This manuscript used the novel methodological approach of classification trees to explore distinct pre- and post-legalization risk profiles for cannabis use in a large sample of youth. Previous research has shown that classification trees are superior to logistic regression with respect to identifying correlates of cannabis use (Dell et al., 2022). Our results highlight an increase in reports of cannabis never use and a slight decline in current cannabis use in our samples. While this is contrary to evidence of a plateau in cannabis use among youth during the early stages of the COVID-19 pandemic (Leatherdale et al., 2021), it may indicate that the declines observed here (a few years out from the pandemic-related restrictions) may not be a result of the pandemic but more likely due to regulations association with legalization and/or changes in social norms. That being said, the more novel finding is that our classification trees identified how the relative ranking of risk factors for predicting current cannabis use changed considerably from pre- to post-legalization. It does appear that there is a meaningful shift in both cannabis use and the predictors of use among youth that occurred over this period of time.

As stated, while many risk factors were common across years, there were also substantial changes in the risk factors identified. These results presented clearly identify that the risk profiles for current cannabis use among youth are not fixed in stone but rather more fluid; it is reasonable to assume they are constantly changing as social norms change over time. As such, consistent with previous cannabis control recommendations (Wellman et al., 2023), prevention efforts may benefit from also adapting over time to target the relevant risk factors associated with cannabis use. Ongoing and timely surveillance efforts are necessary to inform leverage points at a particular period of time.

These results provide robust new evidence that mental health related variables examined have increasing importance with respect to current cannabis use post-legalization relative to pre-legalization. For instance, in the pre-legalization model, the only mental health-related variable with a meaningful relative variable importance score was depression. Whereas, in the post-legalization model, depression had the largest relative variable importance score, followed by the emergent mental health factors of emotional regulation and anxiety. Moreover, in the pre-legalization classification tree (Fig. 1), there were no mental health factors associated with the high-risk groups for current cannabis use. However, in the post-legalization classification tree, both depression and anxiety are characteristics associated with high-risk current cannabis use groups. While this finding is novel, it does align with research suggesting that during this same period of time (where there was also the onset of the COVID-19 pandemic in 2020), population-level changes in mental health occurred (Statistics Canada, 2020). Evidence suggests Canadian students with poor mental health commonly reported using cannabis as a coping mechanism for dealing with the societal changes associated with the early stages of the COVID-19 pandemic (Romano et al., 2021). This is also supported by evidence showing that the COVID-19 pandemic did not have an impact on rates of youth cannabis use among Canadian youth (Leatherdale et al., 2021). As such, it appears that mental health risk factors during the post-legalization period became more important contributors to cannabis use.

Unlike the emergence of mental health variables, there are some characteristics identified in both models as being important correlates of current cannabis use over both time periods examined. Indicators of academic disengagement, such as low value on getting good grades, minimal time spent on homework, lack of school connectedness, and low educational aspirations following high school were key risk factors for cannabis use pre-legalization and remain important post-legalization. While the results suggest their relative importance has decreased somewhat pre- to post-legalization (possibly due to the emergence of the mental health variables within the post-legalization sample), they remain important. This is consistent with research showing that school factors, such as school connectedness, can impact the likelihood of cannabis use. For instance, students with higher school connectedness consistently have a lower likelihood for current cannabis use (Wellman et al., 2023, Varatharajan et al., 2024, Clements-Nolle et al., 2022, Matteau-Pelletier et al., 2020). Evidence also suggests that students with worse than average school performance are more likely to use cannabis (Wellman et al., 2023, Gerra et al., 2020), whereas students who report feeling supported at school are less apt to use cannabis (Moore et al., 2018). While we are unable to determine causality in this study, for schools who want to take action, our results suggest that school-based efforts to improve modifiable risk factors associated with the school context (such as school connectedness, educational aspirations, and valuing homework) among the student population may be effective for reducing youth cannabis use and warrant evaluation.

In terms of sociodemographic predictors, we only observed differences by age and ethnicity among certain subgroups post-legalization. Interestingly, we identified that there were no differences by sex either before or after legalization, while this differs from past research showing that males are more likely to use cannabis then females (Leatherdale et al., 2021, Butler et al., 2023), within the Canadian context post cannabis legalization, there appears to be a narrowing of sex differences in cannabis use among youth (Varatharajan et al., 2024). Moving forward, prevention efforts should be designed to address the needs of both male and female youth.

Key strengths of this study include the very large population-based samples, and the availability of pre- and pot-legalization data. While the COMPASS study is based on self-reported data, which can be prone to recall and social desirability bias, it uses passive consent which is essential for producing robust results that limit self-selection and response bias, particularly for measures of substance use behaviours (Rojas et al., 2008, White et al., 2004). Critically, student names are not required, helping to preserve anonymity and maximally accurate reporting. Possible limitations include COMPASS being based on a convenience sample of participating schools, so the results may not be generalizable to all Canadian youth. However, research has shown that the COMPASS data are comparable to nationally representative data from the Health-Behaviour of School-aged Children Survey (HBSC) (Butler et al., 2023), especially with respect to cannabis use. Moreover, this manuscript did not explore co-occurring substance use as it is already well established that students who use other substances (e.g., alcohol, cigarettes) are substantially more likely to use cannabis (Wellman et al., 2023, Varatharajan et al., 2024). COMPASS does not include measures of peer or parental cannabis use which is an important risk factor for individual cannabis use (Moore et al., 2018). These data also span the early stages of the COVID-19 pandemic, which was previously shown to impact youth mental health and substance use (Romano et al., 2021); however, since research has also shown that the COVID-19 pandemic did not have a substantive impact on rates of youth cannabis use among Canadian youth (even when taking into consideration the transition from paper-based to online survey implementation over the study period) (Leatherdale et al., 2021), it seems reasonable to assume the changes in cannabis use post-legalization are not directly attributable to COVID-19 pandemic impacts. Finally, as with any statistical model, decision trees are subject to limitations, one being the tendency to overfit to sample data relative to other machine learning approaches such as random forests. We partially mitigate this through methodological choices such as employing cross-validation to prune trees and capping the number of levels within this manuscript.