Understanding Patterns of Heterogeneity in Executive Functioning during Adolescence: Evidence from Population-Level Data

Abstract

Executive functioning (EF) is fundamental to positive development. Yet, little is known about how to best characterize constellations of EF skills that may inform disparate associations between EF and behavior during adolescence. In the current study, cross-validated latent profile analysis (LPA) was used to derive profiles of EF based on measures of inhibitory control, working memory, and cognitive flexibility using data from 11,672 youth (52.2% male, mean age = 9.91 years) in the Adolescent Brain and Cognitive Development study. Four meaningful EF profiles emerged from the data representing Average EF, High EF, Low Inhibitory Control, and Low EF. Boys, youth from low-income households, and early developing youth were more likely to be in profiles distinguished by lower EF. Profile membership also predicted differences in externalizing, internalizing, and other problem behaviors assessed one year later. Findings indicate that youth may have distinct constellations of EF skills, underscoring the need for person-centered approaches that focus on patterns of individual characteristics.

Executive functioning (EF) describes a set of higher-order cognitive abilities such as inhibitory control, working memory, and cognitive flexibility that guide goal-directed behaviors (Miyake et al., 2000). EF continues to develop over adolescence and is fundamental to positive youth development (Blakemore & Choudhury, 2006). Indeed, EF predicts a host of behavioral and clinical outcomes including academic achievement (Miller et al., 2012), social skills (Miller & Hinshaw, 2010), and wellbeing (Cotrena et al., 2016). Deficits in adolescent EF have also been linked to substance use (Kim-Spoon et al., 2020), depression (Melendez-Torres et al., 2016), and conduct disorders (Nigg & Huang-Pollock, 2003) and are considered a key transdiagnostic risk factor for the development of psychopathology (Nolen-Hoeksema & Watkins; Snyder et al., 2015)

Although research on EF has flourished in recent years, fundamental questions about the measurement and modeling of adolescent EF remain. For instance, most studies report response latencies (i.e., response times), accuracy, or number of errors. Even though these metrics are correlated, they are also dissociable and may predict different outcomes (Chaytor et al., 2006). Further, evidence suggests there is notable heterogeneity in EF, leading to significant individual differences at any given point in development (Chaku et al., 2021; Vaidya et al., 2020; Vandenbroucke et al., 2018). To date, researchers have used several methodological approaches to investigate this heterogeneity. For example, variable-centered approaches (e.g., factor analysis) suggest that multiple EF dimensions can be represented by one to three latent factors reflecting the conceptual overlap between different EF skills and the empirical distinction between the EF tasks measuring them (Miyake et al., 2000). However, EF factor structures have not always replicated, especially in adolescence (Karr et al., 2018). The structure of adolescent EF is complex and multidimensional, reflecting the rapid and complex biological, cognitive, and social maturation occurring in the second decade of life (Blakemore & Choudhury, 2006; Crone & Dahl, 2012).

Efforts to identify “at-risk” youth with poor EF or to develop effective EF interventions for adolescents may be hampered by current variable-centered approaches which assume normative development, obscure individual differences, and may not align with other neurodevelopmental processes (Beltz, 2018; Crone & Elzinga, 2015). Person-centered approaches can be used to fill this knowledge gap, providing insight into how different patterns of development co-occur, uniquely characterize individuals or subpopulations of individuals, and predict outcomes (Bauer, 2008). The current study used latent profile analysis (LPA) to identify and characterize constellations of EF skills in a population-level sample of early adolescents and then examined how these emergent profiles were associated with sociodemographic characteristics and behavioral outcomes assessed a year later. The LPA solution was then compared to a variable-centered approach (i.e., factor analysis) to examine the predictive validity and added utility of the LPA approach.

Variable-Centered Approaches to the Study of Adolescent Executive Functioning

Confirmatory factor analysis (CFA) is the most common measurement model of adolescent EF (Camerota et al., 2020). CFA and other similar latent variable approaches are designed to elucidate the underlying structure of observed phenomena by explaining patterns of covariation among observed (i.e., directly measured) variables (Brown & Moore, 2012). Within the EF literature, CFAs are used to extract common variance across EF tasks which then defines the latent construct (Harrington, 2009). Thus, what is shared across all EF tasks is considered the measure of underlying EF ability and task-specific variance is considered measurement error (e.g., Huizinga et al., 2006). Seminal work within the CFA framework suggests that EF is comprised of three fundamental skills: inhibitory control (i.e., the ability to resist prepotent responses), working memory (i.e., the ability to update and maintain information), and cognitive flexibility (i.e., the ability to shift between alternative sets of mental operations) which load onto a common EF factor (Miyake et al., 2000). A key strength of this approach is that higher-order EF constructs tend to have better predictive validity than single EF measures particularly for broad constructs such as academic achievement or mental health (Brown & Moore, 2012; Miyake & Friedman, 2012). This factor structure reliably emerges across different sample populations (e.g., normative and clinical samples; Gioia et al., 2002), study designs (Wiebe et al., 2008) and assessment methods (de Frias et al., 2006; Salthouse et al., 2003), and has led to important insights on the nature and development of EF in adolescence (Chen et al., 2013; Prencipe et al., 2011; Thompson et al., 2019).

Despite the popularity of CFAs and other latent variable approaches, however, concerns persist. First, factor structures do not always replicate across different adolescence samples. In a recent review, Karr et al. (2018) re-analyzed summary data from 46 studies (n = 9,756), finding that most adolescent samples supported a two-factor model consisting of inhibition and working memory (without a differentiated cognitive flexibility factor) or nested factor structures (where EF processes loaded onto a common factor), but that no model consistently converged or met fit criteria. This could suggest that typically assessed EF tasks (e.g., inhibition, working memory, cognitive flexibility) may have conceptually and empirically distinct contributions to any individuals’ EF ability. Indeed, EF tasks often exhibit weak to moderate correlations with each other, suggesting that EF tasks may provide unique information and the underlying latent construct may not appropriately index individual differences in EF (see Camerota et al., 2020; Dang et al., 2020; Hedge et al., 2018).

Second, EF tasks are often complex and multidimensional, integrating many lower-level cognitive processes such as processing speed, word-reading, or visual tracking abilities, and often requiring multiple EF processes. For example, although the Dimensional Card Sorting Task (DCST) is primarily a test of cognitive flexibility, it may require participants to engage several different EF processes including inhibition and working memory (Miyake & Friedman, 2012). These task-specific effects or “task impurity” are disregarded as “noise” or measurement error in CFA models, but in some cases, noise may actually be signal (Morrison & Grammer, 2016; Nigg, 2017). On the Flanker task, for instance, participants are asked to push a button corresponding to the direction of an arrow while ignoring whether surrounding arrows are congruent (i.e., pointing in the same direction) or incongruent (i.e., pointing in different directions). Higher error rates are typically interpreted as a failure to inhibit a prepotent response. However, a higher error rate could also be caused by a failure of identification (e.g., did not correctly identify the direction the arrow was pointing), a perseverative response (e.g., when two or more incongruent trials are followed by a congruent trial, youth may fail to update rules), or differing cognitive strategies. Thus, there is conceptual ambiguity in how behavior exhibited in a particular task should be interpreted that may also have implications for how task performance relates to “real world” behavior (e.g., McCoy, 2019).

Finally, the validity of quantitative comparisons across individuals (i.e., whether differences in average EF performance are meaningful?) rests on assumptions of measurement invariance - that the meaning and measurement properties of EF tasks and their associations with each other are equivalent across all individuals in a sample (Hartung et al., 2020; Willoughby et al., 2012). Although these types of structural differences can be assessed across common subgroupings such as age, sex, or socioeconomic status (SES), even normal or mixed clinical populations may espouse underlying differences due to unassessed or “hidden” extraneous factors that emerge from the data due to multiple co-occurring developmental processes expressed over time (Vandenbroucke et al., 2018; Zelazo, 2020). Thus, the organization of EFs and the meaning of different factors may differ across subgroups in a population. Further, in the absence of equivalent factor structures, mean difference comparisons may be compromised, leading to divergent findings in EF literature.

Person-Centered Approaches to the Study of Adolescent Executive Functioning

Despite advances in our understanding of adolescent EF, there is wide heterogeneity in the measurement and conceptualization of EF leading to disparate associations between EF and outcomes (Morrison & Grammer, 2016; Nigg, 2017). Even when EF is measured well, however, averages may not accurately describe individuals. Individual variation in biological, psychological, and contextual factors may lead to meaningful variation in EF not captured by variable-centered approaches. For example, heterogeneity in EF is exacerbated for individuals with attention-deficit-hyperactivity disorder (ADHD) and other clinical disorders, with studies suggesting that different EF mechanisms (e.g., problems with inhibitory control versus problems with cognitive flexibility) may contribute to similar symptom severity (Cordova et al., 2020).

Additionally, a small, but converging body of evidence suggests that variation in pubertal development may be associated with variation in EF processes through multiple pathways (Chaku & Hoyt, 2019; Crone, 2009; Laube et al., 2019). For instance, sex steroids (primarily estradiol in girls and testosterone in boys) increase at the start of puberty, contributing to the reorganization of the adolescent brain, and to on-going maturation of the prefrontal cortex (where EF skills reside; Blakemore et al., 2010). Youth who enter puberty earlier than their peers have longer exposure to these hormones, potentially contributing to differences in plasticity or learning potential (Laube et al., 2019). Alternatively, youth who enter puberty earlier than their peers face a variety of psychosocial stressors which may impact the development of EF skills. Early maturing youth often engage in more risk-taking, substance use problems, and mental health problems, all of which may disrupt regulatory systems associated with EF (Hoyt et al., 2020).

Contextual factors such as SES have also been associated with individual differences in EF (Blair, 2016). Generally, youth from higher SES families have more opportunities to develop EF skills (Lawson et al., 2018), but SES is a multifaceted construct that represents access to material (usually indexed by income) and non-material (such as parental education, occupational prestige, neighborhood quality) resources (Duncan & Magnuson, 2001). These aspects of SES may be associated with EF skills in different ways. For instance, low family income, which may contribute to negative environmental exposures, is associated with inhibitory control (e.g., Deater-Deckard et al., 2019; Moilanen et al., 2009) whereas higher parent education, which is associated with more sensitive parenting behaviors, may be a better predictor of working memory or cognitive flexibility (e.g., Vrantsidis et al., 2020).

Although these demographic and contextual differences can be captured, in part, through multigroup analyses, many of these factors co-occur. For instance, early poverty is associated with both pubertal development and low EF abilities (Deater-Deckard et al., 2019; Stumper et al., 2020), suggesting that examining co-occurring developmental patterns that emerge from data may be a better approach to assessing EF abilities in a population. Empirically identifying these subgroupings in adolescence may have important implications in clinical or educational settings where decisions are made and services are provided based on individual deviation from a standardized assessment or norms that may not apply to a certain subpopulation (Snyder et al., 2015). For example, Cordova and colleagues (2020) suggest that although ADHD symptom severity may be similar across different subpopulations, underlying differences in EF abilities within a population could have implications for optimal intervention efforts. Further, population-level averages may not accurately reflect constellations of individual and cultural attributes that influence EF abilities, potentially leading to inaccurate characterizations of different subpopulations, especially for youth who face challenges due to racism, poverty, or other adversities or contextual factors (Ellefson et al., 2017; Haft & Hoeft, 2017).

Person-centered approaches - which are promising but underutilized in the study of EF - can be used to capture complex associations between EF skills and assess unique constellations of EF skills across individuals. Briefly, person-centered approaches such as growth mixture models, latent profile analysis (LPA), or latent class analysis are data-driven approaches for clustering individuals based on patterns of covariation in observed data (i.e., individuals within a class are more similar than individuals between classes; Bauer & Shanahan, 2007). Although person-centered approaches such as LPA often approximate dimensional reduction solutions like factor analysis, a key strength of this approach, especially for understanding EF, is that LPA can be used to understand non-additive interactions between EF skills (Brand et al., 2005). That is, EF skills can make unique and substantial contributions to an individual’s EF profile.

A limited body of work has found meaningful profiles of EF in clinical and mixed community samples of children, adolescents, and adults. For example, Vaidya et al. (2020) found three transdiagnostic EF profiles in a clinical sample of 8 – 14 year olds that were typified by weaknesses in cognitive flexibility, inhibition, or working memory. Similarly, other studies have found profiles of youth with high and low EF skills, as well as profiles reflecting “discordance” or mixed performance across certain EF tasks (Baez et al., 2019; Chaku et al., 2021; Litkowski et al., 2020). Further, previous studies have illuminated social and demographic factors that predict membership in different profiles, finding some evidence of sex and SES differences (Cassidy, 2016; Molitor et al., 2018) Critically, most previous studies were unpowered (Tein et al., 2013), conflated children and adolescents (limiting our understanding of adolescent-specific processes; see Thompson et al., 2019), or utilized a primarily clinical sample necessitating further exploration of EF profiles.

The Current Study

The measurement and modeling of EF is a developmental question, but previous studies have relied on variable-centered statistical approaches (e.g., regressions, factor analyses) which assume that associations between EF variables are stable across individuals. Further, most studies lack large, representative samples that allow us to examine quantitative and qualitative differences in EF across individuals. The Adolescent Brain Cognitive Development (ABCD) study, an on-going, longitudinal study of over 11,000 youth, provides an unprecedented opportunity to investigate the underlying structure of EF in a population-level sample. Thus, the goal of the current study is to empirically derive constellations of EF skills and explore how profiles of EF skills are related to sociodemographic characteristics and trait-level vulnerabilities in internalizing and externalizing behaviors compared a variable-centered approach (i.e., CFA). Specifically, we aimed to:

1. To identify whether unique EF profiles exist among early adolescents in a population-level dataset using LPA;

   H1. We hypothesized that LPA would identify meaning profiles based on responses to neurocognitive tasks, including a ‘typical’ (i.e., most prevalent) EF profile as well as multiple less prevalent profiles. Emergent profiles could represent generally higher (i.e., better), lower (i.e., worse), or discordant (i.e., mixed) performance across all EF tasks. Profiles that were discordant could be characterized by higher performance on certain EF skills (e.g., inhibition) and lower performance on other EF skills (e.g., working memory) or vice versa.

2. To evaluate sex, SES, and pubertal differences across EF profiles;

   H2. We hypothesized that youth from lower income households or from families where mothers’ had less education would be more likely to be in profiles characterized by lower overall EF; and that youth with more advanced pubertal status (i.e., earlier pubertal timing) would be more likely to be in profiles characterized by higher overall EF. EF profile membership would not be associated with biological sex.

3. To examine longitudinal associations between EF profile membership and behavioral outcomes such as internalizing and externalizing behaviors and, as an exploratory aim, compare these associations to those between a latent factor of EF (derived using CFA) and the same behavioral outcomes.

   H3. We hypothesized that profiles characterized by higher overall EF would be characterized by better outcomes (i.e., lower internalizing and externalizing behaviors), whereas profiles characterized by discordant performance would have mixed outcomes (i.e., may have fewer internalizing behaviors, but more externalizing behaviors as well). Profiles characterized by worse overall EF will be characterized by worse general functioning.

Methods

Data for these analyses were drawn from the ABCD Study (3.0 Data Release), an ongoing, longitudinal study of youth ages 9 – 10 recruited from 21 data collection sites across the United States (U.S.). Information on recruitment efforts, enrollment, and study retention efforts are outlined elsewhere (Ewing, Chang, et al., 2018; Garavan et al., 2018) and detailed information on the ABCD study design and experimental procedures can be found at abcdstudy.org. Institutional Review Board approval was obtained for each site before data collection. All parents provided written informed consent and all youth provided assent. Procedures for data access and analysis were implemented as approved by the Institutional Review Boards at the University of Michigan, and in agreement with the sensitive data security plan approved by ABCD study data managers.

The current study research hypotheses and analytic plan were registered following Open Science Framework (OSF) and ABCD Study recommendations. The initial sample consisted of 11,878 youth (52.16% male). Participants were excluded from the analytic sample if they were missing any EF variables (i.e., Flanker, List Sorting Task, and Card Sorting Task). Although only 2% of the sample (n = 206) were missing one or more EF variables; attrition analysis suggested that there were differences in missingness by race/ethnicity, household income, mother’s education, and site. Specifically, Black youth (χ² = 13.80, p = .008; Cramer’s V = .10), youth from households earning less than $50,000 (χ² = 6.26, p = .04, Cramer’s V = .04), and youth from households where the mothers’ did not obtain a high school diploma (χ² = 10.44, p = .02, Cramer’s V = .02) were more likely to be missing at least one EF variable, but these effects were consistently small. Missingness by site ranged from 0% – 10.68% (M = 4.54%). There were no significant differences in missingness by biological sex.

The final sample (N = 11,672) was 52.16% male with a mean age of 9.91 years (SD = .62). About half the sample (52.18%) reported White race/ethnicity with 20.28% reporting Hispanic race/ethnicity, 14.9% reporting Black race/ethnicity, 2.16% reporting Asian race/ethnicity, and 10.5% reporting another race/ethnicity. Over half the sample (59.6%) came from families where the mother obtained a college degree or higher and 38.53% of the sample came from households with incomes greater than $100,000 a year. Detailed information about sample characteristics are presented in Table 1.

Table 1.

Full Sample Characteristics and Descriptive Information (N = 11,672)

	M (SD)	N (%)
Age	9.91 (.62)	-
Puberty
Male	1.65 (.5)	-
Female	1.69 (.54)	-
Biological Sex
Male	-	6,088 (52.2%)
Female	-	5,584 (47.8%)
Income
Less than $50,000	-	3,154 (29.5%)
> $50,000 & <$100,000	-	4,497 (42.1%)
> $100,000	-	3,025 (28.3%)
Mother’s Education
Did not complete high school	-	574 (4.9%)
Completed high school/GED	-	1,107 (9.5%)
Some college	-	3,029 (26%)
Bachelors/Post-graduate degree	-	6,649 (57%)
Race/Ethnicity
White	-	6,090 (52.2%)
Hispanic	-	2,367 (20.3%)
Black	-	1,739 (14.9%)
Asian	-	252 (2.2%)
Other	-	1,222 (10.5%)

Notes: M = mean, SD = standard deviation. Puberty (1 = pre-pubertal, 2 = early puberty, 3 = mid puberty, 4 = late puberty, 5 = post puberty). Income = household combined income from all sources. GED = general education degree.

Measures

Executive Functioning (EF) Measures

The EF tasks were measured in the ABCD study using the National Institutes of Health (NIH) Toolbox (https://nihtoolbox.desk.com) which consists of several tasks that evaluate memory, executive function, language, and other higher-order cognitive processes (Bleck et al., 2013; Gershon et al., 2013; Hodes et al., 2013). An overview of the neurocognitive battery is described elsewhere (Luciana et al., 2018), but each task used in the subsequent analysis is summarized below.

Inhibitory control was measured using the Toolbox Flanker Task, a variation of the Eriksen Flanker task (Eriksen & Eriksen, 1974). The task assessed the degree to which participants’ responses were influenced by surrounding stimuli when trying to identify the target during congruent and incongruent trials. In each trial, four stimuli (two arrows on the outer left and two arrows on the outer right) either faced the same way (congruent trial) or to the left or right (incongruent trial) of the middle arrow. Adolescents had to push the arrow that corresponded to the middle arrow. The ABCD study provided a composite measure of the Flanker that incorporated both speed and accuracy that was used for analysis. Uncorrected scores were used for analysis. This task has been validated in adolescent samples with good test-retest reliability (ICC = 0.92; Zelazo et al., 2014).

Working memory was assessed using the Toolbox List Sorting Working Memory Test, which is a variant of the letter-number sequencing test (Gold et al., 1997) that uses pictures, rather than words or letters (Tulsky et al., 2014; Tulsky et al., 2013). The task required participants to sequence stimuli based on category membership. Adolescents were presented with a series of pictures of animals and foods of different sizes on an iPad and then had to vocally repeat back the items in order of smallest to largest to the experimenter. The trials started by presenting one category (e.g., animals) with two items and if the adolescent was correct the items increased one at a time up to seven items. This process was repeated with a secondary category (i.e., animals and food) where they listed the items in one category and then the next in order of smallest to largest. The ABCD study provided a sum score of total correct responses across the two list sorting tasks which comprised the “total score” used in the current analysis. The List Sorting task has been validated with adolescents and demonstrated good test-retest reliability (ICC = 0.86; Tulsky et al., 2013).

Cognitive flexibility was assessed using the Toolbox Dimensional Change Card Sort Task (Zelazo, 2006). In this task, adolescents were presented with two objects and then had to sort a third object based on either color or shape according to the rule presented (Zelazo, 2006; Zelazo et al., 2014). Participants completed three blocks of trials where they first sorted objects based on one dimension, then another, and then finally, on a random order of dimensions. The ABCD study provided a composite measure of the Card Sorting task that incorporated both speed and accuracy. Uncorrected scores were used for analysis. Validation testing with adolescents demonstrated that the Card Sorting task showed good test-retest reliability (ICC = 0.92; Zelazo et al., 2014).

Demographic and Biological Predictors

Biological sex (1 = boys; 0 = girls); household income, mother’s education level, and pubertal development were assessed at baseline. Parent-reported yearly household income was categorized by ABCD as: 1 = < $50,000, 2 = > $50,000 and < $100,000, and 3 = > $100,000 which broadly corresponds with lower, median, and higher income levels reported for a family of four in the United States (U.S.) census (Guzman, 2017). Maternal education (1 = < high school diploma; 2 = high school diploma/General Education Degree (GED); 3 = Some college; 4 = Bachelor’s Degree/Post Graduate Degree) was also collected. Pubertal timing was assessed using the Pubertal Development Scale (PDS; Crockett & Petersen, 1987). The PDS is comprised of five items (three sex-neutral items and two sex-specific items) assessing common pubertal milestones (e.g., breast development, voice changes) rated on a four-point scale (1 = pre-pubertal; 2 = early puberty; 3 = mid-puberty; 4 = late puberty; 5 = post-puberty). Items were averaged to create a composite score. Then following previous research (Mendle et al., 2019), the composite score was standardized within sex and regressed onto age to create an indicator of pubertal timing. Negative residual scores suggest later timing (i.e., less physically mature) compared to same-aged, same sex peers while positive residual scores suggest earlier timing (i.e., more physically mature) compared to same-aged, same sex peers.

Behavioral Outcomes

The Child Behavior Checklist (CBCL) from the year 1 follow-up assessment was utilized to explore how EF profiles were associated with a wide range of behaviors. Parents completed the CBCL, a 119-item survey about aspects of the child’s behavior across the past six months (Achenbach & Ruffle, 2000). Items (rated on a three-point scale from 0 = Not True to 2 = Very Often or Always True) were combined to create eight syndrome scales: anxious/depressed, withdrawn/depressed, somatic complaints, delinquent behavior, aggressive behavior, attention problems, thought problems, and social problems. The anxious, withdrawn, and somatic complaints subscales are thought to comprise a broader internalizing composite whereas the delinquent behavior and aggressive behaviors subscales comprise a broader externalizing composite. The other three subscales (attention problems, thought problems, social problems) were classified as “other problem behaviors” following previous research (Dekker et al., 2002; Schroeder et al., 2010). Raw scores were used for analysis.

Analytic Plan

Data was first assessed for normality and outliers in Stata Version 16.1 (StataCorp, 2007). Outliers were determined using the bacon command, which is appropriate for use with multivariate data (Weber, 2010). Then descriptive statistics and correlational analysis were conducted to describe the sample and report correlations between EF variables and all behavioral outcomes.

LPA was then conducted to identify profiles characterized by distinct patterns of executive functioning skills using MPlus Version 8.0 (Muthén, 2007). Based on current recommendations and to minimize overfitting (Hickendorff et al., 2018; Spurk et al., 2020), a split-half cross validation was utilized to select the optimal profile solution. To do so, the sample was randomly split in half, creating a “training sample” (n = 5,836) and a “validation sample” (n = 5,836). The LPA was first run in the training sample. Training models were run iteratively, starting with a one profile solution, and then adding additional profiles until the sample-adjusted Bayesian Information Criteria (saBIC) increased across two models in a row or until one of the profiles in a given solution represented less than 5% of the sample.

The final training sample solution was selected based on fit indices (lower Akaike Information Criterion/AIC, BIC, and saBIC scores; a significant Lo-Mendell-Rubin Test), theoretical relevance (e.g., Have expected profiles emerged from the data?) and parsimony (e.g., Does the addition of a profile add significant interpretability?). Once the final training sample solution was selected, a sensitivity analysis was run to assess whether the profiles were robust to the addition of participants who had completed only one or two EF tasks (n = 11,729) in comparison to those who completed all three EF tasks (n = 11,672). If robust to the sensitivity analysis, the final selected solution was then tested in the validation sample (n = 5,836) and then applied to the full sample (N = 11,672). All models controlled for age and the mixture complex command was used to account for the nested structure of the data (individuals nested within families; families nested within site) and derive robust standard errors. Means and variances were freely estimated, and EF variables were allowed to freely correlate with each other within the profile.

Once the final LPA solution was selected, the most likely latent profile membership was exported to Stata as an observed variable using the estimated posterior probabilities. Chi squares and ANOVAs were used to compare profile membership by demographic factors (biological sex, SES, pubertal timing). Profile comparisons were Bonferroni-corrected and significant omnibus effects were followed up with post-hoc analyses to assess for significant pairwise differences. Effect sizes for chi-square tests were assessed with Cramer’s V where .1 is considered a weak association, .4 is considered a medium association, and .5 is considered a strong association (Acock & Stavig, 1979). Effect sizes for ANOVAs were assessed with a partial eta squared (η²) where .01 is considered a small effect size, .09 is considered a medium effect size, and .25 is considered a large effect size (Richardson, 2011).

Then, the three-step Bolck, Croon, and Hagenaars (BCH) procedure in MPlus was utilized to explore associations between profile membership and distal outcomes assessed one year later. As recommended by Asparouhov (2014), after the estimation of the final LPA solution (step 1), the most likely profile membership was calculated taking classification uncertainty into account (step 2), and applied to a secondary model where profile membership was used to assess mean differences on distal outcomes (step 3). Distal outcomes were adjusted for biological sex, mother’s education, income, and race/ethnicity given well-known sociodemographic differences in these outcomes (Gross et al., 2006). The statistical significance of profile-specific mean differences were evaluated with Wald tests of mean equality, which are appropriate for continuous distal outcomes (Asparouhov, 2014). Additionally, an effect size (Cohen’s d) was calculated by dividing the mean difference score between two profiles by the distal outcome’s standard deviation (SD) where 0.2 is considered a small effect, 0.5 is considered a medium effect and 0.8 is considered a large effect size (Fritz et al., 2012). Full information maximum likelihood (FIML) was used to address missing data.

Finally, to compare the LPA findings to more traditional, variable-centered approaches, a one factor CFA model was estimated in MPlus to represent EF as a single latent factor. The CFA was evaluated with several model fit indices including the comparative fit index (CFI), the Root Mean Square Error of Approximation (RMSEA), the Standardized Root Mean Square (SRMS), and chi-square test where a CFI above .90, a RMSEA below .06, a SRMS value below .08 , and an insignificant chi-square value were indicators of good fit (Perry et al., 2015). Once fit to the data, the latent factor was regressed onto the distal outcomes controlling for biological sex, mother’s education, household income, and race/ethnicity. Although not directly comparable to effect sizes, standardized beta coefficients were reported to increase interpretability and facilitate comparisons between the CFA and the LPA. See supplemental materials for all annotated syntax.

Results

Means, standard deviations, and correlations for all study variables are presented in Table 2. All EF variables demonstrated adequate normality and the bacon command suggested that there were no outliers. As expected, the Flanker, List Sorting, and Card Sorting tasks demonstrated weak to moderate correlations with each other (rs = .30 - .45, ps < .001), but exhibited insignificant or weak correlations with internalizing behaviors, externalizing behaviors, and other problem behaviors. All outcomes demonstrated moderate - high correlations (rs = .27 – .65, ps < .001) with each other.

Table 2.

Descriptive Statistics and Correlational Analyses for Study Variables

	1	2	3	4	5	6	7	8	9	10	11
1. Inhibitory control	94.02 (9.11)
2. Working memory	.30***	96.65 (12.08)
3. Cognitive flexibility	.44***	.33***	92.54 (9.50)
4. Anxious/Depressed	−.01	.01	−.01	2.53 (3.07)
5. Withdrawn/Depressed	−.04***	−.02	−.03***	.57***	1.11 (1.77)
6. Somatic complaints	−.01	−.01	−.01	.46***	.40***	1.45 (1.95)
7. Delinquent behavior	−.07***	−.10***	−.10***	.36***	.40***	.27***	1.11 (1.80)
8. Aggressive behavior	−.07***	−.08***	−.08***	.53***	.49***	.37***	.73***	3.04 (4.17)
9. Attention problems	−.11***	−.13***	−.14***	.46***	.44***	.32***	.56***	.62***	2.84 (3.42)
10. Thought problems	−.04***	−.02*	−.05***	.58***	.51***	.44***	.50***	.59***	.63***	1.61 (2.22)
11. Social problems	−.11***	−.12***	−.11***	.60***	.54***	.41***	.54***	.65***	.63***	.62***	1.50 (2.17)

Notes:

^***p < .001

^**p < .01

^*p < .05

Means and standard deviations (in parentheses) are presented on the diagonal

LPA Model Development and Characterization

Table 3 presents the fit statistics for the LPA. Profile solutions (up to five) were run in the training sample (n = 5,836). The four-profile solution had a lower AIC. BIC, and saBIC value compared to the one - three profile solutions and the LMR test was significant, suggesting a better fit compared to the model with one fewer class. Although the AIC, BIC, and saBIC continued to decrease in the five-profile solution, this solution contained a profile representing less than 5% of the sample, which was considered spurious. An investigation of the four-profile solution also revealed profiles that were theoretically expected, conceptually interesting, and non-redundant compared to the three-profile and five-profile solution (see Figure 1A). The selected four-profile solution was robust to the inclusion of individuals with less data (see sensitivity analysis in supplemental materials, Figure S1). The four-profile solution was then applied to the validation sample (n = 5,836), demonstrating good fit to the data and similar profiles (see Figure 1B). Once validated, the four-profile solution was applied to the full sample (N = 11,679) to obtain profile membership for subsequent analyses.

Table 3.

Latent Profile Analysis Fit Statistics by Training Sample, Validation Sample, and Full Sample

Model	LL	Par	Entropy	AIC	BIC	saBIC	LMR (p value)	Class Size
Training Sample (n = 5,836)
One Profile	−69810.17	11	-	139642.34	139715.73	139680.78	-	5836
Two Profiles	−63400.57	20	.60	126841.15	126974.59	126911.03	< .001	4723	1113
Three Profiles	−63222.10	31	.51	126506.19	126713.02	126614.51	< .001	2795	2487	554
Four Profiles	−63151.79	42	.45	126387.58	126667.80	126534.34	< .001	2574	1743	1296	481
Five Profiles	−63119.07	53	.51	126344.15	126697.95	126529.33	< .001	2762	1401	1028	504	141
Validation Sample (n = 5,836)
Four Profiles	−63288.63	42	.54	126661.26	126941.48	126808.01	.04	3328	1294	777	437
Full Sample (N = 11,679)
Four Profiles	−126467.23	42	.47	253018.46	253327.79	253194.32	< .001	5679	3570	1520	903

Notes: LL = log likelihood. Par = parameters. AIC = Akaike Information Criterion. BIC = Bayesian Information Criterion. saBIC = sample-adjusted Bayesian Information Criteria. LMR = Lo-Mendell-Rubin Test.

Figure 1.

Auxiliary variables in mixture modeling: Using the BCH method in Mplus to estimate a distal outcome model and an arbitrary secondary model

Figure 1C shows the pattern of model-estimated mean values of EF indicators for each profile in the full sample and Table 4 presents items means, variances, and correlations for each EF variable by profile. The largest profile (Average EF) contained 49% of the sample and was characterized by moderate scores on the Flanker, List Sorting, and Card Sorting tasks, suggesting average inhibitory control, working memory, and cognitive flexibility abilities. In this profile, all three EF variables demonstrated weak to moderate correlations with each other (rs = .25 - .42, ps < .001) and exhibited less residual variance compared to the other profiles. The second profile (High EF) represented 31% of the sample and was characterized by the best scores across all three tasks, suggesting relatively high inhibitory control, working memory, and cognitive flexibility abilities. The High EF profile demonstrated weak correlations between EF variables: The Flanker and List Sorting tasks were not significantly correlated and other EF variables demonstrated weak correlations (rs = .10 - .35, ps < .001). The High EF profile also demonstrated the least residual variance within class.

Table 4.

Executive Functioning Item Means, Variances, and Correlations by Profile

	Profile 1 Average EF			Profile 2 High EF			Profile 3 Low IC			Profile 4 Low EF			Effect Size (Cohen’s d)
Items Means	M (SD)			M (SD)			M (SD)			M (SD)			1 vs. 2	1 vs. 3	1 vs. 4	2 vs. 3	2 vs. 4	3 vs. 4
Inhibitory control (Flanker)	95.072***, 3***, 4*** (.42)			100.351***, 3***, 4*** (.34)			86.461***, 2*** (1.21)			84.341***, 2*** (.72)			−.58	.94	1.18	1.52	1.76	.23
Working memory (List Sorting)	95.232***, 3***, 4*** (.58)			102.271***, 3***, 4*** (.41)			98.231***, 2***, 4*** (.82)			83.421***, 2***, 3*** (.90)			−.58	−.25	.98	.33	1.56	1.23
Cognitive Flexibility (Card Sorting)	92.062***, 4*** (.31)			98.36 1***, 3***, 4*** (.35)			92.142***, 4*** (.50)			78.271***, 2***, 3*** (.79)			−.66	−.01	1.45	.65	2.11	1.46
Variances
Inhibitory control (Flanker)	31.20 (2.04)			26.17 (1.63)			77.48 (7.60)			158.33 (6.64)
Working memory (List Sorting)	122.69 (4.75)			85.89 (4.47)			102.76 (6.94)			211.53 (11.52)
Cognitive Flexibility (Card Sorting)	46.74 (2.06)			51.78 (2.38)			44.54 (4.53)			155.45 (8.23)
Correlations	1	2	3	1	2	3	1	2	3	1	2	3
1. Inhibitory Control	-			-			-			-
2. Working Memory	.11**	-		−.01	-		.36***	-		.19***	-
3. Cognitive Flexibility	.33***	.10***	-	.35***	.10***	-	.43***	.25***	-	−.04	.03	-

Notes:

^***p < .001

^**p < .01

^*p < .05

M = Mean. SD = Standard deviation. IC = inhibitory control. Subscripts denote significant differences between two profiles (e.g., 2 = significantly different from Profile 2). Effect sizes were calculated by dividing the mean difference between two variables by that variables’ raw standard deviation.

The third profile (Low IC) represented 18% of the sample and was characterized by moderate to high scores on the List Sorting and Card Sorting tasks and relatively lower performance on the Flanker, suggesting low inhibitory control (IC). Like the Average EF profile, all three EF variables demonstrated weak to moderate correlations with each other (rs = .25 - .43, ps < .001), but the Low IC profile exhibited more residual variance than either the Average EF or High EF profiles. The final profile (Low EF) represented 10% of the sample and was characterized by the worst scores across all three tasks, suggesting relatively low inhibitory control, working memory, and cognitive flexibility abilities. The Low EF profile also demonstrated the smallest correlations between EF variables: The Card Sorting task was not significantly associated with any EF variables and the Flanker and List Sorting task were weakly correlated (r = .19, p < .001). The Low EF profile also demonstrated the most residual variance.

Overall, the profiles demonstrated large effect size differences in performance on the Flanker, List Sorting, and Card Sorting task with two exceptions. First, the Low IC and Low EF had similar scores on the Flanker and the Low IC and Average EF had similar scores on the Card Sorting Task. Second, although the High EF, Average EF, and Low EF profiles were primarily distinguished by differing levels of performance, the Low IC profile exhibited discordance with a much lower score on the Flanker, the second highest score on the List Sorting task, and a moderate score on the Card Sorting task.

Demographic and Biological Predictors

Significant differences by biological sex (χ² = 26.38, p < .001, Cramers’ V = .05), household income (χ² = 547.84, p < .001, Cramers’ V = .16), mother’s education (χ² = 602.83, p < .001, Cramers’ V = .13), and pubertal timing (F = 20.16, p < .001, Cramers’ V = .01) were observed across profiles (Table 5). Relative to youth in the Average EF and Low IC profile, youth in the Low EF profile were significantly more likely to be boys. Further, youth in the Average EF and High EF profile were more likely to come from households where the reported family income was greater than $100,000 whereas youth in the Low IC and Low EF profiles were more likely to come from households where the reported family income was less than < $50,000. A similar pattern emerged for mother’s education: Youth in the Average EF and High EF profiles were most likely to have mothers’ who had earned a higher degree whereas youth in the Low EF profile were most likely to have mothers’ who had not completed high school. Generally, these effects were small, but larger for household income and mother’s education than for sex. Pubertal timing indices suggested that youth in the Average EF and High EF profile had later or on-time development compared to individuals in the Low IC and Low EF profiles, F(3, 9418) = 20.16, p < .001. Although these effects were small, individuals in the Average EF profile (M = −.06, SD = 1.00) demonstrated the latest timing whereas those in the Low EF profile (M = .21, SD = 1.08) demonstrated the earliest timing.

Table 5.

Mean Differences between Demographic Characteristics by EF Profiles

	Profile 1 Average EF	Profile 2 High EF	Profile 3 Low IC	Profile 4 Low EF	Test Statistic	Effect Size
	%	%	%	%	χ2	Cramers’ V
Male	50.6%4	53.9%	50.3%4	58.5%1, 3	26.38	.05
Income
< 50,000	29.9%2, 3, 4	20.0%1, 3, 4	34.8%1, 2, 4	59.2%1, 2, 3	547.84	.16
> 50,000 & < 100,000	29.5%4	28.0%4	28.7%4	21.7%1, 2, 3
> 100,000	40.7%2, 3, 4	52.0%1, 3, 4	36.6%1, 2, 4	19.2%1, 2, 3
Mother’s Education
Did not complete high school	5.2%2, 4	2.3%1, 3, 4	6.4%2, 4	11.2%1, 2, 3	602.83	.13
Completed high school/GED	9.7%2, 4	5.4%1, 3, 4	11.6%2, 4	21.3%1, 2, 3
Some college	25.9%2, 4	22.3%1, 3, 4	28.2%2, 4	37.0%1, 2, 3
Bachelors	59.1%2, 3, 4	70.1%1, 3, 4	53.9%1, 2, 4	30.5%1, 2, 3
	M (SD)	M (SD)	M (SD)	M (SD)	F	η 2
Puberty	−.06 (1.00)2, 3. 4	.02 (.97)1, 4	.03 (.99)1, 4	.21 (1.08)1, 2, 3	20.16	.01

Notes. M = Mean. SD = Standard deviation. IC = inhibitory control. Subscripts denote significant differences at p < .05 between two profiles (e.g., 2 = significantly different from Profile 2).

Male = biological sex (1 = male, 0 = female). Bachelors = Bachelors’ degree or post-graduate degree. GED = general education degree. Puberty (1 = pre-pubertal, 2 = early puberty, 3 = mid puberty, 4 = late puberty, 5 = post puberty).

Links to Behavioral Outcomes

Mean profile differences in internalizing behaviors, externalizing behaviors, and other problem behaviors are presented in Table 6 along with associated effect sizes. All outcome means were adjusted for biological sex, mother’s education, household income, and race/ethnicity (see supplemental materials for covariate parameter estimates, Table S1). There were limited findings for internalizing behaviors. Youth in the Low EF profile (M = 2.95, SD = .19) had more anxious/depressive symptoms compared to those in the High EF profile (M = 2.47, SD = .09) and youth in the Low IC profile (M = 1.00, SD = .09) had more withdrawn/depressive symptoms than those in the Average EF profile (M = .77, SD = .05). The profiles did not differ on somatic complaints. Generally, effects for internalizing behaviors were small.

Table 6.

Associations between Latent EF and Child Behavior Checklist (CBCL) Syndrome Scale compared to Latent Profiles of EF

	Latent EF Factor	Profile 1: Average EF	Profile 2: High EF	Profile 3: Low IC	Profile 4: Low EF	Effect Size
						Latent EF Factor	Latent Profiles (Cohen’s d)
	B (SE)	M (SD)	M (SD)	M (SD)	M (SD)	β	1 vs. 2	1 vs. 3	1 vs. 4	2 vs. 3	2 vs. 4	3 vs. 4
Internalizing Symptoms
Anxious/Depressed	−.03*** (.01)	2.59 (.10)	2.474** (.09)	2.78 (.14)	2.952** (.19)	−.05	.04	−.06	−.12	−.10	−.16	−.06
Withdrawn/Depressed	−.01 (.004)	.773* (.05)	.88 (.05)	1.001* (.09)	.87 (.11)	−.02	−.06	−.13	−.05	−.07	−.01	.08
Somatic Complaints	−.003 (.01)	1.36 (.06)	1.34 (.06)	1.41 (.09)	1.32 (.12)	−.01	−.03	−.03	−.02	−.04	.01	.05
Externalizing Symptoms
Delinquent Behavior	−.02*** (.01)	.584** (.06)	.464*** (.05)	.064* (.08)	.911**, 2***, 3* (.13)	−.08	.07	−.01	−.18	−.08	−.25	−.18
Aggressive Behaviors	−.06*** (.01)	2.312*, 4** (.13)	1.941*, 4*** (.12)	2.344* (.19)	3.211**, 2***, 3* (.29)	−.08	.09	−.004	−.22	−.10	−.31	−.21
Other Problem Behaviors
Attention Problems	−.11*** (.01)	2.152*, 3***, 4*** (.11)	1.461*, 3***, 4*** (.09)	2.231*, 2***, 4*** (.16)	3.601*, 3***, 4*** (.23)	−.19	−.09	−.18	−.23	−.39	−.44	−.21
Thought Problems	−.03*** (.01)	1.402*, 4** (.07)	1.191*, 3*, 4*** (.06)	1.452*, 4* (.11)	1.811**, 2***, 3* (.15)	−.08	.10	−.02	−.18	−.12	−.28	−.17
Social Problems	−.06*** (.01)	1.262***, 4*** (.07)	.851***, 3***, 4*** (.06)	1.292***, 4*** (.11)	1.951***, 2***, 3*** (.15)	−.15	.19	−.01	−.32	−.20	−.50	−.30

Notes.

^***p < .001

^**p < .01

^*p < .05

B = unstandardized coefficient. SE = standard error. β = standardized coefficient. M = Mean. SD = Standard deviation. IC = inhibitory control. Subscripts denote significant differences between two profiles (e.g., 2 = significantly different from Profile 2). Effect sizes were calculated by dividing the mean difference between two variables by that variables’ raw standard deviation.

For externalizing behaviors, a similar pattern of results was observed across delinquent and aggressive behaviors: Youth in the Low EF profile had more delinquent (M = .91, SD = .13) and aggressive behaviors (M = 3.21, SD = .29) compared to youth in all other profiles. Youth in the High EF profile (M = .1.94, SD = .12) also demonstrated significantly fewer aggressive behaviors compared to youth in the Average EF profile (M = 2.31, SD = .13). Compared to internalizing behavior, effects for externalizing behaviors were larger, demonstrating generally moderate effect sizes.

Significant differences in profile means also emerged for other problem behaviors. Specifically, youth in the Low EF profile had the most attention problems (M = 3.60, SD = .23), followed by youth in the Low IC profile (M = 2.23, SD = .16), then youth in the Average EF profile (M = 2.15, SD = .11) and finally, youth in the High EF profile (M = 1.46, SD = .09). Youth in the Low EF profile also had significantly more thought problems (M = 1.81, SD = .15) and social problems (M = 1.95, SD = .15) compared to all other youth whereas youth in the High EF problems had significantly fewer thought problems (M = 1.19, SD = .06) and social problems (M = .85, SD = .06) compared to all other youth. Effects for other problem behaviors were generally moderate in size, with the largest effects emerging between the Low EF profile and other profiles.

Finally, a CFA with one latent factor was run and compared to the latent profile analysis. The CFA had good fit to the data, CFI = .96, RMSEA = .00, SRMR = .05, and all items demonstrated moderate to high loadings onto the latent factor (βs = .45 - .69). Internalizing, externalizing, and other problem behaviors were then regressed onto the latent EF factor, controlling for biological sex, mother’s education, household income, and race/ethnicity. Higher EF was associated with fewer internalizing, externalizing, and other problem behaviors, with small effect sizes (see Figure 2).

Figure 2.

Visualization of Effect Sizes using Traditional Latent Factor Analysis (Latent EF Factor) and Latent Profile Analysis.

Notes. EF =executive functioning. Standardized beta weights are reported for the latent factor analysis. Effect sizes (Cohen’s d; mean differences divided by the standard deviation of the raw outcome variable) are reported for the latent profile analysis.

Discussion

The modeling of EF is essential to developmental science inquiries, but the field has focused primarily on variable-centered approaches, instead of emphasizing the individual. The purpose of this study was to identify profiles of EF characterized by performance on common and well-validated tasks of inhibitory control, working memory, and cognitive flexibility; to examine how profiles were associated with sociodemographic characteristics and pubertal timing; and to explore how LPA profiles predicted behavioral outcomes assessed a year later compared to a latent EF factor model.

Latent Profiles of EF in Adolescence

Using LPA, we identified four meaningful EF profiles: Average EF (49%), High EF (31%), Low IC (18%), and Low EF (10%). Aligned with previous findings (Baez et al., 2019; Cordova et al., 2020; Litkowski et al., 2020), profiles were primarily distinguished by their level of performance across each task. However, the Low IC profile also exhibited some discordance between tasks, with relatively low performance on the inhibitory control task and relatively high performance on the working memory task. This suggests that specific tasks may contribute unique – that is, nonadditive – information about an individual’s EF ability. Further, profiles exhibited varying amounts of variance around their mean EF score. Youth in the High EF profile demonstrated the least heterogeneity across all EF tasks whereas youth in the Low EF profile demonstrated the most heterogeneity. This could be an analytic artifact: Clark et al. (2013) suggests that individuals with higher scores on a given measure may exhibit less variance as they are performing at ceiling. Alternatively, the EF tasks could be indexing differences in ability better for those in the Low EF profile, leading to more heterogeneity across individuals in those profiles (Dang et al., 2020).

Correlations between tasks also differed within each profile: The Average EF and Low IC profile demonstrated weak to moderate correlations between inhibitory control, working memory, and cognitive flexibility. The High EF profile demonstrated low to moderate correlations between cognitive flexibility and the other EF variables, but not between inhibitory control and working memory. In contrast, the Low EF profile demonstrated a weak correlation between inhibitory control and working memory, but not between cognitive flexibility and the other EF variables. These findings generally replicate those found in previous research (Camerota et al., 2020; Dang et al., 2020), but add significant nuance to our understanding of underlying EF structures in adolescence. Although the structure of EF in the Average EF and Low IC profiles was generally aligned with that reported in (Miyake et al., 2000) and other CFAs (Gioia et al., 2002; Thompson et al., 2019; Wiebe et al., 2008), the High EF and Low EF profile results are aligned with Karr and colleagues’ (2018) meta-analysis, which reported generally inconsistent EF factor structures among adolescent samples. These preliminary findings suggest that while most youth may be characterized by a typical factor structure (where multiple EF tasks load onto a single latent factor), some subpopulations may be better characterized by unique combinations of EF skills including those typified by only one or two EF abilities. These differences may have arisen, in part, due to different antecedents and developmental processes (e.g., Blair, 2016; Chaku & Hoyt, 2019; Stumper et al., 2020), but suggest the need for future research on EF factor structures with explicit measurement invariance testing across large, representative subpopulations.

Demographic and Biological Correlates of EF Profiles

EF profiles were differentiated by sociodemographic and biological predictors. First, the results suggested that boys are more likely to be the Low EF profile compared to girls. Sex differences in EF are generally small, inconsistent, and depend on developmental and chronological age as well as testing domain (Grissom & Reyes, 2019; Weiss et al., 2003). Although some emerging research suggests sex differences in EF may first emerge in adolescence (Ahmed et al., 2021), these effects are rarely tested in longitudinal data. Therefore, more research is needed to elucidate sex differences across multiple indicators of EF and examine whether these differences persist across adolescence.

Second, as expected, the Average EF and High EF profiles were typified by youth who came from higher income families or from families with more maternal education whereas the Low EF profile was typified by youth from lower income families or from families with less maternal education. This finding is generally aligned with previous research (Raver et al., 2013; Willoughby et al., 2012) and highlights the potentially long-term impacts of income and parent educational level on EF skills. However, while the Low IC profile reported generally lower income levels, it also contained a higher proportion of youth who came from families with more maternal education, perhaps representing a subpopulation of families with higher education levels, but lower annual income (e.g., recent college graduates, parents with more college debt). It could be that higher parental education is particularly protective for working memory or cognitive flexibility (but not inhibitory control) in lower income households that are likely experiencing both economic and psychological stressors, but more work is needed to explicate the interactions between different indices of SES and EF skills and understand how they persist over time.

Finally, pubertal timing also differed by profile membership with the youth in the Low EF profile having the earliest timing and youth in the Average EF profile having the latest timing. Although the effect size of this finding was small, it adds to emerging research about the role of puberty on EF skills (Crone, 2009; Goddings et al., 2014; Satterthwaite et al., 2013). Indeed, variation in pubertal development could be important to consider when developing EF interventions or considering other applied approaches (Laube et al., 2019). For example, Chaku et al. (2019) found that early developing girls and boys had worse initial EF skills, but girls only experienced faster growth in EF skills over time. Future research, with more time points, could replicate these findings using, for example, growth curve modeling to examine how timing and growth over time predict membership into different EF profiles.

Comparing Person-Centered and Variable-Centered Approaches to Understanding EF: Links to Prospective Outcomes

EF profiles were more likely to predict differences in externalizing behaviors compared to internalizing behaviors, and were most predictive of delinquent behaviors, attention problems, and social problems. This generally replicates findings in other research that suggests EF is robustly associated with externalizing behaviors (Ogilvie et al., 2011; Schoemaker et al., 2013; Sulik et al., 2015), but inconsistently linked with internalizing behaviors (McTeague et al., 2016).

Using the ABCD dataset, Thompson and colleagues (2018) also found small to moderate associations between cognition, externalizing, and internalizing behaviors. Our CFA generally replicated those findings, suggesting that in a dimensional solution, youth with the lowest EF scores reported the most behavioral issues and youth with the highest EF scores reported the least behavioral issues. For the most part, the EF profiles were consistent with these findings, but there were a few key differences between the LPA and CFA results. Namely, youth in the Low IC profile had generally higher withdrawn/depressive symptoms and other problem behaviors (i.e., attention, social problems, and thought problems), but did not have significantly different externalizing symptoms compared to all other profiles. This suggests that while the LPA and CFA were primarily complementary, the LPA uncovered additional, specific relations between EF skills and behavior, especially when task performance was discordant (as in the Low IC profile). Further, although not directly comparable, the meaningful differences in the LPA were larger than those in the CFA, suggesting that person-centered approaches may add additional predictive validity over and above variable-centered approaches, better addressing differences between subpopulations. Thus, examining how performance on different tasks co-occur might better inform our understanding of EF ability: Considering the overall shape of a profile as well as the contribution of each task may provide unique information about EF and its structure that may be more predictive of outcomes (and even trajectories) over time.

Implications for Person-Centered EF Research

The results detailed here highlight the utility of person-centered approaches for the study and modeling of EF processes in adolescence. One of the primary strengths of this paper is its use of population-level data, which is powered for the detection of meaningful EF profiles and small effects. Indeed, future work with ABCD data could explore additional EF tasks measured after baseline that “come online” later in development (Luciana et al., 2018; Snyder et al., 2015) or explore even smaller EF profiles which may be masked in community samples, but benefit the most from person-centered research and targeted intervention efforts (Howard & Hoffman, 2018). Further, although LPAs and many other person-centered analyses are data-driven, the use of a split-half cross-validation approach illustrated the reliability of the final profile solution, suggesting that these profiles may emerge in other adolescent data, even those with smaller samples. Future research should transition into confirmatory hypothesis testing within the ABCD dataset and in other smaller datasets to further test and validate the four profiles presented here.

Critically, the ABCD study is on-going and will span early adolescence to young adulthood, opening many exciting avenues for future person-centered research. For example, the age differential hypothesis suggests that EFs continue to differentiate over time, potentially loading onto multiple factors as youth age (Xu et al., 2013). Growth mixture modeling could be used to model growth in EF changes across different subpopulations over time and test whether different subpopulations evince the factor structure at different points of development. Similarly, latent transition analysis (LTA) could also be used to assess change and stability in profile membership over adolescence as well as derive information about which variables substantively predict change from one profile into another. For example, LTA could be used to assess whether youth from low-income household have unique EF trajectories over adolescence, or delayed trajectories (e.g., are low-income youth in the Low EF profiles going to move into the average EF profile, “catching up” with their high-income peers?).

There are also opportunities to understand transactional relationships between EF and other variables of interest and to look at a wide range of outcomes over time with a diverse, population-level sample of adolescents that have been selected for representativeness (Garavan et al., 2018). Indeed, there is a critical need to understand how EF skills map onto both specific clinical symptoms and general functioning (Snyder et al., 2015). Studies have found general effects of EF on broad diagnostic categories (i.e., externalizing and internalizing disorders), but inconsistent associations when examining specific domains of EF (i.e., inhibitory control) across diagnoses (Bloemen et al., 2018). LPA allows for a nuanced examination of the role of specific cognitive processes across disorders that may be masked when utilizing typical factor analytic approaches. Thus, better understanding the structure of EF early in adolescence, will allow for better identification and treatment of psychopathology and improve long term outcomes.

Finally, additional assessments will bring these EF profiles further in line with other neurocognitive and developmental models of EF which suggest that EF may vary along an affective continuum from “cold” to “hot” (Zelazo & Carlson, 2012). Cold EF skills operate in affectively neutral situations, requiring primarily logic-based skills whereas hot EF skills operate in emotionally and motivational-relevant situations, requiring the integration of cognitive and affective processes (Prencipe et al., 2011). Developmental models of adolescent EF (Crone & Dahl, 2012) suggest that while cold EF skills mature relatively early in adolescence, hot EF skills, “come online” later in puberty, following a protracted development across adolescence that may have unique implications for adolescent health and wellbeing (Poon, 2017). Although only non-affective EF tasks were included in the current LPA, future work could also incorporate hot EF tasks measured in later assessments (e.g., Emotional Stroop, Delay Discounting task) and assess antecedents and consequences of hot EF-specific deficits or strengths.

Limitations

There are several limitations to consider when interpreting these findings. First, there were a limited number of EF tasks at baseline, and they primarily assessed core EF skills such as inhibitory control, working memory, and cognitive flexibility. Only one performance metric was assessed per task, although many others (e.g., response time, error rate, variability across trials) have demonstrated unique relations with outcomes. Further, although these skills are considered “foundational”, Karr et al. (2018) noted that the field has been limited by focus on these three skills and that research would benefit from an evaluation of other EF constructs such as planning, decision making, or future orientation. Second, effect sizes were generally small (albeit larger in the LPA compared to the variable-centered approach). Although this may suggest that the effect of modeling choices is minimal, small effect sizes may have important implications for different populations (Ewing, Bjork, et al., 2018; Matthay et al., 2021). Indeed, Thompson et al., 2019 notes that cumulatively, small effect sizes may explain a large proportion of variance in outcomes and translate to larger, population-level effects.

Third, we used household income and maternal education to assess SES, but ABCD contains a number of other sociodemographic indicators including family type, parent employment status, measures of economic insecurity, and census region (see Barch et al., 2018), which could be used to provide a richer picture of contextual factors linked to EF in future research. Further, although race/ethnicity was included as a covariate in the regression analyses given previous work reporting race/ethnic differences in behavioral outcomes (DeSteno et al., 2013), race/ethnic differences were not interpreted in the LPA and race/ethnicity was not explored by profile. No substantive theoretical perspectives on race/ethnic differences in EF exist that can be investigated in the current analyses (see Coll et al., 1996; Suzuki et al., 2021 for further elaboration). Future research, perhaps within the ABCD dataset, should explore race/ethnic differences in EF that are tied to systemic inequalities (e.g., structural racism, discrimination, task development and norms for different race/ethnic groups) but outside the scope of this study (Ford & Airhihenbuwa, 2010).

Conclusions

The purpose of the current study was to highlight the utility of person-centered approaches for EF research. LPA was used to derive meaningful profiles of adolescent EF which varied primarily in performance level, but also demonstrated discordance. Profiles demonstrated qualitatively and quantitatively different EF structures and were associated with differences by biological sex, household income, mother’s education, and pubertal timing, providing some preliminary evidence of what developmental processes may precede different EF structures. EF profiles also provided additional information about distal behavioral outcomes compared to a factor analytic approach, highlighting the utility of person-centered approaches for future, longitudinal assessments of adolescent EF.

Highlights.

1. Latent profile analysis was used to describe profiles of executive functioning (EF) in a population-level sample of early adolescents

2. Heterogenous constellations of EF were captured by four profiles, distinguished primarily by differences in performance level, but also discordance across tasks.

3. Biological sex, socioeconomic status, and pubertal timing predicted most likely profile membership

4. Profile membership predicted externalizing, internalizing, and problem behaviors assessed a year later.

Data Availability Statement:

The data are from the ABCD Study Curated Annual Release 3.0 and are available on request from the NIMH Data Archive (https://data-archive.nimh.nih.gov/abcd).