Introduction

Adolescence is a life stage characterized by increased prevalence of risk behaviors that can lead to morbidity and mortality. Recent theories based on the neurobiology of adolescent brain development attribute this rise to an imbalance between early-maturing limbic motivational neurocircuitry and later-maturing cognitive control neurocircuitry (Shulman et al. 2016; Somerville and Casey 2010). This developmental imbalance is proposed to be greater during adolescence than during either childhood, when these systems are still developing, or adulthood, when cognitive control circuitry is fully on board and reward-sensitive motivation has returned to pre-adolescent levels (Casey 2015).

Although this model is appealing and has received widespread attention (Shulman et al. 2016; Steinberg et al. 2017), an alternative model known as the Lifespan Wisdom Model (Romer et al. 2017) challenges the proposition that brain development during adolescence entails a normative, universal imbalance between reward and cognitive control systems. Consistent with recent research focused on differentiating different types of adolescent risk-taking (e.g., Maslowsky et al. 2011), the Lifespan Wisdom Model distinguishes between adaptive/exploratory risk-taking (characterized by sensation seeking) and maladaptive forms of risk-taking (characterized by poor impulse control). Maladaptive risk-taking includes behaviors that are more likely to be associated with avoidable and unhealthy outcomes, such as substance dependence and unintended pregnancy. The Lifespan Wisdom Model builds on evidence suggesting that only a sub-group of adolescents engage in such maladaptive risk-taking (Bjork and Pardini 2015; Romer 2010). These adolescents often have early behavioral control difficulties (e.g., disruptive and aggressive tendencies) that can be exacerbated in the context of heightened dopamine expression during adolescence. In support of these claims, many longitudinal studies have found that adolescents with early indicators of behavioral control difficulties are more likely to engage in maladaptive forms of risk-taking with long-term health consequences than those without weakness in behavioral control (Bjork and Pardini 2015; Iacono et al. 2008; Moffitt et al. 2011).

Empirical tests of imbalance models use self-report measures of sensation seeking as a behavioral marker of reward motivation, and find that this biological drive shows a universal peak during adolescence (Duell et al. 2016; Harden and Tucker-Drob 2011; Quinn and Harden 2013; Shulman et al. 2014; Steinberg et al. 2008). The Lifespan Wisdom Model interprets this trend as a rise in an exploratory drive that is developmentally normative and necessary for developing wisdom/gist-based reasoning by trying out and learning from novel experiences (Romer et al. 2017). Furthermore, because sensation seeking tends to be positively related to indicators of cognitive control such as working memory (Khurana et al. 2012), risk-taking driven by sensation seeking is not necessarily characterized by an imbalance between the control and reward systems. Indeed, sensation seeking does not predict maladaptive risk-taking in adolescents controlling for its positive association with impulsivity (Boyer and Byrnes 2009; Khurana et al. 2015a; Magid et al. 2007). The Lifespan Wisdom Model proposes that maladaptive risk-taking in adolescence is attributable to high levels of impulsivity that tend to be clustered in youth with early weakness in cognitive control.

To test the predictions of the Lifespan Wisdom Model requires a longitudinal design in which unique trajectories of sensation seeking and different dimensions of impulsivity can be modeled across adolescence and examined in relation to baseline differences in cognitive control and long-term associations with maladaptive risk-taking outcomes. Previous studies modeling trajectories of these dimensions have either been limited to cross-sectional cohort data (e.g., Steinberg et al. 2008) or have relied on datasets that did not have strong measures of these variables (e.g., Quinn and Harden 2013). This research presents the findings of a study that followed a community cohort of adolescents starting at an average age of 11 through age 18 with repeated assessments of sensation seeking and different forms of impulsivity that permitted the identification of distinct trajectory groups that were examined in relation to baseline differences in working memory (an indicator of cognitive control) and long-term predictive associations with a maladaptive risk-taking outcome, namely substance use disorder.

Of the various dimensions of impulsivity (Whiteside and Lynam 2001), acting-without-thinking and delay discounting, have consistently been linked with maladaptive risk-taking outcomes during adolescence (Romer et al. 2016), including substance abuse and dependence (De Wit 2009; Reynolds 2006). Acting-without-thinking reflects motor impulsivity or “impulsive action”, i.e., acting on the spur of the moment without adequate consideration of the consequences, and is generally assessed using self-report scales like the Barratt (Patton et al. 1995) or Eysenck (Eysenck et al. 1984). Delay discounting is reflective of “impulsive choice” in the context of known risks and rewards, and is assessed using behavioral tasks that measure the ability to choose between two competing rewards, a smaller reward received immediately vs. another that is larger but received with varying delays (Green et al. 1994; Madden and Bickel 2010). Both of these forms of impulsivity are inversely related to indicators of cognitive control, such as working memory (Khurana et al. 2013; Shamosh et al. 2008). Acting-without-thinking is also positively associated with sensation seeking as impulsive action is often driven by rewarding urges (e.g., using an addictive substance) (Khurana et al. 2012; Romer et al. 2011). Delay discounting, however, is not as sensitive to individual differences in sensation seeking given that both choices in the delay discounting task include a reward (Romer 2010; van den Bos et al. 2015; Wilson and Daly 2006).

Current Study

In this study, developmental trajectories of sensation seeking, acting-without-thinking, and delay discounting were modeled to test four critical predictions stemming from the Lifespan Wisdom Model. First, only a subset of youth will exhibit a peak in impulsive action (acting-without-thinking) during adolescence. This sub-group is expected to enter adolescence with pre-existing weaknesses in cognitive control making them more vulnerable to impulsive action in the context of a rising exploratory drive. As a consequence, adolescents in this sub-group are also more likely to engage in maladaptive risk-taking, such as early and progressive drug use that results in substance use disorder. There is mixed support for the prediction of a peak in acting-without-thinking, with some studies observing a peak (Collado et al. 2014; Kasen et al. 2011; Shulman et al. 2016; White et al. 2011) and others observing a decline (Duell et al. 2016; Harden and Tucker-Drob 2011; Quinn and Harden 2013; Steinberg et al. 2008). However, no study to date has examined heterogeneity in adolescent trajectories of acting-without-thinking that might help to identify those at greatest risk for maladaptive outcomes. Given its positive association with sensation seeking and negative association with indicators of cognitive control, the acting-without-thinking dimension of impulsivity distinctly captures the imbalance between the reward and control systems. However, the Lifespan Wisdom Model proposes that a developmental peak in this propensity will only be exhibited by a sub-set of adolescents, especially those who enter adolescence with preexisting weakness in cognitive control. Those without preexisting cognitive control deficits will exhibit low levels of acting-without-thinking throughout adolescence (without any peaks); will not experience an imbalance between the reward and control systems; and will not be at greater risk for substance use disorder.

Second, most adolescents are expected to exhibit a peak in sensation seeking due to the normative rise in dopamine expression during adolescence (Wahlstrom et al. 2010). Previous studies have documented such a peak (Duell et al. 2016; Harden and Tucker-Drob 2011; Quinn and Harden 2013; Romer and Hennessy 2007; Shulman et al. 2014; Steinberg et al. 2008). Nevertheless, qualitative trends in sensation seeking have not been previously explored, likely due to its biological basis. Given its positive correlation with acting-without-thinking, we explored any potential heterogeneity in sensation seeking trajectories in our sample.

Third, delay discounting is expected to decline for most adolescents given the developmental improvements in cognitive control during these years. Since delay discounting is not correlated with sensation seeking, it is unlikely to peak during adolescence. Past studies have reported a decline in delay discounting from childhood to adulthood (Green et al. 1994; Romer et al. 2010; Steinberg et al. 2009; van den Bos et al. 2015), however, these studies are not conclusive owing to their cross-sectional nature. Individual differences in delay discounting are expected in our sample, but these are likely to be present prior to entry into adolescence. Youth with high levels of delay discounting will be at greater risk for substance use disorder (Reynolds 2006). Given lack of empirical or theoretical evidence, we do not have a-priori hypothesis about the presence or absence of distinct trajectory groups for delay discounting.

Finally, based on the Lifespan Wisdom Model, we predict that controlling for its association with acting-without-thinking, sensation seeking is unlikely to predict substance use disorder. These predictions were tested in a study of 387 adolescents followed over a period of eight years from early adolescence (Mean age = 11 ± 0.46 years) to late adolescence (Mean age = 18 ± 0.46 years). We modeled unique trajectories of sensation seeking, acting-without-thinking, and delay discounting, and examined them in relation to baseline differences in cognitive control (assessed using working memory performance) and predictive associations with substance use disorder as a maladaptive risk-taking outcome.

Methods

Present study used data from 387 adolescents recruited as part of the Philadelphia Trajectory Study (PTS) in 2004–2005 and assessed over a period of eight years, including five annual assessments from 2004–2010 (wave 1–5; mean baseline age = 11.41 ± 0.88 years) and a final follow-up after a gap of two years (wave 6; mean age = 18.41 ± 0.64 years). Data from all six waves was included in present analyses. The sample was recruited primarily from schools in the Philadelphia area, and included 52% females, 56% Non-Hispanic Whites, 26% non-Hispanic Blacks, 9% Hispanics, and 9% other races primarily Native American and Asian. Majority of the participants came from low-middle SES backgrounds as assessed by the Hollingshead Two-Factor Index of Social Status (M = 47.0 ± 15.8; reverse scored). Two-thirds of the sample (66%) was from two-parent households with a median parental education of 14 years.

There was 25% attrition across the six waves, with 13% loss to follow-up over the first five waves, and an additional 12% attrition from wave 5 to 6. Missingness was unrelated to participant demographics or key study variables and was handled using Full Information Maximum Likelihood which yields reliable estimates when data are missing at random (Schafer and Graham 2002). The study was approved by the Institutional Review Board of the Children’s Hospital of Philadelphia. Further details about sample recruitment can be found elsewhere (Romer et al. 2009).

Measures

Acting without thinking

Acting without thinking was assessed using a 9-item self-report measure adapted from the Junior Eysenck Impulsivity Scale (Eysenck et al. 1984) that assesses predisposition towards rapid, unplanned reactions to impulsive urges without thinking through the consequences (e.g., do you usually do or say things without thinking?) with binary (Y/N) response options. Responses on the nine items were averaged to create a composite score ranging from 0–1 at each of the six waves. Cronbach α across the 6 waves was 0.74, 0.77, 0.79, 0.80, 0.79, and 0.82, respectively.

Delay discounting

Delay discounting was assessed using a hypothetical monetary choice task where the participant is asked in the context of payment for a job to select an amount between $10 and $90 that if received immediately would be equivalent to receiving $100 six months later (Green et al. 1994). Respondents are initially asked if they would accept an immediate payment of $50. Using an iterative procedure, those who accept/reject this offer are asked if they would accept an amount lower/higher than $50 in $10 decrements. Scores on this variable ranged from 10–100, which were reverse-scored such that higher scores were indicative of greater discounting. The test-retest reliability of this measure across the assessments was high (r’s = 0.42–0.52, p < 0.001). Delay discounting was assessed starting at wave 3 of the study; thus, we were only able to use the last four waves (waves 3–6) of data to model trajectories.

Sensation seeking

Sensation seeking was assessed using respondents’ level of agreement with four items (e.g., “I like to do frightening things”), each on a Likert-type scale ranging from 1 (strongly disagree) to 4 (strongly agree). These items were taken from the Brief Sensation Seeking Scale (Hoyle et al. 2002) and represented each of the four dimensions (i.e., experience seeking, boredom susceptibility, thrill/adventure seeking, and disinhibition) of the original sensation seeking scale (Zuckerman 1971). Cronbach α across the 6 waves was 0.74, 0.73, 0.76, 0.79, 0.81, and 0.79, respectively. Participants’ scores on the four items were averaged to create a composite index ranging from 1 to 4 for each of the six waves.

Working memory

Individual differences in cognitive control were assessed at wave 1 using performance on the following working memory tasks that were largely nonverbal and thus not dependent on differences in reading comprehension: (1) Digit span backwards (2) Corsi-block tapping (3) Letter-two-back, and (4) a spatial working memory task. The four tasks loaded significantly on a single latent factor, with loadings ranging from 0.40–0.60. All four tasks (described below) have been linked to activation in executive control brain regions and are reliable indicators of individual differences in cognitive control abilities (Romer et al. 2009). Cognitive control is typically defined as the ability to engage in top-down control over behavior and involves active maintenance of goal-relevant information. As such, individual differences in working memory performance are a sensitive indicator of cognitive control abilities (Engle 2002; Miller and Cohen 2001). Other executive functions like inhibitory control (measured using Go/No-Go, Stop signal tests) are also used as indicators of cognitive control, but they tend to be highly correlated with working memory measures because successful inhibition relies on active maintenance of what is goal-relevant (Munakata et al. 2011). We chose working memory because it is a more sensitive indicator of the ability to actively maintain goal relevant information while ignoring distractions and suppressing irrelevant information. The latent working memory factor was included as a predictor of sensation seeking, acting without thinking, and delay discounting trajectories.

Digit span

This task tests the auditory-verbal working memory of participants by having them repeat back in reverse order, sequences of digits to the experimenter. The test was administered in standard format according to the procedures listed in the Wechsler Intelligence Scale for Children–Fourth Edition (WISC-IV) manual (Wechsler 2003).

Corsi-block tapping

This task is a non-verbal variant of the digit span task (Milner 1971). Participants view a set of identical blocks that are spatially dispersed on the screen, and are individually lit up in a random sequence. Participants are asked to tap each box in the reverse order of the sequence of lit boxes. This task assesses spatial working memory as the visual sequence must be maintained and reversed in working memory in order to guide the response.

Letter two-back

This task involves monitoring a series of letters for a repeat “two-back.” Letters are presented for 500 ms each, separated by a 1 s interval. Participants must continually update their working memory in order to compare the current letter to the letter shape presented two trials back. This task was adapted for children by Casey et al. (1995).

Spatial working memory

This self-directed computerized task requires the participant to search for hidden tokens one at a time within sets of four to eight randomly positioned boxes. Working memory skills are tapped as the participant while searching must hold in working memory the locations already checked and as tokens are found, must remember and update information about the locations of those tokens (Owen et al. 1990). Between-search errors are made if the participant returns to a box where a token had already been found during a previous search sequence, and was used as a measure of working memory performance.

Substance use disorder (SUD) severity

SUD severity was assessed at wave 6 using indicators of abuse and dependence from the DSM-4 (SAMSHA 2011). Participants who reported use of a specific drug in the past year were asked questions pertaining to abuse (e.g., did you continue to drink alcohol even though you thought your drinking caused problems with family or friends?) and dependence (e.g., did you need to drink more alcohol than you used to in order to get the effect you wanted?) related to that specific drug from the National Survey on Drug Use and Health (NSDUH). NSDUH items are based on the DSM-4 criteria. DSM-5 has replaced the abuse and dependence classifications with SUD. We matched the DSM-5 criteria to questions from the NSDUH to generate SUD criterion scores for each substance (alcohol, marijuana, tobacco). DSM-5 defines mild disorder as meeting 2–3 criteria. Given the relatively low criterion scores in our sample (only about 25.2% of our sample reported sufficient criteria for a mild SUD), we used the continuous criterion scores for alcohol, marijuana, and tobacco as indicators for the latent SUD factor. Given the co-morbidity in early-onset SUD (Jackson et al. 2000) and evidence of a single underlying factor for consumption, dependence, and abuse (Jackson et al. 2014), we used the latent SUD severity factor as our main outcome variable. Confirmatory factor analyses revealed that the SUD criterion scores for the three drugs did indeed load on a single latent factor (Khurana et al. 2017), with loadings ranging from 0.31–0.65.

Analytic technique

Latent Growth Class Analysis (LGCA) was used to examine potential heterogeneity in the developmental trajectories of sensation seeking, acting without thinking, and delay discounting (Muthén and Muthén 2000). Several indices of model fit were employed to compare between models with increasing number of classes, including the Bayesian information criterion (BIC), Akaike information criterion (AIC), Lo–Mendell–Rubin likelihood ratio test (LMR-LRT), bootstrap LRT, and entropy. Final decisions were made based on evaluation of model fit criteria as well as the interpretability of the latent classes. If there was no evidence of heterogeneity, developmental trends were modeled using latent growth curve modeling (LGCM).

To test predictions regarding the trajectory groups, we examined their relation with working memory and SUD in a structural equation modeling framework. If there was evidence of significant heterogeneity in personality dimensions, then we used class membership for that particular dimension as the predictor variable. Participants were assigned to the latent class to which they had the highest likelihood of belonging based on their posterior probabilities. If there was no significant heterogeneity, then latent intercept and slope factors were used as predictors of SUD.

Results

Mean scores of sensation seeking, acting without thinking, and delay discounting across all waves are provided in Table 1, and correlations in Table 2. Mean values of acting without thinking showed an increase from waves 1–4, with a decline thereafter. Delay discounting showed a gradual decline on average across the last 4 waves, with a steeper drop from wave 5 to 6. Mean sensation seeking scores increased from wave 1–5, followed by a decline from wave 5 to 6. Overall, the average trends were similar for boys and girls, although boys reported significantly higher acting without thinking at wave 4, and significantly greater sensation seeking across all waves, as compared to girls. At a bivariate level, SUD risk was greater among males as compared to females, B(SE) = −0.18 (0.07), p = 0.02.

Table 1 Mean (SD) scores of AWT, SS, and DD across the six waves
Full size table
Table 2 Correlation matrix of repeated assessments of sensation seeking and impulsivity
Full size table

Latent Growth Modeling Results

Acting without thinking

Analysis of latent trajectories of acting without thinking yielded clear evidence of heterogeneity with a two-class solution providing the best fit. Compared to the 3-class solution, this model had smaller log-likelihood value (−355.46 vs. 3.16) and a non-significant bootstrap likelihood ratio test (p = 1.0), suggesting that the more parsimonious 2-class solution provided a better fit to the data. Visual inspection of the observed and estimated trajectories also suggested that there were only two qualitatively distinct patterns. The latent classes were well defined and distinguishable based on their unique change patterns as well as the average posterior probabilities of class membership (0.94, 0.96), and high entropy scores (0.83). The first class (44%; n = 171) was labeled as “High Increasing” given that members of this group had higher scores at baseline, with a quadratic growth pattern that peaked around wave 4 and declined thereafter (see Fig. 1). The “Low stable” group (56%; n = 216) was characterized by lower levels at baseline and a relatively flat trajectory across the six waves. Females were less likely to be in the “High increasing” acting without thinking class as compared to males, but this difference was not significant at p < 0.05, B (SE) = −0.39 (0.23), p = 0.09. Adolescents from high SES backgrounds were significantly less likely to be in the “High increasing” acting without thinking class as compared to their peers from low SES backgrounds, B (SE) = −0.03 (0.01), p = 0.001. Adolescents in the “High increasing” acting without thinking class had lower working memory score, B (SE) = −0.08 (0.03), p = 0.02, and higher SUD risk, B (SE) = 0.12 (0.04), p = 0.001, as compared to those in the “Low stable” group.

Fig. 1
figure 1

Latent trajectory classes of acting without thinking (AWT) from waves 1–6

Full size image

Delay discounting

Analysis of latent trajectories of delay discounting using four waves of data (waves 3–6) yielded evidence for significant heterogeneity. Specifically, the 4-class solution provided the best fit to the data, as indexed by lower AIC values (3 vs. 4 class = 12459 vs. 12439) and lower BIC values (3 vs. 4 class = 12505 vs. 12498), significant LMR-LRT (4 vs. 3 class = p < 0.05; 3 vs. 2 class = p < 0.001), significant bootstrap LRT (4 vs. 3 class = p < 0.001; 3 vs. 2 class = p < 0.001). The average posterior probabilities of class membership (0.83, 0.91, 0.78, 0.81) and entropy score (0.69) were reasonably high. The latent classes were well defined and had distinct change patterns. The first class (34%; n = 125) was defined as “Low Stable” given that participants in this class had low rates of delay discounting at baseline, with flat trajectories across the 4 waves. The second class (14%; n = 52) was labeled “High Stable” given that they had high delay discounting scores at baseline that remained stable across the 4 waves. The third class (14%; n = 52) was characterized by an increasing trend of delay discounting across the 4 waves and was labeled as “Low Increasing”. Finally, the fourth class (38%; n = 138) was characterized by high delay discounting scores at baseline with a gradual decline over time and was labeled as “High Declining” (See Fig. 2). There were no significant gender differences in delay discounting class membership. However, SES effects were significant, with adolescents from high SES backgrounds being less likely to be in the “High stable” group, B (SE) = −0.07 (0.02), p = 0.001, and less likely to be in the “High declining” group, B (SE) = −0.03 (0.01), p = 0.004, as compared to the “Low stable” delay discounting group. Further, adolescents in the “High stable” group had lower working memory score, B (SE) = −0.09 (0.03), p < 0.001, and higher SUD risk, B (SE) = 0.07 (0.03), p = 0.001, as compared to those in the “Low stable” group.

Fig. 2
figure 2

Latent trajectory classes of delay discounting (DD) from waves 3–6

Full size image

Sensation seeking

There was no significant heterogeneity in sensation seeking trajectories modeled using six waves of data, suggesting a more universal developmental progression in this dimension. The average trend modeled using LGCM was characterized by a quadratic curve, with a gradual increase in sensation seeking across the early years, a peak around wave 5, and a declining trend thereafter. The mean trajectory depicted in Fig. 3 had significant mean intercept, linear and quadratic slope factors (MInt = 2.10, MSlp = 0.14, Mquad = −0.01; ps < 0.001) and variances (VarInt = 0.31, VarSlp = 0.06, ps < 0.001; Varquad = 0.001, p = 0.001). The intercept and slope factors were significantly correlated, r = −0.30, p < 0.001. Gender was a significant predictor of individual differences in the intercepts, B(SE) = −0.19 (0.05), p < 0.001; β = −0.23, and slopes, B(SE) = −1.18 (0.37), p = 0.001; β = −0.24, with females reporting lower levels than males. SES differences were not significantly associated with sensation seeking intercepts and slopes. Further, at a bivariate level, sensation seeking intercepts were positively associated with working memory, B(SE) = 0.09 (0.04), p = 0.03, and SUD risk, B(SE) = 0.17 (0.07), p = 0.01.

Fig. 3
figure 3

Latent trajectories of sensation seeking (SS) from waves 1–6

Full size image

Final Model with Working Memory and SUD

In the final model, we examined associations of observed trajectory groups of acting-without-thinking and delay discounting, and latent intercepts and slopes of sensation seeking with (a) baseline differences in working memory (as predictor) and (b) rates of SUD at final follow-up (as outcome). As shown in Fig. 4, adolescents in the “High increasing” acting without thinking group, B(SE) = −0.43 (0.18), p = 0.02, and “high stable” delay discounting group, B(SE) = −0.80 (0.23), p = 0.001, had lower levels of working memory (at wave 1) suggesting that adolescents who experienced a peak or high levels of impulsivity had early indicators of weak cognitive control. Individual differences in latent intercepts of sensation seeking were positively related to working memory, B(SE) = 0.09 (0.04), p < 0.05; but not the latent slopes of sensation seeking, B(SE) = 0.003 (0.01), p = 0.78. Controlling for gender, the association between sensation seeking intercepts and working memory dropped in significance, B(SE) = 0.08 (0.04), p = 0.06 (see Fig. 4). Due to the high correlation between the linear and quadratic slope factors of sensation seeking (r = −0.89, p < 0.001), and related model convergence issues, the variance in the quadratic factor was fixed to 0.

Fig. 4
figure 4

Final model showing significant pathways of influence with WM and SUD in relation to latent trajectories of AWT, DD, and SS. Note. WM working memory, AWT acting without thinking, DD delay discounting, SUD substance use disorder. *p < 0.05, †p = 0.06. Sensation seeking slope was included in model but it was not related to working memory or SUD risk

Full size image

In the same model, the latent SUD factor was regressed on acting without thinking class membership (2 classes with “Low stable” as reference group), delay discounting class membership (4 classes with “Low Stable” as reference group), and latent intercepts and slopes of sensation seeking. Adolescents in the “High increasing” acting without thinking class were more likely to develop SUD at wave 6 as compared to those in the “Low Stable” class, B(SE) = 0.45 (0.22), p = 0.04. Similarly, adolescents in the “High Stable” delay discounting class had significantly higher rates of SUD as compared to the “Low Stable” reference class, B(SE) = 0.96 (0.40), p = 0.02. The “Low Increasing” and “High Declining” delay discounting groups were not at significantly greater risk for developing SUD as compared to the “Low Stable” reference class. We also explored delay discounting trajectory class differences by defining the “High Declining” class as the reference group considering that 38% of our sample belonged to this class, and prior research suggests a gradual decline in delay discounting trajectories. When doing so, the “High Stable” delay discounting group was still a significant predictor of SUD, B(SE) = 0.71 (0.33), p = 0.03, as compared to the “High Declining” reference class. The remaining delay discounting classes were not significant predictors of the SUD outcome. Individual differences in latent intercepts and slopes of sensation seeking did not predict the SUD outcome, B(SE) = 0.40 (0.28), p = 0.14 and B(SE) = 0.87 (1.67), p = 0.60, respectively. Working memory did not have a significant direct effect on the SUD outcome, B(SE) = −0.16 (0.16), p = 0.32. Model effects depicted in Fig. 4 were significant controlling for gender, race-ethnicity, and SES, all of which were non-significant except that Black adolescents in our sample had significantly lower SUD risk as compared to their White counterparts, B(SE) = −0.77 (0.26), p = 0.003.

We also tested alternate models where we regressed the residuals of individual drug criterion scores for alcohol, marijuana, and tobacco, on the trajectory classes for acting without thinking, delay discounting, and latent intercepts and slopes of sensation seeking. The rationale for testing these effects was that there could be unique associations of the individual criterion scores (i.e., alcohol, marijuana, tobacco criterion scores) with impulsivity and sensation seeking trajectories that were not captured by the latent factor. There were no significant associations between the impulsivity and sensation seeking trajectories and criterion score residuals suggesting that whatever longitudinal associations were present between the impulsivity trajectory classes and SUD are being captured by the latent factor. There was no left-over variance related to impulsivity or sensation seeking trajectories that was unique to a particular drug.

Discussion

Imbalance models of adolescent risk-taking (e.g., the dual systems model of Steinberg et al. 2008 and maturational imbalance model of Casey et al. 2008) attribute the rise in adolescent risk-taking to a universal imbalance between heightened reward sensitivity (associated with an early maturing limbic brain system) and weak cognitive control (associated with protracted development of the prefrontal brain system). Despite the popularity and heuristic appeal of these models, a number of critiques have challenged their overly-simplistic interpretations of adolescent risk taking (e.g., Pfeifer and Allen 2012; Romer 2010). The present study empirically tested predictions of an alternative Lifespan Wisdom Model (Romer et al. 2017), which emphasizes the exploratory nature of adolescent risk taking and challenges imbalance accounts of adolescent risk-taking by predicting that only a sub-set of adolescents experience an imbalance that is characterized by heightened reward sensitivity and weak cognitive control. For these youth, heightened sensation seeking during adolescence, a form of reward sensitivity, in combination with early weakness in cognitive control results in a peak in one type of impulsivity (acting without thinking). Another form of impulsivity (e.g., delay discounting) that is not as sensitive to heightened sensation seeking is predicted not to exhibit a peak during adolescence. Regardless, both forms of impulsivity reflect weakness in cognitive control, and it is this underlying weakness that is expected to predict maladaptive risk-taking outcomes in adolescence and beyond (Moffitt et al. 2011; Romer et al. 2017). Although sensation seeking is expected to peak during adolescence due to the rise in dopamine expression (Wahlstrom et al. 2010), this peak is not characteristic of an imbalance and is not expected to predict maladaptive risk-taking outcomes in adolescents apart from high levels of impulsivity.

To test this model, we identified developmental trajectories of sensation seeking and two dimensions of impulsivity (acting without thinking and delay discounting) using six waves of data from early to late adolescence and examined these trajectories in relation to baseline differences in cognitive control and a maladaptive risk outcome (substance use disorder) at final follow-up. Overall, we found significant heterogeneity in developmental trajectories of acting without thinking and delay discounting, with only some adolescents experiencing a peak or persistent high levels of impulsivity. Sensation seeking, on the other hand, showed an average curvilinear trend for all adolescents, with a peak around mid-adolescence. Sensation seeking intercepts were positively correlated with baseline working memory scores. In contrast, trajectory groups with high levels of acting without thinking and delay discounting were negatively associated with working memory and were the only groups at increased risk for substance use disorder.

Our findings are consistent with the predictions of the Lifespan Wisdom Model. In case of sensation seeking, we found a universal peak around mid-adolescence, with only quantitative deviations from the average trend. For acting without thinking and delay discounting, sub-groups with distinct change patterns were observed. Some adolescents had low levels at baseline and stayed low throughout adolescence, while others evidenced high baseline levels and escalating or peaking trajectories.

For the most part, delay discounting scores either declined or remained stable across adolescence (72%). This is consistent with the predictions of the Lifespan Wisdom Model in that the ability to engage in rational decision-making in the context of known risks and rewards shows a gradual increase across adolescence (Defoe et al. 2015). The “high stable” and “low stable” delay discounting groups may reflect early levels of impulsivity that continue to persist (Casey 2015; Casey et al. 2011). The high stable delay discounting group was the only group at greater risk for substance use disorder. The high declining group was the largest sub-group comprised of 38% of the sample. Adolescents in this group showed a gradual decline in delay discounting, consistent with the maturation of cognitive control. The presence of a low increasing group is a novel finding—the adolescents in this group showed an increased preference for smaller immediate rewards across the waves. It is possible that their choices are not an indication of their cognitive maturity, but instead reflect choices that are rational given their contexts. For instance, Kidd and colleagues (2013) and McGuire and Kable (2013) have found that for children whose contexts are unpredictable or risky, choosing the smaller immediate reward is the rational option.

A sub-set of youth (44%) exhibited a peak in acting without thinking around mid-adolescence. This group had preexisting weakness in cognitive control at baseline, which likely made it more challenging for them to engage in impulse control (hence the peak) in the context of rising dopamine levels. This group also reported significantly higher levels of substance use disorder. The remaining 56% of the sample had low levels of acting without thinking throughout adolescence. As such, these adolescents did not experience a propensity to engage in impulsive action driven by an imbalance between the reward and control systems. They were also not at higher risk for substance use disorder. These findings are in line with the propositions of the Lifespan Wisdom Model which argues that the imbalance that characterizes maladaptive risk-taking is only experienced by a sub-set of adolescents (see also Bjork and Pardini 2015). It is not a universal phenomenon. Our longitudinal design allowed us to test these predictions that may not always appear in studies that only examine aggregate trends. The present findings demonstrate that it is misleading to characterize adolescent risk-taking tendencies based on aggregate age trends, especially those that focus on maladaptive outcomes.

Although sensation seeking was positively related to acting without thinking, individual variation in trajectories of sensation seeking did not predict substance use disorder risk apart from acting without thinking, and baseline differences in sensation seeking were positively related to working memory. This finding is consistent with the interpretation of sensation seeking as an indicator of rising dopamine expression that leads to heightened reward sensitivity (Wahlstrom et al. 2010) but that is also a critical modulator of working memory performance (Cools and D’Esposito 2011). These patterns are also consistent with the Lifespan Wisdom Model, which interprets the rise in sensation seeking as an indicator of developmentally normative and adaptive exploration that characterizes the adolescent period (Romer et al. 2016). Although sensation seeking can be positively associated with risk taking, such as experimentation with drugs, it tends not to lead to maladaptive outcomes like substance use progression (Khurana et al. 2015a) or dependence (Khurana et al. 2017) apart from its association with acting without thinking.

Our findings in regard to acting without thinking diverge from some studies that have reported a decline in this tendency during adolescence (Harden and Tucker-Drob 2011; Steinberg et al. 2008). This discrepancy may be due to the use of limited measures of acting without thinking that may not have adequately captured this dimension, or reliance on a cross-sectional cohort design. For instance, although Harden and colleagues (2011) analyzed a national longitudinal data set, their sample over-represented older youth, which could be another reason why they failed to adequately capture the variability in this impulsivity dimension at younger ages. Research using longitudinal community samples (similar to ours) with robust measures has reported a quadratic trend with a peak around mid-adolescence (Collado et al. 2014; Kasen et al. 2011; White et al. 2011).

By exploring heterogeneity in sensation seeking, acting without thinking, and delay discounting change during adolescence and examining these trajectories in relation to cognitive control and a maladaptive risk outcome, our findings suggest that imbalance models overgeneralize patterns of poor impulse control that only characterize a subset of youth (see also Bjork and Pardini 2015). Furthermore, it is common to attribute imbalance to the rise in sensation seeking, but as we find, this behavioral indicator is not related to weakness in cognitive control and the rise in sensation seeking by itself does not predict substance use disorder or other negative health outcomes (Khurana et al. 2013, 2015a, b, 2017).

It is also important to note that we did not find gender differences in trajectory groups of acting without thinking and delay discounting, although we did find that males had higher rates of acting without thinking at some assessment time points. In case of sensation seeking, males reported higher intercepts and faster growth in sensation seeking than females. Other studies have similarly found higher rates of sensation seeking among males than females (Cross et al. 2011; Romer and Hennessy 2007; Shulman et al. 2014), and higher levels of impulsivity among males than females (Côté et al. 2002; Shulman et al. 2014). Still others using community-based samples of adolescents have found no gender differences in trajectories of sensation seeking and impulsivity (Collado et al. 2014), or faster growth in sensation seeking among females than males (Littlefield et al. 2016). These inconsistencies could be due to the variability in measures used across studies, or sample-specific effects. In terms of risk for substance use disorder, male and female adolescents tend to have similar rates (Wagner and Anthony 2007), as was the case in our sample where we found that the effect of gender on substance use disorder risk dropped to non-significance, B(SE) = −0.16 (0.23), p = 0.48, controlling for other predictors.

The following limitations should be noted when interpreting current results. First, although we assessed a large and diverse community sample of adolescents from mean ages 11 to 18, our sample may not be representative of all youth, especially those from higher SES backgrounds. Second, we used substance use disorder as the risk behavior outcome given its public health significance; as such our findings cannot be generalized to other outcomes even though similar associations of sensation seeking and impulsivity have been reported with progressive substance use, and risky sexual behaviors (Khurana et al. 2015a, b). Finally, we could only model trajectories for delay discounting using four waves of data, given that this measure was introduced at wave 3 of the study. Nevertheless, even if there was an earlier peak in delay discounting that we did not observe, only one of the delay discounting trajectory groups predicted substance use disorder risk, suggesting that peaks in this form of impulsivity (if any) are unlikely to be a major source of later maladaptive health. Indeed, the dominant pattern of change in delay discounting is a gradual decline from childhood to early adulthood (Green et al. 1994; Steinberg et al. 2009; van den Bos et al. 2015).

Conclusion

Developmental imbalance models attribute the rise in risk-taking during adolescence to a universal imbalance between rising reward sensitivity and lagging cognitive control. An alternate Lifespan Wisdom Model distinguishes between exploratory/adaptive (e.g., sensation seeking) and maladaptive (e.g., acting-without-thinking, delay discounting) risk-taking propensities and attributes the latter to a sub-set of youth with weak cognitive control. Present findings support the Lifespan Wisdom Model’s hypothesis that the risk-taking characterized by an imbalance between the reward and control systems is a phenomenon restricted to a subset of youth. Further, these youth appear to possess this vulnerability of weak top-down behavioral control prior to adolescence (cf. Iocono et al. 2008; Moffitt et al. 2011) and continue to do so beyond this age period. The finding that only adolescents with high levels of impulsivity were at risk for substance use disorder also suggests that individual differences in cognitive control are likely to get amplified during adolescence (given the concomitant rise in reward seeking) and may play a more important role in predicting maladaptive risk-taking. Although a developmental peak in sensation seeking was observed in our sample, it was unrelated to maladaptive risk outcomes. Taken together, our results suggest that it is possible to identify youth with either emergent or early forms of impulsivity so that interventions may be targeted to the groups most at risk for negative outcomes.