The sample was recruited from 68 high schools in Quebec within the context of the evaluation of the new approaches new solutions (NANS) dropout prevention program (2002–2008) (47). Participants attended secondary schools in disadvantaged communities of the province of Quebec (Canada). NANS schools were selected using stratified random sampling to be representative of all schools in disadvantaged areas of Quebec in terms of geographical location, size, and language (47). Data were obtained via self-reported questionnaires administered in class by teachers supervised by trained and supervised experimenters. Seventy-seven percent of eligible participants provided free and informed consent to participate in the study. All procedures were approved from the Arts and Science Faculty Ethical Review Board at University of Montreal. Participants for this study were a cohort assessed annually from grade 7 to grade 11 (2003–2008). The sample for the present study included all participants who provided information on cannabis use in grade 11 (N = 1618). Participants were mostly Quebec-born Caucasians (93%). Other participants were from a diversity of ethnicities. The sample included slightly more females (53%) than males (47%).

Self-reported substance-use behaviors were collected in grade 10. Predictors were considered in grade 7 and 8 and outcomes in grade 11. Available data for outcomes in grade 11 were 61%). Rates of available data for predictors in grades 7–8 ranged from 80 to 99%.

Latent class analysis was used to identify subgroups of cannabis users. This statistical method aims to identifying the most parsimonious classification of individuals into latent classes by maximizing homogeneity within, and heterogeneity between classes. In order to determine the optimal number of classes, different number of latent classes was modeled starting from 1 (e.g., only one class of cannabis users), then 2, and so on until we reach an optimal solution. Different criteria were used to select the most appropriate model (59). These criteria included the following information criteria: deviance, the Akaike information criterion (AIC) (60), the Bayesian information criterion (BIC), and the sample-size adjusted Bayesian information criterion (SSBIC) (61), to compare the relative fit of solutions. Better fitting solutions are reflected in lower values on the indices. We also considered likelihood ratio tests, including the Vuong–Lo–Mendell–Rubin and Lo–Mendell–Rubin adjusted likelihood ratio tests – ALRTs (62). ALRT tests are adequate for non-nested mixture models and test the significance of the difference in fit between two models with a one class difference. We also considered the recommended Bootstrapped Likelihood Ratio Test [BLRT; (63)]. The criterion for significance was α < 0.05. We also relied on entropy, which is indicative of the degree of homogeneity within and independence between classes (60). Elevated scores of entropy indicate high independence and little spillover between classes. Furthermore, we examined the substantive interest of each model by evaluating how solutions compare with theoretical and empirical knowledge. Finally, although we selected a solution based primarily on unconditional models, we also investigated all solutions with predictors to determine whether all classes could be meaningfully differentiated (59). All models were estimated using maximum likelihood, and multiple initial values (5000 starts; 100 optimizations) were used to avoid local maxima. We imputed five datasets with an EM technique in SPSS (version 20.0) and replaced missing values by the mean of all imputed values. Mplus (version 7.0) software (64) was used for the LCA (65, 66).

After selecting a solution with an optimal number of classes, the obtained classes were compared on sociodemographic and psychosocial predictors in grade 7–8 as well as on substance-related problems the following year (grade 11). We evaluated the association between classes and each predictor with all predictors simultaneously in the model. Predictors were linked to class group membership using multinomial regression. For the outcome (attributed substance-related problems), we compared the means of the class model using equality of means test across classes based on posterior probability-based multiple imputation [AUXILIARY option in Mplus; (64)].

Means and standard deviation for continuous variables as well as percentages for categorical variables are presented in Table 1. Missing data ranged from 1 (age) to 776 (48%) (outcome) with a mean of 205 (12.7%).

Comparisons of entropy and spillover indices, fit indices (deviance, AIC, BIC, SSBIC), and likelihood ratio tests for the one to six class LCA models suggested that the four-class model provided the best fit (see Table 2). As can be seen, model fit on all indices tended to improve as the number of classes increased, but the rate of improvement started to diminish around a four-class model. This solution had close to the lowest BIC and adjusted BIC scores with the highest entropy value of 0.83 (60). Likelihood ratio tests suggest few incremental validity beyond a four-class model. Models with 5 and 6 classes did not significantly improve model fit over models with fewer classes. The removal of covariates and outcome did not result in a change to the four-class solution, contrary to other solutions, indicating that the assumption of local independence was not violated. The four-class model also appears better than simpler models and more clinically significant. The four classes are distinct and each represents a significant number of participants. And as we will see below, classes can be discriminated by their association with predictors and outcome. We thus selected a four-class model.

Four distinct classes based on use patterns were identified. These classes were labeled Late-Light Use (1; N = 454, 28%), Late-Heavy + Polydrug Use (2; N = 222, 14%), Early-Moderate Use (3; N = 526, 33%), and Early-Heavy + Polydrug Use (4; N = 416, 26%) (see Figure 1). There are significant differences at the 0.05 level between all classes on all items except alcohol use precocity for comparisons with the Early-Heavy + Polydrug Use class, which had no variance on this item. There are differences between almost all items’ categories. Late Onset/Light Users had the lowest levels of use on each substance-related indicator. Early-Heavy + Polydrug Users had the highest levels of use on most indicators. The other two classes fell in between. Tobacco use was the highest in the Early-Heavy + Polydrug Users, the lowest in Late-Light Users, and was similar between the two other classes. For alcohol use indicators (frequency, binge, typical quantity), Early-Heavy + Polydrug Users and Late-Heavy + Polydrug Users are at similar levels despite early-onsetters showing slightly heavier patterns. Once again, Late-Light Users showed the lowest levels with Early-Moderate Users falling in between. In terms of cannabis use indicators (frequency, typical quantity) as well as of stimulant/hallucinogens and of cannabis use and alcohol polyuse, we observe very similar patterns.

As shown in Table 4, on a mean scale of the 14 substance-related harm items, scores were respectively 0.09, 0.26, 0.17, and 0.36 for each class and were all mutually statistically different. As expected, the Early-Heavy + Polydrug Use class had the highest levels of problems (M = 0.36, SD = 0.012) and the Late-Light Use the lowest (M = 0.09; SD = 008). Notably, the Early-Moderate Use class had a lower level of problems (M = 0.17; SD = 009) than the Late-Heavy + Polydrug Use (M = 0.26; SD = 0.017).

As shown in Tables 3 and 4, significant differences were found between all classes. Odds ratios are calculated for a one standard deviation variation in predictors. Classes Late-Heavy + Polydrug Use and Early-Heavy + Polydrug Use were similar in terms of concurrent use in grade 10. However, compared to the Late-Heavy + Polydrug Use class, Early-Heavy + Polydrug Users had more problems in grade 11, and were older, had an earlier substance use onset (alcohol and cannabis), had a higher proportion of boys, more delinquent behaviors, deviant peers, and conflict with parents and were less monitored by them. Early-Moderate Use and Early-Heavy + Polydrug Use classes both are early-onsetters, but compared to the Early-Heavy + Polydrug Use class (and Late-Heavy + Polydrug Use), the Early-Moderate Use class had more moderate patterns of use as reflected in less problems than in Late-Heavy + Polydrug Use or Early-Heavy + Polydrug Use classes. The Early-Moderate Use class had the highest proportion of female. It has lower peer deviancy proportions than Early-Heavy + Polydrug Use class, but higher than Late-Light Use and Late-Heavy + Polydrug Use classes. It also has comparable levels of delinquent behaviors and parent monitoring with the Early-Heavy + Polydrug Use class, which are at more problematic levels than in Late-Light Use and Late-Heavy + Polydrug Use classes. The Late-Light Use class had, in addition to the lowest level of problems, the lowest level of substance use as well as the lowest level of risk. This class is younger than all other three, it had higher levels of school bonding than both Late-Heavy + Polydrug Use and Early-Heavy + Polydrug Use classes, and it had the lowest peer deviancy but had similar levels of delinquent behaviors with the Late-Heavy + Polydrug Use class, which were lower than in Early-Moderate Use and Early-Heavy + Polydrug Use classes. A similar pattern emerged regarding parental monitoring. The Late-Light Use and Late-Heavy + Polydrug Use classes had similar levels, which were higher than in Early-Moderate Use and Early-Heavy + Polydrug Use classes. The Late-Light Use class had also lower levels of conflicts than the Early-Heavy + Polydrug Use.

The results have many implications. First, they discriminate two different types of early as well as late onset cannabis users. They do so by shedding light on their distinctive relationships with risk factors from multiple dimensions as well as substance-related problems. Furthermore, our results, taken together, also shed light on the fact that not all early-onsetters are at elevated risk and experience a high level of substance-use related problems and that they are even at a lower level of risk than some late-onsetters. These results should be used to more meaningfully target and inform effective interventions toward users experiencing elevated levels of risks and harms. Moreover, a typology provides a useful heuristic for clinicians conducting assessment or screening with cannabis-involved adolescents. Our results suggest that screening and intervention for risky cannabis use among adolescents should focus on school bonding in order to discriminate late-onset classes and on peer delinquency in order to discriminate early-onset classes. Intervention should be prioritized for the Early-Heavy + Polydrug use and Late-Heavy + Polydrug use classes. School bonding and peer deviancy seem to be good targets for intervention with either Early- or Late-Heavy + Polydrug Users and parental monitoring and conflict with parents seem to be further good targets for intervention with Early-Heavy + Polydrug users. In both cases (Early- or Late-Heavy + Polydrug), binge drinking, cannabis use frequency and alcohol and cannabis simultaneous use seem to be the most important substance-use behaviors to target in interventions. In the first case, working simultaneously on these use patterns, in addition to stimulants/hallucinogens use frequency, and psychosocial risk factors, with demand and harm reduction interventions, would probably be a good strategy whereas intervention with the latter group should focus primarily on use patterns. Indeed, it is noteworthy and important to take into account that the Late-Heavy + Polydrug use class is mostly constituted of females with lower levels of risk. These cannabis users are more difficult to predict, but they have important intervention needs. The Early-Moderate Users would on their part benefit from early intervention strategies in order to prevent their use to shift from moderate to heavy as well as to prevent it to become more problematic. Overall, other than substance use behaviors, the main factors to target would generally be school bonding, delinquency, peer delinquency, and parental monitoring. In terms of policy implications, the current legal framework in Canada and elsewhere is characterized by the criminalization of all use; any cannabis use is defined as problematic (68). This approach differs from the one prevailing for alcohol, which has evolved to a public health framework (69, 70). Rather than focusing on use per se, priority is given to the risks and harms associated with problematic patterns of use (e.g., drunk driving). This way, targeted interventions may be applied to relevant behaviors (71). In our study, this could mean targeting binge drinking frequency and substance mixing behaviors as well as other substance use. Harm reduction strategies also seem to be potential useful tools in order to reduce cannabis-related problems.

This study has multiple strengths, including the simultaneous consideration of substance use severity indicators, predictors, and outcomes as well as their multidimensionality, and use of a large prospective community-based sample. However, this study is not without limitations. First, despite the fact that confidentiality was assured, response bias and common method variance could have influenced our results. Fortunately, the validity and reliability of self-reported data on substance use have been established (72–74), but this has not been proved for self-report of problems. In addition, the sample comes from deprived areas, which is a limitation to the generalization of results. However, even if schools from deprived areas were sampled, individual scores of familial adversity vary and include participants from low familial adversity. Another limitation is related to the large amount of missing data and potential attrition bias. Also, the results do not provide information on the sequence of problem as well as the subgroup development over time. Finally, the inclusion of age of onset in the typology, while substance use indicators have been selected from grade 10 is another potential limitation to the current study.

Future prospective studies should examine factors that explain transitions across these subtypes in time. This would however be complex because age of onset is included in the typology. Another important area of development is in the study of specific harm categories (relational, health, school, etc.) related to different patterns of use in order to better inform prevention and treatment efforts to target specific harms. Indeed, if different outcomes are related to different classes, intervention should not only target specific factors related to specific patterns but also focus on specific problems related to each. Which are the most important problems related to each class? Which classes are disproportionally represented for each problem? Another potential improvement over the current study is the use of more specific items for each other drugs than alcohol, Tobacco, and Cannabis (e.g., ecstasy, LSD, Speed, GHB, Ketamine, etc.) as well as substance use motives. Finally, a nationally representative sample would also improve the external validity of the typology.