Engagement patterns among users of a digital intervention to reduce cannabis use: a latent class analysis

Abstract

Background

Understanding and optimizing engagement in digital interventions remains a significant challenge. This study aims to identify different patterns of engagement among users of a digital intervention to reduce cannabis use.

Method

We conducted a secondary analysis of engagement data from the intervention group of a study on the effectiveness of a digital cannabis intervention (ICan). Engagement patterns were identified through a latent class analysis using eight engagement indicator variables. The bias-adjusted three-step approach was used to examine differences in baseline characteristics across classes and associations with average change-from-baseline scores in cannabis use frequency and quantity.

Results

Three latent classes were identified: Class 1 (32%), ‘non-engagers’, showed minimal to no engagement; Class 2 (41%), ‘shorter-term engagers’, showed moderate engagement in the intervention, but with a shorter duration than recommended; Class 3 (27%), ‘long-term engagers’, showed high engagement in the intervention with a duration meeting the recommendations. The proportion of males was significantly higher in the ‘non-engagers’ class compared to the others. The ‘long-term engagers’ class reported fewer tobacco use days at baseline compared to the others. No differences were found between classes regarding average change-from-baseline scores in cannabis use frequency and quantity.

Conclusions

Users of digital interventions show distinct engagement patterns, and characteristics such as male gender and tobacco use may predict sub-optimal engagement. Importantly, also sub-optimal exposure to a digital intervention may be associated with changes in cannabis use, as higher engagement does not necessarily lead to greater effectiveness.

Highlights

• •Users of a digital cannabis intervention show different engagement patterns.
• •Latent class analysis identified non-, shorter-term, and long-term engagers.
• •Male gender and tobacco use are associated with suboptimal engagement.
• •Higher engagement does not necessarily lead to greater reductions in cannabis use.

1. Background

Engagement is a significant challenge for digital substance use interventions with limited amounts of guidance (Linardon and Fuller-Tyszkiewicz, 2020). Several (systematic) reviews have been conducted to identify socio-demographic and clinical characteristics that predict engagement with digital mental health interventions (Beatty and Binnion, 2016; Borghouts et al., 2021; Boucher and Raiker, 2024; Lipschitz et al., 2023; Zainal et al., 2025). Across these reviews, female gender and higher level of education seem to be fairly consistently associated with higher levels of engagement (Beatty and Binnion, 2016; Borghouts et al., 2021; Boucher and Raiker, 2024; Lipschitz et al., 2023; Zainal et al., 2025).

Findings regarding employment status are less consistent. Some evidence suggests that people who work fulltime show higher levels of engagement than people who are unemployed or retired (Borghouts et al., 2021), while another study found no association between employment status and engagement (Beatty and Binnion, 2016). However, there is also evidence that suggest that employment status is only associated with certain types of engagement, namely intervention completion (Zainal et al., 2025). For other characteristics, such as age, ethnicity and symptom severity of the condition targeted by the intervention results seem even more ambiguous (Beatty and Binnion, 2016; Borghouts et al., 2021; Boucher and Raiker, 2024; Lipschitz et al., 2023; Sinadinovic et al., 2020; Zainal et al., 2025).

The relationship between engagement and intervention outcome also appears to be complex. While it seems logical to assume that greater engagement leads to better outcome, previous studies suggest it is not that simple. Gan et al. (2021) conducted a meta-analysis to examine the relationship between engagement with digital mental health interventions and mental health outcomes. Greater engagement appeared to be significantly associated with improvements in mental health. Donkin et al. (2011) previously conducted a similar meta-analysis to investigate the relationship between engagement and outcome in digital therapy. For interventions targeting physical health (e.g. smoking), engagement was positively associated with outcome. For interventions targeting anxiety and depression however, only the number of modules completed appeared to be associated with outcome; other engagement measures (the number of log-ins, self-reported activities, time spent online, and number of pages opened) were not associated with outcome. Linardon and Fuller-Tyszkiewicz (2020) also found mixed results in their systematic review in which they examined the relationship between engagement with smartphone delivered interventions for mental health problems and outcome. Some of the included studies found that more frequent intervention usage was associated with better outcome, while other studies did not find such an association. Thus, the relation between engagement and outcome appears to be a complex phenomenon and not just a simple case of the more the better. Perhaps some users derive enough support from a few essential aspects of the digital intervention, making the rest of the content unnecessary for them, while other users need all the content to achieve the desired result. If we would be able to gain more insight into patterns of engagement and their association with user characteristics and intervention outcomes, we could utilize this information to target and enhance engagement among those for whom it appears most beneficial.

Dacosta-Sánchez et al. (2023) identified latent classes of cannabis users based on their adherence to face-to-face treatment and abstinence during treatment as measured by negative urine tests. Three latent classes were identified. Class 1 had low adherence and low abstinence rates. This class was characterized by a younger age, lower level of education, highest initial frequency of cannabis use, a high likelihood of smoking cannabis as route of administration, higher initial cocaine use and were more likely to be referred by family members. Class 2 had high adherence and high abstinence rates. This class had a higher level of education, a lower level of unemployment, a lower frequency of initial cannabis use and was more likely to be referred by legal services. Class 3 had low adherence but high abstinence rates. The socio-demographical characteristics of this class were quite similar to those of class 1 (lower level of education), while the main method of referral was more similar to class 2 (referral by legal services).

Bell et al. (2020) used k-modes clustering to explore patterns of engagement in Drink Less, a behavior change app that helps people to reduce their alcohol use. Based on whether participants opened the app or not on each day from day 1 to day 30 after downloading, three clusters were identified. The largest cluster was named ‘fast disengagers’ (57%), the second largest cluster was named ‘engagers’ (24%) and the smallest cluster ‘slow disengagers’ (19%). The ‘engagers’ were more likely to be male, were older, and had lower AUDIT baseline scores than the ‘slow disengagers’ and ‘fast disengagers’. Bell et al. did not examine the relationship between the distinct clusters and the efficacy of the Drink Less app in reducing alcohol use.

The study by Bell et al. (2020) demonstrates that log data from an app can provide valuable insights into different patterns of engagement in digital interventions. In the current study, we used data from a randomized controlled trial that was conducted to evaluate the effectiveness of the digital cannabis reduction intervention app ICan compared to four online modules of educational information on cannabis (control condition) (Olthof et al., 2023). ICan is based on the Screening Brief Intervention and Referral to Treatment (SBIRT) approach and contains elements of cognitive behavioral therapy and motivational interviewing. ICan users receive adherence focused guidance, that consists of a weekly personalized WhatsApp message from a coach. The RCT showed no differential effect of ICan on the reduction in cannabis use days, which was the primary outcome measure. The study did show a favourable 3-month effect of ICan on the reduction in the number of grams of cannabis used compared to the control condition. In the current study we are interested in exploring engagement with the ICan intervention in different subgroups and the association of this engagement with its effectiveness.

The aim of the current study was threefold:

• 1)Identify different patterns of engagement through a latent class analysis using a range of engagement indicator variables. Based on the studies by Dacosta-Sánchez et al. (2023) and Bell et al. (2020) we expected to find three latent classes reflecting different levels of engagement.
• 2)Examine how these patterns of engagement are associated with baseline user characteristics. Based on previous studies we expected female gender, high level of education and fulltime employment status to be associated with higher levels of engagement. We also expected that age, ethnicity (approximated using country of birth), baseline symptom severity, baseline cannabis use, cannabis route of administration and baseline use of other substances to be associated with engagement. However, given the ambiguous results or limited evidence in previous reviews we did not have a clear hypothesis for the direction of the association between these variables and levels of engagement.
• 3)Examine how these patterns of engagement are associated with efficacy of the ICan app (reducing cannabis use).

2. Materials and methods

2.1. Participants

The current study was a secondary analysis of data obtained from the intervention group of the randomized controlled trial that evaluated the effectiveness of the digital ICan intervention (Olthof et al., 2023). Approval for the RCT was obtained from the Medical Research Ethics Committees United (NL67449.100.18). Inclusion criteria were aged ≥18 years, using cannabis on ≥3 days/week in the past 3 months, having the desire to reduce/quit cannabis use and having a smartphone available. The final sample used in the current study consisted of the 188 participants allocated to the intervention condition, of whom 133 (70.7%) were male. The mean age of the sample was 27.5 years (SD = 8.6).

2.2. The ICan intervention

The ICan intervention starts with four self-tests based on validated screening instruments, including the Alcohol, Smoking and Substance Involvement Screening Test (ASSIST) and the Cannabis Use Disorders Identification Test Revised (CUDIT-R) (Adamson et al., 2010; WHO ASSIST Working Group, 2002). After completing the self-tests, participants can develop a personalized plan to reduce or quit their cannabis use (reduction plan) through six modules 1) motivation to change 2) setting a goal 3a) how to achieve your goal 3b) optional module: how to solve problems 4) social support 5) withdrawal, craving and relapse 6) summary: your personal plan. After creating the reduction plan, participants are encouraged to monitor their cannabis use on a daily basis using the cannabis consumption diary. A forum allows participants to share experiences with peers, fostering social support. Finally, the ICan intervention includes a treatment page offering comprehensive information on additional resources and specialized treatment services for reducing or stopping cannabis use.

2.3. Procedures

Participants were recruited between December 2019 and June 2020 through paid advertisement campaigns on Facebook and Instagram. After providing informed consent and completing the online baseline questionnaire, participants were granted access to the digital ICan intervention. They were encouraged to use the intervention for at least 4 weeks. After six weeks, three months and six months, participants received email invitations to complete the online follow-up questionnaires.

2.4. Measures

The baseline questionnaire and follow-up questionnaires included a range of measures that are described in Supplement 1. Table 1 shows the variables and instruments that were used in the current study. Cannabis use frequency (the number of cannabis use days in the preceding 7 days) was assessed using the time-line follow-back method (Sobell and Sobell, 1990). Cannabis use quantity (the number of grams of cannabis used in the preceding 7 days) was assessed based on the number of joints made from one gram of cannabis and the number of joints consumed. Besides the questionnaire data, we also collected data regarding the participants' engagement to the ICan intervention. The software program Matomo was used to collect the logdata. Logdata refers to records of time-stamped actions performed by individual participants (Sieverink et al., 2017). After extracting the logdata, R was used to compute the following eight variables that reflect the participants' engagement with the ICan intervention:

• Screening: the screening variable was categorized as ‘low’ if the self-tests were not completed and categorized as ‘high’ if the self-tests were completed.

• Reduction plan: the reduction plan variable was categorized as ‘low’ if none of the modules to create the plan were completed, ‘medium’ if one or two modules were completed, and ‘high’ if three or more modules were completed.

• Treatment page: the treatment page variable was categorized as ‘low’ if the treatment page was not visited and as ‘high’ if the page was visited.

• Cannabis consumption diary: participants were advised to register their cannabis use in the cannabis consumption diary on a daily basis for a minimum of 4 weeks (28 days). Based on this, <28 registrations were categorized as ‘low’, 28–42 registrations were categorized as ‘medium’ and > 42 registrations were categorized as ‘high’. It should be noted that participants can enter one registration per calendar day. However, entries may be backfilled for previous days, up to a maximum of seven days retrospectively.

• Forum: the forum variable was categorized as ‘low’ if the forum was not visited, ‘medium’ if the forum was visited without any interaction and ‘high’ if there was interaction on the forum (i.e. by posting and/or responding).

• Logins: participants were advised to use the ICan intervention on a daily basis for a minimum of 4 weeks (28) days. Based on this, <28 logins were categorized as ‘low’, 28–42 logins as ‘medium’ and > 42 logins as ‘high’.

• Actions: the average number of actions per login divided into tertiles (‘low’, ‘medium’ and ‘high’).

• Number of weeks begin first and last login: based on the recommendation to use the ICan app for at least 4 weeks, <4 weeks was categorized as ‘low’, 4–6 weeks as ‘medium’ and > 6 weeks as ‘high’.

Table 1.

Variables used to examine potential differences across the identified classes.

Variable	Operationalization
Baseline user characteristics
Gender	1 = male 2 = female
Age
Level of education	1 = low
	2 = medium
	3 = high
Employment status	1 = full-time (≥35 h/week)
	2 = part-time (<35 h/week)
	3 = unemployed
Ethnicity	1 = Born in the Netherlands
	2 = Born outside the Netherlands
Symptom severity	Cannabis Use Disorders Identification Test (CUDIT)
Cannabis use frequency	No. of cannabis use days in the preceding 30 days
	No. of cannabis use days in the preceding 7 days
Cannabis use quantity	No. of grams of cannabis used in the preceding 7 days
Cannabis route of administration	Smoking cannabis in a joint mixed with tobacco
	1 = always
	2 = never to most of the time
Alcohol use	No. of alcohol use days in preceding 30 days
Tobacco use	No. of tobacco use days in preceding 30 days
Cocaine use	No. of cocaine use days in preceding 30 days
Ecstasy use	No. of ecstasy use days in preceding 30 days
Change-from-baseline-scores
Cannabis use frequency (3 m)	No. of cannabis use days in the preceding 7 days, difference between baseline and 3-month follow-up
Cannabis use frequency (6 m)	No. of cannabis use days in the preceding 7 days, difference between baseline and 6-month follow-up
Cannabis use quantity (3 m)	No. of grams of cannabis used in the preceding 7 days, difference between baseline and 3-month follow-up
Cannabis use quantity (6 m)	No. of grams of cannabis used in the preceding 7 days, difference between baseline and 6-month follow-up

2.5. Statistical analysis

LatentGOLD (version 6.0) was used to conduct the Latent Class Analysis. Missing data were not imputed. There were (almost) no missing values for indicator or baseline characteristic variables. Missing data occurred only for the cannabis-related outcome variables (change-from-baseline-scores). We decided to include all available cases in the analyses in order to maximize the use of the available information. Cases with missing values on an outcome variable did not contribute to the estimation of that outcome, but were retained in the analysis. The eight engagement variables described above served as the indicator variables and were treated as nominal variables. We estimated models ranging from 2 to 8 classes. The following goodness of fit indices were used to select the model that effectively described the data, while maintaining the highest level of parsimony: Akaike information criterion (AIC), Bayesian information criterion (BIC), the sample size adjusted BIC (aBIC), entropy, and the bootstrap likelihood ratio test (BLRT). After selecting the most optimal model, we used the bias-adjusted three step approach to examine potential differences in specific baseline user characteristics across the classes. We also used the bias-adjusted three step approach to examine whether the classes are associated with the average change-from-baseline score in cannabis use frequency and quantity. To assess whether missing outcome data due to study attrition may have biased the change-from-baseline results, we conducted a sensitivity analysis using the multiply imputed dataset from the main RCT (Olthof et al., 2023). The primary generalized linear mixed-model analyses were repeated within the intervention group, replacing treatment condition (ICan or control) with most likely class membership as predictor of cannabis use over time.

Table 1 shows the variables used and how they were operationalized. The hypotheses and analysis plan were preregistered in OSF registries, doi:10.17605/OSF.IO/HMW7X. In response to peer-review comments, three deviations from the preregistered analysis plan were made: the operationalization of the forum engagement indicator was revised, the operationalization of the ethnicity variable was revised, and the additional sensitivity analysis using the multiply imputed dataset was added.

3. Results

Table 2 shows the fit indices for the estimated models. The BIC and entropy value indicated that the 3-class solution best fitted the data, while the AIC suggested that the 5-class solution best fitted the data and the sample size adjusted BIC favoured the 4-class solution. Based on our initial hypothesis, we selected the 3-class solution as the base case model and included the 4-class solution and the corresponding analyses as an alternative model in a supplement (Supplement 2, Supplement 3).

Table 2.

Fit indices for two- to eight-class models (nominal).

Classes	LL	AIC	BIC	aBIC	Entropy	BLRT
2	−949.17	1956.34	2050.20	1958.34	0.92	0.00
3	−802.40	1692.80	1835.20	1695.83	0.98	0.00
4	−779.02	1676.05	1876.00	1680.12	0.95	0.01
5	−763.90	1675.81	1915.31	1680.91	0.95	0.02
6	−751.53	1681.07	1969.11	1687.20	0.94	0.04
7	−741.58	1691.15	2027.74	1698.33	0.94	0.03
8	−733.80	1705.60	2090.73	1713.80	0.94	0.05

Note: LL = log-likelihood; AIC = Akaike information criterion; BIC = Bayesian information criterion; aBIC = sample size adjusted BIC; BLRT = bootstrap likelihood ratio test. The selected model is shown in bold.

Table 3 shows the item response probabilities for each latent class. Class 1 (non-engagers, 32%), the second largest class, demonstrated low probabilities of completing the self-tests and reduction plan, as well as visiting the treatment page and forum. The class had high probabilities of having a low number of logins and a short duration between first and last login. Class 2 (shorter-term engagers, 41%), the largest class, demonstrated high probabilities of completing the self-tests and at least some modules of the reduction plan. The class showed moderate probabilities of visiting the treatment page and forum, but low probabilities regarding a high number of registrations in the consumption diary and logins. The class also had low probabilities of a long duration between first and last login. The class had the highest probability of having a high average number of actions per login, indicating that although they logged in infrequently, they had a relatively large number of actions during each login. Class 3 (long-term engagers, 27%), the smallest class, had high probabilities of completing the self-tests and the reduction plan, as well as visiting the treatment page and forum. This class also had high probabilities of a high number of logins, registrations in the consumption diary and a long duration between first and last login.

Table 3.

Item response probabilities for the 3-class model.

	Class 1 Non-engagers	Class 2 Shorter-term engagers	Class 3 Long-term engagers	Overall
Class size	0.32	0.41	0.27
Screening
Self-tests not completed	0.99	0.00	0.02	0.32
Self-tests completed	0.01	1.00	0.98	0.68
Reduction plan
0 modules completed	1.00	0.18	0.04	0.40
1 or 2 modules completed	0.00	0.38	0.08	0.18
3 or more modules completed	0.00	0.44	0.88	0.41
Treatment page
Treatment page not visited	1.00	0.51	0.31	0.61
Treatment page visited	0.00	0.49	0.69	0.39
Consumption diary
<28 registrations	1.00	0.91	0.05	0.71
28–42 registrations	0.00	0.09	0.12	0.07
>42 registrations	0.00	0.00	0.84	0.22
Forum
Forum not visited	0.98	0.66	0.27	0.66
Forum visited w/o interaction	0.02	0.21	0.29	0.17
Forum visited with interaction	0.00	0.13	0.44	0.17
Logins
<28 logins	1.00	1.00	0.10	0.76
28–42 logins	0.00	0.00	0.28	0.07
>42 logins	0.00	0.00	0.62	0.16
Actions
<33%	0.86	0.00	0.22	0.34
33%–66%	0.02	0.28	0.78	0.33
>66%	0.12	0.71	0.00	0.34
Weeks between first and last login
<4 weeks	1.00	0.54	0.00	0.54
4–6 weeks	0.00	0.13	0.06	0.07
>6 weeks	0.00	0.34	0.93	0.39

Table 4 shows the associations between the classes, baseline user characteristics and change-from-baseline scores. A significant difference in gender was observed across classes (Wald = 7.39, p = .025). Paired comparisons revealed that the proportion of males was significantly higher in the ‘non-engagers’ class (83.3%) compared to both the ‘shorter-term engagers’ class (68.1%) and the ‘long-term engagers’ class (59.8%). There was no significant variation in country of birth, level of education and employment status across classes.

Table 4.

Associations between the classes, baseline user characteristics and change-from-baseline scores.

Variable		Class 1 Non-engagers	Class 2 Shorter-term engagers	Class 3 Long-term engagers	Total	p value	Paired comparison
Gender, % male		83.3%	68.1%	59.8%	70.8%	.03	1 > 2 & 3
Age, mean		27.92	27.25	27.26	27.47	.88
Level of education						.13
	Low	23.3%	15.0%	12.5%	17.0%
	Medium	46.6%	43.7%	41.9%	44.2%
	High	30.1%	41.3%	45.6%	38.8%
Employment status						.08
	Unemployed	33.3%	27.4%	36.3%	31.7%
	Part-time	38.3%	45.9%	56.4%	46.2%
	Full-time	28.4%	26.7%	7.4%	22.1%
Ethnicity						.36
	Born in NL	95.0%	96.2%	89.9%	94.1%
	Born outside NL	5.0%	3.8%	10.1%	5.9%
CUDIT, mean		18.55	19.41	18.56	18.91	.67
Cannabis use days/past 30 days		25.00	25.14	22.48	24.39	.08
Cannabis use days/past 7 days		5.85	6.10	5.75	5.93	.40
Grams of cannabis used/past 7 days		5.43	5.69	4.60	5.32	.48
Route of administration joint, always		0.78	0.83	0.66	0.77	.09
Alcohol use days/past 30 days		6.16	6.01	6.98	6.31	.69
Tobacco use days/past 30 days		18.38	18.70	10.07	16.30	<.001	1 & 2 > 3
Cocaine use days/past 30 days		0.51	0.42	0.26	0.40	.39
XTC use days/past 30 days		0.32	0.20	0.24	0.25	.55
Difference 3 mnths use days/past 7 days		−2.04	−1.38	−1.58	−1.65	.59
Difference 6 mnths use days/past 7 days		−2.45	−2.11	−1.72	−2.11	.56
Difference 3 mnths grams/past 7 days		−1.48	−1.87	−2.03	−1.79	.85
Difference 6 mnths grams/past 7 days		−0.65	−2.34	−2.32	−1.79	.12

A significant difference in the mean number of tobacco use days was observed across classes (Wald = 15.94, p < .001). Paired comparisons showed that the mean number of tobacco use days in the past 30 days was significantly lower in the ‘long-term engagers’ class (M = 10.07) compared to both the ‘non-engagers’ class (M = 18.38) and the ‘shorter-term engagers’ class (M = 18.70). No difference was observed between classes with regard to baseline symptom severity (CUDIT score), cannabis use frequency, cannabis use quantity, cannabis route of administration, alcohol use, cocaine use and ecstasy use.

No differences were observed between the classes with regard to the average change-from-baseline score in cannabis use frequency and quantity. Sensitivity analyses conducted to assess whether missing outcome data due to study attrition may have biased the change-from-baseline results yielded similar results.

The 4-class solution showed results that were fairly similar to the 3-class solution. As in the 3-class solution, one class displayed a pattern of little to no engagement, the ‘non-engagers’, and one class showed a pattern of high engagement sustained over a long period, the ‘long-term engagers’. In addition, there was a class that demonstrated a moderate level of engagement for a moderate duration the ‘moderate engagers’ and a class that completed the screening and (part of) the reduction plan but appeared to disengage quickly after that, the ‘short-term engagers’. Regarding differences between these classes, we observed results comparable to those in the three-class model, with significant differences across classes in gender and tobacco use days. In contrast to the 3-class model, we also found a significant difference in cannabis use between the classes. Short-term engagers appeared to score higher on the number of cannabis use days compared to the other classes.

4. Discussion

In this study, we aimed to identify different patterns of engagement among users of a digital intervention to reduce cannabis use. The latent class analysis found three latent classes based on eight engagement indicator variables. Class 1 (32%), the ‘non-engagers’, showed a pattern of minimal to no engagement in the intervention. Class 2 (41%), the ‘shorter-term engagers’, showed a pattern of moderate engagement in the intervention, with moderate to high probabilities of fewer logins, fewer registrations in the consumption diary and a shorter duration between first and last login than recommended. Class 3 (27%), the ‘long-term engagers’, showed a pattern of high engagement in the intervention with high probabilities of meeting or exceeding the number of logins, registrations in the consumption diary and duration between first and last login that was recommended.

We also examined how the identified classes were associated with baseline user characteristics and change-from-baseline scores in cannabis use. The proportion of males was significantly higher in the ‘non-engagers’ class compared to both the ‘shorter-term engagers’ class and the ‘long-term engagers’ class. This finding is in line with previous research and our hypothesis that female gender is associated with higher levels of adherence (Beatty and Binnion, 2016; Borghouts et al., 2021; Lipschitz et al., 2023; Zainal et al., 2025). In contrast to our hypothesis, we did not find an association between class membership and employment status. Although not statistically significant, the association between engagement and employment status seems the opposite to what was expected. The percentage of full-time employees is the lowest in the in the ‘long-term engagers’ class and the percentage of unemployed individuals the highest. Perhaps individuals who are unemployed simply have more time to engage with the digital intervention. No differences were found with regard to country of birth, cannabis use frequency, cannabis use quantity and baseline symptom severity (CUDIT).

The mean number of tobacco use days (not in combination with cannabis) was significantly lower in the ‘long-term engagers’ class compared to the other classes. Although there was not a statistically significant difference in route of administration, the percentage of individuals who always smoke their cannabis in a joint mixed with tobacco appeared to be the lowest in the ‘long-term engagers’ class. These findings are in line with previous studies that show that co-use of cannabis and tobacco is associated with higher levels of dependence and poorer treatment outcomes (Hindocha et al., 2015; McClure et al., 2020). In a recent review on this topic, Nguyen et al. (2024) highlight the urgent need for treatments addressing the co-use of tobacco and cannabis.

We did not find any differences between the classes with regard to the average change-from-baseline score in cannabis use frequency and quantity. This suggests that the intervention was not more effective in reducing cannabis use among the classes that engaged more with it compared to the class that had minimal to no engagement with the intervention. However, the sample size was relatively limited, which may have reduced statistical power to detect differences between classes. Although the differences were not statistically significant, the mean reduction in cannabis quantity ranged from −0.65 g in the ‘non-engagers’ class to −2.32 and − 2.34 g in the more engaged classes. Future research with larger samples is needed to determine whether these differences reflect a potentially meaningful association between engagement and intervention outcomes.

This study has some limitations that need to be addressed. First, not all of the fit indices were the most optimal for the 3-class model, some indicated that the 4-class model or the 5-class model was most optimal. However, the 4-class model led to fairly similar results (Supplement 2, Supplement 3). Second, the sample size was relatively limited, which may have reduced the statistical power to detect differences between the classes. Third, we did not find an association between ethnicity and class membership. However, ethnicity was approximated using country of birth, which may not adequately capture cultural factors that could be relevant to intervention engagement.

All in all, the current study demonstrates that users of digital cannabis interventions show different patterns of engagement. The current study provides insights into user characteristics associated with limited or no engagement, such as male gender and tobacco use. Perhaps we could use these insights to try to improve engagement among these specific groups. However, importantly, also sub-optimal exposure to a digital intervention may be associated with changes in cannabis use, as higher engagement does not necessarily lead to greater effectiveness.

Declaration of Generative AI and AI-assisted technologies in the writing process

During the preparation of this work the author(s) used Microsoft Copilot in order to support language editing. After using this tool/service, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the content of the published article.

Funding

The ICan study is funded by the Dutch Ministry of Health, Welfare and Sport. The funder had no role in the design, data collection, analysis and writing of this manuscript.

Declaration of competing interest

The authors declare that they have no competing interests.

Appendix A. Supplementary data

Supplement 1

Measurements and instruments used in the study.

Supplement 2

Item response probabilities for the 4-class model.

Supplement 3

Associations between the classes, baseline user characteristics and change from baseline scores.

Data availability

The data that support the findings of this study are available from the corresponding author upon reasonable request.