Predictors of Substance Use Initiation by Early Adolescence

Abstract

Objective:

Substance use initiation during early adolescence is associated with later development of substance use and mental health disorders. This study used various domains to predict substance use initiation, defined as trying any non-prescribed substance (e.g., alcohol, tobacco, cannabis), by age 12 using a large longitudinal dataset.

Methods:

Substance-naïve youths from the Adolescent Brain Cognitive Development Study (ages 9–10; N=6,829) were followed for 3 years. A total of 420 variables were examined as predictors of substance use initiation, using a penalized logistic regression with elastic net; domains spanned demographic characteristics, self and peer involvement with substance use, parenting behaviors, mental and physical health, culture and environment, hormones, neurocognitive functioning, and structural neuroimaging.

Results:

By age 12, 982 (14.4%) of children reported substance initiation, with alcohol being the most common. Models with only self-report predictors had similar prediction performance to models adding hormones, neurocognitive factors, and neuroimaging predictors (AUC_test=0.66). Sociodemographic factors were the most robust predictors, followed by cultural and environmental factors, physical health factors, and parenting behaviors. The top predictor was a religious preference of Mormon (coefficient=−0.87), followed by a religious preference for Jewish (coefficient=0.32), and by Black youths (coefficient=−0.32).

Conclusions:

Sociodemographic variables were the most robust predictors of substance use initiation. Adding resource-intensive measures, including hormones, neurocognitive assessment, and structural neuroimaging, did not improve prediction of substance use initiation. The application of these large-scale findings in clinical settings could help to streamline and tailor prevention and early intervention efforts.

Introduction

Substance use initiation often begins during adolescence. In 2022, 23% of 13-year-olds reported lifetime use of alcohol, followed by vaping nicotine (17%) and vaping cannabis (8%) [1]. Younger age at substance use initiation is associated with a more rapid transition to substance use disorders (SUDs) [2] and higher rates of psychiatric disorders [3, 4]. The harmful neurobiological, cognitive, and behavioral effects of early substance use, coupled with greater vulnerability for SUDs and psychiatric disorders in adulthood, highlights the need to identify risk for and protective factors of early substance use to inform prevention and early intervention efforts.

Sociodemographic, cultural, and relationship factors have individually been examined as predictors of substance use initiation among adolescents. Biological sex and income are the most commonly examined sociodemographic variables, and their effects on substance use initiation have been mixed [5–8]. Racial and ethnic differences have been observed in the rates and types of substances initiated, with some racial and ethnic groups showing both lower and higher rates of substance use, depending on the type of substance [9]; related factors, such as perception of racism [10] and neighborhood disadvantage [11], may better explain some of these relationships. Other cultural and relationship factors may exert a protective effect, such as familism among Hispanic youths [12]. Conversely, lower quality of parent-child relationships [13, 14], deviant peer associations [15], peer substance use [16], and positive expectations of substance use [17] appear to increase risk for substance use initiation. Comprehensive assessment of sociodemographic factors beyond those commonly collected (e.g., religion) may also be important to understand in a clinical setting.

Various mental health and physical health factors have also been examined in relation to substance use initiation. Higher externalizing symptoms, reward sensitivity, sensation seeking, and impulsivity have been associated with initiation of substance use [18–20]. Findings for internalizing symptoms are mixed and may depend on contextual factors, such as mental healthcare accessibility in rural areas [20, 21]. Among physical health factors, more physical activity has been associated with increased alcohol use and decreased nicotine use [22], indicating substance-specific effects. Other important health factors found in the literature include sleep impairments prior to alcohol use initiation and prenatal exposure to alcohol, which prospectively predict future substance use [23, 24]. Together, a wide range of mental and physical health variables may interact in a multitude of ways to influence early substance use initiation.

Finally, neurobiological and neurocognitive indices have been a focus of adolescent substance use research. Increasing evidence indicates preexisting alterations in brain structure and function, and impaired cognition, may predate substance use initiation and escalation. Lower gray matter volume in areas of the brain involved in reward processing and cognitive control, as well as reduced white matter integrity, have been shown to predict earlier alcohol and cannabis use initiation [25–27]. Similarly, deficits in cognitive domains, including inhibition and working memory, have been shown to predict substance use initiation and escalation [27]. However, these findings are mixed, and the field has yet to identify specific neural regions and cognitive domains that consistently and reliably predict substance use initiation. Hence, the clinical application of neurobiological and neurocognitive indices is unclear.

These mixed results underscore complex associations between factors at the individual, microsystem, and macrosystem levels, which all interact and contribute to the risk for substance use initiation. Dynamic associations between multiple predictors are often ignored with typical research designs in traditional statistical frameworks, creating the need for novel statistical approaches that account for potential multicollinearity and noise predictors. For example, penalized regression approaches reduce the potential for overfitting and can be used as a variable selection approach to build more parsimonious models [28, 29].

Our aim in the present study was to use sophisticated statistical approaches to examine predictors of substance use initiation in early adolescence across multiple domains, using data from the Adolescent Brain Cognitive Development Study (ABCD), a multisite longitudinal study following 9- and 10-year-olds into early adulthood. The ABCD study includes multiple domains of assessment, and we selected predictors from nearly all domains assessed. Specific constructs were included that have exhibited potential clinical relevance to substance use initiation. While previous studies have examined predictors of substance use initiation in the ABCD dataset [8, 30], they focused primarily on self-report measures. In the present study, we used a hierarchical model building approach with blocks of variables entered on the basis how resource-intensive they were to collect, specifically (1) self-report measures, (2) hormones and neurocognitive factors, and (3) neuroimaging. All participants included in the analyses were substance-naïve at baseline (ages 9–10). Substance use initiation was defined as trying any substance (e.g., alcohol, tobacco, cannabis, non-prescribed medications) across a 3-year follow-up period to ages ~12–13.

Methods

Participants

Participants are from the ABCD study (N=11,868) and were recruited between ages 9 and 10 from 21 sites across the United States. We used data from the ABCD 5.1 data release. Inclusion criteria for this analysis were no lifetime substance use at baseline, including experimentation (trying a sip of alcohol, a puff of cigarette or vape pen, etc.); completion of the 3-year follow-up visit and being age 12 or older at that time; sufficient data to determine substance use; and structural and diffusion tensor imaging (DTI) neuroimaging variables that met the ABCD quality control inclusion criteria and a majority of variables present for the gray and white matter metrics selected at baseline (for additional detail on inclusion criteria, see Table S1 in the online supplement). Of the 11,868 youths enrolled in the study at baseline, 6,829 met our inclusion criteria. Figure 1 is a flowchart illustrating how the final sample size was determined and providing a breakdown of substance use at each follow-up time point.

Figure 1.

Flowchart for inclusion criteria and substance use rates at follow-up in a study of predictors of substance use initiation by early adolescence ^a

a See the online supplement for additional details on selected variables that were included for use of nicotine, cannabis, and other substances. DTI=diffusion tensor imaging; GHB=gamma-hydroxybutyric acid; MDMA=3,4-methylenedioxymethamphetamine; NA=not administered; OTC=over-the-counter.

b Participants were missing gating criteria of “heard of” for at least one substance at baseline only or at baseline and all follow-up time points.

c Alcohol use outside of a religious context. This item was not administered at the 30-month follow-up.

d Missing selected age (and selected substance use variables) from longitudinal substance use questionnaire used in the present study at the 3-year follow-up.

Measures

A total of 420 predictors collected at baseline were considered, including a combination of youth-reported and parent-reported items. Domains and questionnaires in each domain were chosen based on maximizing use of data collected within ABCD and on the preexisting literature. For additional detail calculating summary scores, subscales, and references for each questionnaire, see Table S3–S12 in the online supplement. Below is a broad overview of variables in each category and number of variables, accounting for dummy codes with categorical variables.

Demographic characteristics (36 variables)

For an overview of demographics, physical health, and mental health variables collected at baseline, see [31]. Demographic variables consisted of age at baseline, sex, race, ethnicity, parental income, child religious preference, and parent education.

Self and peer involvement with substance use (eight variables).

An overview of substance use assessment methods may be found in [32]. Youths completed the Peer Group Deviance Questionnaire, assessing peer use of alcohol, nicotine, cannabis, other substances, and peers who have problems with substances. Youths also completed the Intent to Use questionnaire, which assessed intent to use alcohol, nicotine, and cannabis.

Parenting behaviors (five variables).

Parents completed the Community Risk and Protective Factor survey, capturing substance use availability, and Parent Rules questionnaire capturing family rules for alcohol, nicotine, and cannabis. Youth report of degree of acceptance by parents was captured by the Child’s Report of Parental and Behavioral Inventory acceptance subscale, and perception of being monitored by a parent was measured with the Parental Monitoring survey. The degree of conflict was captured by the conflict subscale of the Family Environment Scale.

Mental health (22 variables).

Parents completed the Child Behavior Checklist (CBCL) to assess youth psychopathology. The present study included precalculated raw scores for the 11 subscales. Parents also completed the Family History Assessment to capture a family history of alcohol or other substance use (combined into one variable). Psychosis symptoms were captured by total number of positive symptoms endorsed by youths who completed Prodromal Psychosis Scale (PPS). Five facets of impulsivity were captured by the UPPS-P, and four subscales of inhibition and reward seeking were assessed with the Behavioral Inhibition/Behavioral Approach questionnaire.

Physical health (16 variables).

Parents completed the Sports and Activities Involvement questionnaire to capture child engagement in recreational activities. Physical activity was assessed with exercise-related questions from the Youth Risk Behavior Survey. Weekly screentime was captured by a parent-report Screen Time questionnaire. Sleep disturbances were captured by the parent-report Sleep Disturbance Scale. The Puberty Development Scale was used to assess pubertal status. Prenatal substance use exposure before and after knowledge of pregnancy was captured with a Developmental History Questionnaire and assessed for the following substances: caffeine, alcohol, tobacco, cannabis, cocaine, heroin/morphine, oxycodone, and a broad category for any other substances. Prenatal caffeine use was considered separately from a combined prenatal substance use variable that captured all other use. The Ohio State Traumatic Brain Injury Screen was completed by parents to capture previous head injury.

Culture and environment (13 variables).

For an overview of questionnaires assessing cultural and environmental factors, see [33]. The Youth Acculturation Survey measured bilingual status and Youth Neighborhood Safety/Crime Survey captured perceived safety of their neighborhood. Average grades over the past year, history of detention or suspension, and special education services were captured with background items administered as part of the Schedule for Affective Disorders and Schizophrenia for School-Aged Children based on DSM-5 criteria, completed by parents. Lastly, youths completed the School Risk and Protective Factor Survey, from which the environment, involvement, and disengagement subscales were used.

Biospecimens (two variables).

Pubertal hormone levels were collected via saliva sample, and the average level of testosterone and dehydroepiandrosterone across two replications of saliva sample processing were included in the present analyses.

Neurocognition (15 variables).

A detailed overview of the baseline neurocognitive measures included in the ABCD Study is provided elsewhere [34]. Visual processing and abstract, spatial perception were captured by the Matrix Reasoning Task as part of the WISC-V and the Little Man Task. The Rey Auditory Verbal Learning Test (RAVLT) was administered to assess short-term memory, working memory, and long-term memory. The Cash Choice Task was used to capture delay discounting. The present study included age corrected scores of the seven primary tests included in the NIH Toolbox tasks.

Structural Neuroimaging (303 variables).

Details on the scanning parameters and MRI acquisition have been published elsewhere [35]. The present study used post-processed structural data to examine gray matter cortical measures for 68 regions of interest for cortical thickness, sulcul depth, cortical surface area, and cortical volume based on the Desikan-Killiany brain registration atlas (a total of 272 gray matter variables) [36]. White matter microstructural DTI measures included fractional anisotropy for 31 regions of interest labeled using AtlasTrack.

Analytic Approach

The primary outcome was a child self-reported binary indicator of substance use initiation, defined as trying any nonprescribed substance (i.e., any substance use, including using any medication in a way a physician did not instruct), across a 3-year follow-up period. Of note, alcohol initiation was defined as alcohol sipping outside of a religious context. We constructed three hierarchical penalized logistic regression models using the elastic net penalty: 1) a self-report- only model that consisted of all self-report questionnaires (100 variables); 2) all variables in model 1 plus pubertal hormones and neurocognitive predictors (117 variables); and 3) all variables in model 2 plus neuroimaging variables (420 variables). The order in which blocks of variables were entered was determined by the amount of resources needed to collect the data, with the least resource-intensive variables entered in the first block. Figure 2 provides an overview of the domains in each model.

Figure 2.

Overview of models used in a study of predictors of substance use initiation by early adolescence ^a

^a DHEA=dehydroepiandrosterone; ROI=region of interest. Figure created in BioRender (biorender.com).

We tested a series of models, including random forest, logistic regression, and penalized logistic net regression to build a prediction model identifying features that were relevant to our outcome. Each model was then examined in the test data set by calculating area under the curve (AUC) to capture the models’ predictive performance. Prediction performance for our final model was higher with penalized logistic regression compared to traditional logistic regression (AUC_test=0.54) and random forest (AUC_test=0.59), and therefore the results we present here focus on the penalized logistic regression. (See Tables S16–S22 in the online supplement for results of logistic regression and random forest.) Our final model consisted of a penalized logistic regression, specifically a repeated, 10-fold, cross-validated, multiple- imputation, stacked adaptive elastic net regression. This approach was most appropriate for our data given the degree of correlation, potential for noise variables given the number of predictors examined, and use of multiple imputation for missing predictors. Furthermore, elastic net regression can shrink coefficients to zero, serving as a variable selection approach, retaining predictors most predictive of the outcome [37].

To implement our penalized logistic regression with elastic net, the data were separated such that 75% were used as the training data set (N=5,121) for model building and 25% as the test data set (N=1,708) for model testing. Equal proportions of youths who initiated substance use (14.4%) were included in both the test and training data sets, and therefore the class imbalance was maintained in both the test and training data sets to mirror our whole sample. To account for variability in prediction due to randomly splitting the data, we repeated this process 10 times resulting in 10 different training and test datasets. Nonnumeric variables were then dummy coded for use in the stacked adaptive elastic net regression. Numeric variables were centered and scaled. Of note, the test data set was centered and scaled based on the training data set. We cross-validated penalty parameters lambda and alpha, examining the following alpha elastic net penalty values for each model: 0 (ridge penalty), 0.2, 0.4, 0.6, 0.8, and 1.0 (lasso penalty). Our final model was selected based on the alpha elastic net penalty value for the model with the lowest 10-fold cross-validation error. Nonzero beta coefficients were used to determine which variables were retained in the model. The magnitudes of these coefficients were used to assess variable importance. Results include interpretation for variables that were retained in more than five splits, and additional detail on all variables retained, including those retained in only one split, is provided in Table S27–S32 in the online supplement. Positive coefficients indicate greater likelihood of substance initiation, and negative coefficients indicate lower likelihood of initiation. All analyses were performed in R, version 4.3.0. The mice package was used for multiple imputation, glm for logistic regression, and miselect for elastic net regression. We created 10 different splits of our training and test data set, then imputed on each different training and test split separately to avoid data leakage. Additional details on analytic procedures, including on the multiple imputation and reference group selection, are provided in the online supplement. Reference groups for all categorical variables constituted the largest category in the observed sample, apart from a binary indicator of recreational activities. The syntax is available on Github (https://github.com/greenrej/ABCD-SU-Initiation).

Results

Sample Characteristics

Of the 6,829 youths included in the final sample, 982 (14.4%) reported substance use initiation (Figure 1). Alcohol was the most frequent substance to be initiated within the 3-year follow-up period, followed by cannabis and nicotine. The sample’s demographic characteristics are summarized, by substance use initiation, in Table 1 (see Table S13 in the online supplement for non-demographic predictors split by substance use initiation). Age was similar in both groups, with those who progressed to substance use initiation slightly older on average at baseline (no initiation: 9.95 years [SD=0.69]; initiation: 10.08 years [SD=0.69]). The groups exhibited similar distribution across sex (no initiation: 51% male and 49% female versus initiation: 54% male and 46% female). Across both groups, majority of youths identified as White and non-Hispanic.

Table 1.

Sample characteristics by substance use initiation at age 12 in a study of predictors of substance use initiation (N=6,829) ^a

	No Substance Use Initiation (n=5,847)			Substance Use Initiation (n=982)			Test for Difference
Variable	n	Mean	SD	n	Mean	SD	T-statistic	p
Age (years)	5847	9.95	0.69	982	10.08	0.69	−5.40	<0.01
Sexb	n	%		n	%		χ2=4.64	0.03
Female	2888	49.4		448	45.62
Male	2959	50.6		534	54.38
Race	n	%		n	%		χ2=27.24	<0.01
White	3781	65.63		687	70.82
Black	895	15.54		90	9.28
Asian	136	2.36		19	1.96
Multiracial / Other	949	16.47		174	17.94
Ethnicity	n	%		n	%		χ2=1.01	0.32
Hispanic	1162	20.13		181	18.68
Non-Hispanic	4610	79.87		788	81.32
Income	n	%		n	%		χ2=27.45	<0.01
Less than $5,000	180	3.33		21	2.31
$5,000 through $11,999	170	3.15		19	2.09
$12,000 through $15,999	138	2.55		11	1.21
$16,000 through $24,999	254	4.70		28	3.08
$25,000 through $34,999	301	5.57		48	5.29
$35,000 through $49,999	451	8.35		82	9.03
$50,000 through $74,999	749	13.87		127	13.99
$75,000 through $99,999	831	15.38		131	14.43
$100,000 through $199,999	1732	32.06		306	33.70
$200,000 and greater	596	11.03		135	14.87
	n	%		n	%		Fisher’s Exact Test: p < 0.01
Religionc
Mainline Protestant	653	11.65		117	12.45
Evangelical Protestant	457	8.15		69	7.34
Historically Black Church	85	1.52		12	1.28
Roman Catholic	1045	18.65		203	21.60
Jewish (Judaism)	82	1.46		31	3.30
Mormon (Church of Jesus Christ of Latter Day Saints / LDS)	712	12.71		27	2.87
Jehova’s Witness	35	0.62		4	0.43
Muslim (Islam)	49	0.87		3	0.32
Buddhist	19	0.34		5	0.53
Hindu	15	0.27		1	0.11
Orthodox Christian	29	0.52		9	0.96
Unitarian (Universalist)	57	1.02		11	1.17
Other Christian	909	16.22		160	17.02
Atheist (do not believe in God)	107	1.91		22	2.34
Agnostic (not sure if there is a God)	107	1.91		27	2.87
Something Else	77	1.37		15	1.60
Nothing in Particular	1166	20.81		224	23.83
Parent Education	n	%		n	%		χ2=16.05	0.01
Less than HS Degree or GED Equivalent	362	6.20		35	3.56
HS Graduate or GED Equivalent	570	9.76		80	8.15
Some College or Associate’s Degree	1683	28.82		284	28.92
Bachelor’s Degree	1723	29.51		303	30.86
Master’s Degree	1153	19.75		209	21.28
Professional or Doctoral Degree	348	5.96		71	7.23

^aDetails on the proportion of missing data for each variable are provided in Table S13 in the online supplement.

^bOne participant indicated inter-sex male and was coded as male.

^cFisher’s exact test was used for religion because of small cell size.

Penalized logistic regression with elastic net

For model 1, the average AUC across 10 splits of the training data set was 0.704 (95% CI= 0.684–0.723). A slightly lower AUC, 0.666 (95% CI=0.631–0.701), was obtained for the test data set. For model 2, there was a similar average AUC for the training (AUC=0.710; 95% CI=0.691–0.730) and test (AUC=0.669; 95% CI=0.634–0.704) data set. For our final model, with all 420 predictors, the average AUC was 0.723 (95% CI=0.703–0.742) for the training data set and 0.659 (95% CI=0.624–0.694) for the test data set (see Figure 3 for variables retained in more than five splits). Additional detail on specific variables retained across all splits, sensitivity and specificity metrics, and AUC per split is provided in Table S24–S30 and Figure S1–S13 in the online supplement. Our AUCs were largely consistent across splits, further increasing the stability of our results (see Table S26 in the online supplement).

Figure 3.

Model 3 – Coefficients from Penalized Elastic Net Regression^a

^a Results are presented for variables that were retained in more than five splits. Positive coefficients indicate greater likelihood of substance use initiation, and negative coefficients indicate lower likelihood of substance use initiation. CBCL=Child Behavior Checklist; CRPBI=Child’s Report of Parental Behavioral Inventory; CRPF=Community Risk and Protective Factors survey; FA=fractional anisotropy; FES=Family Environment Scale; PPS=Prodromal Psychosis Scale; RAVLT=Rey Auditory Verbal Learning Test; SDS=Sleep Disorder Scale; SRPF=School Risk and Protective Factor Survey.

Top predictors of substance use initiation in the final model.

In our final model, when examining variables that were retained in more than five splits, the largest coefficients, on average, across splits were exhibited by variables in the sociodemographic, culture and environment, physical health, and parenting behaviors domains (Figure 3). In total, 31 self-report variables were retained in more than five splits from the following domains: seven in demographic characteristics, four in self and peer substance use, three in parenting behaviors, eight in mental health, six in physical health, and three in culture and environment. The strongest predictor was religion; specifically, youths with parent report of a religious preference of Mormon were less likely to initiate substance use (coefficient=−0.87, N_splits=10). The next strongest predictor was a religious preference of Jewish, where youths were more likely to initiate (coefficient=0.32, N_splits=10). The third top predictor was race; specifically, Black youths were less likely to initiate (coefficient=−0.32, N_splits=8). Our fourth top predictor was a religious preference of Muslim (Islam), where youths were less likely to initiate (coefficient=−0.25, N_splits=6), followed by a household with an income in the range of $12,000–$15,999, where youths were less likely to initiate (coefficient=−0.22, N_splits=6). Following the top five predictors, our next set of predictors were a history of detention or suspension (coefficient=0.20, N_splits=9), prenatal exposure to substance use availability (coefficient=0.12, N_splits=10), parents with less than a high school degree or GED (coefficient=−0.14, N_splits=8), substance use availability (coefficient=0.12, N_splits=10), and age at baseline (coefficient=0.12, N_splits=10).

Secondary predictors in the final model: self and peer involvement with substance use and parenting behavior.

Extending beyond our top 10 predictors, we found positive associations with intent to use alcohol and nicotine (coefficient=0.09 and 0.07, N_splits=10) and peer use of alcohol and nicotine (coefficient=0.06 and 0.04, N_splits=10). For parenting behaviors, greater reported parent-youth conflict (coefficient=0.03, N_splits=8) was associated with an increased likelihood of initiation, while greater acceptance by parents was associated with a decreased likelihood of initiation (coefficient=−0.02, N_splits=7).

Secondary predictors in the final model: mental health, physical health, and culture and environment.

In the mental health domain, the greatest effect size was observed for the UPPS-P sensation seeking subscale (coefficient=0.10, N_splits=10), with greater sensation seeking being associated with an increased likelihood of initiation. This was followed by three CBCL subscales: positive associations for rule-breaking behavior (coefficient=0.07, N_splits=10) and attention problems (coefficient=0.04, N_splits=8) and a negative association for anxiety / depression (coefficient=−0.06, N_splits=9). Two additional UPPS-P subscales exhibited positive associations with substance use initiation: lack of planning (coefficient=0.06, N_splits=10) and lack of perseverance (coefficient=0.03, N_splits=7). A negative association was observed for number of positive psychotic-like experiences endorsed (coefficient=0.03, N_splits=6), and a positive association for family history of substance use (coefficient=0.02, N_splits=6).

For physical health indicators, positive associations were observed with number of recreational activities (coefficient=0.02, N_splits=10), days of 60-minute physical activity (coefficient=0.07, N_splits=10), and days of strength building activities (coefficient=0.03, N_splits=7). Prenatal exposure to caffeine was also associated with an increased risk of initiation (coefficient=0.04, N_splits=9). Lastly, having greater sleep breathing disorders was associated with a decreased likelihood of initiation (coefficient=−0.02, N_splits=7).

The only additional cultural and environmental factors retained were a positive association with greater bilingual status (coefficient=0.02, N_splits=8) and a negative association with the School Risk and Protective Factor Survey school involvement subscale (coefficient=−0.06, N_splits=10).

Secondary predictors in the final model: hormones and neurocognitive factors.

No hormones were retained in more than five splits. Of the neurocognitive factors, greater scores on the NIH ToolBox Oral Reading Recognition Task (coefficient=0.06, N_splits=10), Picture Vocabulary Task (coefficient=0.04, N_splits=9), and Flanker Inhibitory Control and Attention Task (coefficient=0.04, N_splits=8) were associated with increased likelihood of initiation. Higher overall learning scores on the RAVLT were also positively associated with initiation (coefficient=0.02, N_splits=6).

Secondary predictors in the final model: structural imaging and DTI.

Of the 272 neuroimaging variables examined, 13 were retained in more than five splits. Of these, six were cortical thickness variables, four sulcal depth, and three surface area. No volume indices were retained. The direction of effect varied by region and metric (see Figure 3.); however, the structural MRI variable with the largest effect size was a negative association observed for cortical thickness of the left caudate anterior cingulate cortex (coefficient=−0.06, N_splits=10). Only one DTI variable was retained in more than five splits, fractional anisotropy in the right superior corticostriate (coefficient=−0.02, N_splits=6).

Discussion

Early initiation of substance use increases the risk of developing a substance use disorder and psychosocial problems in adulthood [38, 39]. In this study, we examined a multitude of sociodemographic, neurocognitive, and neuroimaging predictors of substance use initiation by age 12 using a statistical approach that accounts for the complexity of the data. Substance use initiation was defined as trying any non-prescribed substance (e.g., alcohol, tobacco, cannabis, non-prescribed medications). By age 12, approximately 14.4% of the 6,829 youths had initiated substance use. The most commonly reported substance was alcohol, followed by cannabis and nicotine, which aligns with epidemiological data [1].

When comparing hierarchical models, our results suggest that resource-intensive methods (e.g., hormonal, neurocognitive, and neuroimaging data) do not improve our ability to predict substance use initiation over and above self-reported data in this age range. This has important implications for future research and clinical applications focusing on substance use initiation, particularly in low-resource settings. For understanding the risk for substance use initiation, a more prudent allocation of resources may focus instead on the collection of detailed data related to self-, peer-, and familial-related factors, which were more informative in predicting substance use initiation during late childhood and early adolescence in the present study.

Results from this study indicate that parental endorsement of a religious preference, without additional details concerning degree of religiosity, may affect the risk of substance use initiation. This finding complements previous work in this area [40, 41] and highlights an easily collectable sociodemographic data point that can impact clinical care. Notably, religion exerted a strong protective effect particularly for Mormon youths, while Jewish youths were more likely to initiate substance use. Varying substance-specific norms within religions may influence the degree of protection or risk for substance use initiation. Religiosity has previously been identified as a protective factor for internalizing and externalizing behaviors [42], as well as substance use, suggesting a complex interaction between sociodemographic, mental health, and substance use outcomes. Future research could focus on identifying which components of religiosity impact substance use initiation, and subsequently develop and incorporate these aspects into prevention and intervention programs.

Other sociodemographic factors that were also consistently among the top predictors across models include race and income. Specifically, Black youths were less likely to initiate compared to White youths. These findings align with other reports that have noted lower rates of alcohol and nicotine use among Black youth compared to White youth [5, 43]. We also observed lower income as a consistent top predictor. While income has been shown to have mixed findings [6, 7], this may be influenced by context, as low income has been hypothesized to serve as a protective factor in rural communities and a risk factor in urban communities for alcohol initiation among Black youths [11].

Prenatal exposure to substance use was a robust predictor from the physical health category, which aligns with previous findings specific to substance use initiation [24]. Translational research should be done to determine the degree to which prenatal substance exposure incurs risk for earlier substance use initiation and the underlying mediators of such risk (e.g., altered brain development, differences in parental practices) to inform clinical procedures. While it may be difficult for mothers to discuss prenatal substance exposure in a clinical setting, providing specific contextual reasons for reporting (i.e., increased risk for early substance use) may promote honest communication and lead to better clinical care.

Lastly, within culture and environment domain, history of detention and suspension was a consistent top predictor. Early school behavior, including history of suspension, has been associated with problematic substance use in adulthood [44]. The present results demonstrate the more immediate negative impact that detention and suspension can have on substance use initiation, providing additional support for alternative diversion programs to alter this modifiable source of risk [45]. Prevention and early intervention efforts may also benefit from attending to more modifiable risk factors to delay the onset or prevent substance use initiation during adolescence. Our results indicate that greater sensation seeking, substance use availability, peer use of alcohol and nicotine, and rule breaking behavior increased the likelihood of substance use initiation. These results align with studies focused on predictors of binge drinking among adolescents, specifically greater sensation seeking and impulsivity as risk factors for binge drinking at age 14 [46]. These modifiable risk factors may be the target of prevention programs that aim to delay early onset of substance use. While some factors may be largely modifiable through macrostructural, or policy-based, interventions (e.g., the availability of substances in the area in which one lives), others may be more amenable to individual-level preventive interventions (for example, prevention programs that focus on the role of peers may be particularly useful [47]). Accounting for modifiable risk factors may improve prevention and early intervention outcomes by targeting factors that are most salient to early substance use behavior. While previous research has shown poorer cognition is a risk factor for heavy and problematic substance use [48], our findings indicate that better cognitive functioning in select domains (i.e., cognitive control, attention, and language ability) are associated with greater likelihood of initiation, which is typically a precursor to heavier substance use.

This study must be interpreted in light of its strengths and weaknesses. Strengths included a large sample size, predictors that span across multiple domains that include both parent and youth report, and an analytic approach to minimize the potential for different selection of variables with each imputed dataset. The development of our predictive model across sites and in a large and diverse sample means that we have appropriately considered geographic and demographic factors during the model development phase. This model is therefore an excellent candidate for future external validation studies focused on fairness and bias across geographical locations and key demographic subgroups. A primary limitation is the imbalanced outcome groups of those who initiated substance use. Further studies are needed to reexamine whether these risk factors for initiation remain consistent or change over time during mid to late adolescence, when we expect to see greater rates of substance use initiation. While religious categories were frequently among our top predictors, there were cases in which the sample size for selected religious categories was very small; further research will be needed to determine whether these findings replicate across samples. One possible explanation for the lack of improvement in performance when neuroimaging variables were added is that multiple regions of interest in isolation may only have a small effect; however, networks of brain regions may be more informative predictors of substance use initiation at this age. Finally, our analytic approach did not account for multi-level data within the context of site and families.

Findings from this study suggest that factors across multiple domains are predictive of substance use initiation. The use of resource-intensive measures, such as neurocognitive tests and neuroimaging, did not improve our ability to predict substance use initiation compared to self-report measures. Among our self-report measures, the most robust factors in our sample were religious preference, race, and income. Modifiable risk factors such as sensation seeking, substance use availability, and peer use of alcohol and nicotine also predicted greater likelihood of substance use initiation. Our approach goes beyond previous studies by comparing domains directly against each other in terms of prediction ability to identify the most critical risk factors. Our findings suggest a comprehensive assessment of sociodemographic factors and modifiable risk factors may be necessary to optimally provide information on risk for substance use initiation.

Previous presentation:

Portions of this paper were presented at the Research Society on Alcohol (RSA) annual meeting in Orlando, Florida (June, 2022), American Academy of Child and Adolescent Psychiatry (October, 2022), and the Association for Clinical and Translational Science (ACTS) annual meeting in Toronto, Canada (April, 2023).