A canonical trajectory of executive function maturation from adolescence to adulthood

The study addresses when and how executive functions (EF)—including response inhibition, working memory, switching, and planning—mature from adolescence into adulthood, and whether EF development reflects a unitary domain-general process or multiple specific subcomponents. Prior theories propose prolonged EF maturation into late adolescence or even emerging adulthood, potentially explaining adolescent risk-taking and vulnerability to psychopathology. Empirical work has shown adolescents perform worse than adults on EF tasks, but the precise maturational timing, developmental shape (linear vs non-linear), and plateau toward adult levels remain unresolved due to small samples, narrow task batteries, and analytic approaches that assume fixed developmental forms. This work aims to define the canonical trajectory and maturational timing of EF using large-scale, multi-assessment, multi-dataset analyses and to test whether EF development is predominantly domain-general.

Prior research documents age-related increases in EF accuracy and faster response speed across adolescence on tasks such as working memory, inhibition, switching, and planning. However, studies were often limited by small sample sizes (N ≤ 200) with broad tasks or larger samples (N ≥ 1000) with narrow EF assessments, reducing generalizability and preventing precise maturation estimates. Common analytic approaches (linear, quadratic regressions; age-group comparisons) do not directly quantify timing or plateaus and cannot capture non-linear developmental shapes suggested by theory. The unity/diversity framework in adults indicates both commonality and separability across EF tasks, but adolescent EF development has rarely been examined across a broad, standardized set to adjudicate domain-general vs domain-specific trajectories. Emerging methods (e.g., GAMs) and large public datasets enable precise, non-parametric estimation of EF developmental timing and reproducibility across cohorts.

Design and samples: Aggregated four independent US-based datasets spanning ages 8–35 (N=10,766; total visits=13,817): two longitudinal (Luna: N=196, 666 visits; NCANDA: N=831, 3412 visits) and two cross-sectional (NKI: N=588; PNC: N=9151). Samples were community-based, sex-balanced, and broadly consistent with national race/ethnicity patterns. Inclusion focused on datasets with EF tasks covering late childhood through adulthood. Measures: 23 EF measures from 17 tasks targeting inhibition (e.g., Antisaccade, Stroop), working memory (e.g., Spatial Span, N-Back), planning (e.g., Stockings of Cambridge, D-KEFS Tower), attention and performance monitoring (e.g., Continuous Performance Test), and standardized batteries (Penn CNB, D-KEFS, CANTAB). Both accuracy and latency metrics were included where applicable. Preprocessing and quality control: Ensured valid IDs, ages, visit indices; removed task sessions with excessive missingness (e.g., >30% dropped eye-tracking trials), outliers based on residuals from initial GAM/GAMM fits (>4 SD), and multivariate outliers via Mahalanobis distance (>4 SD). Statistical analyses:

• Developmental trajectories: For each measure, fit non-linear age functions using GAMs (cross-sectional) and GAMMs (longitudinal) with penalized splines (mgcv). Longitudinal models included smooth visit effects and random intercepts/slopes.
• Periods of significant change and maturation: Computed first derivatives of age functions in 0.1-year bins; generated simultaneous 95% confidence intervals via posterior simulation (10,000 iterations) to identify ages with significant change (derivative CI excluding zero). Aggregated across measures/datasets using pointwise three-level meta-analysis (metafor).
• Interdependence and factor structure: Calculated between-person and within-person correlations among EF measures; performed exploratory factor analyses with bifactor rotation to assess dimensionality (parallel analysis, optimal coordinate, acceleration factor, Kaiser rule).
• Domain-general vs specific processes: For each measure, compared models with age predicted by a domain-general composite (leave-one-task/domain-out) vs the specific measure to partition age-related deviance into domain-general and measure-specific components.
• Normative templates and basis function regression: Created canonical EF trajectories (accuracy and latency) via out-of-sample meta-analytic GAM/GAMM fits and smoothing. Used these as single-parameter basis functions in GLM/GLMM to predict age-related change in left-out datasets, compared against standard age models (linear, inverse age, quadratic) using multiple fit/complexity metrics (R², adjusted R², ICC, RMSE, sigma, AIC, BIC). Sensitivity analyses: Examined robustness across sex, socioeconomic indicators (parental education, income), culturally acquired knowledge (verbal reasoning, vocabulary), mental health inclusion/exclusion, and practice effects (modeled visit effects in longitudinal data).

• Canonical non-linear trajectory: 20/23 measures showed significant age-related effects (corrected p < 0.004), with rapid improvement from ages 10–15, smaller but significant changes from 15–18, and stabilization to adult levels by approximately 18–20 years. Accuracy increased, latency decreased across tasks and datasets. Mean total age-related change across measures was large (mean 1.38 SD units).
• Periods of significant change: Derivative analyses showed significant per-year changes at 10–15 years (mean z change per year: accuracy +0.142; latency −0.175). From 15–18 years, effects were smaller yet often significant. After 18 years, few measures showed significant change; aggregate accuracy changes from 18–20 were minimal (mean +0.023 z/year, about one-fifth of the 10–15 rate). For measures with significant overall age effects, on average 95.0% (accuracy) and 99.7% (latency) of detectable change between 8–35 occurred before age 18.
• Domain generality: EF measures were positively intercorrelated within accuracy and latency. Factor analyses across datasets supported a dominant domain-general factor. Model comparisons indicated that a single composite domain-general EF metric accounted for most age-related variance in individual measures (meta-analytic mean deviance explained by domain-general processes: 79.3% for accuracy, 70.6% for latency), with some variability by dataset and measure.
• Normative templates and predictive modeling: A simplified single-parameter data-driven age basis function (derived out-of-sample) outperformed or matched standard age models in cross-validation. It was the top model 55.6% of the time (vs quadratic 37.3%, inverse age 7.03%, linear 0%); significantly better than inverse age (χ²=21.6, p<0.001) and linear (χ²=30.6, p<0.001), and comparable to quadratic (χ²=2.29, p=0.130). Results generalized even between datasets with no shared measures (best 69.2% between Luna and NKI). Offsetting the basis function reduced performance, underscoring the importance of accurate maturational timing.
• Robustness: Primary findings held across sex, socioeconomic covariates, culturally acquired knowledge measures, and mental health inclusion thresholds; practice effects were modeled and results replicated in cross-sectional datasets where practice cannot occur.

The findings provide convergent, reproducible evidence that EF develops rapidly in early to mid-adolescence, slows in mid to late adolescence, and reaches adult-like stability by approximately 18–20 years. This supports heuristic models positing adolescence as a period of ongoing EF maturation and clarifies its approximate closure before or near age 20, earlier than accounts proposing substantial EF development through emerging adulthood (18–25). The predominance of a domain-general driver suggests a common system of goal-directed cognition underpinning improvements across varied EF tasks, consistent with unity frameworks and potentially linked to global inhibitory control and common neural circuitry. Establishing a canonical EF trajectory enables developmentally informed modeling in new samples via basis function regression, improves reproducibility across instruments, and offers normative templates for identifying deviations relevant to health and disease. These results refine neurodevelopmental theories, inform clinical and policy considerations about adolescent capacities, and set the stage for translational applications using EF growth-charting approaches.

This work aggregates four large datasets and a broad EF battery to define a canonical, non-linear EF developmental trajectory from late childhood through adulthood, with most change occurring by late adolescence and stabilization around ages 18–20. EF development is largely driven by domain-general processes, and a simple data-driven basis function derived from this trajectory generalizes across datasets and tasks, often outperforming standard age models. These contributions provide reproducible normative templates to guide future research, clinical assessment, and policy. Future directions include: expanding EF batteries (including affective EF tasks), model-based computational measures, disentangling age from pubertal effects, increasing cultural and international generalizability, leveraging dense longitudinal and multimodal (e.g., neuroimaging) designs, and integrating normative templates into translational studies to quantify deviations linked to psychopathology and social determinants.

• Task scope: Focused on commonly used outcome measures; did not exhaustively cover all EF domains. Affective EF tasks were largely absent, limiting conclusions about EF under emotional contexts.
• Puberty vs age: Did not disentangle chronological age effects from pubertal development due to methodological constraints and large cross-sectional age effects.
• Generalizability: Predominantly US-based cohorts; need multinational, multicultural replication.
• Normative focus: Emphasized average trajectories; individual and dataset-level variability exists.
• Composite precision: Domain-general estimates depend on the breadth and precision of available measures; datasets with fewer measures showed more variability.
• Practice effects: Although modeled, longitudinal visit effects may persist; mitigated by replication in cross-sectional data.