Uncovering Hidden Risk Profile Phenotypes of Substance Use Among Canadian Adolescents via Cluster Analysis of COMPASS Data

doi:10.21203/rs.3.rs-2837559/v1

Download PDF

Article

Uncovering Hidden Risk Profile Phenotypes of Substance Use Among Canadian Adolescents via Cluster Analysis of COMPASS Data

https://doi.org/10.21203/rs.3.rs-2837559/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

This study utilized a longitudinal health survey to identify sub-phenotypes of behavioral risk profiles related to youth substance use. We analyzed 8824 Canadian secondary school students who completed COMPASS questionnaires in 2016/17 and were followed up until 2018/19. The fuzzy clustering algorithm was used to identify four sub-phenotypes of risk profiles, including low-risk, medium-low-risk, medium-high-risk, and high-risk subgroups. The study found that the mean scores of health risk behaviors within each subgroup increased across the three waves, indicating a rising risk of health behaviors over time. The results confirm heterogeneity in the prevalence and characteristics across the subgroups. These findings can help school program managers and policymakers develop targeted interventions to reduce substance use among at-risk individuals and improve youth health behaviors. By identifying at-risk groups, stakeholders can develop effective preventive measures against addictive behaviors in school settings. The study results offer evidence to support the development of risk reduction interventions for youth substance use and ultimately improve youth health behaviors.

Health sciences/Risk factors

Health sciences/Health care/Public health

Substance use

health risk behaviors

phenotype

cluster analysis

Canadian youth

COMPASS study

Adolescence is a critical period marked by increased risk-taking behaviors, particularly in mid-adolescence, which encompasses the 9th -10th grades and is a time when many young people begin experimenting with risky behaviors. Substance use, including alcohol, cigarettes, marijuana, and other drugs, is a significant risk behavior that many mid-adolescents engage in. Research has shown that substance use during mid-adolescence is associated with a higher likelihood of substance use disorders in adulthood¹. While numerous studies have explored the use patterns and risk factors associated with youth substance use, there remains a need to examine these risk factors within the specific context of mid-adolescence. Research has suggested that the risk factors associated with substance use may differ across developmental stages, meaning that the factors that influence substance use in early adolescence may not be the same as those that influence use in later adolescence or adulthood^{2,3,4,5,6,7,8}. Understanding the unique risk factors and profiles of substance use specific to mid-adolescence is crucial for developing more targeted prevention and intervention strategies.

Mid-adolescence is a time of significant changes and challenges, including an increase in autonomy and independence, changes in peer relationships, and increased exposure to risky situations^9,10,11,12. Several studies have highlighted the importance of examining substance use risk factors during mid-adolescence. For example, a longitudinal study found that alcohol use initiation during mid-adolescence was associated with a greater risk for developing alcohol use disorders in adulthood compared to those who initiated use later in adolescence¹³. Similarly, a study examining cigarette smoking patterns among youth found that mid-adolescents were more likely to report regular smoking compared to early and late adolescents¹⁴. Examining the risk factors associated with substance use during mid-adolescence is important for understanding the factors that contribute to the initiation and maintenance of substance use behaviors¹⁵. The insights gained from exploring the risk factors linked to substance use in mid-adolescence can be leveraged to create tailored prevention and intervention strategies aimed at reducing substance use among young individuals.

Age, sex, and ethnicity are well-established individual-level risk factors that impact adolescent substance use^{16,17,18,19,20,21,22}, but other factors such as depression and sedentary behavior also play a significant role^22,23,24. Additionally, eating habits, social connectedness, family and peer influence^25,26, and low social connectedness²⁷ may influence the risk of youth substance use. School connectedness or engagement is another factor that has been linked to substance use among youths^16,28,29,30. Environmental factors such as living in a non-urban setting¹⁸ and socioeconomic status also influence youth substance use^{16,18,28,31,32,33}. While there is existing knowledge on the risk factors linked to substance use among youth, further research is necessary to identify specific risk profiles in order to develop targeted prevention and intervention strategies. Understanding these profiles can aid in identifying opportunities for intervention, evaluating existing policies and practices, and determining overall risk evolution³⁴.

To date, most studies have focused on identifying individual risk factors associated with substance use among youth, and few have taken a data-driven approach to examine sub-phenotypes of risk profiles. The COMPASS longitudinal study aims to address these gaps by collecting comprehensive data about youth health behaviors from multiple sources, examining the relationship between school environmental characteristics and youth health behaviors across several Canadian provinces³⁵. The study collects hierarchical (student-level and school-level) health data via anonymous COMPASS Questionnaires (Cq) each school year, asking student participants about their health behaviors, including healthy eating, physical activity, smoking, alcohol and drug use, school connectivity, and mental health^35,36. The aim of this study was to identify distinct sub-phenotypes of risk profiles associated with substance use among Canadian youth, with the goal of informing the development of targeted risk-reduction interventions. Specifically, the study aimed to provide evidence to health practitioners working in school settings to assist in decision-making around prevention and intervention efforts aimed at reducing substance use among youth.

Descriptive statistics

Out of the 9307 linked samples, only 8824 students were included in the current analyses due to their regular patterns of grade advancement across each school year. At Waver I (W_I, 2016/17), over half (50.9%) of the students were in grade 9, 34.2% were in grade 10, and 14.9% were grades 7–8 (secondary I-II) students in Quebec. Fifty-five percent were female, 74.2% were White, and 67.5% were from Ontario. Fifty-four percent resided in large urban, followed by small urban (30.9%), and 46.6% of students’ household incomes ranged between $50K and $75K. Descriptive statistics of three annual waves of the linked samples were published in another paper²².

Factors associated with youth substance use

Table 1 lists the top factors associated with substance use for each wave, with the importance scores in brackets. The number of smoking friends, number of classes skipped, and weekly allowance was the top three factors consistently appearing across the three waves. Other correlates ranked differently by waves (Fig. 1). The total importance scores for each factor can be seen in the last column of Table 1. Detailed descriptions of each factor were published in another paper²².

Table 1

**Top factors associated with substance use by waves. # of smoking friends**: coded as 0 = “None,” 1 = “1 friend,” 2 = “2 friends,” 3 = “3 friends,” 4 = “4 friends,” and 5 = “5 or more friends.” **# of classes skipped**: coded as 1 = “0 classes,” 2 = “1 or 2 classes,” 3 = “3 to 5 classes,” 4 = “6 to 10 classes,” 5 = “11 to 20 classes,” and 6 = “More than 20 classes.” **Weekly allowance**: coded as 0 = “I do not know how much money I get each week,” 1 = “Zero,” 2 = “$1-$20,” 3 = “$21-$100,” and 4 = “$100+.” **Sedentary time**: range of [0, 2925] minutes. **CESD**: the Center for Epidemiological Studies-Depression, range of [0, 30]. **School connectedness**: range of [6, 24], higher scores indicate higher connectedness. **Eating breakfast**: coded as 0 = “No” and 1 = “Yes.”
Ranking	Wave I (2016/17)	Wave II (2017/18)	Wave III (2018/19)	Voting (Total Score)
1	# of smoking friends (47·92)	# of smoking friends (60·05)	# of smoking friends (56·90)	# of smoking friends (164·87)
2	# of classes skipped (38·01)	# of classes skipped (43·76)	# of classes skipped (54·12)	# of classes skipped (135·89)
3	Weekly allowance (17·48)	Weekly allowance (31·07)	Weekly allowance (37·66)	Weekly allowance (86·21)
4	School connectedness (16·65)	Sedentary time (20·12)	Sedentary time (20·26)	Sedentary time (53·14)
5	Grade/Age (15·58)	CESD (15·87)	Eating breakfast (19·03)	CESD (42·74)
6	CESD (14·19)	School connectedness (14·29)	# of physically active friends (17·41)	School connectedness (42·51)
7	Province (13·38)	Urbanity (13·95)	English marks (15·16)	Eating breakfast (38·29)
8	Sedentary time (12·76)	Eating breakfast (13·32)	Urbanity (14·95)	Grade/Age (38·05)

Characteristics of sub-phenotypes

Four sub-phenotypes (SP1 through SP4) of risk profiles were identified. i.e., low-risk (SP1), medium-low-risk (SP2), medium-high-risk (SP3), and high-risk (SP4). This was determined by the average group score of substance use measures and the included risk factors.

At W_I, SP1, the largest sub-phenotype, included 2799 (32.5%) students in the study cohort. These individuals were characterized by having the lowest mean scores (denoted by $Mean\pm SD$) of substance use (${S}_{Cigarette}=0.06\pm 0.29$, ${S}_{E-Cigarette}=0.14\pm 0.44$, ${S}_{Alcohol}=0.39\pm 0.68$, ${S}_{Marijuana}=0.06\pm 0.30$), the lowest mean scores of CESD (the Center for Epidemiologic Studies Depression, ${S}_{CESD}=7.08\pm 5.39)$, the lowest sedentary time (170$\pm$1.3 minutes), the fewest smoking friends (0.19$\pm$0.65), the fewest classes skipped (1.20$\pm$0.58), the fewest weekly allowance (1.70$\pm$1.14), the highest mean scores of school connectedness (${S}_{School Connectedness}=19.8\pm 2.78$), and the majority eating breakfast ($N=1831$ (65.4%)). Compared to the other three sub-phenotypes, students in SP1 had a low risk of substance use.

SP2 was the second-largest sub-phenotype at W_I, consisting of 2712 (31.5%) students in the cohort. These students had second-lowest mean scores of substance use (${S}_{Cigarette}=0.11\pm 0.38$, ${S}_{E-Cigarette}=0.22\pm 0.55$, ${S}_{Alcohol}=0.57\pm 0.77$, ${S}_{Marijuana}=0.12\pm 0.40$), the second-lowest mean scores of CESD (${S}_{CESD}=8.34\pm 5.89)$, the second-lowest sedentary time (345$\pm$50.7 minutes), the second-fewest smoking friends (0.31$\pm$0.81), the second-fewest classes skipped (1.23$\pm$0.59), the second-fewest weekly allowance (1.81$\pm$1.11), the second-highest mean scores of school connectedness (${S}_{School Connectedness}=19.1\pm 2.66$), and the second-largest sub-phenotype eating breakfast ($N=1400$ (51.6%)). SP2 was defined as the medium-low risk profile.

As the second-smallest sub-phenotype at W_I, SP3 included 2113 (24.5%) students in the cohort, whose mean scores of substance use were higher than SP2 (${S}_{Cigarette}=0.17\pm 0.47$, ${S}_{E-Cigarette}=0.31\pm 0.63$, ${S}_{Alcohol}=0.65\pm 0.82$, ${S}_{Marijuana}=0.21\pm 0.54$). They had higher mean scores of CESD (${S}_{CESD}=9.30\pm 6.04)$, longer sedentary time (563$\pm$83.3 minutes), more smoking friends (0.46$\pm$1.03), more classes skipped (1.31$\pm$0.70), more weekly allowance (1.82$\pm$1.13), lower mean scores of school connectedness (${S}_{School Connectedness}=18.6\pm 2.81$) than SP2. There were less students eating breakfast ($N=896$ (42.4%)) than SP2. SP3 was defined as the medium-high risk profile.

SP4, the minor sub-phenotype at W_I, consisted of 986 (11.5%) students with the highest mean scores of substance use (${S}_{Cigarette}=0.25\pm 0.56$, ${S}_{E-Cigarette}=0.35\pm 0.65$, ${S}_{Alcohol}=0.71\pm 0.82$, ${S}_{Marijuana}=0.29\pm 0.63$). Students in this sub-phenotype had the highest mean scores of CESD (${S}_{CESD}=10.8\pm 6.70)$, longest sedentary time (1012$\pm$269 minutes), largest number of smoking friends (0.46$\pm$1.03), largest number of classes skipped (1.31$\pm$0.70), largest weekly allowance (1.82$\pm$1.13), lowest mean scores of school connectedness (${S}_{School Connectedness}=18.6\pm 2.81$) among the four sub-phenotypes. Only 315 (31.9%) students in SP4 reported eating breakfast. SP4 was considered the high-risk profile.

Higher scores on substance use measures indicate more frequent use of the corresponding substance. At W_I, the magnitudes comparing SP4 vs. SP1 were 4.83 (marijuana), 4.17 (cigarette), 2.5 (e-cigarette), and 1.82 times (alcohol). The longitudinal evidence showed that, in general, the scores of these indicators increased across the three-wave, indicating that with time, students engaged in higher use of each substance across the risk profiles (SP1 through SP4). Note that from W_I◊W_III (Wave III, 2018/19), the incremental rate of marijuana consumption in SP1 was the largest (6.17) among the four substances, followed by e-cigarettes (4.71), cigarettes (3.5), and alcohol (1.64).

It is observed that similar trends of risk profiling appear throughout the three-wave, showing consistent risk profiles over time (Table 2). By examining the longitudinal prevalence, we found that the number of students in SP1 increased from W_I◊W_III. In contrast, the number of students decreased with time in SP3. Interestingly, the prevalence of students in SP2 increased W_I◊W_II (Wave II, 2017/18) and decreased by W_III.

Table 2

**Sub-phenotypes (SP1-SP4) of risk profiles with associated characteristics.** Mean (SD) of the average group scores of substance use measures and the risk factors; count (%) of grades and the “eating breakfast” factor. **Substance use measures**: cigarette, e-cigarette, alcohol, and marijuana, were coded as 0 = ”Never use,” 1 = ”Occasional use,” and 2 = ”Current use.” Important factors concerning group comparison (e.g., SP4 vs. SP1) are highlighted in bold font. A. Wave I (2016/17), B. Wave II (2017/18), C. Wave III (2018/19).
A. Wave I (2016/17)
	SP1 (Low risk)	SP2 (Medium-low risk)	SP3 (Medium-high risk)	SP4 (High risk)
	N = 2799 (32·5%)	N = 2712 (31·5%)	N = 2113 (24·5%)	N = 986 (11·5%)
Cigarette (SD)	0·06 (0·29)	0·11 (0·38)	0·17 (0·47)	0·25 (0·56)
E-Cigarette (SD)	0·14 (0·44)	0·22 (0·55)	0·31 (0·63)	0·35 (0·65)
Alcohol (SD)	0·39 (0·68)	0·57 (0·77)	0·65 (0·82)	0·71 (0·82)
Marijuana (SD)	0·06 (0·30)	0·12 (0·40)	0·21 (0·54)	0·29 (0·63)
# of Smoking Friends (SD)	0·19 (0·65)	0·31 (0·81)	0·46 (1·03)	0·64 (1·21)
# of Classes Skipped (SD)	1·20 (0·58)	1·23 (0·59)	1·31 (0·70)	1·38 (0·81)
Weekly Allowance (SD)	1·70 (1·14)	1·81 (1·11)	1·82 (1·13)	1·90 (1·15)
Sedentary Time (minutes, SD)	170 (61·3)	345 (50·7)	563 (83·3)	1012 (269)
CESD (SD)	7·08 (5·39)	8·34 (5·89)	9·30 (6·04)	10·8 (6·70)
School Connectedness (SD)	19·8 (2·78)	19·1 (2·66)	18·6 (2·81)	18·0 (3·35)
Eating Breakfast	1831 (65·4%)	1400 (51·6%)	896 (42·4%)	315 (31·9%)
Grade 7	370 (13·2%)	170 (6·27%)	107 (5·06%)	33 (3·35%)
Grade 8	299 (10·7%)	176 (6·49%)	100 (4·73%)	44 (4·46%)
Grade 9	1377 (49·2%)	1430 (52·7%)	1088 (51·5%)	486 (49·3%)
Grade 10	753 (26·9%)	936 (34·5%)	818 (38·7%)	423 (42·9%)
B. Wave II (2017/18)
	SP1 (Low risk)	SP2 (Medium-low risk)	SP3 (Medium-high risk)		SP4 (High risk)
	N = 3022 (34·7%)	N = 3082 (35·4%)	N = 1920 (22·0%)		N = 684 (7·9%)
Cigarette (SD)	0·14 (0·43)	0·22 (0·52)	0·29 (0·60)		0·45 (0·72)
E-Cigarette (SD)	0·35 (0·68)	0·52 (0·80)	0·61 (0·83)		0·67 (0·85)
Alcohol (SD)	0·73 (0·80)	0·91 (0·84)	0·96 (0·84)		1·05 (0·86)
Marijuana (SD)	0·17 (0·48)	0·33 (0·64)	0·43 (0·72)		0·57 (0·78)
# of Smoking Friends (SD)	0·28 (0·79)	0·44 (1·01)	0·64 (1·21)		0·94 (1·47)
# of Classes Skipped (SD)	1·28 (0·65)	1·37 (0·76)	1·40 (0·75)		1·62 (1·06)
Weekly Allowance (SD)	2·06 (1·31)	2·17 (1·32)	2·24 (1·28)		2·24 (1·31)
Sedentary Time (minutes, SD)	193 (64·2)	385 (60·0)	647 (101)		1176 (308)
CESD (SD)	7·31 (5·54)	8·33 (5·80)	9·37 (5·96)		11·5 (7·07)
School Connectedness (SD)	19·4 (2·97)	18·7 (3·01)	18·0 (3·15)		17·3 (3·54)
Eating Breakfast	1865 (61·7%)	1443 (46·8%)	718 (37·4%)		201 (29·4%)
Grade 8	336 (11·1%)	231 (7·50%)	93 (4·84%)		25 (3·65%)
Grade 9	302 (9·99%)	207 (6·72%)	87 (4·53%)		24 (3·51%)
Grade 10	1415 (46·8%)	1578 (51·2%)	1035 (53·9%)		390 (57·0%)
Grade 11	969 (32·1%)	1066 (34·6%)	705 (36·7%)		245 (35·8%)
C. Wave III (2018/19)
	SP1 (Low risk)	SP2 (Medium-low risk)	SP3 (Medium-high risk)		SP4 (High risk)
	N = 3385 (38·9%)	N = 2847 (32·7%)	N = 1776 (20·4%)		N = 686 (7·9%)
Cigarette (SD)	0·21 (0·50)	0·30 (0·59)	0·40 (0·66)		0·53 (0·74)
E-Cigarette (SD)	0·66 (0·84)	0·87 (0·90)	0·95 (0·91)		1·05 (0·90)
Alcohol (SD)	1·03 (0·83)	1·14 (0·82)	1·16 (0·82)		1·19 (0·82)
Marijuana (SD)	0·37 (0·66)	0·53 (0·75)	0·68 (0·82)		0·83 (0·87)
# of Smoking Friends (SD)	0·32 (0·86)	0·46 (0·99)	0·65 (1·19)		0·94 (1·50)
# of Classes Skipped (SD)	1·47 (0·83)	1·56 (0·88)	1·69 (1·03)		1·87 (1·21)
Weekly Allowance (SD)	2·47 (1·38)	2·57 (1·36)	2·58 (1·33)		2·50 (1·32)
Sedentary Time (minutes, SD)	209 (68·0)	403 (58·2)	653 (94·2)		1191 (347)
CESD (SD)	8·02 (5·57)	9·16 (5·95)	10·0 (6·15)		12·2 (6·87)
School Connectedness (SD)	18·9 (3·08)	18·4 (3·04)	17·7 (3·32)		17·0 (3·81)
Eating Breakfast	1878 (55·5%)	1195 (42·0%)	593 (33·4%)		170 (24·8%)
Grade 9	363 (10·7%)	202 (7·10%)	99 (5·57%)		20 (2·92%)
Grade 10	346 (10·2%)	179 (6·29%)	75 (4·22%)		22 (3·21%)
Grade 11	1611 (47·6%)	1441 (50·6%)	976 (55·0%)		392 (57·1%)
Grade 12	1065 (31·5%)	1025 (36·0%)	626 (35·2%)		252 (36·7%)

Clustering validity

The internal measures of clustering performance measured by the average silhouette width ranged from 0.51 to 0.55 across three waves with different clustering algorithms. Comparing the external estimates amongst the clustering algorithms experimented with, the FANNY and PAM clustering algorithms achieved the highest degree of agreement on cluster membership, as opposed to the comparison with hierarchical clustering (Table 3). Further assessment of the average silhouette width showed that the value for the FANNY algorithm was slightly higher (SFANNY = 0.52) than that of the PAM algorithm (SPAM = 0.51). The results showed that FANNY is a robust clustering method on the linked COMPASS data for reporting sub-phenotypes of risk profiles.

Table 3

**Comparison of clustering validity for each pair of clustering algorithms. ARI**: adjusted Rand index (the range of ARI is between − 1 and 1, representing no agreement to perfect agreement); VI: variation index; **FANNY**: Fuzzy ANalYsis; **PAM**: Partitioning Around Medoid.
Clustering Algorithm	Index	Wave I (2016/17)	Wave II (2017/18)	Wave III (2018/19)
FANNY vs PAM	ARI	0·9698	0·7676	0·6452
FANNY vs PAM	VI	0·1154	0·6023	0·7621
PAM vs Hierarchical	ARI	0·4905	0·5093	0·4898
PAM vs Hierarchical	VI	0·8621	0·9181	0·9315
FANNY vs Hierarchical	ARI	0·4736	0·5449	0·6651
FANNY vs Hierarchical	VI	0·9241	0·9106	0·7029
A. Wave I (2016/17)
B. Wave II (2017/18)
C. Wave III (2018/19)
A. Wave I (2016/17)
B. Wave II (2017/18)
C. Wave III (2018/19)

The silhouette values for each sub-phenotype ranged from S_WI=[0.31, 0.57], S_WII=[0.35, 0.58], and S_WIII=[0.30, 0.56] (Fig. 2). The average silhouette widths were S_WI=0.52, S_WII=0.53, and S_WIII=0.53. Except for the minority sub-phenotype (SP4) with the lowest silhouette value, the average silhouette width for all the other three sub-phenotypes was positive. Kaufman et al. (2009) proposed that the silhouette values between 0.71 and 1 indicate a strong structure for that cluster³⁷. The reasonable cluster had a silhouette value between 0.51 and 0.70 and a weak structure value below 0.50³⁷. The two clusters with the second largest and largest sample size had reasonable structures, whereas the other two with the smallest and the second smallest sample size had weak structures at W_I and W_II. As for W_III, three clusters had reasonable structures, and only one cluster with the smallest sample size had a weak structure. These clustering results support the applicability of sub-phenotypes of risk profiles.

This study breaks new ground by utilizing cluster analysis to identify sub-phenotypes of behavioral risk profiles related to youth substance use on population-based longitudinal health surveys. Traditional studies in this field have focused on identifying individual risk factors associated with substance use among youth and investigating the statistical power of each factor. Some studies have also used clustering algorithms to identify patterns of substance use. However, no existing studies have assessed risk profile phenotypes of youth substance use involving substance use indicators and associated health risk behaviors with a data-driven approach. As the first study to ascertain sub-phenotypes of risk profiles among Canadian youth, our research sheds light on risk profiling with cross-sectional and longitudinal evidence. Our findings are consistent with previous studies that employ statistical modelling approaches in determining risk factors^{16,17,18,19,31,36,38}.

Our study results showed that the number of smoking friends, number of classes skipped, and weekly allowance were consistently among the top factors associated with substance use among Canadian youth. These findings are consistent with previous studies that have identified peer influence, academic problems, and access to money as key risk factors for substance use among young people^39,40. Studies have shown that peers play a significant role in youth substance use. Adolescents with more friends who use substances are more likely to use substances themselves³⁹. This can be attributed to peer pressure, social norms, and the influence of social networks on behavior⁴⁰. In the context of smoking, one study found that the number of smoking friends was a strong predictor of adolescent smoking⁴¹. Skipping classes is also associated with increased substance use among youth. Studies have shown that students who skip classes are more likely to use substances, including alcohol and marijuana^42,43. This could be due to a lack of supervision, boredom, or increased opportunities to use substances when not in school. Finally, the association between weekly allowance and substance use is complex. While some studies have found that higher levels of allowance are associated with increased substance use⁴⁴, others have found no significant association⁴⁵. It is possible that the relationship between allowance and substance use may be mediated by other factors, such as peer influence or parental monitoring.

Our finding agrees with Halladay et al. (2020), who revealed an association between substance use and mental health⁴⁶. Many studies have found that individuals in the multi-use group report higher psychiatric symptoms, including depression and anxiety than in the single-use group^23,24. Radloff (1977) states that a CESD score ≥ 10 represents clinically relevant depressive symptoms⁴⁷. In our study at W_I, students in SP4 had a CESD score of 10.8 ± 6.70, suggesting that, on average, this group of individuals already experienced clinically relevant depressive symptoms. In contrast, those in SP1 had the smallest CESD score (7.08 ± 5.39). On average, individuals in SP3 started to have clinically relevant depressive symptoms at W_III. This was determined by their average CESD score of 10.0 ± 6.15.

A sedentary lifestyle was associated with a high risk of substance use. Our finding agrees with West et al. (2020) that sedentary behavior was positively associated with adolescent drinking and marijuana consumption⁴⁸. At W_I, the difference between the students who were in SP1 had, on average, the smallest sedentary time (170 ± 61.3 minutes) compared to their peers in SP4 (1012 ± 269) by a magnitude of 5.95 times. Comparing SP4 vs. SP1 for the last two years, the sedentary time differences were similar at WI, 6.09 and 5.70 times for W_II and W_III, respectively.

Similarly, eating breakfast was associated with a low risk of substance use among Canadian youth. Our result is consistent with the literature about youth substance use and the correlation of nutrition-related attitudes, e.g., substance users were more likely to be at greater risk of poor eating habits, including not eating breakfast²⁶. The prevalence of eating breakfast decreased while the risk level increased from SP1 to SP4. The longitudinal evidence suggests that the prevalence of the students eating breakfast decreased across three years in the SP1 group. The same trend was observed across three years among the other risk profile groups (SP2 to SP4).

Utilizing the top factors correlated with youth substance use, this study identifies four sub-phenotypes of risk profiles, SP1 through SP4, with escalating risk by groups. These sub-phenotypes provide a more comprehensive overview and quantification of the prominent characteristics for each risk level of engaging in substance use among youth. Note that marijuana consumption had increased by 3.98 times from 4.0% at W_I to 15.9% at W_III. The highest magnitude of marijuana use amongst other substances may be partly due to the legalization of the recreational use of marijuana in October 2018.

This study has implications from public health and health policy perspectives. Our findings suggest that the correlates of youth substance use are multifaceted, concerning individual-level factors, peer influence, and environmental effects. Understanding risk profiles related to youth substance use will help school program managers/policymakers identify and characterize valuable measures to evaluate risk reduction interventions. From the population level, differences between sub-phenotypes of risk profiles may have essential effects on subjective youth behavioral and mental health. This can be further conceptualized by having different preventive capabilities against addictive behaviors in school settings. The diverse associations between youth substance use and multifaceted health-related behaviors should be considered for decision-makers who want to invest in risk reduction interventions targeting multiple health risk behaviors among youth.

Cluster analysis is a valuable tool for this study, identifying hidden patterns of risk profile related to youth substance use. This approach enabled the identification of subgroups of individuals with similar patterns of risk factors, providing a more nuanced understanding of the complex factors associated with substance use. By understanding the unique risk profiles of different subgroups of adolescents, interventions can be tailored to address the specific needs of each group, leading to more effective prevention and intervention strategies. For example, a prevention program that addresses peer influence may be particularly effective for a subgroup of youth who are heavily influenced by their smoking friends, while a program that focuses on academic support may be more effective for a sub-cohort who are skipping classes and struggling with their schoolwork.

Cluster analysis has advantages and potential limitations; for example, the various clustering algorithms usually provide very different results due to the distinct criteria for merging clusters. Although cluster analysis has unique advantages for revealing “hidden” patterns and unexpected associations in variables, no backward option can be made in earlier steps due to the hierarchical nature of the analysis. Therefore, we implemented various clustering algorithms to mitigate these limitations, including fuzzy clustering, partitioning-based, and hierarchical-based. The last two types of clustering methods are hard clustering algorithms that assign data elements to one cluster. Unlike hard clustering, fuzzy clustering algorithms are soft-clustering methods, assigning membership coefficients of objects to all clusters. A focus was given to fuzzy clustering algorithms, considering the overlapping nature of risk profiles observed on the longitudinal samples of COMPASS data. Our study results show that fuzzy clustering outperforms hierarchical and partitioning clustering on COMPASS data, demonstrating its appropriateness for unearthing the overlapping nature of risk profile phenotypes related to youth substance use. Furthermore, we performed data visualization via the t-SNE algorithm to project high-dimensional data to lower-dimensional space. In addition, silhouette plots demonstrate the internal index of clustering validation, and different risk levels with associated characteristics represent risk profiling.

The major strengths of this study derive from COMPASS data and methodologies. COMPASS data are high-dimensional population-level health surveys with large sample sizes and reliable data quality, using national surveillance instruments-based measurements⁴⁹. The Cq employs the active-information passive-consent protocols, which help achieve high participation rates and reduce sampling bias while preserving student confidentiality⁵⁰. COMPASS methodologies are sufficiently robust given the delicate balance of data accuracy and participant anonymity in longitudinal studies concerning youth health behaviors⁵¹. The multiple sources of COMPASS data bring new insights into strengthening the cross-sectional and longitudinal evidence of risk profiling. Another strength is the applied methods, undertaking comprehensive data preprocessing steps and employing advanced unsupervised machine learning (ML) models such as fuzzy clustering to discover hidden patterns. The ML pipeline developed in this study can be used in real-world decision support, making a clinical tool determining different risk profile phenotypes from real-world raw data towards the final analysis results that are interpretable to stakeholders.

This study has certain limitations²². First, it is preferable to have external validation data to validate the identified clusters. However, the subsequent data collection cycles were challenging due to the COVID-19 pandemic. Second, COMPASS data are subject to self-reported bias and may lead to inaccurate measures. Third, many participating schools in the COMPASS study are convenience samples that do not cover all Canadian provinces. The sampling of schools may not be genuinely representative, limiting the results’ generalizability. Moreover, omitting participants with irregular patterns in their grades may also compromise generalizability. Future work is warranted to evaluate whether there were any systematic differences between those retained in the analyses and those excluded. Last, as with any large-scale health survey, non-responses introduce missing values. We applied various imputation techniques to impute missing values.

In conclusion, this study provides novel insights into the identification of sub-phenotypes of risk behavioral profiles related to substance use among Canadian youth using cluster analysis. Our findings have practical implications for practitioners in school settings who can use the evidence provided to develop targeted risk reduction interventions for at-risk adolescents. Furthermore, this study provides valuable information for stakeholders to guide the implementation of school policies and procedures to improve youth health behaviors. This research contributes to a better understanding of the complex relationships between substance use and associated health risk behaviors among adolescents and can inform the development of effective prevention strategies.

Data source and preprocessing

This retrospective cohort study used three annual waves of COMPASS data. The study cohort consisted of 9307 Canadian youth in grades 9–10 from 76 secondary schools across four provinces in 2016/17. The participating students were followed up for three years until they reached grades 11–12 in 2018/19. Using the Cq, the study assessed substance use measures including cigarette, e-cigarette, alcohol, and marijuana use. In addition to substance use, participating students’ demographic, personal information, and responses to their health behaviors and school-level data were available for analysis. The study obtained informed parental/guardian consent for all student participants. The COMPASS host study has received ethics clearance from the University of Waterloo Office of Research Ethics (ORE 30118). Detailed information about data sources and preprocessing steps are published in another paper²².

Cluster analysis and validation

Cluster analysis, a class of multivariate techniques, has been applied to investigate addictive behaviors of substance and non-substance use⁴⁶. Clustering algorithms, primarily k-means and hierarchical clustering, have been employed to examine the patterns of youth substance use⁴⁶. Elliott et al. (2019) applied cluster-type analyses using a dendrogram to visualize the relationships between substance types. The results show significant distinctions in the use patterns of substances and polysubstance between States in the U.S. Interventions to address clinical challenges of multi-use of substances are essential for individuals who engage in concurrent use of other substances with binge drinking⁵². Implementing hierarchical and k-medoids clustering algorithms, Sun et al. (2012) identified five homogenous inherited patterns of opioid use, varying by use levels, onset time, and comorbidity⁵³.

This study implemented various clustering algorithms, including fuzzy clustering as the primary focus and PAM and hierarchical clustering for validation. A well-known fuzzy clustering algorithm, FANNY (Fuzzy ANalYsis)³⁷ was applied. FANNY minimizes the objective function

$$\sum _{v=1}^{k}\sum _{i,j=1}^{n}\frac{{\text{u}(\text{i},\text{v})}^{\text{r}} {\text{u}(\text{j},\text{v})}^{\text{r}} \text{d}(\text{i},\text{j}))}{\left(2 \sum _{j}{\text{u}(\text{j},\text{v})}^{\text{r}}\right)}$$

where $n$ is the sample size, $k$ is the number of clusters, $r$ is the membership exponent, and $d(i,j)$ is the dissimilarity between subjects $i$ and $j$. Increasingly crisper clustering can be achieved when $r$ is close to 1 and complete fuzziness when $r$ approaches infinity³⁷. Further note that even the default value, $r$ = 2, will lead to complete fuzziness, that is, the degree of membership $u(i,v) = \frac{1}{k}$. FANNY is more robust to non-spherical clusters than other fuzzy clustering algorithms, accepting a proximity matrix $d(i,j)$ instead of estimating central values, i.e., Euclidean distances as in the FCM algorithm⁵⁴.

The optimal number of clusters was determined by implementing multiple validation indices available in the NbClust package⁵⁵. An automatic voting scheme by implementing all available indices was applied in this study to avoid bias towards a specific criterion, such as the silhouette index, or GAP statistics alone, as each stopping criterion favors one validation method. Across three waves, most of the clustering validity indices voted for four clusters, which were selected for the study cohort to make the clustering results consistent and easier for interpretation. Although the computational complexity is much higher than a single index, this method provides an unbiased approach to selecting the most appropriate clusters on COMPASS data.

Given the mixed type of COMPASS data and most variables are categorical, different dissimilarity matrices have experimented with, and the clustering results were compared. Firstly, as part of the clustering process, grade-of-membership (GOM) analysis assigns two or more latent subgroups for each object based on the probabilities of their cluster membership. One specific distance metric, the Gower distance, was implemented. Gower distance is calculated as the average partial dissimilarities between data elements, depending on the evaluated variable types⁵⁶. Each partial dissimilarity ranges between 0 and 1. Secondly, the ordinal data were treated as continuous, so the commonly used Euclidean distance measure can be implemented. The Euclidean distance measure is appealing because the distance between two objects can be interpreted as physical distance obtained from multivariate used for clustering. The optimal cluster analysis was proposed from the results obtained from all cluster combinations, distance measures, and clustering algorithms.

Two types of validity indices, i.e., external and internal measures for clustering validation, were used to assess clustering results. External indices use the adjusted Rand index (ARI) and Melia’s variation index (VI) to measure the consistency between the partition clusters and the external reference⁵⁷. Among internal indices, the silhouette coefficient is the most common measure for evaluating clustering results, assessing the applicability of assigning a subject to one cluster rather than another, and considering cluster compactness and separation⁵⁸. Besides the total silhouette value, the silhouette coefficient for each cluster was also calculated.

The Boruta algorithm⁵⁹ was applied to obtain the most representative subset of characteristics for clustering. Each variable importance score was summed up across three waves. The total score was sorted from largest to smallest. The same procedure of feature selection was repeated for each wave. Preliminary clustering results demonstrated that the number of features around eight leads to the most meaningful solutions. Including the four substance use indicators, all variables were standardized to ensure equal weighting before clustering.

We employed t-Distributed Stochastic Neighbor Embedding (t-SNE)⁶⁰ for data visualization. An essential step to using the t-SNE algorithm effectively is to tune hyperparameters, such as perplexity and the number of iterations. As a smooth measure of the adequate number of neighbors, Van der Maaten suggests setting the typical range of perplexity values between 5 and 50 to achieve a relatively robust performance of t-SNE⁶⁰. For a larger dataset, the rule of thumb is to select a larger perplexity value⁶¹. Oskolkov proposed the optimal perplexity approximates $\sqrt[2]{N}$, which is analytically derived from the power law, given that t-SNE is based on the minimization of Kullback-Leibler (KL) divergence⁶¹. In this study, a range of perplexity was experimented with from 5 to 100. Perplexity = 100, which approximates $\sqrt[2]{N}$, is the optimal value for the linked COMPASS data.

For consistency and robustness, results from each applied clustering method were compared with the indices discussed in this section. Data analyses were performed using the R language, open-source software for statistical computing and graphics⁶².

Burke, P. J., O'Sullivan, J., & Vaughan, B. L. (2005). Adolescent substance use: brief interventions by emergency care providers. Pediatric emergency care, 21(11), 770–776.
Johnston, L., Miech, R., O'Malley, P., Bachman, J., Schulenberg, J., & Patrick, M. (2020). Monitoring the Future national survey results on drug use, 1975–2019: Overview, key findings on adolescent drug use.
Chen, P., & Jacobson, K. C. (2012). Developmental trajectories of substance use from early adolescence to young adulthood: Gender and racial/ethnic differences. Journal of adolescent health, 50(2), 154–163.
Kandel, D. B. (Ed.). (2002). Stages and pathways of drug involvement: Examining the gateway hypothesis. Cambridge University Press.
Ellickson, P. L., Tucker, J. S., & Klein, D. J. (2003). Ten-year prospective study of public health problems associated with early drinking. Pediatrics, 111(5), 949–955.
Brown, S. A., McGue, M., Maggs, J., Schulenberg, J., Hingson, R., Swartzwelder, S., … Murphy, S. (2008). A developmental perspective on alcohol and youths 16 to 20 years of age. Pediatrics, 121(Supplement_4), S290-S310.
Grant, B. F., & Dawson, D. A. (1997). Age at onset of alcohol use and its association with DSM-IV alcohol abuse and dependence: results from the National Longitudinal Alcohol Epidemiologic Survey. Journal of substance abuse, 9, 103–110.
O’Malley, P. M., Johnston, L. D., & Bachman, J. G. (1998). Alcohol use among adolescents. Alcohol health and research world, 22(2), 85.
Dahl, R. E. (2004). Adolescent brain development: a period of vulnerabilities and opportunities. Keynote address. Annals of the New York Academy of Sciences, 1021(1), 1–22.
Gardner, M., & Steinberg, L. (2005). Peer influence on risk taking, risk preference, and risky decision making in adolescence and adulthood: an experimental study. Developmental psychology, 41(4), 625.
Steinberg, L. (2005). Cognitive and affective development in adolescence. Trends in cognitive sciences, 9(2), 69–74.
Romer, D. (2010). Adolescent risk taking, impulsivity, and brain development: Implications for prevention. Developmental Psychobiology: The Journal of the International Society for Developmental Psychobiology, 52(3), 263–276.
King, K. M., Meehan, B. T., Trim, R. S., & Chassin, L. (2006). Marker or mediator? The effects of adolescent substance use on young adult educational attainment. Addiction, 101(12), 1730–1740.
National Institute on Drug Abuse. (2019). Substance Use in Women. Retrieved from https://www.drugabuse.gov/publications/research-reports/substance-use-in-women/sex-gender-differences-in-substance-use.
Canadian Centre on Substance Use and Addiction. (2018). Substance use among middle and high school students in Canada. Retrieved from https://www.ccsa.ca/
Zuckermann, A. M., Williams, G. C., Battista, K., Jiang, Y., de Groh, M., & Leatherdale, S. T. (2020). Prevalence and correlates of youth poly-substance use in the COMPASS study. Addictive behaviors, 107, 106400.
Banks, D. E., Rowe, A. T., Mpofu, P., & Zapolski, T. C. (2017). Trends in typologies of concurrent alcohol, marijuana, and cigarette use among US adolescents: An ecological examination by sex and race/ethnicity. Drug and alcohol dependence, 179, 71–77.
Silveira, M. L., Green, V. R., Iannaccone, R., Kimmel, H. L., & Conway, K. P. (2019). Patterns and correlates of polysubstance use among US youth aged 15–17 years: wave 1 of the Population Assessment of Tobacco and Health (PATH) Study. Addiction, 114(5), 907–916.
Rose, R. A., Evans, C. B., Smokowski, P. R., Howard, M. O., & Stalker, K. L. (2018). Polysubstance use among adolescents in a low income, rural community: Latent classes for middle-and high‐school students. The Journal of Rural Health, 34(3), 227–235.
Connell, C. M., Gilreath, T. D., & Hansen, N. B. (2009). A multiprocess latent class analysis of the co-occurrence of substance use and sexual risk behavior among adolescents. Journal of studies on alcohol and drugs, 70(6), 943–951.
Gilreath, T. D., Astor, R. A., Estrada, J. N., Johnson, R. M., Benbenishty, R., & Unger, J. B. (2014). Substance use among adolescents in California: A latent class analysis. Substance Use & Misuse, 49(1–2), 116–123.
Yang, Y., Butt, Z. A., Leatherdale, S. T., Morita, P. P., Wong, A., Rosella, L., & Chen, H. H. (2022). Exploring the dynamic transitions of polysubstance use patterns among Canadian youth using Latent Markov Models on COMPASS data. The Lancet Regional Health-Americas, 16, 100389.
Patton, G. C., Coffey, C., Carlin, J. B., Degenhardt, L., Lynskey, M., & Hall, W. (2002). Cannabis use and mental health in young people: cohort study. Bmj, 325(7374), 1195–1198.
Mathers, M., Toumbourou, J. W., Catalano, R. F., Williams, J., & Patton, G. C. (2006). Consequences of youth tobacco use: a review of prospective behavioural studies. Addiction, 101(7), 948–958.
Lesjak, V., & Stanojević-Jerković, O. (2015). Physical Activity, Sedentary Behavior and Substance use Among Adolescents in Slovenian Urban Area/Telesna Aktivnost, Oblike Sedečega Vedenja In Uživanje Psihoaktivnih Snovi Med Mladostniki V Slovenskem Urbanem Okolju. Slovenian Journal of Public Health, 54(3), 168–174.
Isralowitz, R. E., & Trostler, N. (1996). Substance use: toward an understanding of its relation to nutrition-related attitudes and behavior among Israeli high school youth. Journal of adolescent health, 19(3), 184–189.
White, J., Walton, D., & Walker, N. (2015). Exploring comorbid use of marijuana, tobacco, and alcohol among 14 to 15-year-olds: findings from a national survey on adolescent substance use. BMC public health, 15(1), 1–9.
Jongenelis, M., Pettigrew, S., Lawrence, D., & Rikkers, W. (2019). Factors associated with poly drug use in adolescents. Prevention Science, 20, 695–704.
Su, J., Supple, A. J., & Kuo, S. I. C. (2018). The role of individual and contextual factors in differentiating substance use profiles among adolescents. Substance use & misuse, 53(5), 734–743.
Pettigrew, S., Jongenelis, M., Lawrence, D., & Rikkers, W. (2017). Common and differential factors associated with abstinence and poly drug use among Australian adolescents. International Journal of Drug Policy, 50, 41–47.
[16] Tomczyk, S., Isensee, B., & Hanewinkel, R. (2016). Latent classes of polysubstance use among adolescents—a systematic review. Drug and alcohol dependence, 160, 12–29.
[17] Valente, J. Y., Cogo-Moreira, H., & Sanchez, Z. M. (2017). Gradient of association between parenting styles and patterns of drug use in adolescence: A latent class analysis. Drug and alcohol dependence, 180, 272–278.
Tomczyk, S., Hanewinkel, R., & Isensee, B. (2015). Multiple substance use patterns in adolescents—a multilevel latent class analysis. Drug and alcohol dependence, 155, 208–214.
Lau, E. Y., Faulkner, G., Qian, W., & Leatherdale, S. T. (2016). Longitudinal associations of parental and peer influences with physical activity during adolescence: findings from the COMPASS study. Health promotion and chronic disease prevention in Canada: research, policy and practice, 36(11), 235.
Leatherdale, S. T., Brown, K. S., Carson, V., Childs, R. A., Dubin, J. A., Elliott, S. J., … Thompson-Haile, A. (2014). The COMPASS study: a longitudinal hierarchical research platform for evaluating natural experiments related to changes in school-level programs, policies and built environment resources. BMC public health, 14(1), 1–7.
Morean, M. E., Kong, G., Camenga, D. R., Cavallo, D. A., Simon, P., & Krishnan-Sarin, S. (2016). Latent class analysis of current e-cigarette and other substance use in high school students. Drug and alcohol dependence, 161, 292–297.
Kaufman, L., & Rousseeuw, P. J. (2009). Finding groups in data: an introduction to cluster analysis. John Wiley & Sons.
Choi, H. J., Lu, Y., Schulte, M., & Temple, J. R. (2018). Adolescent substance use: Latent class and transition analysis. Addictive behaviors, 77, 160–165.
Kandel, D. B., Yamaguchi, K., & Chen, K. (1992). Stages of progression in drug involvement from adolescence to adulthood: further evidence for the gateway theory. Journal of studies on alcohol, 53(5), 447–457.
Brook, J. S., Brook, D. W., Arencibia-Mireles, O., Richter, L., & Whiteman, M. (2001). Risk factors for adolescent marijuana use across cultures and across time. The Journal of genetic psychology, 162(3), 357–374.
De Vries, H., Engels, R., Kremers, S., Wetzels, J., & Mudde, A. (2003). Parents’ and friends’ smoking status as predictors of smoking onset: findings from six European countries. Health education research, 18(5), 627–636.
Kuo, M., Adlaf, E. M., Lee, H., Gliksman, L., Demers, A., & Wechsler, H. (2002). More Canadian students drink but American students drink more: comparing college alcohol use in two countries. Addiction, 97(12), 1583–1592.
Sussman, S., Skara, S., & Ames, S. L. (2008). Substance abuse among adolescents. Substance use & misuse, 43(12–13), 1802–1828.
Perkins, H. W., Haines, M. P., & Rice, R. (2005). Misperceiving the college drinking norm and related problems: a nationwide study of exposure to prevention information, perceived norms and student alcohol misuse. Journal of studies on alcohol, 66(4), 470–478.
Hingson, R., & Kenkel, D. (2004). Social, health, and economic consequences of underage drinking. Reducing underage drinking: A collective responsibility, 351–382.
Halladay, J., Woock, R., El-Khechen, H., Munn, C., MacKillop, J., Amlung, M., … Georgiades, K. (2020). Patterns of substance use among adolescents: A systematic review. Drug and alcohol dependence, 216, 108222.
Radloff, L. S. (1977). The CES-D scale: A self-report depression scale for research in the general population. Applied psychological measurement, 1(3), 385–401.
West, A. B., Bittel, K. M., Russell, M. A., Evans, M. B., Mama, S. K., & Conroy, D. E. (2020). A systematic review of physical activity, sedentary behavior, and substance use in adolescents and emerging adults. Translational behavioral medicine, 10(5), 1155–1167.
Elton-Marshall, T., Leatherdale, S. T., Manske, S. R., & Wong, K. (2011). Research methods of the youth smoking survey (YSS). Chronic diseases and injuries in Canada, 32(1).
Thompson-Haile, A., & Leatherdale, S. T. (2013). Student-level data collection procedures. University of Waterloo: Waterloo, ON, Canada.
Battista, K., Qian, W., Bredin, C., & Leatherdale, S. T. (2019). Student Data Linkage over Multiple Years: Compass Technical Report Series. Waterloo, ON, USA.
Elliott, L., Haddock, C. K., Campos, S., & Benoit, E. (2019). Polysubstance use patterns and novel synthetics: A cluster analysis from three US cities. PLoS One, 14(12), e0225273.
Sun, J., Bi, J., Chan, G., Oslin, D., Farrer, L., Gelernter, J., & Kranzler, H. R. (2012). Improved methods to identify stable, highly heritable subtypes of opioid use and related behaviors. Addictive behaviors, 37(10), 1138–1144.
Landau, S., Leese, M., Stahl, D., & Everitt, B. S. (2011). Cluster analysis. John Wiley & Sons.
Charrad, M., Ghazzali, N., Boiteau, V., & Niknafs, A. (2014). NbClust: an R package for determining the relevant number of clusters in a data set. Journal of statistical software, 61, 1–36.
Gower, J. C., & Legendre, P. (1986). Metric and Euclidean properties of dissimilarity coefficients. Journal of classification, 3, 5–48.
Steinley, D. (2004). Properties of the hubert-arable adjusted rand index. Psychological methods, 9(3), 386.
Rousseeuw, P. J. (1987). Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of computational and applied mathematics, 20, 53–65.
Kursa, M. B., & Rudnicki, W. R. (2010). Feature selection with the Boruta package. Journal of statistical software, 36, 1–13.
Van der Maaten, L., & Hinton, G. (2008). Visualizing data using t-SNE. Journal of machine learning research, 9(11).
Oskolkov, N. (2019). How to tune hyperparameters of tSNE. Towards Data Science, US. Retrieved from https://towardsdatascience.com/how-to-tune-hyperparameters-of-tsne-7c0596a18868
R Core Team, R. (2013). R: A language and environment for statistical computing.

There is NO Competing Interest.

Download PDF

Version 1

posted

You are reading this latest preprint version

Uncovering Hidden Risk Profile Phenotypes of Substance Use Among Canadian Adolescents via Cluster Analysis of COMPASS Data

Status:

Version 1

Abstract

Figures

Introduction

Results

Descriptive statistics

Factors associated with youth substance use

Characteristics of sub-phenotypes

Clustering validity

Discussion

Methods

Data source and preprocessing

Cluster analysis and validation

References

Additional Declarations

Status:

Version 1