The longitudinal structure of cognition in the ABCD study and associations with neural function and psychopathology: A Bayesian probabilistic principal components analysis Authors: Monica Luciana, Kelly A. Duffy, Wesley K. Thompson Published: 2026 ABSTRACT Longitudinal studies of cognition are necessary to estimate individual developmental trajectories and to link these trajectories with neurodevelopment. Short task batteries, if psychometrically sound, may have benefits over lengthy assessments. This paper focuses on the assessment of cognition in the Adolescent Brain Cognitive Development™ (ABCD) Study. Previously, we performed a Bayesian Probabilistic Principal Components Analysis (BPPCA) on ABCD’s baseline data, extracting components representing general ability, executive function and learning/memory. Because ABCD’s cognitive assessments changed over time, a strategy is needed to evaluate the longitudinal stability of these component scores based on the modified batteries, changes in scores over time, and associations with external variables. The current analyses address these gaps using available data from the ABCD baseline, year 2 and year 4 assessments. We extracted components from the modified task batteries, aligning them with those described by Thompson et al. (2019) at baseline. When the battery is reduced from 9 to 7 variables (baseline to year 2) and from 7 to 5 variables (year 2 to year 4), a three-component structure remains a good fit with task loadings similar to those observed for the full baseline battery. There are strong intercorrelations among component scores across 5, 7, and 9 subtest solutions. Using components derived from the 5-subtest solution, age- and practice-related changes are longitudinally evident; longitudinal component scores also differentially associate with psychopathology indices and working-memory task-related MRI imaging activation. Researchers may wish to leverage this approach to assess how cognition relates to neurodevelopment and other sources of individual variation. Introduction With refinements in neuroimaging technology that have emerged in the past twenty-five years, there have been increased efforts worldwide to model, within individuals, the course of brain maturation and associated behavioral processes (Kooijman et al., 2016, Mascarell Maričić et al., 2020, Volkow et al., 2018). Cognitive development is an area of particular interest given extant theories regarding the importance of processes such as executive function (EF) for adaptive behavior (Tervo-Clemmens et al., 2023). Longitudinal studies of cognition and cognitive development are necessary to assess variation in within-person change over time (Grammer et al., 2013, Lerner et al., 2009). Longitudinal studies enable researchers to estimate individual developmental trajectories of cognitive function, to link these trajectories with patterns of neurodevelopment, and to ascertain how various risk factors might contribute to significant deviations in function. The adolescent period has received particular scrutiny given the numerous social, cognitive and affective changes evident during this time (Backes and Bonnie., 2019). Despite its clear importance, quantifying within-person change through longitudinal assessment is fraught with pragmatic challenges, including the need to optimize the assessment approach so that it is not too burdensome as well as assuring that it is feasible, developmentally appropriate across the longitudinal assessment span, and psychometrically sound. Assessment batteries must be designed to reliably and validly capture cognitive constructs of interest. However, batteries that are inordinately lengthy may alienate participants, impacting retention and leading to missing data. Missing data, in turn, impacts the generalizability of findings and introduces numerous challenges for data analysis and interpretation (Caron-Diotte et al., 2023, Jelicić et al., 2009, Matta et al., 2018, Zhang and Sackett, 2023). Thus, short batteries, if psychometrically sound, may have pragmatic benefits over more lengthy assessments (Goetz et al., 2013). Moreover, once a longitudinal study is initiated, the assessment battery or individual measures may need to be shortened for pragmatic reasons (Oswald et al., 2015). The way this process is approached varies considerably. Statistical approaches such as Item Response Theory (Arrington et al., 2020, Ilardi et al., 2022) have been adopted, as have rational approaches, judgment-based strategies, and decisions based simply on practical contingencies (Coste et al., 1997, Kruyen et al., 2013). This paper focuses on the assessment of cognition in the Adolescent Brain Cognitive Development™ (ABCD) Study, comprising 21 data collection sites across the United States. Almost 12,000 youth were enrolled at ages 9–10 and are currently being followed with yearly longitudinal visits to assess patterns of brain development using magnetic resonance imaging (MRI), cognitive development using lab-based tasks, mental health, quality of life, and sociodemographic factors that may impact functioning over time (Garavan et al., 2018, Volkow et al., 2018). As of 2026, the ABCD Study® is in its 9th year of assessment, and to date, over 1200 papers have been published worldwide using the dataset. Approximately 35% of these papers represent studies that have utilized cognitive data in their analyses. The assessment of cognition was achieved at baseline using seven tests from the NIH Toolbox Cognition Battery plus several additional measures (e.g., the Little Man Task; Rey Auditory Verbal Learning Test; Luciana et al., 2018). Cognition has been comprehensively reassessed in the sample every two years since the baseline visit. However, the assessment battery has been modified over time. Specifically, some tasks that were administered at baseline were not administered at year 2 or at year 4 (see Table 1), leading to possible data analytic limitations for researchers interested in longitudinal analyses.Table 1ABCD core cognition measures and longitudinal assessment strategy.Assessment WaveBaselineYear 2Year 4Year 6Year 8Youth Age9–1011–1213–1415–1617–18TB Picture VocabularyXXXXTB Oral ReadingXXXXXTB List Sort Working MemoryXXXXTB Pattern Comparison Processing SpeedXXXXXTB FlankerXXXXXTB Dimensional Change Card SortXXXTB Picture Sequence MemoryXXXXXRey Auditory Verbal Learning TestXXXXLittle Man TaskXXXMatrix ReasoningXXGame of Dice TaskXXXSocial Influence TaskXTB = NIH Toolbox; To date, data from the baseline, year 2 and year 4 assessments are available to researchers for analysis. Year 6 data collection is complete but not yet released as of this work. Year 8 data collection has been initiated. The Matrix Reasoning, Game of Dice, and Social Influence Task data are not included in the PC analyses. Table 1ABCD core cognition measures and longitudinal assessment strategy.Assessment WaveBaselineYear 2Year 4Year 6Year 8Youth Age9–1011–1213–1415–1617–18TB Picture VocabularyXXXXTB Oral ReadingXXXXXTB List Sort Working MemoryXXXXTB Pattern Comparison Processing SpeedXXXXXTB FlankerXXXXXTB Dimensional Change Card SortXXXTB Picture Sequence MemoryXXXXXRey Auditory Verbal Learning TestXXXXLittle Man TaskXXXMatrix ReasoningXXGame of Dice TaskXXXSocial Influence TaskXTB = NIH Toolbox; To date, data from the baseline, year 2 and year 4 assessments are available to researchers for analysis. Year 6 data collection is complete but not yet released as of this work. Year 8 data collection has been initiated. The Matrix Reasoning, Game of Dice, and Social Influence Task data are not included in the PC analyses. ABCD core cognition measures and longitudinal assessment strategy. TB = NIH Toolbox; To date, data from the baseline, year 2 and year 4 assessments are available to researchers for analysis. Year 6 data collection is complete but not yet released as of this work. Year 8 data collection has been initiated. The Matrix Reasoning, Game of Dice, and Social Influence Task data are not included in the PC analyses. In prior work, we used the baseline ABCD data to perform a Bayesian Probabilistic Principal Components Analysis (BPPCA) and found evidence for three principal components (PCs), representing what we termed: 1) general ability; 2) executive function (EF); and 3) learning/memory (Thompson et al., 2019). Follow-up analyses of the baseline sample using ACE models indicated heritabilities of 0.36 for general ability, 0.22 for EF, and 0.29 for learning/memory; thus, for all three components, the influences of shared and unique environment were substantial (Maes et al., 2023). Composite measures, such as the three PCs as well as the fluid and crystallized composites from the NIH Toolbox, yielded higher heritabilities than did individual task measures. Assessment of the longitudinal retest stability of individual differences in ABCD participants’ baseline to year 2 data (Anokhin et al., 2022) indicated that retest stability ranged from fair (for individual tasks such as the Flanker test: r = 0.43) to excellent (NIH Toolbox Crystallized composite: r = 0.82). The ABCD study’s observed stability estimates for the NIH TB are highly consistent with a prior three-year longitudinal study of NIH Toolbox performance in youth aged 9–15 (Taylor et al., 2022). While the NIH Toolbox composite measures appear to have excellent psychometric properties, the BPPCA model suggested that cognition may be structured in a more nuanced way in young adolescents, accounted for by three primary components rather than two (Thompson et al., 2019). That analysis also supported the notion that executive functions are unified and not fully differentiated into relevant sub-components (Miyake and Friedman, 2012) until later adolescence, given that tasks that measured processing speed, inhibitory control, and cognitive flexibility loaded onto the same component in early adolescence. Individual differences in general ability, executive function, and learning/memory, as represented by the BPPCA components, also appear to be associated with variation in neural structure and function. A study that used precision functional mapping techniques to assess brain-behavior associations found that fronto-parietal precision functional mapping topography was associated with individual differences in general ability and, to a lesser extent, executive function and learning/memory in the ABCD baseline sample (Keller et al., 2023). Moreover, analyses have shown that task control-default mode network interconnections are a major locus of individual differences in cognitive functioning, particularly as related to general ability, in early adolescence (Sripada et al., 2020). Concordant with the observation that environmental variation impact neurocognitive function in this sample (Maes et al., 2023), resting state neural connectivity has been shown to be negatively impacted by environmental stressors, including living in resource-poor or disadvantaged environments; the lower within-network functional connectivity that is associated with these disadvantages is, in turn, associated with lower levels of general ability and learning/memory in the ABCD sample (Rakesh et al., 2021). Thus, extracting PCs from the ABCD study’s neurocognitive data has yielded novel insights regarding sociodemographic influences on cognition, the heritability of cognitive functions in early adolescence, and brain-behavior associations. However, because the task battery was shortened at subsequent assessments (ABCD’s year 2 and year 4 follow-up assessments), it is a challenge to evaluate the longitudinal stability of the BPPCA component scores, to evaluate mean and within-person development in these components over time, to assess experience effects, and to assess how correlations among cognition, neural structure and function, and behavior change with maturation. The current analyses address these gaps. Here, we present a strategy for extracting components from the modified battery that match those described by Thompson et al. (2019) at baseline, and we illustrate the reliability and utility of this method for describing cognition as well as its behavioral and neural correlates in the ABCD dataset. Methods We analyzed longitudinal data from the ABCD Study NIMH Data Archive (NDA) Data Release 5.1 (http://dx.doi.org/10.15154/z563-zd24). Within Data Release 5.1, cognitive assessments are available for participants at the study baseline (n = 11,868), year 2 (n = 10,973) and year 4 (n = 4754). Note that only roughly half of the sample had been assessed for the year 4 follow-up visit as of the time of the data freeze for the NDA Data Release 5.1. There were n = 2944 participants who completed visits for all three timepoints and provided complete data for the cognitive measures used in these analyses. Table 2 presents the socio-demographic characteristics of each of these subsamples.Table 2Demographic characteristics of the ABCD sample across waves and selected subsample under study.BaselineYear 2Year 4Longitudinal Samplen11,86810,97347542944Age in Months (Mean (SD))118.98 (7.50)144.32 (8.03)168.98 (8.17)119.60 (7.60)Female (%)5676 (47.8)5216 (47.5)2265 (47.6)1411 (47.9)Race (%)White7515 (64.3)7067 (65.3)3286 (70.0)2021 (69.5)Black1869 (16.0)1637 (15.1)533 (11.4)328 (11.3)Asian275 (2.4)252 (2.3)118 (2.5)68 (2.3)AIAN/NHPI78 (0.7)67 (0.6)35 (0.7)22 (0.8)Other525 (4.5)467 (4.3)188 (4.0)120 (4.1)Mixed1434 (12.3)1330 (12.3)535 (11.4)351 (12.1)Highest Parental Education (%)< HS Diploma593 (5.0)496 (4.5)191 (4.0)131 (4.5)HS Diploma/GED1132 (9.6)991 (9.0)345 (7.3)204 (6.9)Some College3074 (25.9)2769 (25.3)1196 (25.2)731 (24.9)Bachelor3013 (25.4)2857 (26.1)1301 (27.4)819 (27.9)Post-Graduate4041 (34.1)3846 (35.1)1714 (36.1)1055 (35.9)Married Parents (%)7982 (67.8)7525 (69.0)3393 (71.7)2111 (72.0)Household Income (%)< $50 K3222 (29.7)2859 (28.3)1171 (26.6)726 (26.4)$50K-100K3068 (28.3)2893 (28.7)1300 (29.5)818 (29.8)> $100 K4560 (42.0)4342 (43.0)1933 (43.9)1201 (43.8)Note: The age given for the longitudinal subsample (those with complete data at all timepoints) is for their ages at baseline, but note that data points from baseline, year 2, and year 4 are used for this sample Table 2Demographic characteristics of the ABCD sample across waves and selected subsample under study.BaselineYear 2Year 4Longitudinal Samplen11,86810,97347542944Age in Months (Mean (SD))118.98 (7.50)144.32 (8.03)168.98 (8.17)119.60 (7.60)Female (%)5676 (47.8)5216 (47.5)2265 (47.6)1411 (47.9)Race (%)White7515 (64.3)7067 (65.3)3286 (70.0)2021 (69.5)Black1869 (16.0)1637 (15.1)533 (11.4)328 (11.3)Asian275 (2.4)252 (2.3)118 (2.5)68 (2.3)AIAN/NHPI78 (0.7)67 (0.6)35 (0.7)22 (0.8)Other525 (4.5)467 (4.3)188 (4.0)120 (4.1)Mixed1434 (12.3)1330 (12.3)535 (11.4)351 (12.1)Highest Parental Education (%)< HS Diploma593 (5.0)496 (4.5)191 (4.0)131 (4.5)HS Diploma/GED1132 (9.6)991 (9.0)345 (7.3)204 (6.9)Some College3074 (25.9)2769 (25.3)1196 (25.2)731 (24.9)Bachelor3013 (25.4)2857 (26.1)1301 (27.4)819 (27.9)Post-Graduate4041 (34.1)3846 (35.1)1714 (36.1)1055 (35.9)Married Parents (%)7982 (67.8)7525 (69.0)3393 (71.7)2111 (72.0)Household Income (%)< $50 K3222 (29.7)2859 (28.3)1171 (26.6)726 (26.4)$50K-100K3068 (28.3)2893 (28.7)1300 (29.5)818 (29.8)> $100 K4560 (42.0)4342 (43.0)1933 (43.9)1201 (43.8)Note: The age given for the longitudinal subsample (those with complete data at all timepoints) is for their ages at baseline, but note that data points from baseline, year 2, and year 4 are used for this sample Demographic characteristics of the ABCD sample across waves and selected subsample under study. Note: The age given for the longitudinal subsample (those with complete data at all timepoints) is for their ages at baseline, but note that data points from baseline, year 2, and year 4 are used for this sample Cognitive measures At the study baseline, the neurocognitive battery was designed to be completed in 70 min (see Table 1). Participants first completed the Snellen vision chart (Snellen, 1862) as a measure of visual acuity. A brief handedness inventory, consisting of four self-report questions, was also administered (Veale, 2014). The neurocognitive testing battery, comprised of ten measures, was then initiated (Luciana et al., 2018). All tests were administered in person using an iPad with one-on-one monitoring by a research assistant. At years 2 and 4, there was a mixture of in person and remote assessment as a consequence of the COVID-19 pandemic, though this varied by participant and site. NIH toolbox® cognition measures The NIH Toolbox® cognition measures (hereafter referred to as “the NIH Toolbox”) were developed as part of the NIH Blueprint for Neuroscience Research (http://www.nihtoolbox.org). The Toolbox battery, which was normed for samples between the ages of 3 and 85, includes seven tasks that index episodic memory, executive function, attention, working memory, processing speed, and language abilities (Bleck et al., 2013, Gershon et al., 2013b, NIH Blueprint for Neuroscience Research, 2013) and can be administered within approximately 35 min. Despite the availability of a Spanish language version (Casaletto et al., 2016, Flores et al., 2017), the ABCD study has administered only the English language version to youth because English fluency was an inclusion criterion at initial recruitment. The Toolbox Picture Vocabulary Task® (Gershon et al., 2013a) measures language skills and verbal intellect. The Toolbox Oral Reading Recognition Task® is a reading test that asks individuals to pronounce single words (Gershon et al., 2013a). The Toolbox Pattern Comparison Processing Speed Test® (Carlozzi et al., 2015, Carlozzi et al., 2014, Carlozzi et al., 2013) measures rapid visual processing. The Toolbox List Sorting Working Memory Test® requires participants to use working memory to sequence task stimuli based on category membership and perceptual characteristics (Tulsky et al., 2014, Tulsky and Carlozzi, 2013). The Toolbox Picture Sequence Memory Test® was modeled after memory tests asking children to imitate a sequence of actions using props (Bauer et al., 2013, Dikmen et al., 2014). The Toolbox Flanker Task®(Eriksen and Eriksen, 1974) is a response inhibition/conflict monitoring task that measures the ability to modulate responding under congruent versus incongruent stimulus contexts. The Toolbox Dimensional Change Card Sort Task® measures cognitive flexibility (Zelazo et al., 2014, Zelazo et al., 2013). For these measures, uncorrected standard scores were used in our analyses. Rey auditory verbal learning test The Rey Auditory Verbal Learning Test (RAVLT) measures auditory verbal learning and short-term memory (Strauss et al., 2006, Taylor, 1959). A customized automated version, created through the Q-interactive platform of Pearson Assessments (http://www.helloq.com/home.html) was used. This test requires participants to listen to and recall a list of 15 unrelated words over five learning trials. Next, a distractor list of 15 words is presented, and the participant is asked to recall words from this second list. Recall of the initially learned list is then assessed. Recall following a 30-min delay (during which participants engage in non-verbal tasks) permits retention to be assessed. Little man task This task engages visual-spatial processing, specifically mental rotation, with varying degrees of difficulty (Acker and Acker, 1982). The task involves the presentation of a rudimentary male figure holding a briefcase in one hand in the middle of the iPad screen. The figure may appear in one of four positions: right side up vs. upside down and either facing the respondent or with its back to the respondent. The briefcase may be in either the right or left hand. Respondents indicate by button press which hand is holding the briefcase. Accuracy and response latency is recorded. Other measures Two additional cognitive measures were administered to participants at the study baseline but were not included in our analysis. One measure (the Cash Choice Task, Wulfert et al., 2002) is a single-item delay of gratification measure with dichotomous scoring, which was not used after the baseline assessment. The second measure, the Matrix Reasoning Task, from the Wechsler Intelligence Test for Children-V (WISC-V; Wechsler et al., 2014) is a measure of fluid reasoning that was administered at baseline but not at year 2 or 4 using automated technology (Q-interactive; Daniel et al., 2014). Preliminary analyses of the baseline data suggested that the Cash Choice task did not load onto any of the three extracted components. Preliminary analyses further indicated that the Matrix Reasoning Task scores were distributed across all components and that the general BPPCA solution was equivalent with and without inclusion of the task (see Thompson et al., 2019). Child behavior checklist At baseline, year 2 and year 4, externalizing and internalizing behaviors were reported by the parent using an automated version of the Child Behavior Checklist (CBCL: Achenbach, 2009). The CBCL is comprised of 113 items that measure aspects of the child’s behavior across the past six months. Each item was rated using a three-point rating scale (not true, somewhat or sometimes true, very often or always true). Internalizing and externalizing summary scores are derived from the following syndrome scores: anxious/depressed, withdrawn/depressed, somatic complaints, social problems, thought problems, rule-breaking behavior, and aggressive behavior. Similar to Thompson et al. (2019), four CBCL-derived raw score measures were included in these analyses: Internalizing, Externalizing, Stress Reactivity, and Total Problem scores. EN-back functional magnetic resonance imaging (fMRI) task In addition to neurobehavioral measures of cognition, the ABCD study also includes several fMRI tasks that index various components of cognition (Casey et al., 2018). One of these measures, the EN-back task (Barch et al., 2013, Cohen et al., 2016), requires memory as well as emotion regulation for successful completion. We focused on the EN-back task for several reasons. First, based on assessments of the ABCD baseline data, the EN-back task has the best reliability of the study’s task-based fMRI measures (Kennedy et al., 2022). Second, the task is complex, theoretically requiring a range of perceptual and cognitive abilities for successful execution including attention, visuospatial reasoning, memory, cognitive control over the content of working memory, and control over emotional information as information is cognitively processed. We reasoned that this task, relative to the other fMRI measures, might be expected to recruit aspects of general ability, executive function and learning/memory as indexed by the PC scores. In contrast, the Stop-Signal task is more focused on inhibitory control but in the context of some methodological controversy (Garavan et al., 2022), and the Monetary Incentive Delay Task was primarily designed to index aspects of reward processing (Casey et al., 2018). Third, the neural networks recruited by the EN-back have been particularly well characterized in adolescents as well as adults with similar overall patterns observed across development (Chaarani et al., 2021, Satterthwaite et al., 2013, Yaple et al., 2019). Finally, among papers published to date, the literature base is most extensive for the EN-back task (abcdstudy.org). The task includes two runs, each of which includes eight trial blocks. On each of 160 total trials, participants view pictures one at a time and must indicate with button presses whether the currently viewed picture matches or does not match stimuli that were presented previously. Within each run, four task blocks are 0-back conditions during which participants must respond “match” if the current stimulus is the same as a cued target presented at the beginning of the block. The other four task blocks are 2-back conditions during which participants respond “match” if the current picture is the same as the one presented two trials previously. ABCD uses an emotional variant of the task where some stimuli include happy, fearful, and neutral facial expressions; places are also targeted. For each stimulus type (faces and places), accuracy within each of the two memory load conditions (0- and 2-back) is the primary behavioral measure; accuracy measures for each of the two memory load conditions can also be contrasted with one another. Moreover, the EN-back reliably elicits activation within frontolimbic and frontostriatal regions (Casey et al., 2018, O'Brien et al., 2020). Similar to accuracy scores, BOLD activations acquired during each stimulus condition can be measured and contrasted, and these have been reported for ABCD study participants (O'Brien et al., 2020). Based on prior work (Barch et al., 2013, Casey et al., 2018, O'Brien et al., 2020), we selected a priori the average beta weights for BOLD contrasts of 2-back versus 0-back conditions (all runs) within 12 Desikan parcellations that reflect regions associated with N-back working memory performance and associated these with PC scores (see below). Statistical approach Bayesian probabilistic principal components analysis (BPPCA) The Bayesian Probabilistic Principal Components (BPPCA) model (Gelman et al., 2003) was described in detail in Thompson et al. (2019). The BPPCA model applied here is identical to the one described there, with a straightforward extension to handle longitudinal visits by incorporating a random participant effect. This extension is described in the Supplementary Materials section. Briefly, let Yij denote the P-dimensional vector of NIH Toolbox measures for the ith participant and jth visit. In this application, P = 5 (number of NIH Toolbox measures common to all three visits) and j = 1, 2, or 3 depending on if the data are from Baseline, Year-2 Follow-up, or Year-4 Follow-up visits, respectively. We modeled the covariance structure of Yij using a latent variable formulationYij=Λθij+ϵij,where Λ is a P×D matrix, D≤P is the number of latent variables, θij is the D-dimensional vector of latent scores for the ith subject at the jth visit, and ϵij is the P-dimensional vector of residuals. We assumed that θij were marginally normal with zero mean and unit variance, indicated by θij∼ND(0,ID), where ND(μ,Σ) denotes a D-dimensional normal distribution with mean μ and variance-covariance matrix Σ, 0 is a D-dimensional vector of zeros and ID is the D×D identity matrix. We further assumed that the residuals were marginally distributed as ϵij∼NP(0,σϵ2IP). This model forms the basis for Probabilistic PCA (PPCA: Bishop, 1999). By specifying prior distributions on all the parameters to be estimated, this becomes a Bayesian PPCA (BPPCA) model. Relative to standard PCA modeling, BPPCA has several advantages: 1) the Bayesian approach results in principled posterior standard errors of all parameters, including loadings, that would be very difficult to obtain using a non-probabilistic approach; 2) the LOOIC selection criterion provides a more nuanced assessment of model complexity and number of factors to retain; and 3) it allows quantification of the amount of variance explained by family, site, and individual for all PC’s, which is of independent interest and otherwise not easily obtainable from standard methods in one algorithm. We incorporated the study design of ABCD into the analysis via random effects for data collection site, for family (for those with one or more siblings included in the study) and for participant. Random effects thus accounted for correlation among the responses between participants caused by nesting within participant (over time), within family, and within site. Random effects were included in both the latent scores θij and the residuals ϵij, with variances constrained to meet the identifiability assumptions of the BPPCA model. For details on the basic model formulation and implementation, please see Thompson et al. (2019) [Appendix A: Supplemental Data]. This BPPCA model was implemented in stan and the R software environment for statistical computing and graphics (Carpenter et al., 2017, R Core Team, 2024, Stan Development Team, 2023, Stan Development Team, 2025). Follow-up analyses After performing the BPPCA analyses, we extracted the participant latent scores (θij in the notation above). We first placed them as dependent variables in generalized additive mixed models (GAMMs: Wood, 2003, Wood, 2004, Wood, 2011, Wood, 2017; Wood et al., 2016). Age and sex at birth were entered as independent variables. To investigate practice effects, we also entered visit number (1, 2, or 3) as a factor-valued variable (Salthouse, 2010). Note that visit number is distinct from study wave and reflects how many visits had been completed by the participant to date. Site, family, and individual were entered as random effects. GAMMs were run in R Version 4.4.1 using the package mgcv. Next, we used the factor scores θijas independent variables in association analyses with the CBCL symptom measures as dependent variables (Internalizing, Externalizing, Stress Reactivity, and Total Problem R-scores). Because these measures are bounded below by zero and skewed left (see Fig. 1), we implemented this regression using a generalized linear mixed model with a Gamma distribution (log link) function. To do so, we had to first add 1 to each score to avoid taking the logarithm of zero. Because prior research has shown that psychopathology varies as a function of age and sex at birth, we used these two variables as fixed-effect covariates. We also included ABCD site, family and participant as random effects (Ren et al., 2022).Fig. 1Spaghetti plots of raw scores for CBCL Internalizing, Externalizing, Stress Reactivity, and Total Problem Scores as a function of age. Only 250 of the total N = 2944 participants are shown for clarity. Spaghetti plots of raw scores for CBCL Internalizing, Externalizing, Stress Reactivity, and Total Problem Scores as a function of age. Only 250 of the total N = 2944 participants are shown for clarity. Finally, we used the factor scores θij as dependent variables in association analyses with task activation for the emotional N-back fMRI data. Independent variables were task activation contrasts (2- vs. 0-back) in regions of interest identified from both an prior report using baseline ABCD data (O'Brien et al., 2020) as well as the broader literature (Owen et al., 2005). Independent variables of interest were activation contrasts in 12 Desikan parcellations bilaterally including the caudal and rostral middle frontal regions, the caudate, anterior cingulate cortex, parahippocampal gyrus and hippocampus. Age and sex were included as fixed effects, and ABCD site, family and participant were included as random effects. R and stan scripts for the BPCCA algorithm and the follow-up analyses are available within the ABCD GitHub repository (https://github.com/ABCD-STUDY/BPPCA_Neurocognition_Luciana_etal_2026). Scripts for creating the input RDS file from the NDA Data Release 5.1 are available there. The results of this paper should thus be exactly reproducible for any authorized user of the ABCD Study Release 5.1 tabulated data. Results Baseline to year-2 replication of component structure Of the nine neurocognitive tasks administered at the ABCD Study Baseline visit (Luciana et al., 2018, Thompson et al., 2019), two were not included as part of the Year 2 follow-up visit due to concerns about potential redundancy with ABCD’s fMRI measures (NIH Toolbox List Sort Working Memory) as well as concerns about practice effects after learning the task rule (NIH Toolbox Dimensional Change Card Sort). We repeated the BPPCA analysis to evaluate whether the same general principal components structure we observed at baseline using the full nine-subtest battery of tasks (Thompson et al., 2019) would replicate once again using this subset of seven tasks: 1) the NIH Toolbox Picture Vocabulary task; 2) the NIH Toolbox Flanker task; 3) the NIH Toolbox Pattern Comparison Processing Speed task; 4) the NIH Toolbox Picture Sequence Memory task; 5) the NIH Toolbox Oral Reading task; 6) the Little Man Task [LMT]; and 7) the Rey Auditory Verbal Learning Task [RAVLT]) and using the data collected at both the baseline and year 2 follow-up. As there were now repeated measurements from individuals across these two assessment waves, this required adding an individual-level random effect to the BPPCA method, as described in the Methods section. Thus, we ran the BPPCA algorithm, including random effects for individuals nested within families nested within sites, for all participants with complete neurocognitive data at both baseline and year 2 (N = 6985). The BPPCA algorithm was run for D = 1, …, 5, where D is the number of retained PCs. We evaluated the fit of each model by computing the Leave-One-Out Information Criterion (LOOIC; Vehtari et al., 2017) for each of the solutions. We found a substantial improvement moving from one (LOOIC = 260,075, sd = 567) to two (LOOIC = 258,189, sd = 575) components, as well as from two to three (LOOIC = 256,879, sd = 577) components. The improvement moving from three to four (LOOIC = 256,334, sd = 590) components was more modest, and the further improvement from four to five (LOOIC = 256,167, sd = 603) components was nominal. Thus, in the interest of parsimony and for comparability to the previous BPPCA solution, we elected to use D = 3 components. The loadings for the tasks onto this three-PC solution, following a varimax rotation, are shown in Table 3.Table 3Varimax-rotated loadings for the three-component solution based on the reduced 7-subtask battery using baseline and year 2 data.PC 1PC 2PC 3Task0.0250.500.9750.0250.500.9750.0250.500.975Pic Vocab0.7260.7530.7740.1230.1430.1690.1660.1880.216Flanker0.1840.2100.2400.8080.8360.8590.0530.0780.107Pattern0.1470.1710.2010.8870.9130.9350.1030.1250.153Picture0.0480.0740.1040.3080.3340.3630.7710.8020.829Reading0.7830.8070.8290.1560.1840.2060.1040.1240.146RAVLT0.2120.2470.278-0.075-0.058-0.0360.7240.7570.784LMT0.4770.5300.5750.4500.4930.5280.1020.1370.178Table providing the loadings, at the 0.025, 0.50, and 0.975 %iles, of each of the 7 selected tasks from the baseline and year 2 data onto the three components. Tasks that load highly onto each component (above 0.40) have their loadings bolded at the 0.50 %ile only. LMT = Little Man Task, RAVLT = Rey Auditory Verbal Learning Task. Table 3Varimax-rotated loadings for the three-component solution based on the reduced 7-subtask battery using baseline and year 2 data.PC 1PC 2PC 3Task0.0250.500.9750.0250.500.9750.0250.500.975Pic Vocab0.7260.7530.7740.1230.1430.1690.1660.1880.216Flanker0.1840.2100.2400.8080.8360.8590.0530.0780.107Pattern0.1470.1710.2010.8870.9130.9350.1030.1250.153Picture0.0480.0740.1040.3080.3340.3630.7710.8020.829Reading0.7830.8070.8290.1560.1840.2060.1040.1240.146RAVLT0.2120.2470.278-0.075-0.058-0.0360.7240.7570.784LMT0.4770.5300.5750.4500.4930.5280.1020.1370.178Table providing the loadings, at the 0.025, 0.50, and 0.975 %iles, of each of the 7 selected tasks from the baseline and year 2 data onto the three components. Tasks that load highly onto each component (above 0.40) have their loadings bolded at the 0.50 %ile only. LMT = Little Man Task, RAVLT = Rey Auditory Verbal Learning Task. Varimax-rotated loadings for the three-component solution based on the reduced 7-subtask battery using baseline and year 2 data. Table providing the loadings, at the 0.025, 0.50, and 0.975 %iles, of each of the 7 selected tasks from the baseline and year 2 data onto the three components. Tasks that load highly onto each component (above 0.40) have their loadings bolded at the 0.50 %ile only. LMT = Little Man Task, RAVLT = Rey Auditory Verbal Learning Task. Table 3 shows that PC1 has its strongest loadings on the NIH Toolbox (TB) Picture Vocabulary task and the TB Oral Reading task, with a more moderate (but still strong) loading on the Little Man Task. This component is very similar to the “General Ability” component identified in the full-battery baseline analyses, which had the strongest loadings for the TB Picture Vocabulary and TB Oral Reading tasks, and more moderate loadings for the Little Man and TB List Sort Working Memory tasks, the latter of which was dropped from the present analyses. PC2 has its strongest loadings on the TB Flanker and TB Pattern Comparison Processing Speed tasks, with a somewhat more moderate (but still strong) loading from the Little Man Task. This component is very similar to the “Executive Functioning” component identified in the full-battery baseline analyses, which had the strongest loadings for the TB Flanker and Pattern Comparison Processing Speed tasks as well as the TB Dimensional Change Card Sort task, which was dropped from the present analyses. Interestingly, though, the Little Man task loads more strongly onto this component in the present analysis than it did in the full-battery baseline analyses (Thompson et al., 2019). PC3 has its strongest loadings for the TB Picture Sequence Memory task and the RAVLT total number correct. This component is very similar to the “Learning and Memory” component identified in the full-battery baseline analyses, which had the strongest loadings on these same two tasks. Thus, the three-component solution obtained using the reduced battery of seven subtests from the data collected at Baseline and Year-2 Follow-up has essentially the same overall structure as the three components obtained from the full nine-test battery, but with the two dropped tasks (the TB List Sort Working Memory and Dimensional Change Card Sort tasks) no longer loading on their respective components. The only other noticeable difference is that the Little Man task now loads approximately equally strongly onto both the General Ability and Executive Functioning components (PC1 and PC2), whereas previously it loaded more strongly onto the General Ability component. To confirm that the component structure was generally replicated, we correlated the PC scores from the three-component solution extracted from the reduced seven-subtest battery with the scores from the three-component solution extracted from the original full nine-subtest battery within individuals at baseline using Spearman rank-order correlations. These correlations are presented in Table 4. As indicated, at baseline, the correlation between the original full-battery and the reduced battery scores was r = 0.97 for PC1 (General Ability), r = 0.91 for PC2 (Executive Functioning) and r = 0.96 for PC3 (Learning/Memory). These strong correlations demonstrate conclusively that the components from the full nine-subtest battery and the reduced seven-subset battery are capturing essentially the same cognitive structure at baseline.Table 4Correlations among components derived from 9, 7, and 5 subtest solutions.Reduced PC1Reduced PC2Reduced PC3NIHTB-only PC1NIHTB-only PC2NIHTB-only PC3Original PC10.970.130.320.920.110.13Original PC20.160.910.170.100.910.12Original PC 30.260.110.960.280.060.85Spearman rank-order correlations between 3 component PC solution (“Original”) based on 9 subtasks and 3 component solution based on 7 subtasks (“Reduced”) and between the “Original” PC solution and the 5 subtask NIH Toolbox-only (“NIHTB-only”) 3 component solution, at baseline only Table 4Correlations among components derived from 9, 7, and 5 subtest solutions.Reduced PC1Reduced PC2Reduced PC3NIHTB-only PC1NIHTB-only PC2NIHTB-only PC3Original PC10.970.130.320.920.110.13Original PC20.160.910.170.100.910.12Original PC 30.260.110.960.280.060.85Spearman rank-order correlations between 3 component PC solution (“Original”) based on 9 subtasks and 3 component solution based on 7 subtasks (“Reduced”) and between the “Original” PC solution and the 5 subtask NIH Toolbox-only (“NIHTB-only”) 3 component solution, at baseline only Correlations among components derived from 9, 7, and 5 subtest solutions. Spearman rank-order correlations between 3 component PC solution (“Original”) based on 9 subtasks and 3 component solution based on 7 subtasks (“Reduced”) and between the “Original” PC solution and the 5 subtask NIH Toolbox-only (“NIHTB-only”) 3 component solution, at baseline only Longitudinal component structure: Baseline, year-2 follow-up and year-4 follow-up At the year 4 follow-up, the ABCD neurocognitive battery was further reduced (Table 1). Specifically, the RAVLT was not administered at this visit. The component structure was examined again as a function of tasks that were consistently included over time. For this longitudinal examination, we elected to drop the RAVLT as well as the Little Man task, the only non-NIH toolbox task remaining in the battery. Of note is the fact that the Little Man Task is not being administered after the year 4 assessment (e.g., at years 6 or 8), inspiring us to focus on tasks that are central to the neurocognition battery not only from baseline to year 4 but afterwards. Moreover, in addition to the motivation of parsimony, the Little Man Task showed inconsistency in which component it loaded most highly on when separating the observations into distinct age bands as a sensitivity analysis to determine whether the structure was stable across age ranges, meaning that the component loadings are unstable (See Supplementary Material). Thus, the BPPCA analysis was repeated using data from baseline, year 2 and year 4 with only the five TB Tasks that were common to all three assessments (Picture Vocabulary, Oral Reading, Flanker, Pattern Comparison Processing Speed, and Picture Sequence Memory tasks) to determine whether the component solution again replicated with this yet further reduced battery. For this analysis, participants with complete neurocognitive data at the baseline, year 2 and year 4 waves (n = 2944) were included. We again ran the BPPCA algorithm for D = 1, …, 5 and evaluated the LOOIC for each of the solutions. We found a small improvement moving from one (LOOIC = 115,816, sd = 352) to two (LOOIC = 115,369, sd = 362) components, and a substantial improvement moving from two to three (LOOIC = 113,704, sd = 357) components. The improvement moving from three to four (LOOIC = 113,674, sd = 357) components was nominal, and moving from four to five (LOOIC = 113,678, sd = 357) components led to a slight increase in the LOOIC. Thus, the three component solution was selected for the purposes of parsimony and for the greatest comparability to the baseline and year 2 PC solutions. Task loadings for this three-component solution, following a varimax rotation, are shown in Table 5. As indicated, PC1 has the strongest loadings from the Picture Vocabulary and Oral Reading tasks. This component, now best construed as a measure of crystallized cognition, overlaps considerably with the “General Ability” component identified both the full nine-subtest battery from the baseline analyses and the reduced seven-subtest battery from the baseline and year 2 follow-up analyses shown above. The solution based on seven subtests differs from this further reduced solution in that it additionally had a moderate loading from the Little Man task, which was dropped from the present analyses. PC2 has the strongest loadings from the Flanker and Pattern Comparison Processing Speed tasks and is very similar to the “Executive Functioning” component identified in both the full-battery baseline analyses and the reduced seven-subtest battery from the baseline and year 2 follow-up analyses shown above. Once again, the only difference between this component in this further-reduced battery and that based on the seven-subtest solution is the moderate loading for the Little Man task when using seven subtasks. PC3 has strong loadings from the Picture Sequence Memory task. This component is thus similar to the “Learning and Memory” component identified in both the full nine-subtest battery from the baseline analyses and the reduced seven-subtest battery from the baseline and year 2 analyses shown above. Previously, in both solutions, this component also showed a strong loading from the RAVLT total correct score, which was dropped in the current analyses due to its absence from the year 4 assessment.Table 5Reduced five subtest battery: Varimax-rotated loadings for the three-component solution.PC 1PC 2PC 3Task0.0250.500.9750.0250.500.9750.0250.500.975Pic Vocab0.7550.7780.7990.1890.250.2210.1220.1420.161Flanker0.1840.2050.2260.9110.9310.9530.0800.0990.120Pattern0.3040.3330.3630.8540.8790.9050.1300.1540.177Picture0.1530.1650.1760.1350.1470.1590.9710.9861.002Reading0.8040.8240.8460.2380.2560.2740.0810.0990.118Varimax-rotated loadings for the three-component solution based on the reduced 5-subtask battery using Baseline, Year-2 Follow-up, and Year-4 Follow-up data; The table provides the loadings, at the 0.025, 0.50, and 0.975 %iles, of each of the 5 selected tasks from the baseline, year 2, and year 4 data onto the three components. Tasks that load highly onto each component (above 0.40) have their loadings bolded at the 0.50 %ile only. Table 5Reduced five subtest battery: Varimax-rotated loadings for the three-component solution.PC 1PC 2PC 3Task0.0250.500.9750.0250.500.9750.0250.500.975Pic Vocab0.7550.7780.7990.1890.250.2210.1220.1420.161Flanker0.1840.2050.2260.9110.9310.9530.0800.0990.120Pattern0.3040.3330.3630.8540.8790.9050.1300.1540.177Picture0.1530.1650.1760.1350.1470.1590.9710.9861.002Reading0.8040.8240.8460.2380.2560.2740.0810.0990.118Varimax-rotated loadings for the three-component solution based on the reduced 5-subtask battery using Baseline, Year-2 Follow-up, and Year-4 Follow-up data; The table provides the loadings, at the 0.025, 0.50, and 0.975 %iles, of each of the 5 selected tasks from the baseline, year 2, and year 4 data onto the three components. Tasks that load highly onto each component (above 0.40) have their loadings bolded at the 0.50 %ile only. Reduced five subtest battery: Varimax-rotated loadings for the three-component solution. Varimax-rotated loadings for the three-component solution based on the reduced 5-subtask battery using Baseline, Year-2 Follow-up, and Year-4 Follow-up data; The table provides the loadings, at the 0.025, 0.50, and 0.975 %iles, of each of the 5 selected tasks from the baseline, year 2, and year 4 data onto the three components. Tasks that load highly onto each component (above 0.40) have their loadings bolded at the 0.50 %ile only. Based on the component loadings, it appears that the three-component solution obtained using the further reduced battery of the five NIH toolbox only subtests using baseline through year 4 follow-up data captures the same overall cognitive structure as the three components obtained from the full nine-subtest battery at baseline and as the three components obtained from the reduced seven-subtest battery at baseline and year 2. To confirm that the component structure was generally replicated from the original nine-subtest battery solution, we correlated the PC scores from that three-component solution with the PC scores extracted from this reduced NIH Toolbox-only battery within individuals at baseline. These correlations are presented in the right-most columns of Table 4. Using baseline data, the correlation between the original 9 subtest-based full-battery and the reduced 5 subtest-based NIH Toolbox-only battery scores was r = 0.92 for PC1 (General Ability), r = 0.91 for PC2 (Executive Functioning) and r = 0.85 for PC3 (Learning/Memory). These correlations are slightly lower than the correlations seen between the full nine-subtest battery solution and the reduced seven-subtest battery at baseline, but this reduction is to be expected given that fewer tasks are loading onto each component (particularly PC3, the Learning/Memory component, which is largely carried by a single task). These reduced correlations could be due to less precision in component scores with fewer tasks. Nevertheless, the correlations are still quite high, suggesting a general replication of the component structure from the full nine-subtest battery at baseline and the reduced NIH Toolbox-only battery through year 4. The five-subtest-based NIH Toolbox only solution was also examined in relation to the reduced seven-subtest battery solution constructed using the baseline and year 2 follow-up data. First, we correlated the PC scores from the reduced seven-subtest battery solution with the scores from this NIH Toolbox-only battery within individuals at baseline. The results, shown in Table 6, indicate that the highest correlation between the seven-subtest battery solution and the further reduced NIH Toolbox-only battery is for PC1 (General Ability), which has a correlation of r = 0.95 for the observations from baseline. PC2 (Executive Function) has a correlation of r = 0.92, and PC3 (Learning/Memory) has a correlation of r = 0.81 across the two solutions at baseline. As the reduced seven-subtest battery solution was constructed using data from baseline and year 2 follow-up, we also correlated the PC scores from the seven-subtest battery solution with the NIH Toolbox-only solution within individuals using their scores at year 2. These results, shown in the right-hand columns of Table 6, indicate correlations of r = 0.95 for PC1, r = 0.91 for PC2 and r = 0.79 for PC3. In general, these patterns indicate that the three-component PC solution is largely stable and replicable across the two reduced subsets of tasks, and across both the baseline and year 2 follow-up visits. As such, we will consider the solution for this reducted NIH Toolbox-only battery of tasks as the final model of interest, and use these scores to assess relationships between cognition and other measures, and we suggest the use of these scores for longitudinal analysis in future work.Table 6Correlations among components based on the 7 and 5-subtest solutions.BaselineYear 2Reduced PC1Reduced PC2Reduced PC3Reduced PC1Reduced PC2Reduced PC3NIHTB-only PC10.950.050.350.950.040.35NIHTB-only PC20.110.920.060.070.910.07NIHTB-only PC30.150.170.810.140.220.79Spearman correlations between the 3 component solution based on 7 subtasks (“Reduced”) and the 5 subtask NIH Toolbox-only (“NIHTB-only”) 3 component solution values calculated at baseline and at year 2 Table 6Correlations among components based on the 7 and 5-subtest solutions.BaselineYear 2Reduced PC1Reduced PC2Reduced PC3Reduced PC1Reduced PC2Reduced PC3NIHTB-only PC10.950.050.350.950.040.35NIHTB-only PC20.110.920.060.070.910.07NIHTB-only PC30.150.170.810.140.220.79Spearman correlations between the 3 component solution based on 7 subtasks (“Reduced”) and the 5 subtask NIH Toolbox-only (“NIHTB-only”) 3 component solution values calculated at baseline and at year 2 Correlations among components based on the 7 and 5-subtest solutions. Spearman correlations between the 3 component solution based on 7 subtasks (“Reduced”) and the 5 subtask NIH Toolbox-only (“NIHTB-only”) 3 component solution values calculated at baseline and at year 2 Longitudinal change in component scores Here, we retained the NIH Toolbox-only scores θij in performing follow-up analyses, in the sample of n = 2944 participants who had complete data on all five subtests at all three visits. We first fit GAMMs using the θijas dependent variables (one GAMM per PC, for three total). Age was entered as a smooth fixed-effect variable, along with sex-assigned-at-birth and visit. Spaghetti plots of the PC scores as a function of age are given in Fig. 2. Results of the GAMMs are given in Table 7. All three PCs show strong positive associations with age (p’s < 0.001). The mean trajectories are displayed in Fig. 2 (heavy red lines). Likelihood ratio tests of the smooth (non-linear) model vs. a linear model indicates strong evidence for non-linearity for all three PCs (p’s < 0.001). As can be seen in Fig. 2, mean trajectories are still rising but with a slowing rate of growth by the year 4 follow-up, when participants are 150–189 months (12.5–15.8 years) of age. Sex is not a significant predictor of PC scores except for PC3, where male participants show slightly lower scores than females (coef = −0.106, p = 0.001). Interestingly, experience effects (as captured by “visit”) are mixed, with no evidence of experience effects for PC1, evidence of a year 2 experience effect for PC3 (coef = 0.198, p < 0.001) but no year 4 follow up experience effect, and evidence for more robust experience effects for PC2 (Year 2 follow-up: coeff = 0.427, p < 0.001; Year 4 follow-up: coeff = 0.658, p < 0.001). Thus, this index of crystallized cognition seems generally impervious to practice effects, while the component indexing explicit learning/memory shows practice effects after the initial administration. Executive function shows continued gains with practice. Finally, we used the GAMMs to compute the conditional R-squared (given variance accounted for by both fixed and random effects). For PC1 the conditional R-squared= 0.838, for PC2 the conditional R-squared = 0.686, and for PC3 the conditional R-squared = 0.503.Fig. 2Spaghetti plots of PC scores as a function of age. Only 250 of the total N = 2944 participants are shown for clarity. Heavy red lines depict GAMM fits.Table 7Coefficients for GAMM models predicting PC1, PC2, and PC3 using sex assigned at birth, visit, and a smooth effect of age.PC 1PC 2PC 3Coeff.Std. ErrortpCoeff.Std. ErrortpCoeff.Std. ErrortpIntercept-0.0220.071-0.3090.758-0.3510.074-4.7222.37e-06***-0.0970.053-1.8050.071Sex0.0050.0490.1030.918-0.0480.046-1.0410.298-0.1060.033-3.2170.001**Visit 20.0030.0700.0420.9670.4270.0775.5403.12e-08***0.1990.0553.5843.40e-04***Visit 30.1350.1301.0330.3020.6580.1384.7661.91e-06***0.1090.09971.0960.273edfRef dfFpedfRef dfFpedfRef dfFps(Age)4.5864.58683.16< 2e-16 ***3.6493.64988.24< 2e-16 ***2.9422.94244.17< 2e-16 ***Adj. R-squared0.3490.4080.178Note: Asterisks indicate significance of p-values: * corresponds to a p-value of < 0.05, ** corresponds to a p-value of < 0.01, and *** corresponds to a p-value of < 0.001. “s(Age)” indicates that Age is being modeled as a smooth effect, and as such, nonparametric coefficient terms are included. All other terms in the model display their corresponding parametric coefficients. Spaghetti plots of PC scores as a function of age. Only 250 of the total N = 2944 participants are shown for clarity. Heavy red lines depict GAMM fits. Table 7Coefficients for GAMM models predicting PC1, PC2, and PC3 using sex assigned at birth, visit, and a smooth effect of age.PC 1PC 2PC 3Coeff.Std. ErrortpCoeff.Std. ErrortpCoeff.Std. ErrortpIntercept-0.0220.071-0.3090.758-0.3510.074-4.7222.37e-06***-0.0970.053-1.8050.071Sex0.0050.0490.1030.918-0.0480.046-1.0410.298-0.1060.033-3.2170.001**Visit 20.0030.0700.0420.9670.4270.0775.5403.12e-08***0.1990.0553.5843.40e-04***Visit 30.1350.1301.0330.3020.6580.1384.7661.91e-06***0.1090.09971.0960.273edfRef dfFpedfRef dfFpedfRef dfFps(Age)4.5864.58683.16< 2e-16 ***3.6493.64988.24< 2e-16 ***2.9422.94244.17< 2e-16 ***Adj. R-squared0.3490.4080.178Note: Asterisks indicate significance of p-values: * corresponds to a p-value of < 0.05, ** corresponds to a p-value of < 0.01, and *** corresponds to a p-value of < 0.001. “s(Age)” indicates that Age is being modeled as a smooth effect, and as such, nonparametric coefficient terms are included. All other terms in the model display their corresponding parametric coefficients. Coefficients for GAMM models predicting PC1, PC2, and PC3 using sex assigned at birth, visit, and a smooth effect of age. Note: Asterisks indicate significance of p-values: * corresponds to a p-value of < 0.05, ** corresponds to a p-value of < 0.01, and *** corresponds to a p-value of < 0.001. “s(Age)” indicates that Age is being modeled as a smooth effect, and as such, nonparametric coefficient terms are included. All other terms in the model display their corresponding parametric coefficients. Associations of component scores with psychopathology We examined Pearson as well as Spearman rank-order correlations between the principal components and CBCL scores for each of the three calculated solutions using the baseline data (Tables S9 and S10A). As indicated, correlation magnitudes and directions are highly similar. To the extent that there were differences, there is some slight attenuation in correlation magnitudes with the Baseline + Year 2 + Year 4 solution. We then carried the θij in analyses using them as independent variables in generalized mixed effects models (GLMMs) with CBCL raw scores (Externalizing, Internalizing, Stress Reactivity, and Total Problems) as dependent variables. Histograms and spaghetti plots of the CBCL scores as a function of age are given in Fig. 1. Table 8 displays the coefficients of the four models, revealing distinct patterns in the association of CBCL scores with the PCs. For PC1, a significant positive association was observed in the Internalizing model (b = 0.068, t = 4.48, p < 0.001) as well as the Stress Reactivity model (b = 0.028, t = 2.07, p = 0.039). However, PC1 was not significantly related to Externalizing or Total Problems as a main effect. However, there were several interactions by sex for PC1. Specifically, for Internalizing, higher PC1 scores were associated with higher predicted internalizing in both males and females, but the relative increase is steeper in females. For Externalizing, both males and females have lower predicted externalizing scores as the PC1 score increases, but the relative decline in externalizing as PC1 scores increase is steeper in females. We observed a similar pattern for Stress Reactivity. These interactions by sex are presented in a Supplemental FigureTable 8Coefficients for GLMM models predicting CBCL Psychopathology Indices.InternalizingExternalizingStress ReactivityTotal ProblemsEstimatet-valuep-valueEstimatet-valuep-valueEstimatet-valuep-valueEstimatet-valuep-valueAge-1.37e-03-2.6090.009-0.001-2.6850.007**0.00040.9150.360-0.003-5.4405.47e-08***Sex-0.128-4.5375.771e-06***0.1314.4339.41e-06***0.0431.7410.08160.1073.2979.80e-04***PC 10.0684.4797.61e-06***-0.003-0.1990.8420.0282.0670.038*0.0201.1860.236PC 20.0070.5990.576-0.0001-0.0090.9920.00090.0830.9340.0120.8750.382PC 3-0.035-2.4080.016*-0.035-2.4020.016*-0.047-3.6392.75e-04***-0.048-3.0950.002**Sex:PC 1-0.0560-2.7590.006**-0.040-1.9730.049*-0.060-3.3747.45e-04***-0.040-1.8520.064Sex:PC 2-0.026-1.5510.1210.0070.4160.6770.007-0.4620.644-0.029-1.6430.100Sex:PC 30.0090.4530.651-0.005-0.2790.7800.0221.2510.2110.0080.3920.695Note: Coefficients for GLMM models predicting CBCL Internalizing, Externalizing, ADHD, and Total Problems Raw Scores using age, sex assigned at birth, and cognitive principal components; Asterisks indicate significance of p-values: * corresponds to a p-value of < 0.05, ** corresponds to a p-value of < 0.01, and *** corresponds to a p-value of < 0.001. Table 8Coefficients for GLMM models predicting CBCL Psychopathology Indices.InternalizingExternalizingStress ReactivityTotal ProblemsEstimatet-valuep-valueEstimatet-valuep-valueEstimatet-valuep-valueEstimatet-valuep-valueAge-1.37e-03-2.6090.009-0.001-2.6850.007**0.00040.9150.360-0.003-5.4405.47e-08***Sex-0.128-4.5375.771e-06***0.1314.4339.41e-06***0.0431.7410.08160.1073.2979.80e-04***PC 10.0684.4797.61e-06***-0.003-0.1990.8420.0282.0670.038*0.0201.1860.236PC 20.0070.5990.576-0.0001-0.0090.9920.00090.0830.9340.0120.8750.382PC 3-0.035-2.4080.016*-0.035-2.4020.016*-0.047-3.6392.75e-04***-0.048-3.0950.002**Sex:PC 1-0.0560-2.7590.006**-0.040-1.9730.049*-0.060-3.3747.45e-04***-0.040-1.8520.064Sex:PC 2-0.026-1.5510.1210.0070.4160.6770.007-0.4620.644-0.029-1.6430.100Sex:PC 30.0090.4530.651-0.005-0.2790.7800.0221.2510.2110.0080.3920.695Note: Coefficients for GLMM models predicting CBCL Internalizing, Externalizing, ADHD, and Total Problems Raw Scores using age, sex assigned at birth, and cognitive principal components; Asterisks indicate significance of p-values: * corresponds to a p-value of < 0.05, ** corresponds to a p-value of < 0.01, and *** corresponds to a p-value of < 0.001. Coefficients for GLMM models predicting CBCL Psychopathology Indices. Note: Coefficients for GLMM models predicting CBCL Internalizing, Externalizing, ADHD, and Total Problems Raw Scores using age, sex assigned at birth, and cognitive principal components; Asterisks indicate significance of p-values: * corresponds to a p-value of < 0.05, ** corresponds to a p-value of < 0.01, and *** corresponds to a p-value of < 0.001. In contrast, PC2 did not significantly predict any of the outcomes. PC3 showed a robust and consistent pattern of negative associations across all models, with significant effects in the Internalizing (b = −0.035, t = -2.41, p = 0.016), Externalizing (b = −0.035, t = -2.40, p = 0.016), Stress Reactivity (b = −0.0466137, t = -3.64, p < 0.001), and Total Problems models (b = −0.048, t = -3.10, p = 0.002). This pattern suggests that while low learning/memory scores are a consistent predictor of higher symptom levels across all domains, the association of crystallized ability with psychopathology varies by outcome and, surprisingly, executive function does not appear to play a significant role in these models. Associations of component scores with emotional N-back task activations Finally, we used the θij in analyses as dependent variables in mixed-effects models with task MRI activation contrasts (2- vs. 0-back) in twelve Desikan regions of interest (ROIs) identified from the literature as relevant to N-back based working memory function (Barch et al., 2013, O'Brien et al., 2020, Owen et al., 2005). Since 36 total models were fitted (three PC’s as dependent variables and twelve ROIs as independent variables), a Bonferroni adjustment for 36 total comparisons was implemented. Resulting associations are presented in Table 9.Table 9Coefficients for LMM models predicting cognitive PC scores from emotional n-back task fMRI activation values.PC 1PC 2PC 3Estimatet-valuep-valueEstimatet-valuep-valueEstimatet-valuep-valueCaudate Middle Frontal (Left)0.0252.2480.8860.0503.4190.023*0.0443.7510.006**Caudate Middle Frontal (Right)0.0252.2970.7800.0422.8190.1750.0393.3010.035*Rostral Middle Frontal (Left)0.0212.0201.0000.0332.3180.7380.0363.1490.059Rostral Middle Frontal (Right)0.0121.1451.0000.0322.2430.8970.0201.7351.000Caudate Anterior Cingulate (Left)0.0131.2781.0000.0120.8431.0000.0242.1251.000Caudate Anterior Cingulate (Right)0.0141.3271.0000.0070.5081.0000.0161.4191.000Rostral Anterior Cingulate (Left)-0.010-0.9651.000-0.044-3.1150.067-0.028-2.5310.411Rostral Anterior Cingulate (Right)0.0020.1781.000-0.019-1.3271.000-0.018-1.5961.000Parahippocampal Gyrus (Left)-0.005-0.5181.000-0.005-0.3831.000-0.017-1.5401.000Parahippocampal Gyrus (Right)-0.019-1.7891.000-0.025-1.7491.000-0.023-2.0491.000Hippocampus (Left)-0.031-3.0220.091-0.053-3.770.006**-0.042-3.7670.006**Hippocampus (Right)-0.023-2.2350.918-0.03-2.1471.000-0.026-2.3570.665Note: Coefficients for LMM models predicting cognitive PCs 1, 2, and 3 using activation values from various brain ROIs in the emotional n-back fMRI task, controlling for age and sex assigned at birth and including random effects for individuals nested within families nested within sites. As models for each outcome were fit sequentially, all p-values are corrected for multiple comparisons (36 total comparisons) using a Bonferroni correction. Asterisks indicate significance of p-values: * corresponds to a p-value of < 0.05, ** corresponds to a p-value of < 0.01, and *** corresponds to a p-value of < 0.001. Table 9Coefficients for LMM models predicting cognitive PC scores from emotional n-back task fMRI activation values.PC 1PC 2PC 3Estimatet-valuep-valueEstimatet-valuep-valueEstimatet-valuep-valueCaudate Middle Frontal (Left)0.0252.2480.8860.0503.4190.023*0.0443.7510.006**Caudate Middle Frontal (Right)0.0252.2970.7800.0422.8190.1750.0393.3010.035*Rostral Middle Frontal (Left)0.0212.0201.0000.0332.3180.7380.0363.1490.059Rostral Middle Frontal (Right)0.0121.1451.0000.0322.2430.8970.0201.7351.000Caudate Anterior Cingulate (Left)0.0131.2781.0000.0120.8431.0000.0242.1251.000Caudate Anterior Cingulate (Right)0.0141.3271.0000.0070.5081.0000.0161.4191.000Rostral Anterior Cingulate (Left)-0.010-0.9651.000-0.044-3.1150.067-0.028-2.5310.411Rostral Anterior Cingulate (Right)0.0020.1781.000-0.019-1.3271.000-0.018-1.5961.000Parahippocampal Gyrus (Left)-0.005-0.5181.000-0.005-0.3831.000-0.017-1.5401.000Parahippocampal Gyrus (Right)-0.019-1.7891.000-0.025-1.7491.000-0.023-2.0491.000Hippocampus (Left)-0.031-3.0220.091-0.053-3.770.006**-0.042-3.7670.006**Hippocampus (Right)-0.023-2.2350.918-0.03-2.1471.000-0.026-2.3570.665Note: Coefficients for LMM models predicting cognitive PCs 1, 2, and 3 using activation values from various brain ROIs in the emotional n-back fMRI task, controlling for age and sex assigned at birth and including random effects for individuals nested within families nested within sites. As models for each outcome were fit sequentially, all p-values are corrected for multiple comparisons (36 total comparisons) using a Bonferroni correction. Asterisks indicate significance of p-values: * corresponds to a p-value of < 0.05, ** corresponds to a p-value of < 0.01, and *** corresponds to a p-value of < 0.001. Coefficients for LMM models predicting cognitive PC scores from emotional n-back task fMRI activation values. Note: Coefficients for LMM models predicting cognitive PCs 1, 2, and 3 using activation values from various brain ROIs in the emotional n-back fMRI task, controlling for age and sex assigned at birth and including random effects for individuals nested within families nested within sites. As models for each outcome were fit sequentially, all p-values are corrected for multiple comparisons (36 total comparisons) using a Bonferroni correction. Asterisks indicate significance of p-values: * corresponds to a p-value of < 0.05, ** corresponds to a p-value of < 0.01, and *** corresponds to a p-value of < 0.001. After deleting observations with missing task MRI data or not meeting QC standards, there were n = 2, 804 participants in these association analyses. After Bonferroni corrections, there were no significant associations with PC1, two significant associations with PC2, and three significant associations with PC3. The significant associations with PC2 (executive function) were with the left caudal middle frontal region (beta = 0.050, t = 3.42, p = 0.023) and the left hippocampus (beta = −0.053, t = -3.77, p = 0.006). The significant associations with PC3 (learning/memory) were with the left caudal middle frontal region (beta = 0.044, t = 3.751, p = 0.006), the right caudal middle frontal region (beta = 0.039, t = 3.30, p = 0.035) and the left hippocampus (beta = −0.042, t = -3.77, p = 0.006). Associations between PC2 (executive function) and PC3 (learning/memory) and BOLD activation in the caudal middle frontal region were positive, while associations with activation in the left hippocampus were negative, indicating that better cognitive performance outside of the scanner was associated with relative reductions in task-based hippocampal activation. As a posthoc analysis, we assessed whether these associations varied as a function of in-scanner task performance and found no significant interaction effects. Discussion In the course of conducting large-scale longitudinal studies, thoughtful decisions must be made regarding which measures to administer and the schedule for administration over time. Using a data reduction approach applied to three waves of the Adolescent Brain Cognitive Development study, we have shown that a reduced battery, implemented at waves 2 and 4, performs comparably to the full baseline battery in terms of the component structure, how specific tasks load onto each component, and each component’s pattern of association with an external measure, the CBCL. Specifically, at the study baseline, when participants were ages 9–10 years, 9 measures were used to assess various domains of cognition, and a Bayesian principal components analysis revealed that these 9 measures loaded differentially onto three primary components that represent general cognition, executive function and learning/memory, respectively (Thompson et al., 2019). At wave 2, when participants were aged 11–12 years, two of the 9 measures were dropped from the battery and here we have shown that the three-component structure remains a good fit to the baseline data with task loadings that are very similar to what was observed for the larger battery. Moreover, there are strong intercorrelations between components based on 7 versus 9 subtests. Researchers who choose to rely on data from the baseline and year 2 assessment waves can be confident that the integrity of the battery, from a conceptual standpoint, is intact. The changes that were made from baseline to year 2 primarily impact the assessment of executive function given that two EF tasks were eliminated, so it is notable that the correlation between the 9-subtest-based second component and the 7-subtest variant is quite high, well above 0.90. At year 4, when participants were aged 13–14, the cognitive assessment was further adjusted. Five of the original 9 subtests, all from the NIH Toolbox Cognition battery, were consistently administered to participants at baseline, year 2 and year 4. Although there were some shifts in task loadings, the three-component structure remains a good fit to the data when the structure of these five tasks is considered. Moreover, the longitudinal stability of the three components (as derived from the five-subtask battery) is good to excellent as reflected by the conditional R2 values extracted from our GAMM analyses, and particularly so for the component that reflects crystallized cognition. These stability patterns are consistent with those based on individual subtasks as reported by Anokhin et al. (2022) using data from baseline and year 2 as well as data from a three-year longitudinal study of the NIH Toolbox involving youth aged 9–15 (Taylor, 2022). There were relatively minimal shifts in task loadings, particularly for the five tasks consistently administered over time between the 9-subtest-based baseline solution (Thompson et al., 2019) and the 7-subtest and 5-subtest solutions. It is more challenging to speculate on the meanings of changes in loadings for tasks that were inconsistently administered over time. Loadings for the Picture Vocabulary and Oral Reading tasks on the first principal component are nearly identical across solutions. The Flanker and Pattern Comparison Processing Speed tasks load most strongly onto the executive function component across all solutions but loadings increase for each task as the overall number of tasks decreases from 9 to 7 to 5. Similarly, for the Learning/Memory component, the Picture Sequence Memory task loadings increase from a value of 0.86 (based on the 9 subtest battery: Thompson et al., 2019) to 0.99 when a five-subtest battery is implemented. These shifts in task loadings are sensible from a psychometric perspective given that the overall configuration of tasks changed over time. Moreover, these shifts do not generally alter the nature of the overarching constructs reflected by each component based on correlations that we observe with external measures (e.g., the CBCL). Despite changes in task configuration, the second and third principal components appear to consistently index Executive Function and Learning/Memory, respectively. One noticeable difference is that with the reduced seven-task PC solution, the Little Man task now loads approximately equally strongly onto both the General Ability and Executive Functioning components (PC1 and PC2), whereas previously it loaded more strongly onto the PC1 (General Ability) component. Whether this finding is meaningful cannot be easily determined. As we have described above, loadings for the Little Man Task are unstable across age bands, decreasing the likelihood that this shift is meaningful. The change that is most pronounced as we move from a 9- to 7- to 5-subtest solution is that the first principal component, which we initially labeled “general ability” is better construed as reflective of crystallized cognition in the context of the 5-subtest solution. Thus, the very slight changes in the structure of cognition that are observed across solutions are largely driven by psychometrics and are not likely to be conceptually meaningful. Data from ABCD’s year 6 and, to some extent, year 8 batteries will be more informative in addressing whether there are changes in the structure of cognition over time, because most measures from the full baseline battery are repeated at those assessment waves. Overall, these patterns provide specific support for the use of the NIH Toolbox Cognition Battery to index cognitive development during adolescence and support the idea that major domains of cognition can be reliably measured longitudinally using an abbreviated battery, a finding that has both pragmatic and theoretical implications. Neurodevelopmental changes in cognition during adolescence are of particular interest given the educational advances that occur during this time, the increasing demands on youth to be able to problem solve under novel conditions and given the theoretical importance of executive functions for decision-making when youth are faced with risky choices. It is therefore particularly encouraging that the psychometric integrity of the second component (which primarily reflects EF given the nature of task loadings) was maintained as the overall battery was reduced from 9 to 7 to 5 subtests. Developmentally, using the EF component based on the five-subtest solution, we were able to demonstrate continued gains in EF capacities as the ABCD sample matures from ages 9‐10 to ages 13‐14. The age-related increases that we observe (Fig. 1) cohere with the extant literature, which overwhelmingly suggests marked increases in common, as well as specific, aspects of EF during this age range (Tervo-Clemmens et al., 2023, Lozano Wun et al., 2025). Similarly, we observe age-related increases in crystallized cognition as well as learning and memory functions, although age effects were strongest for EF followed by crystallized cognition and learning/memory. Notably, these age-related increases could be differentiated from the effects of practice given that we controlled for each participant’s degree of prior experience with the battery. This differentiation is essential to interpretation given that the ABCD study is following a large cohort that is relatively homogeneous in age within each assessment wave and given that practice-related gains are evident after one’s initial experience with laboratory-based cognitive testing (Elman et al., 2025, Sullivan et al., 2017). These analyses indicate differential effects of practice as a function of which component is being measured. Specifically, for crystallized cognition, there were no effects of practice as individuals moved from 1 to 2 to 3 assessments, a finding that coheres with what has been observed in the NCANDA study (Sullivan et al., 2017); this finding is particularly encouraging given that the NIH Toolbox does not include alternate forms for its Picture Vocabulary and Oral Reading measures. For EF, there were practice-related increments in component scores with each successive assessment, consistent with the idea that task familiarity benefits EF performance (Strobach et al., 2014). For explicit learning and short-term recall, practice-related gains were evident from the first to the second test assessment but were not significant thereafter, consistent with the notion that for some tasks, practice effects may be greatest after the first test administration (Ibrahim et al., 2015, Waber et al., 2012). Notably, the NIH Toolbox Picture Sequence Memory task, which loads heavily onto the learning/memory component, has three alternate forms, which were utilized by ABCD across assessment waves, so the decreasing impact of practice with successive administrations may be due to ongoing changes in item content. Thompson et al. (2019) examined principal component scores derived from 9 subtests at the ABCD baseline in relation to various scores from the Child Behavior Checklist, while controlling for demographic factors. Numerous demographic factors were included as covariates in that study, and the analyses focused cross-sectionally on roughly half of the ABCD baseline sample. Low general ability (PC1) scores were associated with high externalizing and high stress reactivity. Low learning/memory scores (PC3) were associated with high stress reactivity. Executive function did not significantly relate to CBCL-based internalizing, externalizing or stress reactivity. When a similar approach was adopted here using component scores derived from the five NIH Toolbox tasks in a longitudinal sample assessed between the study baseline and year 4, several of the previously-reported associations were replicated and some more nuanced findings emerged. Some patterns did not fully replicate. Consistent with prior findings, the basic patterns of association between the PC scores and CBCL indices of psychopathology were maintained in terms of both the magnitudes of association and direction of effects. In addition, executive function as represented by the second principal component did not relate to CBCL-based indices of psychopathology in the longitudinal data. Low scores on the learning/memory component were once again associated with higher levels of externalizing, higher levels of internalizing, and higher levels of stress reactivity, suggesting that difficulties in this cognitive domain may represent a general marker of psychopathology in the adolescent period, consistent with the idea that these forms of psychopathology are risk factors for poor academic achievement (Van der Ende et al., 2016). A new finding was that high rather than low scores on the first principal component, which represents crystallized cognition, were associated with relatively more internalizing symptomatology as well as more stress reactivity, associations that were more prominent in females. Interactions by sex were not assessed in the prior paper. Though speculative, it may be that females who are relatively high-achieving are more vulnerable to internalizing due to perfectionistic tendencies (Sand et al., 2021). Causal modeling was not utilized in this study, so the direction of influence remains unclear. Unlike the pattern observed by Thompson et al. (2019), scores on the first principal component were not significantly related to externalizing. As we have noted, the first principal component changed with the reduction from nine to five subtests: the five-subtest solution focuses more narrowly on crystallized cognition, while the nine-subtest solution includes additional domains of cognition such as visuospatial reasoning. Nonetheless, it was expected that associations between externalizing symptoms and crystallized cognition would be weak in magnitude and negative in direction based on the extant literature (Hinshaw, 1992) as well as our prior paper (Thompson et al., 2019). The basic correlation between the first principal component score and externalizing is negative in magnitude across all three PC solutions that were evaluated longitudinally, but is slightly attenuated when the PC scores are based on the five subtest solution (Table S10A). An additional methodological complication is that fewer demographic covariates were included in the current longitudinal modeling scheme as compared to the Thompson et al. (2019) analyses, which could account for differences observed across the two studies. Thompson et al. (2019) did not examine associations between PC scores and fMRI task performance at the study baseline; the current analyses extend this prior work and provide further support for distinctions as well as some similarities among the observed components in relation to their external correlates. In their analysis of task-based fMRI data from the ABCD baseline sample, O'Brien et al., (2020)) observed for the EN-back that as memory load increased and regardless of stimulus type, there was increased activation in frontostriatal regions bilaterally, including the dorsolateral prefrontal cortex, caudal anterior cingulate cortex (ACC), the caudate and the putamen, consistent with the idea that these regions are activated as memory demands increase and the task becomes more effortful; increased activations in these frontostriatal as well as frontoparietal regions are commonly observed during N-back working memory performance in adults (Barch et al., 2013, Owen et al., 2005) and in adolescents, including ABCD participants (Rosenberg et al., 2020), although within the literature as a whole, the positive direction of activation has been less consistently observed in children and adolescents (Seidman et al., 2016). With increased memory demand and similar to other reports (Rosenberg et al., 2020), O’Brien et al. observed decreased activations bilaterally in hippocampal and parahippocampal regions, the amygdala, nucleus accumbens, and rostral ACC as EN-back memory load increased; these regions are generally associated with memory (e.g., hippocampus) as well as emotional processing. Our observation that relatively better out-of-scanner performance in executive function and learning/memory, indexed here by two separate principal components that cut across multiple task measures, is associated with increased EN-back-associated bilateral caudal middle frontal activation as well as decreased left hemisphere hippocampal activation, both in the context of increased memory demand, validates the EN-back as a measure of complex domain-general memory-based executive function that can be used to index individual differences in behavior outside of the restrictive fMRI setting. Better performance on composite measures of memory and executive function measured out-of-scanner is associated not only with an in-scanner increased capacity to recruit cognitive control regions as executive demands increase but also with simultaneous decreases in subcortical (hipppocampal) activation. This combination of increased cortical but decreased subcortical activation might be viewed as a signature of more strongly developed cognitive, as well as emotional, regulation skills (Luciana, 2016), consistent with dual systems frameworks of adolescent development (Casey et al., 2018, Lozano Wun et al., 2025). The observation that these patterns hold across the longitudinal assessment span from ages 9–10 (study baseline) to ages 12–13 (year 4 assessment) expands upon prior work with the ABCD sample (O'Brien et al., 2020, Rosenberg et al., 2020), further suggesting that interactions among age-independent individual strengths and vulnerabilities and neurodevelopmental trajectories are important to examine in future studies. Overall, one might question the value of longitudinally examining three components versus five individual tasks. We maintain that using the components has several advantages over use of the individual tasks. First, the PCA approach yields components that identify shared variance across multiple tasks, which aids in interpretation by making the relative contribution of each task onto broader concepts (such as executive functioning or general cognition) more clear. The components are also (roughly) orthogonal, which the individual tasks are not, which further aids in interpretation when all tasks are considered simultaneously. Importantly, because the PCA approach consistently identifies three similar components across differing variations of contributing tasks, age ranges, and subject groups, these three components appear to robustly reflect the actual structure of the data. The PCA approach also leads to components that are likely to have more favorable psychometric properties, such as higher retest stability, as compared to the individual tasks. Finally, although we acknowledge that there is not a great degree of dimension reduction from five tasks to three components, reducing the dimensionality of any large dataset is often considered valuable for subsequent work and will be of use to future researchers who want to simplify their analyses and minimize multiple comparisons. Limitations Several limitations must be acknowledged. Difficult decisions were made by the ABCD Consortium regarding which measures to retain over time as longitudinal assessment progressed. These decisions were made by a committee based on pragmatic considerations and were not necessarily fully informed by psychometric data regarding the integrity of each specific subtest. Measures of EF, a construct that is of great interest to the adolescent research community, were most strongly impacted by the assessment changes that occurred over time. In assessing the coherence of the battery in the context of these reductions, we focused on ABCD’s out-of-scanner behavioral battery and did not include metrics from the functional MRI measures, several of which reflect aspects of EF. Moreover, it must be acknowledged that the Bayesian PCA approach is only one of several possible methods that might have been utilized to examine the structure of cognition based on multiple available subtests. While this technique results in data reduction, it is certainly not the only possible approach for assessing the latent structure of cognition in the sample; future work might focus on techniques such as confirmatory factor analysis with more extensive testing of factor invariance over time. In assessing the intercorrelations of components derived from the full baseline task battery and those derived from the reduced batteries, we used the baseline sample. Because that sample is limited to youth in the 9–10 year-old age range, it is not fully clear that associations of the same magnitude would be observed in adolescents as they age. In that regard, it is encouraging that we were able to detect age- and practice-related changes in performance over time using the five-subtest battery. Future releases of ABCD data will allow for further validation of the methods presented herein given that the baseline battery is being largely repeated at year 8. Conclusion Overall, using longitudinal data from ABCD study participants who completed three waves of assessment, we have demonstrated cohesion of the neurocognitive assessment battery using a Bayesian principal components analysis and have shown that components derived from a reduced battery are sensitive to patterns of age- and practice-related change in the span from late childhood to mid-adolescence. We also demonstrate the utility of the components in describing associations between cognitive performance and psychopathology as measured by the CBCL and with patterns of working memory-based neural activation during functional neuroimaging. Researchers who utilize these same tasks are readily able to replicate this approach to assess how these major domains of cognition relate to markers of neural structure and function, factors such as pubertal development, environmental variables, and other sources of individual variation. CRediT authorship contribution statement Monica Luciana: Writing – review & editing, Writing – original draft, Supervision, Methodology, Funding acquisition, Conceptualization. Wesley K. Thompson: Writing – review & editing, Writing – original draft, Visualization, Supervision, Software, Project administration, Methodology, Funding acquisition, Formal analysis, Data curation. Kelly A. Duffy: Writing – review & editing, Writing – original draft, Visualization, Methodology, Formal analysis, Conceptualization. Declaration of Competing Interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.