A short-term longitudinal study linking adolescents' metacognition, learning, and social friendship networks

The study examines how adolescents’ metacognitive processes—monitoring and control—relate to learning outcomes, and whether classroom friendships shape these processes and outcomes over time. During the transition to secondary school, adolescents face increased demands for self-regulated learning and heightened social dynamics. Prior work indicates on-task measures of procedural metacognition relate more strongly to achievement than self-reports, yet adolescent research often relies on self-reports. Additionally, social network studies typically focus on broad outcomes (e.g., GPA), leaving uncertainty about peer influence on task-specific learning processes. The authors address these gaps by combining on-task metacognition measures with classroom friendship networks over two waves (3 months apart). Hypotheses: H1—higher monitoring accuracy, monitoring-based restudy, and decision accuracy predict higher task scores; H2—adolescents become more similar to friends in metacognition and task performance over time (peer influence), while considering potential friendship selection based on similarity.

Effects of metacognition on learning: Metacognitive monitoring (assessing one’s performance) and control (regulating actions such as restudy and response submission) directly influence learning outcomes (Nelson, 1990; De Bruin & van Gog, 2012). Procedural, on-task measures better predict achievement than declarative, self-reported measures (Muncer et al., 2022; Ohtani & Hisasaka, 2018). Yet, adolescent studies rarely use on-task methods (Gascoine et al., 2017), and often focus only on monitoring rather than both monitoring and control. Metacognitive skills improve across development, but individual differences in translating monitoring into effective control are substantial (Roebers, 2017; van Loon & Roebers, 2017). Experience with specific tasks may refine metacognitive strategies (Bayard et al., 2021; Paulus et al., 2014). Peer influences on metacognition and learning: Adolescents’ behaviors and achievement are shaped by peers via selection (befriending similar others) and influence (becoming more similar over time) (Laninga-Wijnen & Veenstra, 2021). Social network analyses show selection and influence for academic engagement and achievement (Laninga-Wijnen et al., 2019; Wang et al., 2018; Shin, 2018). Metacognitive processes can be visible and discussed (Paulus et al., 2014; Dindar et al., 2019; Hurme et al., 2006), suggesting potential peer influence. However, little is known about whether friends shape on-task monitoring and control processes. This study tests selection and influence on metacognitive components and task performance.

Design: Short-term longitudinal study with two classroom-based waves separated by ~3 months. Participants: T1 N=136 seventh graders (53.8% female; mean age 13.8) from eight Swiss classrooms; T2 N=117 (attrition due to school changes/absence/motivation). Informed consent obtained; those without data consent completed the task but were excluded from stored data and social network survey. None had prior Kanji knowledge. Procedure: At each wave, researchers administered an on-computer Kanji learning task, followed by a paper-based classroom friendship nomination survey. The task included study, self-test, confidence judgments (monitoring), optional restudy selection and retest, and decisions to submit or withdraw responses for scoring. Classes received a monetary reward. Kanji metacognition task: Adapted from van Loon & Oeri (2023) using items from Destan & Roebers (2015) and Roebers et al. (2019). Programmed in Qualtrics. Each wave used different, difficulty-matched sets. Participants learned 24 Kanji across 3 blocks of 8 (randomized order). After each study block: (a) self-test (item scored 0/1), (b) confidence judgment (0–3), (c) optional restudy selection (0/1 per item), then retest and updated response for restudied items. Final phase: decide for each response whether to submit for grading or withdraw. Instructions emphasized points gained for correct submissions and deductions for incorrect submissions. Coding: performance 0/1; confidence 0–3; restudy 0/1; decision accuracy 1=maintain correct or withdraw incorrect; 0 otherwise. Task score (0–24) summed points for submitted responses using initial responses for non-restudied items and updated responses for restudied items. Social network questionnaire: At each wave, students nominated up to seven classroom friends (could nominate none). Nominations were restricted to within-class ties; impossible between-class ties coded as structural zeros for RSiena. Data aggregated into an adjacency matrix (rows=nominators; columns=nominees; 1=friendship nomination, 0 otherwise). Measures of metacognition:

• Monitoring accuracy: Linear mixed modeling (lme4) regressing item-level initial performance on confidence judgments with random intercepts and participant-specific random slopes (Murayama et al., 2014) to derive individualized monitoring accuracy independent of performance.
• Monitoring-based restudy: Mixed models regressing item-level restudy decisions on confidence judgments; participant-specific random slopes index how monitoring informed restudy choices.
• Decision accuracy: Proportion of accurate submit/withdraw decisions (Roebers & Spiess, 2017). Statistical analyses:
• Path analyses (lavaan) at T1 and T2 tested H1: monitoring accuracy, monitoring-based restudy, and decision accuracy predicting task scores while accounting for correlations among predictors.
• Social network analysis (RSiena): Stochastic actor-based models tested co-evolution of friendship networks and each attribute (monitoring accuracy; monitoring-based restudy; decision accuracy; task scores) across T1–T2. Network structure controls: reciprocity, GWESP (triadic closure), squared indegree popularity, squared outdegree activity; gender effects (ego, alter, same). Selection effects: attribute ego/alter and similarity (simX). Influence: average similarity (avSim). Attribute dynamics included linear and quadratic shape terms and controls for indegree/outdegree on attribute change. Convergence and GOF assessed per RSiena guidelines. Missing data handling:
• Attributes: Multiple imputation via Predictive Mean Matching; predictors included gender, class, and task/MC variables at both waves. Five imputations averaged.
• Network ties: MoM hybrid imputation; at T1 missing ties imputed as 0; at T2 last value carried forward; missing ties excluded from target statistics. Categorization for RSiena: Continuous attributes (monitoring accuracy, monitoring-based restudy, task scores) binned into low/medium/high using T2-based cutoffs applied to both waves; decision accuracy dichotomized (low/high) due to ceiling. Attributes treated as ordinal in models. Preregistration and ethics: Study preregistered (OSF: https://osf.io/kqvgu/). Approved by University of Bern IRB. Data and scripts available on OSF.

Descriptives and stability:

• Average number of nominated friends: T1 M=4.38; T2 M=3.91; 74% of ties reciprocated at T1 and 77% at T2; Jaccard index=0.63 (sufficient stability). Moran’s I near zero to slightly negative for all attributes at both waves (little baseline similarity among friends).
• Monitoring accuracy high at both waves: initial performance predicted confidence at T1 (Std. Est.=0.70, SE=0.02, t(2571)=33.66, p<.001) and T2 (Std. Est.=0.71, SE=0.025, t(2557)=28.24, p<.001); no T1–T2 difference (Z=−0.25, p=.81).
• Monitoring-based restudy evident at both waves: higher confidence predicted less restudy at T1 (Std. Est.=−0.44, SE=0.03, t(2571)=−13.55, p<.001) and T2 (Std. Est.=−0.46, SE=0.04, t(2540)=−12.56, p<.001); no T1–T2 difference (Z=0.38, p=.70).
• Stability (T1–T2 Pearson r): monitoring accuracy r=.189 (p=.048), monitoring-based restudy r=.426 (p<.001), decision accuracy r=.349 (p<.001), task scores r=.293 (p=.004). Path models (predicting task scores):
• T1: Decision accuracy strongly predicted task scores (unstd=20.537, SE=2.188; std=0.68, p<.01). Monitoring accuracy and monitoring-based restudy were non-significant.
• T2: Decision accuracy remained a strong predictor (unstd=23.429, SE=3.03; std=0.53, p<.01). Monitoring-based restudy also predicted task scores (unstd=−27.518, SE=5.954; std≈0.31, p<.01), indicating that more strongly restudying low-confidence items related to higher scores. Monitoring accuracy did not predict scores. Model R²≈.48 (T1) and .47 (T2). Social network models:
• Influence: Friends did not significantly influence monitoring accuracy, monitoring-based restudy, or decision accuracy. Friends did significantly influence task scores: the avSim parameter was positive and significant (Est.=2.119, SE=0.757), indicating increasing similarity in performance among friends over time.
• Selection: No significant similarity-based friendship selection on metacognitive attributes or task scores. Two ancillary effects: higher monitoring-based restudy (ego) and lower task performance (ego) associated with sending more nominations. Overall: Metacognitive control—especially decision accuracy—robustly predicted performance; with task experience, strategic, monitoring-informed restudy additionally predicted performance. Peer influence emerged for task outcomes but not for the measured on-task metacognitive processes.

Findings address two central questions. First, how do on-task metacognitive processes relate to adolescent learning outcomes? Decision accuracy—choosing to maintain correct answers and withdraw incorrect ones—was the strongest predictor of task scores at both waves, underscoring the importance of metacognitive control actions. With task experience (T2), monitoring-based restudy also predicted performance, suggesting adolescents increasingly translated monitoring into proactive, goal-directed control by prioritizing low-confidence items. Monitoring accuracy alone did not predict scores in path models, implying that control mediates the link between monitoring and performance; it is not merely knowing how one is doing, but using that knowledge to act adaptively that drives outcomes. Second, do friendships shape metacognition and learning? Longitudinal social network analyses indicated no friendship influence on monitoring or control indices, nor selection based on these attributes. However, friends became more similar in task scores across waves, extending prior peer influence findings from broad achievement indices to an objectively measured, task-specific outcome. The divergence—peer influence on performance but not on on-task metacognition—suggests that other socially shaped processes (e.g., motivation, effort allocation, norms) may underpin convergence in scores. With experience, adolescents appeared to balance proactive (monitoring-based restudy) and reactive (decision accuracy) control, consistent with frameworks positing developmental shifts in control strategies. The study highlights that individual metacognitive control processes primarily drive immediate performance, while social contexts shape overall performance levels via mechanisms beyond the specific metacognitive acts measured during task execution.

This study integrates individual metacognitive process measures with classroom friendship networks to illuminate adolescent learning. It shows that metacognitive control—especially accurate submit/withdraw decisions—strongly predicts task performance, and that with task experience, strategic monitoring-based restudy further enhances outcomes. Friendships did not shape on-task metacognitive processes over 3 months, but friends did become more similar in objectively measured task scores, indicating peer influence on learning outcomes. Implications: Interventions may be more effective when emphasizing the translation of monitoring into adaptive control (e.g., targeted restudy, judicious response submission) rather than improving monitoring accuracy alone. Educational strategies could also leverage positive peer dynamics to support performance. Future directions: Use larger samples with more waves to increase power for detecting social effects on metacognition; examine higher-order, collaborative tasks where metacognitive processes are more observable and socially shared; assess full self-regulated learning cycles (forethought, performance, reflection) and motivational/affective factors; and investigate peer-learning networks and concrete interaction mechanisms (e.g., encouragement, imitation, group norms) that may drive convergence in performance.

• Statistical power: Only two measurement waves and moderate sample size may limit power to detect selection/influence effects in RSiena, despite adequate convergence and fit.
• Measurement constraints: High monitoring accuracy (likely due to recall-based task with delayed tests) reduced variability; decision accuracy and performance showed ceiling tendencies, possibly overestimating metacognitive abilities in more complex contexts.
• Task characteristics: The Kanji memorization task reflects lower-level cognitive processing and may limit visibility of socially influenced metacognitive behaviors; it did not count towards grades and may not represent general academic ability.
• Scope of social data: Focus on classroom friendships excludes broader social influences (e.g., family, out-of-class peers) and may not capture specific peer-learning dynamics most relevant to metacognition.
• Mechanisms unmeasured: The study did not directly measure motivational factors, self-regulated learning phases beyond performance, or interaction processes that could explain peer influence on performance.
• Generalizability: Short-term (3 months), specific task context, and Swiss classrooms may limit generalization to other tasks, contexts, or longer developmental periods.