Developmental and Etiological Patterns of Substance Use from Adolescence to Middle Age: A Longitudinal Twin Study

Abstract

Background:

The common liability to addiction framework suggests the tendency to use substances is largely a general heritable liability, but little is known about how expression of liability varies across development. We evaluated average developmental trajectories and covariation underlying commonly used substances using a genetically informative prospective design spanning three decades.

Methods:

Using a sample of 3,762 twins across seven waves of assessment spanning ages 14-40, we modeled these relationships using two complementary approaches: piecewise latent growth and common factor modeling on four measures of alcohol, tobacco, and marijuana use

Results:

Average use increased across adolescence and either stabilized (alcohol frequency) or declined (all others) in adulthood. Trajectories were heritable (~.35-.75), and genetically correlated with one another (~.40-.80). The random intercepts, centered at age 16, exhibited shared environmental correlations across substances. We found moderate to large phenotypic (r_p~.3-.9) and genetic correlations (r_g~.3-1) among the longitudinally varying common factors loading on use of each substance at each age. The factor loadings declined with age, reflecting waning influence of common etiology in substance use.

Conclusions:

Trajectories of substance use were strongly correlated with each other and influenced primarily by genetic and non-shared environment. A heritable common factor accounted for co-occurring substance use from mid-adolescence to mid-adulthood, and greater substance specificity emerged with maturation. These results extend and reinforce prior work examining consumption and problem use, providing new evidence over a broad age range showing that substance use behaviors are influenced by a more general liability in adolescence and specificity increases across development.

1. Introduction

Substance use typically begins in adolescence, escalates into the early 20s, and declines thereafter (Chen and Jacobson, 2012; Jackson et al., 2002; McGue et al., 2014; Patrick et al., 2021; Richmond-Rakerd et al., 2017). Furthermore, use of one substance is related to use of others, and trajectories of use covary. Like many complex traits, all stages of substance use development are heritable. Various etiological accounts seek to explain this observation, such as the common liability framework, which suggests the tendency for an individual to use and misuse substances is attributed to a highly heritable general liability (Hicks et al., 2011; Iacono et al., 2008; McGue et al., 2014; Vanyukov et al., 2012). Other developmental models, including dual systems theories (e.g., Casey et al., 2008; Iacono et al., 2008), attempt to explain both the developmental trajectories and cross-substance covariation, especially during adolescence. A great deal of support for the common liability model comes from behavioral genetic studies of adolescent substance use, dependence, and disinhibitory behavior, including studies on measures of disordered use (Vrieze et al., 2012) and consumption (Kendler et al., 2008) that demonstrated phenotypic and genetic covariation between substances. That said, common liability does not imply that heritable influences are stable over time, in either source or magnitude.

Less well-characterized than the shared variation are the course and etiology of these developmental trajectories as they extend into middle age, after the rapid physiological and psychosocial maturation during adolescence has completed. Indeed, the magnitude of genetic influence varies as a function of age. In a retrospective recall study, Kendler et al. (2008) found that genetic factors have minimal effects on alcohol, tobacco, and marijuana use around age 14, instead finding prominent familial environmental effects. Genetic factors increased in magnitude across later adolescence through age 35, while familial environmental factors decreased across that same timeframe. Prospective longitudinal work covering a shorter developmental window (Dick et al., 2007; Rose et al., 2001) and meta-analysis of cross-sectional and longitudinal heritability estimates demonstrated consistent results (Bergen et al., 2007). In a study focused on the covariation of tobacco, alcohol, and marijuana dependence, Vrieze et al. (2012) found that genetic correlations decreased across ages 14-29, suggesting that the importance of a common liability to problematic substance use weakened with age. These changing heritabilities and genetic correlations indicate potential changes in etiology across the lifespan, such as a transition from general to specific influences from adolescence to adulthood. Importantly, these results are generated from different phenotypes (consumption and dependence), so further work exploring longitudinal heritability as well as specific and shared genetic effects over time in the same phenotypes is necessary. Additional longitudinal work across older ages will deepen understanding of the common liability model and the degree to which it is relevant across the lifespan.

Here, we expand on this literature in a seven-wave, 30-year, genetically informative study to address three research questions: 1) What are the normative developmental courses of alcohol use frequency and quantity, tobacco use, and marijuana use frequency from adolescence to middle age? 2) How does the covariation between substances change with age? 3) What are the underlying etiological sources of variation and covariation and how do they change from adolescence to middle age? Given prior work on developmental trajectories, we expected to find increases in mean levels of use of each substance across adolescence up to young adulthood, then decreases from young adulthood through mid-adulthood. In line with the common liability framework, we expected to see heritable influences across all ages, and we predicted heritability would increase over time, with decreasing influence of the shared environment as individuals mature and depart the rearing environment. Lastly, prior work on problem use indicates that the common liability to abuse substances is most prominent in adolescence prior to the development of cognitive and social controls in adulthood. Therefore, we predicted covariance between substances would decrease in adulthood.

2. Methods

2.1. Participants

3762 twins (2410 monozygotic, 1352 same-sex dizygotic) were recruited and assessed from cohorts maintained by the Minnesota Center for Twin and Family Research (Iacono and McGue, 2002; Keyes et al., 2009; Wilson et al., 2019). Cohorts include a community sample first assessed at 17, a community sample first assessed at 11, and a sample screened for higher rates of childhood externalizing psychopathology first assessed at 11. Table 1 lists the sample size for each cohort and ages at each wave. Attrition was minimal; participation in completed waves ranged from 81-93% (see Supplemental Table 1)

Table 1:

Total Sample Size and Age Descriptives per Target Age

Target Age	N (Total)	N (OC)	N (YC)	N (ES)	Age Mean	Age Std. Dev.	Age Range
14	2334	MBD	1404	930	14.9	0.6	13.6, 17.0
17	3485	1252	1320	913	17.8	0.6	16.6, 20.3
20	2711	1111	1335	265	21.1	0.8	19.2, 24.3
24	3304	1167	1328	809	24.9	0.9	22.6, 29.3
29	3068	1168	1324	576	29.2	0.9	24.8, 33.2
35	966	MBD	966	MBD	35.0	1.6	32.7, 39.9
40	1243	745	498	MBD	40.4	2.6	35.3, 47.8

Note: MBD = Missing by design; OC = Minnesota Twin Family Study (MTFS) Older Cohort; YC = MTFS Younger Cohort; ES = Enrichment Sample.

2.2. Measures

We used four measures of substance use, all assessed with respect to the previous 12 months: alcohol frequency, alcohol quantity, tobacco use, and marijuana frequency. Response options for all measures are presented in the Supplemental Methods. Alcohol frequency represented the approximate number of occasions per year when the individual used alcohol on an ordinal scale. Alcohol quantity was a binned variable that represented the typical number of standard alcoholic drinks consumed on occasions when the respondent drank. Tobacco use was calculated based on the type of tobacco the participant used most; the majority of the tobacco-using sample reported using cigarettes most. The number of cigarettes, cigars, pipefuls or chews used per day was multiplied by 30 (days in a month) then divided by the number of days per month the individual used, to arrive at a frequency-adjusted quantity. The resulting amount was then binned to represent landmark quantities (ex. one pack of cigarettes per day) for each product. Both alcohol quantity and tobacco use were binned to ensure comparability across assessments from adolescence through adulthood, for which response options slightly varied. Marijuana frequency represented the approximate number of occasions per year when the individual used marijuana on an ordinal scale. Marijuana quantity used was not assessed in this sample due to complications in measuring dose (Cuttler and Spradlin, 2017).

2.3. Analyses

We conducted two sets of analyses: longitudinal confirmatory factor analysis and mixed effects models with random effects of age (latent growth models). This allowed us to characterize individual developmental trajectories and change in use of each substance, as well as how the covariation in use of these substances may change from adolescence to middle age. Analyses were completed in the full sample and stratified by sex to accommodate and model possible sex differences.

Analyses in the full sample included a fixed effect of sex and all analyses included a fixed effect of cohort. Both sets of analyses included a twin-based biometric variance decomposition to partition variation and covariation in the random effects into additive genetic (denoted A), shared environmental (C), and non-shared environmental components (E).

In the longitudinal confirmatory factor analysis, we specified a seven-factor solution with one factor per age of assessment, where all substance use measures within that age loaded onto that age’s factor (See Figure 1, top panel). Residual correlations were freed between alcohol frequency and quantity within each age, and between the same measure across ages (e.g., between tobacco use at one age and tobacco use at all other ages). Model fit was compared via AIC and BIC to a biometric model in which no factor structure was imposed. Weak longitudinal invariance was tested by equating factor loadings between two or more consecutive waves and evaluating change in model fit compared to a model in which all loadings were freely estimated (Widaman et al., 2010).

Figure 1.

Path diagram depicting the common factor model (top) and latent growth models phenotypes (bottom). Dotted lines represent the types of correlations estimated and not all biometric decompositions, correlations, or residuals are presented for clarity. In the top panel, only two of seven ages are presented for clarity, but all seven common factors are modeled simultaneously and freely correlated following the same patterns depicted here (e.g., for all ages common factor A, C, and E are correlated, residual variance in all four manifest variables within a phenotype are correlated across all ages). In the bottom panel, only two of four phenotypes are presented here for clarity, but all four phenotypes are modeled simultaneously and freely correlated following the same patterns depicted here, and all manifest variables have residual variances estimated (only age 14 alcohol frequency shown).

For growth models, we specified a random intercept and random effects of age. The structure of the random effects was determined based on background knowledge and exploratory generalized additive mixed models of mean change (Lin and Zhang, 1999). Generalized additive mixed models were fit using gamm4 v0.2-6 (Wood and Scheipl, 2020) in order to visualize change in each phenotype over time using nonparametric methods. Each phenotype was modeled as a function of age with a smoothing parameter and random effect of family. Plots of resulting trajectories were used to inform the shape of change for latent growth models.

We then estimated linear, quadratic, and piecewise linear random effects of age for each phenotype using latent growth modeling. Intercepts were centered at age 16, selected to represent average age of initiation. Knot points in piecewise models were fixed at expected ages of peak use estimated in independent datasets with similar measures and overlapping birth years to our sample (22 for alcohol, 21 for tobacco and marijuana; Chen et al., 2015; Chen and Jacobson, 2012; Clark et al., 2013; HHS, 2014; Jager et al., 2015; Patrick et al., 2019; Schulenberg et al., 2017). The linear, quadratic, and piecewise models were compared with AIC and BIC, as global fit indices are not appropriate in such models (Grimm et al., 2016; Sterba, 2014; Vrieze, 2012). Given the relatively better fit of the piecewise linear model (Supplemental Table 2), we next used the piecewise approach in a larger model where all four growth models were simultaneously estimated, and the per-phenotype random effects were allowed to correlate (See Figure 1, bottom panel). Residual variances were freely estimated and decomposed for each manifest variable, but residuals were uncorrelated between manifest variables.

Confidence intervals and significance tests for parameter estimates were estimated via bootstrapping with 1000 replicates resampled by twin pair stratified by zygosity. We computed the 2.5% and 97.5% quantiles of each parameter distribution to form 95% confidence intervals. Tests of parameter equivalence (e.g., whether males and females differed on parameter estimates) were conducted by bootstrapping the difference in parameters and computing a 95% confidence interval around that difference. Confidence intervals were computed using two approaches: 1) use of all bootstrap replicates regardless of model convergence, 2) use of only those replicates without optimization problems. No differences were found between models with normal optimization as compared to those with abnormal optimization, and abnormal optimization was rare (Supplemental Table 3). All analyses were conducted in R, with structural equation modeling in OpenMx v2.9.6 (Boker et al., 2011; Neale et al., 2016).

3. Results

Age descriptives are presented in Table 1 with phenotype descriptives presented in Supplemental Tables 4A and 4B. The sample was characterized by relatively low, normative levels of substance use, with means across ages corresponding to approximately 1-3 drinks per month, 1-3 drinks per drinking occasion, 1-2 cigarettes per day, and less than one use per month of marijuana. Attrition was generally unrelated to levels of earlier substance use, with the exception of tobacco use, where participants who did not return used tobacco more frequently than participant who did return (p=0.02; see Supplemental Table 5).

Figure 2 displays the substance use correlations for each target age; there are clear patterns of high correlation for a given phenotype across ages, and for a given age across substances. Twin correlations (Supplemental Table 6) indicated that substance use is heritable with monozygotic cross-twin correlations ranging from .13-.71. and dizygotic cross-twin correlations ranging from −.04-.68. The majority of monozygotic correlations were around ~.4-.6 whereas the majority of dizygotic correlations were around ~.2-.5. Supplemental Table 7 contains univariate biometric variance decompositions at each age.

Figure 2.

Heat map of correlations between substance use phenotypes over time. Blank squares indicate where no correlation could be estimated due to missingness by design. The triangles of correlations under the main diagonal depict clusters of correlations across phenotypes within a given age; these weaken over time. The minor diagonals show that within-phenotype correlations are stronger than correlations across phenotypes.

We fit a common factor model at each age to understand how covariation between all four phenotypes changes over time. An examination of the eigenvalues and scree plots at each age indicated that a single factor was appropriate (Supplemental Table 8), and the factor model fit acceptably compared to the ACE model with no factor structure (Factor model AIC=161,296.8, BIC=165,576.5, estimated parameters=732, ACE AIC=161,855.2, BIC=169,355.8, estimated parameters=1,354; difference in log-likelihood=685.6, df=622, p=.039).

At each age, phenotypes loaded positively on the common factor, with the majority of loadings being between .4-.7. The magnitude of standardized factor loadings was not constant across ages (Figure 3, top panel, and Supplemental Table 9). As fluctuation in a phenotype’s factor loading across ages may reflect fluctuation in sample size and composition, we computed average factor loadings within each age. The average factor loading across all four phenotypes decreased over time between 17 to 29, then stabilized from 29 to 40 (p < .05, Table 2), indicating that individual differences in substance use gradually becomes less attributable to the common factor with age. Tests of longitudinal invariance indicated similarly that the four standardized factor loadings could not be equated across most waves, the exception being between 29-35 and 35-40 (p=.21 and p=.06 respectively; Table 2).

Figure 3.

Line graphs representing the 7-factor model results, including the changes in factor loadings over time (top panel) as well as the standardized variance components over time (bottom panel). Vertical lines represent the width of the 95% confidence interval. The magnitudes of sources of variation (bottom panel) are largely stable over time, with the factor remaining moderately heritable, whereas factor loadings (top panel) generally decrease over time.

Table 2:

Longitudinal Factor Invariance

Ages Equated	Change in Average Factor Loading [95% CI]	χ2 for Test of Weak Invariance	df	p-value
14-17	−0.03 [−0.06, 0.01]	15.8	4	3.3e−3
17-20	−0.16 [−0.18, −0.13]	181.4	4	3.7e−38
20-24	−0.06 [−0.08, −0.03]	32.8	4	1.3e−6
24-29	−0.13 [−0.17, −0.08]	64.7	4	3.0e−13
29-35	0.01 [−0.05, 0.07]	5.9	4	0.21
35-40	−0.02 [−0.10, 0.06]	9.1	4	0.06
29-35-40	NA	17.0	8	0.03

Note: Bolded values indicate values that are significantly different from zero at p < .05.

Figure 3 (bottom panel) presents the standardized biometric variance components for the common factors. Heritability varies between ~.3-.7 but shows little evidence of systematic change with age in part due to uncertainty in estimates especially at age 40 (no significant changes in heritability between consecutive ages, all p > .05). The contributions of shared environment were low to moderate across the seven ages, ranging from ~.3-.5, with limited evidence of fluctuation (no significant changes in shared environmental variance between consecutive ages, all p > .05).

All common factors were positively phenotypically correlated over time, ranging from ~.7-1 between consecutive factors and ~.4-.6 for temporally distant factors (Supplemental Table 10). Genetic correlations between common factors (Supplemental Table 11) followed a similar pattern, with consecutive factors genetically correlating between ~.6-1 and temporally distant factors between .1-.7. Shared environmental correlations between consecutive and temporally distant common factors were strongly positive, ranging between ~.7-1 (Supplemental Table 12). Non-shared environmental correlations ranged between ~.4-.9 for consecutive factors and ~.3-.7 for temporally distant factors (Supplemental Table 13).

The factor model is useful to characterize patterns of covariation across substances as individuals age but does not provide direct information about mean change in use as a function of age. To examine this, we used generalized additive mixed models to visualize trajectories and latent growth modeling to evaluate an average trajectory and variation around it. The final model for each substance was a piecewise model that contained an intercept, adolescent slope, and adult slope. As seen in the generalized additive mixed model in Figure 4, peak use in our sample occurs around age 21 for all substances, indicating that the nationally representative peak values described in Section 2.3 were applicable to our sample.

Figure 4.

Generalized additive mixed models showing predicted substance use over time, accounting for family structure. All four phenotypes show an increase in use from early adolescence to approximately 20 years of age, followed by decreases and/or persistence depending on the phenotype

Table 3 presents maximum likelihood estimates for the growth parameter means. All four intercepts indicate that while some individuals may remain abstinent, on average our sample is using substances at age 16. Each adolescent slope indicates escalation of use across adolescence to young adulthood. The adult slope for alcohol quantity indicates that as individuals progress through young to mid-adulthood, on average alcohol quantity decreases. While this slope is small in magnitude (−.06 units per year), this corresponds to de-escalation of about one point on our 6-point ordinal scale across ages 22-40 (e.g., reducing use by 1-3 drinks). Alcohol frequency, tobacco use, and marijuana frequency adult slopes were all not significantly different from zero, indicating limited evidence for mean change in these phenotypes across adulthood.

Table 3:

Variance Decomposition of Latent Growth Model Parameters

Phenotype	Parameter	Mean [95% CI]	Variance [95% Cl]	A [95% CI]	C [95% CI]	E [95% CI]
Alcohol Frequency	Intercept	0.66 [0.54, 0.79]	.588 [.518, .665]	0.46 [0.28, 0.61]	0.31 [0.18, 0.47]	0.23 [0.17, 0.29]
	Adolescent Slope	0.38 [0.36, 0.41]	.021 [.018, .024]	0.44 [0.22, 0.60]	0.2 [0.06, 0.39]	0.36 [0.27, 0.45]
	Adult Slope	−0.01 [−0.03, 0.01]	.004 [.003, .005]	0.52 [0.32, 0.63]	0.1 [0.01, 0.23]	0.38 [0.28, 0.52]
Alcohol Quantity	Intercept	0.65 [0.53, 0.77]	.433 [.377, .437]	0.45 [0.27, 0.60]	0.3 [0.17, 0.46]	0.25 [0.18, 0.32]
	Adolescent Slope	0.26 [0.24, 0.28]	.012 [.011, .014]	0.51 [0.31, 0.63]	0.06 [0.00, 0.21]	0.43 [0.30, 0.55]
	Adult Slope	−0.06 [−0.08, −0.051]	.002 [.002, .0021]	0.51 [0.30, 0.651]	0.08 [0.01, 0.211]	0.41 [0.27, 0.551]
Tobacco Use	Intercept	0.70 [0.54, 0.86]	.902 [.790, 1.042]	0.4 [0.22, 0.59]	0.45 [0.27, 0.61]	0.15 [0.09, 0.21]
	Adolescent Slop	0.29 [0.25, 0.32]	.049 [.045, .056]	0.55 [0.38, 0.65]	0.06 [0.00, 0.19]	0.39 [0.30, 0.48]
	Adult Slop	−0.01 [−0.04, 0.03]	.004 [.003, .005]	0.42 [0.17, 0.60]	0.11 [0.01, 0.30]	0.46 [0.30, 0.62]
Marijuana Frequency	Intercept	0.45 [0.33, 0.57]	.465 [.364, .594]	0.52 [0.29, 0.71]	0.36 [0.17, 0.56]	0.13 [0.07, 0.20]
	Adolescent Slop	0.21 [0.17, 0.25]	.062 [.054, .073]	0.72 [0.54, 0.83]	0.07 [−0.01, 0.19]	0.21 [0.13, 0.31]
	Adult Slop	0.00 [−0.03, 0.03]	.003 [.002, .005]	0.68 [0.26, 0.81]	0.11 [−0.04, 0.38]	0.21 [0.04, 0.54]

Note: Bolded values indicate the parameter estimate is significantly different from 0 at p < .05.

Variation in intercepts was larger than variation in adolescent slopes, and variation in adolescent slopes was larger than adult slopes. We biometrically decomposed the growth factors (Table 3) to examine the degree to which genetic and environmental sources of variation underpin change in substance use over time. The majority of variation from the growth factors was heritable in nature (~.4-.7), indicating that change in use of these substances is influenced by genetic factors. Growth factor shared environmental effects ranged from ~0-.3, with larger shared environmental influences on the intercepts, which is expected as that reflects use at an earlier age. In contrast, there was minimal residual shared environmental variation (i.e., variation in manifest variables not explained by the growth factors); instead, unique environmental sources made up the majority of residual variation (Supplemental Table 14).

We evaluated the extent to which the random intercept, adolescent slope, and adult slope correlate for a given phenotype. The interpretation of the intercept depends on how age was centered; here we interpret correlations with the intercept as the correlations with model predicted average use at age 16, but these correlations should be interpreted with caution. With that caveat in mind, we found that those individuals using alcohol frequently at age 16 are expected to escalate less than their age mates who were not drinking as much at 16, and likewise are expected to decline more rapidly during adulthood; that is, the correlations between the intercept and both slopes were negative (correlation of −.54 and −.21 respectively). This was true for alcohol quantity as well (correlations of −.49 and −.33). Tobacco use and marijuana frequency showed a slightly different pattern; individuals using more at age ~16 escalated more rapidly (correlation of .27 for tobacco and .35 for marijuana) during adolescence, and then declined faster during adulthood (correlation of −.49 and −.45).

Correlations between the two slopes showed a consistent pattern across all phenotypes. The adolescent slope was negatively correlated with the adult slope, indicating that those who escalate faster during adolescence decline faster in adulthood (−.21, −.42, −.44, and −.36 for alcohol frequency, alcohol quantity, tobacco use, and marijuana frequency, respectively). Biometrically decomposed correlations are presented in Supplemental Table 15.

We evaluated the extent to which the random intercept, adolescent slope, and adult slope were correlated across phenotypes (ex. alcohol frequency and alcohol quantity adolescent slopes) to describe the degree to which changes across the same timespans are related across substances (Table 4). The majority of phenotypic correlations were significant and positive, indicating shared variation underlying the various components of the substance use trajectories across phenotypes. Biometrically decomposed correlations are presented in Supplemental Table 16.

Table 4:

Phenotypic Correlations between Phenotypes for Growth Model Parameters

Intercepts
	Alcohol Frequency	Alcohol Quantity	Tobacco Use	Marijuana Frequency
Alcohol Frequency	1
Alcohol Quantity	0.99 [0.98, 1.00]	1
Tobacco Use	0.85 [0.79, 0.88]	0.84 [0.78, 0.87]	1
Marijuana Frequency	0.74 [0.65, 0.80]	0.72 [0.62, 0.78]	0.83 [0.75, 0.87]	1
Adolescent Slopes
	Alcohol Frequency	Alcohol Quantity	Tobacco Use	Marijuana Frequency
Alcohol Frequency	1
Alcohol Quantity	0.81 [0.76, 0.85]	1
Tobacco Use	0.32 [0.24, 0.38]	0.37 [0.29, 0.44]	1
Marijuana Frequency	0.32 [0.24, 0.39]	0.27 [0.18, 0.34]	0.39 [0.32, 0.45]	1
Adult Slopes
	Alcohol Frequency	Alcohol Quantity	Tobacco Use	Marijuana Frequency
Alcohol Frequency	1
Alcohol Quantity	0.83 [0.74, 0.89]	1
Tobacco Use	0.27 [0.15, 0.37]	0.42 [0.30, 0.54]	1
Marijuana Frequency	0.25 [0.10, 0.39]	0.34 [0.18, 0.49]	0.36 [0.19, 0.52]	1

Note: All correlations significantly different from zero at p < .05

All factor and growth models were additionally fit stratified by sex to evaluate possible sex differences. Broadly speaking, we find minimal evidence for patterns of sex differences in the biometric estimates, common factor loadings, and growth factors; the distribution of p-values is presented in Supplemental Figures 1 and 2. A notable exception are the growth model adolescent slopes. Across all four adolescent slopes, mean estimates for males were larger than those for females, indicating more rapid growth for males (see Supplemental Table 17 for means and p-values).

3.1. Results Summary

The degree to which a common factor explains shared variation in alcohol frequency, alcohol quantity, tobacco use, and marijuana frequency decreased over time. Common factors were heritable at each age and the phenotypic and genetic correlations between ages are moderately to strongly positive. Results from the latent growth model indicated that on average, participants are using substances at age 16 and are escalating their use across adolescence into their early 20s. Following an early 20s peak, individuals reduce their quantity of alcohol consumption, but there are no significant changes observed in alcohol frequency, tobacco use, or cannabis frequency. Intercepts, adolescent, and adult slopes are correlated within and across substances, but correlations with intercepts should be interpreted with caution as they depend on centering.

4. Discussion

Here we evaluated relationships between alcohol quantity, alcohol frequency, tobacco use, and marijuana frequency across adolescence to mid-adulthood in a large twin sample, extending the developmental window across which these relationships have been evaluated. To our knowledge, this study is the first to demonstrate genetic correlations between developmental trajectories of substance use across adolescence and adulthood, using entirely prospective data spanning into middle age.

We fit a single common factor at each age to evaluate the degree to which the four measures of consumption have a shared etiology across ages. Mirroring existing work on substance dependence symptoms (Vrieze et al., 2012), the shared variation underlying measures of consumption weakened between ages 17-24, evidenced by decreasing factor loadings. Our results and those by Vrieze et. al. (2012) indicate developmental patterns for normative measures of consumption that reflect clinical measures of problem use. That is, the factor loadings on average decrease in magnitude, and standardized loadings cannot be equated between most consecutive factors, violating measurement invariance (Table 2). This indicates the factors are not measuring the same constructs at each age and therefore are not directly comparable. This may suggest that the risk underlying substance use is different at varied developmental points; risk may transition from general earlier in development to substance-specific later in life.

Our results were consistent with the common liability to addiction framework, which suggests shared genetic influences underlie substance use and other expressions of disinhibition. Indeed, the common factor was heritable at all seven ages and genetically correlated over time, indicating shared genetic influences across all four phenotypes and across development. Similarly, the growth models resulted in heritable and genetically correlated trajectories. Consistent with previous research, we found moderate shared environmental influences on the development of substance use, particularly at earlier ages (Kendler et al., 2008).

We saw parallels between the longitudinal factor models and growth models in terms of patterns of change. In the common factor approach, factor loadings decrease until age 24 and remain stable between 29-35 and 35-40, indicating declining covariation between substances that stabilizes in mid-adulthood. In the growth model, we see significant increases in use across adolescence to young adulthood, but across young to mid-adulthood, substance use remains relatively stable. These congruent results between means and variances show that adolescence and young adulthood are periods of broader change with regards to substance use, but substance use changes less dramatically or remains stable across mid-adulthood.

Dual systems and common liability provide compatible explanations we can apply to observed trajectories of normative substance use development and patterns of relationships between substances over time. Disinhibited behaviors are highest in adolescence when individuals have high reward sensitivity and low impulse control, but the behaviors desist as impulse control and ability to recognize consequences increase with maturation (Mcclure and Bickel, 2014; Steinberg, 2010). As substance use can be considered a manifestation of disinhibited behavior (Hicks et al., 2004; Iacono et al., 2008; Krueger et al., 2002; Young et al., 2009, 2000), it follows that patterns of substance use should be consistent with the dual systems model. Our results are not a direct test of the dual systems model, but they may be consistent with it, as our results indicate a pattern of escalation and de-escalation and genetic innovation across development. Future work informed by a dual systems perspective could investigate how new sources of genetic variation over time correspond to changes in maturation of impulse control.

4.1. Limitations

Generalizability is a limitation for several reasons. There is a lack of racial and ethnic diversity in the present sample. Additionally, phenotypes are largely based on past-year recall and single items, which limits generalizability to other conceptualizations of substance use. Cannabis use quantity was not assessed, making our cannabis analyses less generalizable than our alcohol and tobacco phenotypes, which incorporated multiple aspects of consumption. As our sample is a community sample characterized by low levels of substance use and misuse, we cannot generalize to samples with higher levels of use. The low level of substance use, particularly at earlier ages, also creates analytic limitations. It is possible that zero-inflation may drive the common factor coherence at young ages, potentially making the factor reflective of substance non-involvement. Lastly, the common factor is not longitudinally invariant, meaning we cannot make direct comparisons of the factor and its meaning over time.

4.2. Conclusions

The present study provides descriptions of normative changes in substance use over time that are consistent with the common liability explanatory model. The factors underlying substance use, parameterized as both a longitudinal common factor model and piecewise latent growth models, are heritable, and there are shared genetic influences across substances and ages. We also found evidence for shared environmental covariation, particularly in adolescence. These results may have implications for treatment and interventions, in that interventions in adolescence may be best targeted to general risk factors, whereas interventions in adulthood could target substance-specific risk factors.

We expand upon existing literature by demonstrating similar developmental patterns for the variation shared across measures of consumption as there are for measures of dependence (Vrieze et al., 2012). Lastly, we extended the developmental window in a genetically informative sample, an area in which there is limited existing research. Future research can directly evaluate the relevance of dual systems theory to change in substance use over time by integrating measures of disinhibition, reward processing, or brain-based measures of maturation. Future research could also incorporate age of initiation, to evaluate the degree to which developmental trajectories and the relationships between them differ for earlier and later initiators.

Highlights.

• On average, individuals increase their substance use during adolescence

• On average, individuals decrease or remain stable in their use across adulthood

• Use of different substances are highly related in adolescence

• Use of different substances becomes less related over time

• Substance use is genetically influenced with genetic overlap between substance