Bidirectional links between adolescent brain function and substance use moderated by cognitive control

Abstract

Background:

No clear consensus exists as to whether neurodevelopmental abnormalities among substance users reflect predisposing neural risk factors, neuroadaptive effects of substances, or both. Using a longitudinal design, we examined developmental patterns of the bidirectional links between neural mechanisms and substance use throughout adolescence.

Method:

167 adolescents (aged 13-14 years at Time 1, 53% male) were assessed annually four times. Risk-related neural processing was assessed by blood-oxygen-level-dependent responses in the insula during a lottery choice task, cognitive control by behavioral performance during the Multi-Source Interference Task, and substance use by adolescents’ self-reported cigarette, alcohol, and marijuana use.

Results:

Latent change score modeling indicated that greater substance use predicted increased insula activation during risk processing, but the effects of insula activation on changes in substance use were not significant. The coupling effect from substance use to insula activation was particularly strong for adolescents with low cognitive control, which supports the theorized moderating role of cognitive control.

Conclusions:

Our results elucidate how substance use may alter brain development to be biased toward maladaptive decision making, particularly among adolescents with poor cognitive control. Furthermore, the current findings underscore that cognitive control may be an important target in the prevention and treatment of adolescent substance use given its moderating role in the neuroadaptive effects of substance use on brain development.

Introduction

Currently, no clear consensus exists as to whether neurodevelopmental abnormalities among substance users reflect predisposing neural risk factors, the neuroadaptive effects of the substances, or both. Available studies suggest that abnormalities in the brain related to risky decision-making may be a predisposing risk factor for developing substance use disorders in the future (Peters, Peper, Van Duijvenvoorde, Braams, & Crone, 2017; Squeglia, Jacobus, Nguyen-Louie, & Tapert, 2014), and that longer and greater substance use during adolescence results in such brain abnormalities (Sagar et al., 2015; Weissman et al., 2015). The lack of prospective longitudinal neuroimaging studies examining bidirectional links between substance use and brain development diminishes our ability to draw directional inferences regarding the brain-substance use association. Limitations of previous research include cross-sectional comparisons of chronic, heavy substance users versus non-users (e.g., Sagar et al., 2015; Squeglia et al., 2014). Even longitudinal studies have used data at only two time points and have examined only one direction rather than bidirectional relations (e.g., Peters et al., 2017; Squeglia et al., 2014), resulting in an inability to disentangle premorbid patterns of brain function from effects of substance use. As such, the present longitudinal study aimed to clarify reciprocal temporal patterns between neural mechanisms and substance use throughout adolescence. Specifically, we expected neural risk processing to predict subsequent substance use behaviors, as well as prior substance use to alter subsequent brain activation during risk processing. We further tested whether these bidirectional links were modulated by cognitive control.

Developmental neuroscience work has suggested that neurobiological pathways underlie substance use behaviors in adolescence (Casey & Jones, 2010; Ernst, Pine, & Hardin, 2006). In particular, individual differences in the subcortical, motivational systems involved in reward processing (e.g., the ventral and dorsal striatum) have been linked to greater risky decision making in adolescence (van Duijvenvoorde et al., 2014). However, value-based decision making involves neural computations of risk (i.e., variances in the outcome), rather than simply being based on the value of rewards (d’Acremont & Bossaerts, 2008; Mohr, Biele, & Heekeren, 2010). A key region consistently implicated in the processing of risk information is the anterior insular cortex, which acts as a signal to guide adolescents towards or away from risky choices aligned with individual preferences for risk (Mohr et al., 2010). Research has shown that adolescents recruit the insular cortex during risky decision-making more than children or adults, and adolescents’ hypersensitivity of the insular cortex to increasing variance of potential outcomes is related to making safer choices (van Duijvenvoorde et al., 2015). Further, insula activation during risk processing has been linked to adolescent reports of real-life health risk behaviors (Kim-Spoon, Deater-Deckard, Lauharatanahirun, et al., 2016).

Irrespective of the temporal patterns, the links between neural risk processing and substance use behaviors are regulated by the cognitive control system, which works to guide goal-directed behavior and inhibit prepotent responses (Ernst et al., 2006; Kim-Spoon et al., 2017). For example, lower levels of self-reported behavioral activation system were predictive of earlier onset of substance use among adolescents with poor cognitive control (Kim-Spoon, Deater-Deckard, Lauharatanahirun, et al., 2016). Similarly, risk-related hemodynamic activity in the anterior insula during the anticipation of uncertain outcomes predicted health risk behaviors, such as substance use and risky sexual behaviors, among adolescents with poor cognitive control (Kim-Spoon, Deater-Deckard, Holmes, et al., 2016). In the present investigation, we consider that these neural systems are critical to adolescent risky decision making, proposing that prospective bidirectional links between risk-related neural activation and substance use behaviors are moderated by cognitive control.

Three models address the temporal association between substance use and brain development: the exposure model, the premorbid model, and the bidirectional model (Windle et al., 2018). The exposure model posits that substance use alters subsequent brain development, illustrated by neuroadaptations (Robinson & Berridge, 1993). That is, incremental changes to the brain, as a result of use, render the brain hypersensitive to substances and their related cues. Alternatively, the premorbid model proposes that individual differences in the brain confer subsequent vulnerability for substance use. Finally, the bidirectional model proposes bidirectional relations between premorbid brain function and substance use exposure that interface across time.

A small number of neuroimaging studies provide preliminary evidence for an association between variance in brain function and adolescent substance use. Cross-sectional studies comparing heavy substance users and nonusers revealed differences in blood oxygenation level-dependent (BOLD) responses in the dorsolateral and medial prefrontal cortices, inferior frontal gyrus, and insula while performing inhibition tasks (Galván, Poldrack, Baker, McGlennen & London, 2011; Feldstein Ewing, Houck, & Bryan, 2015; Tapert et al., 2007). In addition, a longitudinal study showed that binge-drinking adolescents showed significantly lower BOLD responses in the left cerebellum, but not in the ventral striatum, during reward receipt compared to non-users, after controlling for baseline activity two years earlier (Cservenka, Jones, & Nagel, 2015). Similarly, cross-sectional studies examining substance use onset demonstrated that earlier onset is related to aberrant resting-state connectivity between reward processing and cognitive control networks (Weissman et al., 2015) as well as altered patterns of inhibition-related activation and functional connectivity (Lopez-Larson, Rogowska, & Yurgelun-Todd, 2015; Sagar et al., 2015). This “age of onset” effect has been interpreted as support for the exposure model, such that longer exposure to substance use is responsible for altered subsequent brain function over time. However, those findings demonstrate neural correlates of substance use based on cross-sectional data. Thus, whether substance use indeed affects brain development or, conversely, whether early-onset users show qualitatively different brain function, remains unclear.

Turning to evidence for the premorbid model, abnormalities in brain function may be a predisposing risk factor for adolescent substance use. A couple of longitudinal studies have found that lower inhibitory-related BOLD responses (including dorsolateral and medial prefrontal cortices and supplementary motor area) are predictive of later alcohol use (Norman et al., 2011; Wetherill, Squeglia, Yang & Tapert, 2013). These studies, however, tested only the direction from earlier brain function to later substance use. Although studies testing the bidirectional model are rare, one study did test reciprocal temporal patterns, showing that reduced resting-state connectivity between the amygdala and the orbitofrontal cortex was related to greater alcohol use two years later in a sample of adolescents and young adults, without evidence for the reverse relation (Peters et al., 2017).

In the current longitudinal study, we investigated bidirectional links between neural risk processing and substance use behaviors across adolescence and tested whether these associations were moderated by cognitive control. Given the consistent and robust results implicating the insular cortex as a key region involved in risk processing (Mohr et al., 2010) and the critical role that the insula plays in addiction (Naqvi & Bechara, 2009), we focused on examining insular activation during risk processing. Specifically, we hypothesized that bidirectional links between changes in neural patterns of risk processing and changes in substance use would be strongest for adolescents with low cognitive control.

Method

Participants

The sample included 167 adolescents (53% males) who participated in annual assessments across four years. Adolescents were 13 to 14 years of age at Time 1 (M = 14.07, SD = 0.54 for Time 1, M = 15.05, SD = 0.54 for Time 2, M = 16.07, SD = 0.56 for Time 3, and M = 17.01, SD = 0.55 for Time 4). About 82% of adolescents identified as Caucasian, 12% African-American, and 2% other. Median annual family income was in the $35,000-$50,000 range, with varying levels of family economic status (50% “poor/near poor” and 50% “non-poor” according to income-to-needs ratio). Among the primary caregivers (137 mothers, 21 fathers, and 9 others), 34% had a high school degree or less, 24% some college education, 24% bachelor’s degree, and 18% graduate degree. Inclusion criteria included being ages 13 to 14 at Time 1 with vision corrected to be able to see the computer display clearly. Exclusion criteria were claustrophobia, history of head injury resulting in loss of consciousness for >10 minutes, orthodontia impairing image acquisition, severe psychopathology (e.g., psychosis), and other contraindications to magnetic resonance imaging.

At Time 1, 157 adolescents participated. At Time 2, 10 adolescents were added (to offset annual attrition) for a final sample of 167 (150 at Time 2, 147 at Time 3, and 150 at Time 4). Across all four years, 24 adolescents did not participate at all four time points for reasons including: ineligibility for tasks (n = 2), declined participation (n = 17), and lost contact (n = 5) during the follow-up assessments. Rate of participation was not significantly predicted by age, income, sex, race or study variables (ps > .15).

Procedures

Adolescent participants and their primary caregivers were recruited via email, recruitment letters, and flyers. Participants visited the laboratory to complete behavioral measures and MRI scans at four annual time points and were compensated for their participation. All procedures were approved by the institutional review board of the university and written informed consent (from parents) and assent (from adolescents) were received.

Measures

Cognitive control.

Cognitive control was measured with the Multi-Source Interference Task (MSIT; Bush, Shin, Holmes, Rosen, & Vogt, 2003). Adolescents were presented with sequences of three digits, two of which were identical, and were instructed to indicate the identity (but not the position) of the unique, target digit. In the neutral condition, target digits were congruent with position. In the interference condition, target digits were incongruent with position. To assess task performance, we used intraindividual variability in reaction time, indexed as intraindividual standard deviations (MacDonald, Karlsson, Rieckmann, Nyberg, & Backman, 2012) for correct responses in the interference condition. We reverse scored these data such that higher scores represented better cognitive control.

Substance use.

Adolescents self-reported cigarette, alcohol, and marijuana use using a substance use index adapted from Wills, Yaeger, and Sandy (2003). This index consists of three items asking typical frequency (i.e., which is the most true for you about using alcohol/smoking cigarettes/using marijuana?), using a 6-point response scale ranging from 1 (never used) to 6 (usually use every day). A maximum substance use score was used across cigarette, alcohol, and marijuana use, with higher scores indicating greater use, (α = .61 to .75 for Times 1-4).

Neural risk processing.

Adolescents engaged in a modified economic lottery choice task (Holt & Laury, 2002) while their BOLD responses were recorded. On each trial, adolescents were asked to choose between two gambles, where one gamble was always riskier (higher coefficient of variation; CV) than the other (see Figure 1). The CV was computed by dividing the standard deviation of a gamble by the expected value (i.e., probability-weighted outcome) of that gamble (see Appendix S1). For each gamble, there was a high and low monetary outcome, each associated with a specific probability that varied across a total of 72 trials (approximately 25 minutes to complete). To incentivize performance, participants were compensated based on their winnings from four randomly selected trials.

Figure 1.

a) In the Lottery Choice Task, adolescents were asked to choose between pairs of uncertain gambles. For each pair of gambles, one option was always riskier (higher variance; e.g., $3.61 vs $0.09) than the other (lower variance; e.g., $1.88 vs $1.50). The associations between outcomes and probabilities are represented with corresponding colors (orange or blue). b) Each trial consisted of a decision phase, a fixation phase, an outcome phase, and an inter-trial-interval (ITI).

Imaging acquisition and analysis.

Functional neuroimaging data were acquired on a 3T Siemens Tim Trio scanner with a 12-channel head matrix coil. High-resolution structural images were acquired using the following parameters: repetition time (TR) = 1200 ms, echo time (TE) = 2.66 ms, field of view (FoV) = 245x245 mm, and 192 slices with the spatial resolution of 1x1x1 mm. Echo-planar images were collected using the following parameters: slice thickness = 4mm, 34 axial slices, FoV = 220x220mm, TR = 2 s, TE = 30 ms, flip angle = 90 degrees, voxel size = 3.4x3.4x4 mm, 64x64 grid, and slices were hyperangulated at 30° from anterior-posterior commissure. There were 141 valid scans obtained at Time 1, 136 at Time 2, 126 at Time 3, and 129 at Time 4, due to not meeting MRI safety criteria, excessive head motion (>3 mm), technical errors, or imaging artifacts.

The following procedure was used for preprocessing of imaging data: data were corrected for excessive head motion using a six-parameter rigid body transformation and realigned, the mean functional image was co-registered to the anatomical image, the anatomical image was segmented, functional volumes were normalized and registered to the Montreal Neurological Institute (MNI) template, and then smoothed using a 6mm full-width-half-maximum Gaussian filter. Analysis of neuroimaging data was conducted using SPM8 (Wellcome Trust Neuroimaging Center, University College London).

Whole-brain general linear model (GLM) analysis was conducted to assess neural computations of risk. At the subject level, the following regressors were included: decision event (4 s), parametric regressor of decision event for magnitude of risk (coefficient of variation; CV), outcome event (2 s), and a parametric regressor of outcome event for high and low outcomes. In addition, regressors of no interest (the button press and six motion regressors) were included. All regressors included in the model were convolved with a canonical hemodynamic response function. At the group level, to assess neural responses of risk processing, a separate one-sample t-test was conducted to determine how the CV for chosen gambles was related to BOLD response.

All statistical inferences were made at a cluster-corrected threshold of p < .05 with a Family-Wise Error (FWE) correction to account for multiple comparisons with an initial cluster-forming uncorrected threshold of p < .001. Region of interest (ROI) analyses focused on the insular cortex because of its key role in risk processing (Mohr et al., 2010). Eigenvariate values from the left and right insula cortex were extracted using a 6mm sphere around the peak voxel coordinates of each region (see Figure 2 and Tables S1–S4). We created bilateral insula composite scores by averaging left and right insula eigenvariate values (r = .74-.87 between left and right insula values) that were extracted using the aforementioned procedures for each wave. The bilateral insula composite score was used as the neural risk-processing variable, with higher scores indicating greater activation.

Figure 2.

Increased Activation in the Bilateral Anterior Insular Cortex (INS) during Riskier Gambles across Times.

Data analytic plan

Correlations and descriptive statistics for all composite variables used in the Structural Equation Model (SEM) are presented in Table 1 (see Table S5 for descriptive statistics of the raw scores). We winsorized outliers (19 scores, deviating > 3.29 SD from the mean) resulting in all variables with acceptable skewness and kurtosis (< 3 and < 10, respectively). Multivariate GLM analyses indicated that demographic covariates (sex, race, and family income) were not significant predictors of study variables (ps > .08) and were therefore not included in the main analyses.

Table 1.

Descriptive Statistics and Bivariate Correlations for Insula Activation, Substance Use, and Cognitive Control

	1	2	3	4	5	6	7	8	M (SD)	Min	Max
1. Bilateral insula Time 1	-								0.04 (0.05)	−0.08	0.22
2. Bilateral insula Time 2	.33**	-							0.56 (0.72)	−1.09	3.00
3. Bilateral insula Time 3	.25**	.32**	-						0.61 (0.84)	−1.51	3.56
4. Bilateral insula Time 4	.27**	.39**	.46**	-					0.79 (1.16)	−3.23	4.81
5. Substance use Time 1	−.04	−.07	−.10	.07	-				1.48 (0.75)	1.00	4.00
6. Substance use Time 2	.00	−.05	−.07	.09	.70**	-			1.75 (1.03)	1.00	5.00
7. Substance use Time 3	.06	.03	−.03	.13	.52**	.67**	-		2.19 (1.22)	1.00	6.00
8. Substance use Time 4	−.04	−.04	−.06	.02	.49**	.63**	.79**	-	2.67 (1.40)	1.00	6.00
9. Cognitive control grand meana	.19*	.26**	.20*	.22*	−.08	−.09	−.14	−.19*	−0.21 (0.04)	−0.35	−0.12

Note.

^aReverse coded.

^*p < .05,

^**p < .01

We used Latent Change Score (LCS) modeling (McArdle, 2009) in Mplus (Muthén & Muthén, 1998-2017) with full information maximum likelihood (FIML) estimating model parameters using all available data regardless of the patterns of missing data. LCS modeling provides a superior framework to evaluate dynamic longitudinal changes by testing whether change in one variable may depend on any prior state within the system over time, a prior state of another variable, or both. In particular, LCS modeling can evaluate directional reciprocal effects between intraindividual changes (latent change factors) via cross-lagged dynamic coupling. We first fit a series of univariate dual process LCS models to describe basic patterns of changes in neural risk processing and substance use behavior. We then tested bidirectional effects between insula activation and substance use using bivariate LCS models. To use cognitive control as a time-varying moderator in the bivariate LCS models, we tested a growth mixture model that identifies discrete latent classes based on longitudinal trajectory patterns. We regarded the model as having a good fit when RMSEA values were less than .08 (Browne & Cudeck, 1993) and CFI values were greater than .90 (Bentler, 1990).

Results

Latent change score models of neural risk processing and substance use

We fit a series of univariate dual process LCS models to describe basic patterns of constant and time-dependent changes. Dual process LCS models decompose changes into proportional change, which represents the influence of the variable itself at the previous measurement occasions, and a constant slope, which represents the underlying rate of linear change. For neural risk processing, model fit was good (χ² = 8.73, df = 5, p = .12, RMSEA = .07, CFI = .94). The mean (M = 0.04, SE = 0.004, p < .001) and variance (σ² = 0.003, SE < .001, p < .001) of the intercept factor, and the mean (M = 0.48, SE = 0.07, p < .001) and variance (σ² = 0.16, SE = 0.04, p < .001) of the slope factor were significant. Thus, insula activation demonstrated positive constant growth with significant individual differences in initial levels as well as the rate of growth. Furthermore, significant negative proportional changes (b = −0.64, SE = 0.15, p < .001) between time points indicated that adolescents with higher insula activation scores on the previous occasion changed less than those with lower insula activation scores (when accounting for average change), resulting in a slight deceleration of insula activation change over time.

For substance use, fit was good (χ² = 7.05, df = 4, p = .13, RMSEA = .07, CFI = .99). However, there was no evidence of significant proportional change. Removing proportional effects did not significantly degrade model fit (Δχ² = 1.10, Δdf = 1, p = .30) and the resulting model fit was good (χ² = 8.15, df = 5, p = .15, RMSEA = .06, CFI = .99). In this final model, the mean (M = 1.46, SE = 0.06, p < .001) and variance (σ² = 0.51, SE = 0.07, p < .001) of the intercept factor and the mean (M = 0.39, SE = 0.03, p < .001) and variance (σ² = 0.14, SE = 0.02, p < .001) of the slope factor were significant. The results indicated positive constant growth with significant individual differences in the baseline and the rate of growth, but no evidence for proportional effects.

The univariate LCS analyses indicated that latent changes in insula activation were characterized as a function of two forces, a constant change process and an autoregressive mechanism, whereas latent change in substance use was mainly explained by a constant change process.

Growth mixture modeling of cognitive control

Given that we had cognitive control data assessed over four years, we explored cognitive control as a time-varying moderator. We estimated discrete latent classes based on longitudinal trajectories of cognitive control using a growth mixture model (Muthén & Muthén, 2000). Results indicated that the one-class model provided the best fit, indicating that growth trajectories of cognitive control were homogenous (see Appendix S2; also see spaghetti plot in Figure S1). Therefore, we calculated a grand mean by averaging across four time points to represent each individual’s general level of cognitive control.

Bivariate latent change score model moderated by cognitive control

Next, in a series of bivariate LCS models, we tested bidirectional effects between insula activation and substance use (see Figure 3). We tested whether adolescents with low cognitive control showed stronger coupling effects between brain function and substance use, compared to those with middle-to-high levels of cognitive control (i.e., low cognitive control as a vulnerability factor). Alternatively, we tested whether adolescents with high cognitive control showed weaker coupling effects between brain function and substance use, compared to those with low-to-middle levels of cognitive control (i.e., high cognitive control as a protective factor). For testing group differences, we imposed equality constraints to test numeric invariance with respect to (1) the effects of insula activation on changes in substance use, and (2) the effects of substance use on changes in insula activation. If model fit was significantly degraded by imposing equality constraints, the results indicated significant differences between differing levels of cognitive control. As suggested by Flora, Khoo, and Chassin (2006), we used longitudinal levels of cognitive control (moderator) to create groups for testing the moderator effect via a multiple group modeling method. In doing so, we created three groups following an extreme group approach to maximize individual differences (McClelland & Judd, 1993): those in the lowest 25% of cognitive control ability (“low cognitive control” group, n = 41), those in the highest 25% of cognitive control ability (“high cognitive control” group, n = 41), and the 50% in-between (“middle cognitive control” group, n = 84). Model fit for the three-group SEM was good (χ² = 66.62, df = 58, p = .20, RMSEA = .05, CFI = .98).

Figure 3.

Bivariate Latent Change Score Model of Insula Activation (Ins) and Substance Use (SU).

Notes. dy = change score for Ins; dx = change score for SU; T1 = Time 1; T2 = Time 2; T3= Time 3; T4 = Time

Table 2 presents results examining bidirectional relations between insula activation and substance use across the three cognitive control groups. As for bivariate associations, we found significantly stronger positive coupling effects for adolescents with low cognitive control compared to adolescents with middle or high cognitive control (Wald test χ² = 5.33, df = 1, p = .02). Specifically, substance use was a leading indicator for subsequent changes in insula activation for adolescents with low cognitive control, indicating that greater substance use predicted increased insula activation. In contrast, there were no significant effects of insula activation on changes in substance use, with no significant differences between the cognitive control groups (Wald test χ² = 0.001, df = 1, p = .97).

Table 2.

Estimates from Bivariate Latent Change Score Models of the Insula Activation-Substance Use Association in the Three Cognitive Control Groups

	Low Cognitive Control		Middle Cognitive Control		High Cognitive Control
Parameter	Insula	Substance Use	Insula	Substance Use	Insula	Substance Use
Proportional effects	−1.78*** (0.22)	-	0.01 (0.38)	-	−0.95*** (0.19)	-
Coupling effects
Ins → SU	0.08 (0.47)		0.02 (0.16)		0.10 (0.14)
SU → Ins	0.37* (0.15)		−0.25 (0.20)		0.17 (0.26)
Latent Covariances
Ins int ↔ Ins slope	0.00 (0.01)		0.00 (0.00)		0.02* (0.01)
SU int ↔ SU slope	0.00 (0.07)		−0.03 (0.04)		0.02 (0.04)
Ins int ↔ SU int	0.01 (0.01)		−0.01 (0.01)		0.01 (0.01)
Ins slope ↔ SU slope	−0.10 (0.07)		0.03 (0.03)		−0.06 (0.06)
Ins int ↔ SU slope	0.00 (0.01)		0.00 (0.00)		−0.01 (0.01)
SU int ↔ Ins slope	−0.03 (0.12)		0.18 (0.12)		−0.14 (0.13)
Intercept mean	0.04*** (0.01)	1.54*** (0.11)	0.04** (0.01)	1.51** (0.09)	0.06***(0.01)	1.32*** (0.10)
Slope mean	−0.14 (0.26)	0.48*** (0.13)	0.71* (0.30)	0.38*** (0.07)	0.54 (0.37)	0.28** (.09)
Intercept variance	0.00*** (0.00)	0.50** (0.14)	0.00*** (0.00)	0.57*** (0.11)	0.00*** (0.00)	0.37*** (0.08)
Slope variance	0.24** (0.09)	0.16** (0.06)	0.13 (0.07)	0.15*** (0.03)	0.51* (0.20)	0.10** (.03)
Error variance	0.00 - 0.73	−0.00 - 1.00	0.00 - 0.47	0.08 - 0.40	0.00 - 1.51	0.00 - 0.40

Note. Ins = insula activation; SU = substance use; int = intercept.

^*p < .05,

^**p < .01,

^***p < .001.

As for within-construct changes in insula activation, both the low and the high cognitive control groups exhibited significant, negative proportional (self-feedback) effects with non-significant constant slopes, suggesting that changes in insula activation were explained primarily by year-to-year changes, rather than constant growth. Specifically, higher insular activation had dampening effects on subsequent changes over a year. In contrast, for the middle cognitive control group, changes in insular activation were explained primarily by positive constant growth, rather than year-to-year changes. For within-construct changes in substance use, all three groups exhibited similar patterns of positive constant change without proportional effects.

Covariances between growth factors were largely non-significant for all three groups, suggesting that intercept and growth in insula activation and substance use were not strongly tied together. The only exception was the significant covariance between the intercept and the slope of insula activation for the high cognitive control group, suggesting that higher insula activation at baseline was related to more positive growth over the next four years. Finally, results of the alternative tests contrasting the high cognitive control group with the other groups revealed non-significant group differences for the substance use → insula activation coupling and for the insula activation → substance use coupling (Wald tests p = .46 and .73 respectively).

Inferences for the detected significant effects of substance use on insula activation would be stronger if insula activation was first assessed prior to any substance use, thereby ruling out the possibility that premorbid insula deviations predicted substance initiation. In supplemental analyses, we tested the hypothesized model using only adolescents who were substance use naïve at Time 1. Consistent with results using the full sample, the effects of substance use on insula activation were significant only for the low group but not for the middle or high cognitive control group (see Appendix S3)

Discussion

While research identifying neural correlates of substance use based on cross-sectional data is prevalent, this is the first study investigating dynamic bidirectional links between brain function and substance use across adolescence using prospective longitudinal data at four time points. We directly tested competing models of the brain-substance use association, including the exposure, premorbid, and bidirectional models. Methodologically, we used a novel approach of latent change score modeling to examine both systematic growth and cross-lagged interplay between the brain and substance use.

We found that substance use was the leading indicator of insula activation changes, consistent with the exposure model. Specifically, substance use → brain coupling was significant among adolescents with low cognitive control, but brain → substance use coupling was not significant regardless of cognitive control levels. The current results represent a unique quantitative test of the cross-domain temporal hypothesis within a person over time, demonstrating neuroadaptive effects of substance use on brain function. These dynamic results are broadly consistent with previous cross-sectional studies suggesting a link between heavy substance use and alteration in brain function among adolescents (Feldstein Ewing et al., 2015; Sagar et al., 2015; Weissman et al., 2015).

We note that higher levels of insular activation during risk processing were predictive of higher substance use for adolescents with low cognitive control abilities. This finding appears to be inconsistent with previous research demonstrating that heightened insular activation is related to risk avoidance in adolescents (van Duijvenvoorde et al., 2015). However, a recent study demonstrated a significant association between high insula activation and high substance use among adolescents with low cognitive control (Kim-Spoon et al., 2019). The insular cortex, while known to be a key region in the neural representation of risk, also plays an important role in translating interoceptive signals into conscious feelings and behavioral biases during risk/reward-related decision making (Naqvi & Bechara, 2009). Specifically, the insula mediates anticipation for rewards by translating sensory signals into the emotional experience of desires to engage in the behavior. Because most drugs exert the interoceptive effects the insula translates, higher insula activation is related to conscious urges across different substances (Naqvi & Bechara, 2009). Our finding implies that repeated exposure to drugs may result in heightened sensitization towards reward-related cues so that increases in reward availability (with increasing variance in potential outcomes) elevate insula activation. Further, adolescents with low cognitive control may be particularly susceptible to heightened arousal in the insula in response to conscious urges, negative affect, and other somatic sensory input, instead of encoding the potential negative outcomes of their actions, because of their compromised cognitive control which would normally modulate such arousal.

We found different patterns of year-to-year changes in insula activation during risky decision making, depending on the level of cognitive control. For low cognitive control, changes in insula activation were predicted by both prior insula activation and prior substance use, but not maturational change. That is, there existed self-feedback influences from prior states of insula activation to subsequent changes in insula activation (thus adolescents who started lower changed more), in tandem with an amplifying force from prior substance use to subsequent changes in insula activation (thus adolescents with higher substance use increased more). At average levels of cognitive control, the overall change in insula activation appeared to reflect maturational change, but not depend upon prior insula activation or substance use. At high cognitive control, yearly changes in insula activation were related to prior states of insula activation in the absence of maturational growth, and substance use did not have a significant time-lagged effect.

Such findings support the theorized moderating role of cognitive control in adolescent risk taking (Ernst et al., 2006; Kim-Spoon et al., 2017) and extend previous cross-sectional studies by demonstrating how low cognitive control could serve as a vulnerability factor to neuroadaptive (i.e., exposure) effects of substance use. For example, self-reported reward sensitivity and punishment sensitivity have been linked to substance use outcomes particularly in conjunction with poor cognitive control (Kahn et al., 2018; Kim-Spoon, Deater-Deckard, Holmes, et al., 2016). Similar patterns have been observed on a neurobiological level, such that insula activation during risk processing predicted health risk behaviors, but only among adolescents with poor cognitive control (Kim-Spoon, Deater-Deckard, Lauharatanahirun, et al., 2016). Whereas previous studies tested only the interaction effects between cognitive control and risk/reward processing to predict substance use outcomes, the current study demonstrated the moderating role of cognitive control in the neuroadaptive effects of substance use on adolescent brain development.

The findings of this study should be interpreted in light of some limitations. First, because this study used a quasi-experimental research design, the effects detected were statistical and cannot be presumed to be causal in nature. Second, although our sample was economically diverse, it was representative of the region where we recruited adolescents with respect to ethnicity/race. Accordingly, further research is warranted for evaluating generalizability of our findings to more ethnically/racially diverse samples. Relatedly, as a neuroimaging study, this study involved relatively large numbers of participants and unprecedented four yearly repeated measures. However, replications of our findings with larger samples are needed to confirm the robustness of the findings given the complexity of the model relative to the sample size. Third, we focused on insular functioning during risk processing. Future research may benefit from examining multiple affective and cognitive networks as well as frontolimbic connectivity. Fourth, we did not measure subjective risk perception during this task (i.e., how risky each choice felt). Whether such subjective perception of risk may uniquely predict substance use, above and beyond the objective outcome variability, is a valuable question to examine in future research. Finally, the found neuroadaptive effects shown by changes in insula activation may reflect not only the neurotoxic effects of substances but also learning processes in which the adolescent’s neural capacities became entrained with maladaptive environments laden with social stressors (Lewis, 2018). As such, fruitful directions for future research include clarifying whether the brain-substance use association may vary depending on the type of substance and elucidating how the interaction between neurobiological and social-environmental contexts is related to substance use. For example, contextual factors such as parent substance use and parenting behavior may moderate the brain-substance use association either for better or for worse.

To our knowledge, this is the first study testing time-lagged bidirectional links between substance use and brain function throughout adolescence. Our prospective longitudinal neuroimaging analyses illustrate that substance use behavior is a temporal leading indicator of insular changes in a sample of typically developing adolescents. Such neuroadaptive effects of substance use were prominent among vulnerable adolescents with poor cognitive control, lending support for the theoretical model that highlights the moderating role of cognitive control (Kim-Spoon et al., 2017). The findings have implications for preventive intervention approaches: Given that the level of cognitive control seems to set the stage for the impact of substance use on the brain across adolescence, the promotion of cognitive control abilities may help to mitigate the detrimental neuroadaptive effects of substance use on brain development.

Supplementary Material

supinfo

Appendix S1. Information on the coefficient of variation (CV).

Table S1. Parametric regressor of coefficient of variation for chosen options during the decision phase at Time 1.

Table S2. Parametric regressor of coefficient of variation for chosen options during the decision phase at Time 2.

Table S3. Parametric regressor of coefficient of variation for chosen options during the decision phase at Time 3.

Table S4. Parametric regressor of coefficient of variation for chosen options during the decision phase at Time 4.

Table S5. Descriptive statistics for raw study variables.

Appendix S2. Growth mixture modeling of cognitive control.

Figure S1. Spaghetti plot of cognitive control data across four waves.

Appendix S3. Bivariate latent change score model of insula activation and substance use using adolescents who were substance use naïve at Time 1.

Key points.

• Differences in brain function have been shown between substance users and non-users.

• Whether neurodevelopmental abnormalities among substance users reflect predisposing neural risk factors or the neuroadaptive effects of the substances is not clearly understood.

• Latent change score modeling analyses suggest that greater substance use predicted increased insula activation, but the effects of insula activation on changes in substance use were not significant.

• The effect of substance use to subsequent changes in insula activation was particularly strong for adolescents with low cognitive control supporting the theorized moderating role of cognitive control.

• Cognitive control may be an important target in the prevention and treatment of adolescent substance use given its modulating role in the neuroadaptive effects of substance use on brain development.