Socioeconomic Risk for Adolescent Cognitive Control and Emerging Risk-Taking Behaviors

Abstract

This study examined whether cognitive control mediated the association between socioeconomic status (SES; composite of income-to-needs ratio and parent education) and changes in risk-taking behaviors. The sample included 167 dyads of adolescents (53% male; Mage = 14.07 years at Time 1) and their parents, assessed annually across four years. Parents reported socioeconomic variables at Time 1. Adolescents reported risk-taking behaviors at Times 1 and 4, and completed a functional magnetic resonance imaging (fMRI) cognitive control task at Times 2 and 3. Lower SES was associated with lower behavioral (but not neural) cognitive control, which was associated with increases in risk-taking behaviors. The findings suggest that elevated socioeconomic risk may compromise cognitive control which can cascade into maladaptive behaviors in adolescence.

Socioeconomic status (SES) is best understood as the overall social standing of an individual or a family, and includes factors such as income, educational attainment, and social status (McLoyd, 1998). Research has shown that individual differences in these factors contribute to cognitive development; specifically, cognitive development among individuals from lower SES families may be altered as a result of growing up in environments that are often laden with stress or lack sufficient stimulation and enrichment (Ursache & Noble, 2016). Relatedly, human neuroimaging work has illustrated the impact of poverty on both structure (Noble, Houston, Kan, & Sowell, 2012) and function (Sheridan, How, Araujo, Schamberg, & Nelson, 2013) of the developing brain. However, much of this work has been conducted in early childhood, leaving a gap in our understanding of socioeconomic vulnerability across different developmental periods.

Due to the neurobiological changes that individuals undergo during adolescence, environmental factors such as SES may have pronounced effects during this developmental stage (Belsky & Shalev, 2016; Somerville & Casey, 2010). Indeed, emerging evidence suggests that socioeconomic disadvantage is related to heightened risk-taking behaviors in adolescence (Delker, Bernstein, & Laurent, 2018). Thus, understanding the mechanisms that explain the effects of socioeconomic disadvantage during adolescence is critical for pinpointing protective factors that prevent development of risk-taking behaviors. The present study examined whether cognitive development may explain the link between socioeconomic disadvantage and risk-taking behaviors in adolescence.

As economic hardship continues to affect many families in the United States, theoretical frameworks have been developed in an effort to better understand how and why SES impinges on child development. Conger, Conger, and Martin (2010) propose an interactionist perspective on SES and family processes which accounts for the role of both social causation (i.e., socioeconomic conditions affecting development) and social selection (i.e., individual traits that affect SES). Here, we focus primarily on the role of social context by considering how socioeconomic conditions impact adolescent cognitive development and adjustment. Within this framework, it has been proposed that parent behavior and functioning associated with family stress (Conger & Conger, 2002) and parent resource investment (Bradley & Corwyn, 2002) may explain why SES affects child and adolescent outcomes. For example, parents in lower SES families may engage in more conflict or provide fewer learning opportunities at home which ultimately compromise adaptive cognitive development. In this way, the consequences of low SES may accumulate over time, affecting adjustment across multiple domains. The developmental cascade approach is a useful methodological framework for modeling these types of cumulative effects in developing systems (Masten & Cicchetti, 2010). For example, this approach may illustrate how low SES contributes to deficits in cognitive development, which then “cascade” into maladaptation in other areas of adjustment (e.g., risk-taking behaviors).

The effects of socioeconomic disadvantage are of particular interest during adolescence, as it is a developmental period marked by heightened engagement in risk taking (Steinberg, 2008). Risk taking includes behaviors that have potential consequences for individual health and well-being, such as substance use, risky sexual behaviors, violence, and delinquency (Dahl, 2004). Conceptually, these behaviors overlap considerably with externalizing problems, such as aggression and rule-breaking. Socioeconomic disadvantage may exacerbate these types of behaviors. For example, lower SES has been consistently linked to the development of externalizing behaviors across childhood and adolescence (Li, Johnson, Musci, & Riley, 2017; Schonberg & Shaw, 2007) and a meta-analysis on the association between SES and antisocial behaviors (including rule-breaking, delinquency, and fighting) demonstrated robust effects in childhood and adolescence (Piotrowska, Stride, Croft, & Rowe, 2015). In addition, lower family income was associated with longitudinal increases in adolescent smoking (Cambron, Kosterman, Catalano, Guttmannova, & Hawkins, 2017) and economic hardship during childhood predicted risk-taking behaviors (including drug and alcohol use, sexual activity, and aggressive behaviors) in adolescence (Delker et al., 2018; Holmes, Brieant, Kahn, Deater-Deckard, & Kim-Spoon, 2019).

The propensity for adolescents from disadvantaged backgrounds to engage in greater risk-taking behaviors may be explained in part by the impact of SES on neurocognitive development. During adolescence, neurodevelopment is marked by the protracted development of prefrontal regions of the brain as well as the earlier and more rapid development of limbic systems (Casey, Jones, & Hare, 2008). The joint development of these two systems and their imbalance is proposed to underlie risk taking behaviors in adolescence. In particular, maturation in prefrontal regions allows for cognitive control, which is the flexible regulation of behavior to override inappropriate responses (Casey et al., 1997). Cognitive control involves both working memory and attention shifting (Carter & Krus, 2012) and is closely related to other constructs of self-regulation such as executive functioning. Functional task-based neuroimaging research has implicated a number of different brain regions that are involved in cognitive control. For example, the anterior cingulate cortex is often activated during cognitive interference, serving to monitor conflicts in information processing (Banich et al., 2000; Botvinick, Nystrom, Fissell, Carter, & Cohen, 1999; Kerns et al., 2004). Similarly, when both conflict monitoring and response inhibition are required, increased activity is observed in the dorsal anterior cingulate cortex, medial prefrontal cortex, and bilateral inferior frontal gyrus (Deng, Wang, Wang, & Zhou, 2018). The right inferior frontal gyrus also supports cognitive control through its involvement in response inhibition (Aron, Robbins, & Poldrack, 2014; Hampshire, Chamberlain, Monti, Duncan, & Owen, 2010).

These emerging neurobiological changes and associated neurocognitive abilities may be affected by environmental conditions such as SES, especially given adolescents’ heightened susceptibility to contextual influences (Blakemore, 2008). Resource scarcity in low SES households may compromise prefrontal brain functioning (e.g., higher-order cognitive processing, planning, and attention) by taxing available cognitive resources, thereby limiting these cognitive control capacities (Mani, Mullainathan, Shafir, & Zhao, 2013). Indeed, lower SES has consistently been linked to lower executive functioning among children (Lawson, Hook, & Farah, 2017). Similarly, in adolescence, poverty has been shown to predict lower cognitive control performance (Lambert, King, Monahan, & McLaughlin, 2017). Together, these findings indicate that family resources such as income and education have important implications for higher-order cognition across development.

A recent meta-analysis examined structural and functional brain correlates of SES, illustrating that lower SES is associated with reduced gray matter volume in regions associated with cognitive control and hypoactivation in executive control networks (Yaple & Yu, 2019). However, there is a dearth of research on the association between SES and cognitive control brain functioning using task-based fMRI in adolescence. Spielberg and et al. (2015) reported that low SES was related to longitudinal increases in cognitive control measured by no-go vs. go activation in the anterior cingulate cortex among early-adolescent females (but not males). These findings suggest that females from lower SES homes require greater compensatory recruitment of the anterior cingulate cortex while engaging in cognitive control. Studies on working memory, which is involved in cognitive control, have demonstrated that individual differences in SES are associated with differences in activation in the prefrontal cortex during working memory tasks (Finn et al., 2017; Rosen, Sheridan, Sambrook, Meltzoff, & McLaughlin, 2018; Sheridan, Peverill, Finn, & McLaughlin, 2017). Thus, there is initial evidence that facets of socioeconomic disadvantage may impact brain functioning related to cognitive control. However, it remains unclear whether these associations then translate to later behavioral outcomes in adolescence.

Since prefrontal functioning is a strong predictor of adolescent risk taking (Steinberg, 2008), it follows that cognitive control may explain the association between socioeconomic risk and risk-taking behaviors. Empirical research on the association between cognitive control and adolescent risk-taking behaviors is still limited, yet a couple of available studies provide initial insight into the nature of these associations. One behavioral study found no direct evidence linking behavioral indices of cognitive control to risk-taking behaviors in pre-adolescence (Romer et al., 2009). In contrast, a neuroimaging study reported that longitudinal increases in neural activation during cognitive control, but not behavioral performance, were correlated with increases in adolescents’ risk taking (McCormick, Qu, & Telzer, 2016). Thus, variation in adolescents’ neural recruitment during cognitive control may have important implications for adolescent risk-taking behaviors.

Taken together, there is considerable evidence for theoretical perspectives that illustrate the influence of social conditions related to SES on adolescent psychosocial adjustment (Conger et al., 2010). However, there is limited empirical evidence demonstrating how socioeconomic conditions impact cognitive control during adolescence, and whether these effects may contribute to the development of maladaptive behaviors. Thus, the current study sought to investigate whether cognitive control accounts for the association between socioeconomic risk and adolescent risk-taking behaviors. Accordingly, we used prospective longitudinal data and a developmental cascade model (Masten & Cicchetti, 2010) to test our hypothesis that lower SES would predict lower cognitive control performance, which in turn would predict longitudinal increases in risk-taking behaviors. We also tested whether neural indices of cognitive control would similarly mediate SES effects on risk taking behaviors. Finally, in light of results from Spielberg et al. (2015) indicating sex differences in the association between SES and cognitive control, we explored differences in these associations between male and female adolescents.

Method

Participants

The current study includes four waves of data collected as part of an ongoing longitudinal study. The sample includes 167 adolescents (53% males) and their primary caregiver (82% biological mothers, 13% biological fathers, 2% grandmothers, 1% foster, 2% other). Adolescents were 13–14 years of age at Time 1 (M = 14.07, SD = 0.54), 14–15 years of age at Time 2 (M = 15.05, SD = 0.54), 15–16 years of age at Time 3 (M = 16.07, SD = .56), and 16–17 years of age at Time 4 (M = 17.01, SD = 0.55), with approximately one year in between each time point. Participants were recruited from rural, suburban, and urban communities in the southeastern United States. Adolescents primarily identified as White (78%), 14% as Black or African-American, 6% as more than one race, 1% as American Indian or Alaska Native, and 1% Asian. Participating caregivers also primarily identified as White (89%), 10% as Black or African American, and 2% as more than one race. In the current sample, median annual household income fell between $35,000-$50,000 for all time points (consistent with the median for the region; United States Census Bureau, 2010), ranging from less than $1,000 to greater than $200,000 per year. Among the parent participants, 34% had a high school degree or less, 24% some college education, 24% bachelor’s degree, and 18% graduate degree. About 67% of families had two parents in the home with an average of 2.77 children per family.

At Time 1, 157 families participated. At Time 2, 10 families were added in order to account for participant attrition between Times 1 and 2, yielding a final sample of 167 parent-adolescent dyads. However, 24 families did not participate at all possible time points for reasons including: ineligibility for tasks (n = 2), declined participation (n = 17), and lost contact (n = 5) during the follow-up assessments. We performed attrition analyses using general linear model (GLM) univariate procedure to determine whether there were systematic predictors of missing data. Results indicated that rate of participation (indexed by proportion of years participated to years invited to participate) was not significantly predicted by demographic covariates (i.e., sex, race, age, and income) or study variables at Time 1.

Of the 150 participants who participated at Time 2, 23 adolescents did not have imaging data for the following reasons: either they were not comfortable or did not meet MRI safety criteria (n=10), excessive motion during scanning (>3mm in any direction; n=11), dropped out of the study halfway through the session resulting in incomplete imaging tasks (n = 1) and imaging artifacts (n=1). Of the 147 participants who participated at Time 3, 38 adolescent participants did not have imaging data for the following reasons: either they were not comfortable or did not meet MRI safety criteria (n=18), unable to come into the lab and completed other portions of the study remotely (n=4), excessive motion during scanning (>3mm in any direction; n=13), poor understanding of the task (n=1), imaging artifacts (n=1), and technical error (n=1). After exclusions, a total of 127 participants were included in fMRI analyses at Time 2 and 109 participants at Time 3. Note that missing data were handled with Full Information Maximum Likelihood (FIML) so all participants were retained in final path analyses.

Procedures

Participants were recruited via flyers, recruitment letters, and e-mail from the community in the Southeastern United States, including small cities and rural towns and counties in Appalachia. Data collection occurred at university offices where adolescents agreed to participate via written assent, while parents provided written consent, and were then administered the protocol by trained research assistants. The procedures took approximately five hours total. Adolescents and their parents received monetary compensation for their time.

Measures

Socioeconomic status.

At each time point, parents completed a demographic interview and reported years of education for themselves and their spouse (if applicable) as well as their annual household income and family size. Income was reported on a 15-point scale from “None” to “$200,000 or more”. The current analyses include SES at Time 1; for the 10 families who were added at Time 2, we used their baseline SES as measured at Time 2. For each family, an income-to-needs (ITN) ratio was calculated (i.e., total household income of the family divided by the poverty threshold for a family of that size, as defined by the U.S. Census Bureau). Greater ITN ratios reflect higher socioeconomic status with ITN ≤ 1 being defined as poor and ITN ≤ 2 as near-poor. According to this ratio, 48% of the sample qualified as poor or near-poor. We used confirmatory factor analysis (CFA) to test the factor structure of SES based on the two indicators of parent education and ITN ratio. The model was fully saturated, and both factor loadings were significant (.78 for ITN ratio and .61 for education, ps < .001). Based on this finding, we standardized and averaged these two indicators into a composite score of SES, with higher scores indicating higher SES.

Risk-taking behaviors.

At each time point, adolescents reported engagement in risk-taking behaviors with the Things I Do scale (Conger, Elder, Lorenz, Simons, & Whitbeck, 1994). The current analyses use data from Time 1 and Time 4. The questionnaire asks about risky things that an adolescent may have done in the past year, including alcohol and drug use and delinquency (such as stealing or fighting). There are 19 total items and responses ranged from “0 = not at all”, “1 = once or twice”, or “2 = more than two times”. Sample items include, “Ridden in a car without a seatbelt”, “Had a fist fight with another person”, “Done something dangerous on a dare”, and “Drunk a bottle or glass of beer or other alcohol”. Mean risk scores were calculated, with higher scores reflecting more risk-taking behaviors. The scale demonstrates acceptable reliability in the current sample (α = .74 at Time 1; α = .84 at Time 4).

Cognitive control.

Cognitive control was measured at each time point with the Multi-Source Interference Task (MSIT; Bush & Shin, 2006). The current analyses use data from Time 2 and Time 3. Participants were presented with sequences of three digits, two of which were identical (see Figure 1A). Participants 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 (e.g., “2” is in the second position in the sequence “020”). In the interference condition, target digits were incongruent with position (e.g. “3” was in the second position in the sequence “131”). Four blocks of 24 interference trials and 4 blocks of 24 neutral trials were interleaved with an interstimulus interval of 1.75 seconds. To assess task performance, we used accuracy and intraindividual variability in reaction time, indexed as intraindividual standard deviations (ISDs; MacDonald, Karlsson, Rieckmann, Nyberg, & Backman, 2012) for correct responses in the interference condition. In addition to behavioral performance, we also monitored hemodynamic response during the task. Hemodynamic response and behavioral performance during the MSIT are negatively associated, such that worse performance (e.g., higher reaction time) in the interference condition is associated with higher activation (Bush et al., 2003).

Figure 1.

A) In the multi-source interference task (MSIT), adolescents were asked to identify the digit that differed from two other concurrently presented digits, ignoring its position in the sequence. B) Interference – neutral t-test at the Time 2-Time 3 combined GLM, displayed un-thresholded and without gray matter mask. C) Mask for cluster ROI, z = 55.0.

Imaging acquisition and analysis.

Functional neuroimaging data were acquired on a 3T Siemens Tim Trio MRI scanner with a standard 12-channel head matrix coil. Structural images were acquired using a high-resolution magnetization prepared rapid acquisition gradient echo sequence with the following parameters: repetition time (TR) = 1200 ms, echo time (TE) = 2.66 ms, field of view (FoV) = 245×245 mm, and 192 slices with the spatial resolution of 1×1×1 mm. Echo-planar images were collected using the following parameters: slice thickness = 4mm, 34 axial slices, FoV = 220×220mm, TR = 2 s, TE = 30 ms, flip angle = 90 degrees, voxel size = 3.4×3.4×4 mm, 64×64 grid, and slices were hyperangulated at 30 degrees from anterior-posterior commissure. Imaging data were preprocessed and analyzed using SPM8 (Wellcome Trust Neuroimaging Center). For each scan, data were corrected for head motion using a six-parameter rigid body transformation and realigned. The mean functional image was co-registered to the anatomical image, then the anatomical image was segmented and registered to the MNI template and functional volumes were normalized using parameters from the segmented anatomical image, and were smoothed using a 6mm full-width-half-maximum Gaussian filter.

Following preprocessing, we estimated a General Linear Model (GLM) for each participant combining data from Time 2 and Time 3 into a single first-level two-session model. For each session, interference and neutral trials were each modeled using a boxcar function convolved with a canonical hemodynamic response function. For each session, the six motion realignment parameters were included as nuisance covariates. Two session box-car covariates were added to account for differences in baseline between session. Volumes with framewise displacement (FD) greater than 0.9mm were censored by adding a volume-specific regressor for each scrubbed volume in the GLM (Power, Barnes, Snyder, Schlaggar, & Petersen, 2012; Siegel et al., 2014). FD was computed from the realignment parameters, with rotational displacement converted into millimeters assuming a sphere of 50mm radius. Finally, a high-pass filter was applied with a cutoff of 168 seconds to model out low-frequency noise. The contrast of interest was interference minus neutral equally weighted across the two sessions. These contrasts of interest were entered into a second-level one-sample t-test to obtain whole-brain maps of voxels with significant differences in interference and neutral mean activation (see Figure 1B). Areas of significant activation are presented in Table 1 and areas of relative deactivation are presented in Table 2. After thresholding the resulting group map at T > 10.0, equivalent to approximately p < 1e-11 family-wise error rate, and applying an inclusive gray-matter mask, two clusters with four peaks were identified as cognitive control regions of interest: a cluster with peaks in the left pre-supplementary motor area, left middle frontal gyrus, and the right cingulate gyrus, and a cluster with peak in the right middle frontal gyrus. A single mask comprising these two clusters was made (see Figure 1C), and then, for each participant, the first eigenvariate values were extracted from voxels inside this mask using the SPM toolbox.

Table 1.

Areas of Significant Activation for the Contrast of Interference Minus Neutral Blocks of the Multi-Source Interference Task at Time 2 and Time 3

Time 2 + Time 3 MSIT Interference - Neutral
				Peak MNI Coordinates
Cluster #		Region	Size	x	y	z	T
1		L Middle Occipital Gyrus	732	−39	−70	−5	18.5
		L Inferior Occipital Gyrus		−36	−82	−5	18.48
		L Middle Occipital Gyrus		−30	−88	1	17.7
2		R Middle Occipital Gyrus	744	33	−88	−2	18.07
		R Inferior Occipital Gyrus		39	−82	−8	17.5
		R Inferior Temporal Gyrus		42	−64	−8	16.98
3		L Pre-Supplementary Motor Area	536	−6	14	46	17.12
		L Middle Frontal Gyrus		−24	−4	58	16.33
		R Cingulate Gyrus		9	14	43	14.12
4		L Inferior Parietal Lobule	630	−33	−46	49	16.65
		L Superior Parietal Lobule		−21	−64	49	16.47
		L Inferior Parietal Lobule		−45	−37	46	16.13
5		R Superior Parietal Lobule	314	27	−61	52	16.34
		R Inferior Parietal Lobule		48	−31	46	11.83
		R Inferior Parietal Lobule		39	−37	43	10.71
6		L Insula	93	−27	20	7	14.04
7		L Thalamus	155	−12	−22	16	13.65
		L Thalamus		−15	−10	13	13.08
		Extra Nuclear		−18	−4	25	11.61
8		R Middle Frontal Gyrus	138	27	−1	58	13.58
9		L Midbrain	44	−6	−25	−8	13.37
		L Midbrain		6	−22	−11	11.17
10		R Putamen	84	30	17	7	13.27
		R Insula		39	14	10	12.38
11		L Inferior Frontal Gyrus	59	−45	2	34	12.97
12		Extra-Nuclear	47	24	−31	16	12.81
		R Thalamus		15	−10	13	11.04
13		R Inferior Frontal Gyrus	19	48	8	31	12.28
14		R Caudate	10	21	−1	22	11.72
15		Cerebellum Posterior Lobe	54	6	−70	−17	11.69
		Cerebellum Posterior Lobe		6	−70	−29	11.39
		Cerebellum Posterior Lobe		−6	−73	−20	10.68
16		Cerebellum Anterior Lobe	15	3	−55	−29	11.42
				30	−34	7	11.05
				−3	5	22	10.14

Note: MNI, Montreal Neurological Institute; L, Left; R, right. Size refers to the number of voxels in the cluster. All activations reported here survive whole-brain family-wise error multiple comparisons correction at a threshold of p < .001. Bolded text indicates the regions that were included in the region of interest.

Table 2.

Areas of Significant Relative Deactivation for the Contrast of Interference Minus Neutral Blocks of the Multi-Source Interference Task at Time 2 and Time 3

Time 2 + Time 3 MSIT Neutral > Interference
				Peak MNI Coordinates
Cluster #		Region	Size	x	y	z	T
1		R Precuneus	393	3	−43	40	16.5
2		R Inferior Parietal Lobule	229	48	−67	40	15.87
		R Inferior Parietal Lobule		57	−64	37	15.16
		R Superior Temporal Gyrus		63	−55	16	11.32
3		Medial Frontal Gyrus	480	6	26	−17	14.29
		R Middle Frontal Gyrus		27	29	52	13.94
		R Superior Frontal Gyrus		6	41	49	13.5
4		L Inferior Parietal Lobule	74	−51	−67	43	13.95
		L Parietal Lobe		−45	−79	37	12.56
5		L Superior Temporal Gyrus	136	−60	−64	28	10.72
		Cerebellum Posterior Lobe		−15	−85	−32	13.53
		Cerebellum Posterior Lobe		−39	−73	−38	13.03
6		R Middle Temporal Gyrus	50	63	−19	−11	12.19
		R Middle Temporal Gyrus		72	−37	1	10.68
7		R Inferior Temporal Gyrus	17	60	−19	−23	10.08
8		R Middle Frontal Gyrus	12	42	38	−11	11.43
9		R Inferior Frontal Gyrus	15	54	26	13	11.39
10		R Middle Temporal Gyrus	46	−66	−55	−5	11.07
		R Precuneus		15	−55	19	11.07
11		R Posterior Cingulate	13	9	−52	10	10.62
12		L Superior Frontal Gyrus	55	−18	32	49	10.83
		L Frontal Lobe		−21	26	40	10.14
13		L Middle Temporal Gyrus	8	−63	−31	−8	10.8
14		L Middle Temporal Gyrus	16	−63	−28	−17	10.38
15		L Temporal Lobe	4	−66	−43	4	10.09
16		L Parahippocampal Gyrus	3	−27	−34	−11	10.46

Note: MNI, Montreal Neurological Institute; L, Left; R, right. Size refers to the number of voxels in the cluster. All activations reported here survive whole-brain family-wise error multiple comparisons correction at a threshold of p < .001.

Plan of Analysis

Skewness and kurtosis were examined for all variable distributions and acceptable levels were less than 3 and 10, respectively (Kline, 2011). Prior to analysis, univariate outliers for study variables were identified, defined as values ≥ 3.29 SD from the mean (Tabachnick & Fidell, 2001). In these cases (n = 8), values were winsorized to retain statistical power and attenuate bias resulting from elimination. Multivariate GLM analyses indicated that demographic variables (i.e., sex, age, number of parents in the home, and race) were not associated with endogenous study variables and thus were not included as covariates (ps > .05).

The hypothesized models were tested via Structural Equation Modeling (SEM) using Mplus version 8 (Muthén & Muthén, 1998–2017). RMSEA values of less than .05 were considered a close fit while values less than .08 were considered a reasonable fit (Browne & Cudeck, 1993), and CFI values of greater than .90 were considered an acceptable fit while values greater than .95 were considered an excellent fit (Bentler, 1990). Little’s MCAR test (Little, 1988) indicated that patterns of missing data on demographic and study variables (i.e., SES, behavioral and neural cognitive control, and risk taking at Times 1 and 4) were completely random (χ² = 37.74, df = 40, p = .57); thus, full information maximum likelihood (FIML) estimation was used to handle missing data. FIML uses maximum likelihood estimation based on all available data, and is superior to alternative missing data methods such as listwise deletion or imputation (Enders & Bandalos, 2001). To test significance levels of mediated effects, asymptotic and resampling strategies were used with bootstrapping, with 10,000 iterations with bias-corrected bootstrap estimations of the 95% confidence interval (Preacher & Hayes, 2008).

Results

For the behavioral data, confirmatory factor analysis (CFA) was used to calculate a factor score using standardized accuracy difference scores (i.e., interference condition – neutral condition) and ISD of reaction time (reverse coded). The model was fully saturated and both loadings were significant at both Time 2 (.57, p < .001) and Time 3 (.63, p < .001). Given the stability of the behavioral indices of cognitive control (r = .53), we averaged Time 2 and Time 3 factor scores to represent the process linking Time 1 SES and Time 4 risk taking. This behavioral indicator of cognitive control was negatively correlated with the selected ROI (r = −.27, p = .005), suggesting that higher cognitive control performance was associated with lower BOLD response during the MSIT, consistent with previous reports (Bush et al., 2003).

Behavioral Indices of Cognitive Control

Table 3 includes descriptive statistics and correlations for all study variables. First, we estimated the longitudinal mediation model examining the indirect effect of socioeconomic risk on adolescent risk-taking behaviors via adolescent cognitive control. We began by estimating all possible paths, yielding a saturated model (i.e., χ² = 0.00, df = 0). We then trimmed all paths that were non-significant and not central to study hypotheses in order to promote parsimony and accuracy in model estimation, as recommended by Little (2013). In addition, because the full model included all possible paths and was a just-identified model, it was not possible to evaluate model fits (i.e., degree to which data were consistent with the hypothesized model). Non-significant paths included the effect of risk taking at Time 1 on cognitive control at Times 2–3 (averaged), and the direct effect of SES at Time 1 on risk taking at Time 4. Trimming these paths did not significantly degrade model fit (Δχ² = 2.56, Δdf = 2, p = .278), therefore this more parsimonious model was chosen as the final model. The final model demonstrated good fit (χ² = 2.56, df = 2, p = .278, CFI = .99, RMSEA = 0.04). Results indicated that lower SES at Time 1 was significantly associated with lower cognitive control at Times 2–3 (b = 0.13, SE = 0.06, p = .031). In turn, lower cognitive control predicted higher levels of risk-taking behaviors at Time 4, after controlling for baseline risk-taking behaviors at Time 1 (b = −0.11, SE = 0.03, p = .001). Furthermore, bias-corrected bootstrapped confidence intervals indicated that the indirect effect of socioeconomic risk on risk-taking via adolescent cognitive control was significant (b = −0.01, SE = 0.01, β = −.04, 95% CI [−0.03; −0.001]). Standardized estimates are presented in Figure 2.

Table 3.

Correlations and Descriptive Statistics for Socioeconomic Status, Cognitive Control, and Risk-Taking Behavior

Variable	1	2	3	4	5	6	7	Range	M (SD)
1. Parents’ Education T1								9.00–25.00	14.84 (2.42)
2. Income-to-Needs Ratio T1	.48**							0.00–8.39	2.49 (1.89)
3. Socioeconomic Status Composite T1	.86**	.86**						−1.73–2.54	0.00 (0.86)
4. Risk-Taking Behaviors T1	−.18*	−.18*	−.20*					0.00–0.68	0.24 (0.17)
5. Risk-Taking Behaviors T4	−.21*	−.15	−.21*	.42**				0.00–1.21	0.36 (0.28)
6. Behavioral Cognitive Control T2-T3	.23**	.07	.17*	−.13	−.30			−1.98–1.26	0.02 (0.65)
7. Neural Cognitive Control T2-T3	.04	.20	.13	.12	.08	−.27**		−0.06–1.05	0.36 (0.21)

Note. T1 = Time 1, T2-T3 = Time 2 and Time 3 combined score, T4 = Time 4.

^*p < .05;

^**p < .01

Figure 2.

Standardized estimates of the longitudinal associations among socioeconomic status, behavioral cognitive control, and risk-taking behaviors in adolescence.

*p < .05; **p < .01; *** p < .001

Neural Indices of Cognitive Control

The longitudinal mediation model was also tested with neural indices of cognitive control in order to examine whether the influences of socioeconomic disadvantage on cognitive control were also present on a neurobiological level. The model with all paths estimated was saturated (i.e., χ² = 0.00, df = 0). Non-significant paths that were not part of the mediating process were trimmed, including the effects of risk taking at Time 1 on cognitive control at Times 2–3 and the direct effect of SES at Time 1 on risk taking at Time 4. Trimming these paths did not significantly degrade model fit (Δχ² = 5.82, Δdf = 2, p = .054), therefore this more parsimonious model was chosen as the final model. However, the final model did not demonstrate adequate fit (χ² = 5.82, df = 2, p = .054, CFI = .86, RMSEA = 0.11). This likely resulted from the fact that neural cognitive control was not significantly associated with the predictor or the outcome variables in the model, indicating that the data did not support the hypothesized model.

Supplementary Analyses

Moderating Effects of Sex.

Using two group SEM, we tested whether the association between SES and cognitive control was significantly different between males and females. For behavioral cognitive control, fixing the path between SES at Time 1 and cognitive control at Times 2–3 to be equal across groups did not significantly degrade model fit (Δχ² = 0.64, Δdf = 1, p = .424), suggesting that the difference between groups was not statistically significant. Similarly, equalizing the path between SES at Time 1 and neural cognitive control at Times 2–3 did not degrade model fit (Δχ² = 0.01, Δdf = 1, p = .934). The results suggest non-significant gender differences in the link between SES and behavioral and neural indicators of cognitive control.

Differential Effects of Education and Income to Needs Ratio.

Based on recent work suggesting differential effects of income and education on brain development (e.g., Brito & Noble, 2018; Ellwood-Lowe et al., 2018), we also tested the model including ITN ratio and parent education as separate (but correlated) predictors. For behavioral cognitive control, the model was fully saturated. Non-significant paths (risk taking at Time 1 on cognitive control at Times 2–3 and direct effects of education and ITN ratio at Time 1 on risk taking at Time 4) were trimmed and doing so did not significantly degrade model fit (Δχ² = 2.74, Δdf = 3, p = .434). The final model demonstrated excellent fit (χ² = 2.74, df = 3, p = .434, CFI = 1.00, RMSEA = 0.00). The effect of parent education (b = 0.07, SE = 0.02, p = .001), but not ITN ratio (b = −0.02, SE = 0.03, p = .464), was significantly associated with cognitive control. Furthermore, the indirect effect of parent education on Time 4 risk taking was significant (b = −0.08, SE = 0.003, β = −.07, 95% CI [−0.02; −0.002]).

For neural cognitive control, the model was fully saturated. Non-significant paths that were not part of the mediating process (risk taking at Time 1 on cognitive control at Times 2–3 and direct effects of parent education and ITN ratio at Time 1 on risk taking at Time 4) were trimmed and doing so did not significantly degrade model fit (Δχ² = 6.95, Δdf = 3, p = .073). However, the model demonstrated poor fit (χ² = 6.95, df = 3, p = .073, CFI = .87, RMSEA = 0.09). Similar to the primary model with the SES composite, the finding seems to reflect the discrepancy between the hypothesized model and the data (i.e., neural cognitive control was not significantly associated with the predictor or the outcome variables).

Discussion

Understanding the mechanisms through which SES may contribute to risk-taking outcomes in adolescence is a critical step in identifying intervention targets for at-risk youth. We tested developmental cascade models to elucidate the prospective associations between socioeconomic risk, cognitive control, and risk-taking behaviors in adolescence. Results demonstrated that behavioral indicators of cognitive control mediated the longitudinal association between SES and adolescent risk-taking behaviors. Specifically, lower SES (a composite of ITN ratio and parent education) was associated with diminished cognitive control, which in turn predicted increases in risk-taking behaviors. This indirect effect was specific to behavioral, rather than neural, indices of cognitive control.

Our findings demonstrate that environments characterized by lower income and lower parental education can adversely affect adolescents’ cognitive control development. This pattern corroborates previous work which has illustrated the detrimental effects that socioeconomic disadvantage can have on cognitive development. Studies investigating these associations in adolescence are rare; however, evidence from early childhood indicates that children exposed to chronic poverty demonstrate lower executive functioning (Raver, Blair, & Willoughby, 2012). Initial work that has been done with early adolescents has similarly demonstrated that lower SES is associated with worse behavioral performance and increased activation in the anterior cingulate cortex during an inhibitory control task, albeit only for females (Spielberg et al., 2015). We extend this work by linking SES and changes in cognitive control to changes in risk-taking behaviors across adolescence.

Different aspects of socioeconomic status such as income, parent education, and neighborhood quality have previously been linked to the development of externalizing behaviors (e.g., Li et al., 2017), behaviors which are conceptually similar to risk taking. However, developmental processes underlying this association in adolescence have yet to be identified. Our findings suggest that the impact of socioeconomic risk on cognitive control may explain why socioeconomically disadvantaged youth become more vulnerable to engaging in risk-taking behaviors. Adolescents from families living in poverty are less likely to receive environmental stimulation when their brain is still malleable and plastic (Bradley & Corwyn, 2002; Rosen et al., 2018), and they may allocate substantial cognitive resources to navigating challenges associated with low SES (Mani et al., 2013). As a result, compared to their same-age counterparts, they may be more limited in their abilities to control impulsive behaviors in socially and emotionally salient situations, and more likely to engage in risk-taking behaviors such as substance use, truancy, stealing, or fighting.

Initial evidence demonstrating the differential effects of distinct components of SES on the brain (e.g., Brito & Noble, 2018; Ellwood-Lowe et al., 2018) led us to also consider whether ITN ratio and parent education, when tested separately, would similarly predict cognitive control. These post-hoc results indicated that parent education, but not ITN ratio, was associated with behavioral indices of cognitive control. Our results seem to support emerging theoretical models illustrating that different components of SES may operate through different pathways to influence child outcomes (Noble et al., 2012). Specifically, it has been proposed that education is more closely linked to parent behaviors that stimulate cognitive development, whereas income is more closely linked to family material resources (Brito & Noble, 2014; Duncan & Magnuson, 2012). Indeed, the family stress model (Conger et al., 2010; Masarik & Conger, 2017) posits that socioeconomic factors exert their influence on child outcomes through parent behaviors more broadly. Thus, parents with higher education may promote cognitive development through providing stimulating learning opportunities or using more complex language, which has downstream consequences for the development of adolescent cognitive control. It will be important for future research to further delineate whether different components of SES have unique effects on neurocognitive development, and what factors underlie these different pathways.

While there was a significant indirect effect of SES on changes in adolescent risk-taking behaviors via behavioral cognitive control, the pattern of effects did not manifest on a neurobiological level. Specifically, frontal activation during the cognitive control task was not significantly associated with risk-taking behaviors. The non-significant link between neural activation during cognitive control and risk-taking behaviors appears to be inconsistent with previous findings (e.g., McCormick et al., 2016) showing that neural correlates of cognitive control were significantly associated with risk-taking behaviors, whereas behavioral performance of cognitive control was not. The discrepancy in the findings may be in part due to differences in study design. That is, McCormick et al. (2016) tested the concurrent associations between year-to-year changes in neural activation and year-to-year changes in risk-taking behaviors, whereas we tested a prospective association between earlier brain activation and subsequent levels of risk taking. Additionally, the limited ecological validity of our cognitive control task may have attenuated any associations between neural cognitive control and risk-taking outcomes. That is, the MSIT is a lab-based task primarily focused on cognitive interference, and neural activation during cognitive interference may not be directly related to adolescents’ risk-taking behaviors in their daily lives. Future research may benefit from evaluating whether different aspects of neural correlates of cognitive control (e.g., cognitive interference, response inhibition, conflict monitoring) may differentially predict adolescent risk-taking behaviors.

We did not find moderating effects of sex in our sample of adolescents for the association between socioeconomic risk and cognitive control behavioral performance or BOLD response. Spielberg et al. (2015) explored possible sex differences in the effects of SES on inhibitory control because of previous findings reporting greater activation in the dorsal anterior cingulate cortex in males relative to females, suggesting better inhibitory control among females. Their finding showing significant effects of SES on behavioral and neural inhibitory control only among females was unexpected, and the authors speculated that females from low SES backgrounds may experience different social stressors than males from low SES backgrounds during early adolescence. These sex differences may also be related to heightened susceptibility to stress in females during pubertal maturation. That is, adolescent females demonstrate heightened reactivity in the hypothalamic-pituitary-adrenal (HPA) axis, which has downstream consequences for the development of brain structure and function (see Bale & Epperson, 2015 for a review). Though a few structural neuroimaging studies have tested SES by sex interactions in relation to brain outcomes with mixed results (e.g., Kim et al., 2019; Whittle et al., 2017), the study by Spielberg and et al. (2015) is the only one that tested the effects SES on brain functioning related to cognitive control. Based on the limited availability of evidence as well as the discrepancy in findings, we suggest that future work consider possible moderation effects of sex in the association between SES and brain development, and also examine possible differential pathways through which SES may influence brain development between males and females.

Limitations of the current study offer potential directions for future research. First, in our developmental cascade model, we included socioeconomic disadvantage measured only at Time 1, in order to maintain temporal precedence across the mediation process (Cole & Maxwell, 2003; Kendall, Olino, Carper, & Makover, 2017). However, recent work has demonstrated that longitudinal change in income may also contribute to neurodevelopment in adolescents (Weissman, Conger, Robins, Hastings, & Guyer, 2018). Examining possible differential effects between levels and change in socioeconomic factors is an important direction for future research. Second, although our sample is economically diverse (with 48% qualifying as poor or near-poor), there is limited racial diversity. Thus, we could not address how socioeconomic disadvantage intersects with other demographic characteristics, such as race and ethnicity, to predict adolescent risk taking. Third, the effect sizes of the paths in our model were relatively small; replication of these results will strengthen our conclusions regarding the association between SES and adolescent risk taking mediated through cognitive control. Finally, not all adolescents exposed to socioeconomic disadvantage will demonstrate altered cognitive development; there are likely important moderators beyond the scope of the current study (e.g. parenting, coping styles) that protect vulnerable adolescents against diminished cognitive control and thus prevent the emergence of risk-taking behaviors.

Our prospective longitudinal analysis presents the first evidence revealing a developmental pathway through which family SES is associated with adolescent risk taking development via cognitive control. The results offer insight into why socioeconomically disadvantaged youth may be more likely to engage in risk-taking behaviors and elucidates these processes during a developmental period when individuals are especially susceptible to environmental influences (Somerville & Casey, 2010). Directing intervention and prevention efforts to specific contextual factors associated with socioeconomic disadvantage (e.g. parenting, household chaos, educational opportunities) may disrupt these associations. For example, a family-focused intervention with families living in poverty improved growth in self-regulation and resulted in increased resting-state connectivity between brain regions involved in socioemotional functioning (Hanson et al., 2018). In these ways, there is potential to positively alter the effects of socioeconomic risk on the development of cognitive control, and ultimately mitigate adolescent engagement in behaviors that are detrimental to physical and psychological health.