypotheses, ultimately limiting the pace at which we can advance theoretical knowledge of substance use (see Ryan et al., 2023; Woodward, 2011 for greater disucssion of issues with testing psychological theories).

Given challenges with using existing theoretical models to guide examinations of cannabis use in everyday life, what are steps that researchers can take to increase the precision of clinical assessments? Improving measurement of substance use and related constructs, particularly when assessed using EMA, help improve the clarity in concepts outlined in existing theoretical models (in line with calls from methods-focused researchers, see Bringmann et al., 2022). For example, in this study, we sampled all constructs every few hours; however, some constructs like craving or intoxication likely change at different frequencies or exhibit different associations as a function of timescale (Hopwood et al., 2022; Wright & Hopwood, 2016 for a critical discussion of timescale). Thus, sampling some constructs more frequently (e.g., every 10 minutes) may have allowed us to model some associations with greater precision³. Going forward, hybrid EMA designs that pair event-contingent prompts, alongside fixed interval prompts, may be useful to improve assessment of cannabis related variables at timescales with sufficient frequency and intensity (Pearson et al., 2022). Additionally, meta-science efforts to characterize data collected using different experience sampling protocols strengthens our understanding of variables in everyday life (Votaw & Witkiewitz, 2021; Wycoff et al., 2018). Improved measurement also ensures that advanced statistical methods will be useful in modeling the dynamics that drive and maintain complex behaviors, such as substance use (e.g., see Cui et al., 2023; Driver et al., 2017 for examples of newly developed methods for analyzing complex time-series).

Regarding study-specific limitations, first, the relatively longer reporting period for EMA prompts (e.g., multiple hours versus minutes or a single hour) increased the risk of retrospective reporting bias and may not align with the timescale at which variables occur and fluctuate (see Haslbeck & Ryan, 2022 for disucssion of insufficient sampling frequency). Additionally, we are unable to account for the potential influence of cannabis in our contemporaneous networks (or conduct a rigorous test of affect regulation theory) as this study did not assess variables with close temporal proximity prior to and following cannabis use. These considerations limit the accuracy of inferences from statistical networks (see discussion of model misspecification; Haslbeck & Ryan, 2022).

Second, the use of single items limited our ability to assess item reliability (see Dejonckheere et al., 2022 for two design-related solutions for assessing reliablity of single items). Third, although the average number of within-person observations was consistent with simulation work (Epskamp et al., 2018), our temporal network demonstrated lower stability and missing data were removed via listwise deletion, which lowered the stability of the temporal network further. Finally, the assumptions of network analytic methods used here do not adequately accommodate the complexity of substance use data particularly as measured using (intensive) longitudinal designs (e.g., count structure, non-Gaussian distribution; Atkins et al., 2013). Ultimately, continuous time models or other time-varying methods for modeling dynamic processes around a change point (e.g., an instance of cannabis use) may provide useful alternatives for modeling relationships between maintenance factors of substance use (Driver et al., 2017; Hamaker et al., 2016).

Our work helps to characterize dynamic patterns relevant to the maintenance of problematic cannabis use and delineates the (in)consistency between dynamic data measuring cannabis use and related experiences with existing theory. These findings are critical to refining a more dynamic conceptualization of substance use and provide evidence to help guide assessment and future, adaptive, interventions. Given the strength of associations demonstrated in the contemporaneous network, psychosocial interventions addressing social and environmental triggers may more proximally impact cannabis cravings, use, or intoxication versus addressing affect or coping strategies directly. Additional research is needed to characterize the extent to which theoretical constructs fluctuate and relate to one another across more granular timescales and across longer periods of time, to provide a more fine-grained assessment of temporal processes that maintain problematic cannabis use, as well as the extent to which there is individual-level variation around such pathways. Taken together, these results provide the initial evidence needed to facilitate the development of dynamic, adaptive treatment and prevention approaches for supporting individuals motivated to reduce their cannabis use.