Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 61 of the 61 references it cites were found in a registry or an index; none of the 70 sources its claims cite is checked for support yet.

Distorted learning from local metacognition supports transdiagnostic underconfidence

Open access

Published version of a preprint

Abstract

Individuals experiencing symptoms of anxiety and depression have been shown to exhibit persistent underconfidence. The origin of such metacognitive biases presents a puzzle, given that individuals should be able to learn appropriate levels of confidence from observing their own performance. In two large general population samples ( N = 230 and N = 278), we measure both 'local' confidence in individual task instances and 'global' confidence as longer-run self-performance estimates while manipulating external feedback. Global confidence is sensitive to both local confidence and feedback valence—more frequent positive (negative) feedback increases (respectively decreases) global confidence, with asymmetries in feedback also leading to shifts in affective self-beliefs. Notably, however, global confidence exhibits reduced sensitivity to instances of higher local confidence in individuals with greater subclinical anxious-depression symptomatology, despite sensitivity to feedback valence remaining intact. Our finding of blunted sensitivity to increases in local confidence offers a mechanistic basis for how persistent underconfidence is maintained in the face of intact performance.

Overview

CoChat Summary

TL;DR

In two large general-population samples, asymmetric performance feedback shifted global self-performance estimates (SPEs) without changing objective performance. Individuals with greater subclinical anxious-depression symptoms showed blunted sensitivity to their own higher local confidence when forming global SPEs, while sensitivity to feedback valence remained intact.

Key findings

  • Feedback manipulation strongly shifted global SPEs: positive feedback increased and negative feedback decreased global SPEs relative to baseline, with no significant effect on accuracy (Exp 1: χ² = 273.77, p < 0.001, β = −0.26, Cohen’s d = 1.77; Exp 2: χ² = 332.08, p < 0.001, β = −0.28, Cohen’s d = 1.77; accuracy Exp 1 p = 0.856, BF01 = 9.67; Exp 2 p = 0.642, BF01 = 7.22).
  • Average local confidence predicted global SPEs across blocks (Exp 1: χ² = 48.08, p < 0.001, β = 0.46; Exp 2: χ² = 51.27, p < 0.001, β = 0.46), whereas block-wise performance fluctuations did not after accounting for local confidence.
  • Higher anxious-depression symptoms were associated with lower local confidence and lower global SPEs (Exp 1 GAD-7: β = −0.03, p = 0.002; Exp 2 Anxious-Depression axis: β = −0.35, p < 0.001; global SPEs Exp 1: β = −0.05, p < 0.001; Exp 2: β = −0.29, p < 0.001).
  • The slope relating higher local confidence to global SPEs was blunted in individuals with higher anxious-depression scores (Exp 1 GAD: χ² = 13.66, p < 0.001, β = −0.017; Exp 1 PHQ: χ² = 4.75, p = 0.029, β = −0.009; Exp 2: χ² = 5.39, p = 0.020, β = −0.06), while feedback-valence sensitivity did not significantly interact with symptom scores. Model comparison favored a confidence-distortion-only model (D1), with negative confidence-distortion parameter βc and feedback-distortion parameter βf not differing from 0.
  • In Exp 2, feedback also affected affective self-evaluation: a significant Feedback type × Word valence interaction on self-endorsement change (χ² = 6.08, p = 0.014, β = 0.11), with greater endorsement of negative than positive words after negative feedback (t(593) = 4.21, p < 0.001) but no such difference after positive feedback (t(593) = 0.65, p = 0.516).

Methods

  • Two non-overlapping online general-population samples from Prolific: Exp 1 exploration (N = 230) and Exp 2 preregistered validation (N = 278). Participants completed gamified perception and memory tasks with trial-wise local confidence, block-wise global SPEs, and manipulated intermittent positive or negative feedback in intervention blocks interleaved with no-feedback test blocks; Exp 2 also included a self-referential encoding task before and after feedback.
  • Analyses used linear mixed regression models in R and computational modeling extending a Beta-distribution model of global SPE formation with asymmetry parameters for learning from positive vs. negative feedback (ΔLRf) and high vs. low local confidence (ΔLRc). Models D0–D6 were compared using Deviation Information Criteria, with parameters fit via MCMC sampling in JAGS.
  • Symptom measures: Exp 1 used PHQ-9, GAD-7, and mini-SPIN; Exp 2 used 71 questions yielding transdiagnostic Anxious-Depression, Compulsivity and Intrusive Thought, and Social Withdrawal axes.

Limitations

  • The paper states that intervention trials available to estimate feedback-learning distortions were substantially fewer than trials for confidence-learning distortions, though the authors argue the intact feedback-learning result is robust based on strong feedback effects, pooled feedback trials (4600 in Exp 1 and 5560 in Exp 2), Bayes factors, and parameter-recovery simulations.
  • The paper states that the sample was a general-population online sample; although many participants surpassed clinical cutoffs, further work is needed to test generalization to individuals with clinical anxiety and/or depression diagnoses.
  • The paper reports preregistration deviations: the response-bias model D3 was not preregistered, intercept terms for Models D1 and D2 were initially omitted, and after including intercepts the preregistered βf < 0 effect in Exp 1 was no longer observed, although βc < 0 remained. The text also notes DIC can favor complex models and that the very small fitted response-bias βa values were likely spurious.
  • Evident from the text: Exp 2 is reported as N = 278 in the Abstract and Methods, but Fig. 3d and Fig. 4 label Exp 2 as N = 287.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s41467-025-57040-0

Cites

61 of 61 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1 A schematic depiction of the hypotheses. a Outline of a canonical task in which individual trials are followed by local confidence reports and occasional (veridical) performance feedback. After completing the block, participants provide a global self-performance estimate (SPE) for that block. b, c Schematic illustration of different influences on global confidence within computational accounts of confidence formation. b Global SPEs are expected to increase proportionally with local confidence. c More frequent feedback on correct (respectively incorrect) trials is expected to increase (decrease) global SPEs. d Three candidate computational mechanisms for how the formation of global SPEs differs as a function of anxious-depression symptoms. e–g Visualisation of how the base patterns of global confidence formation depicted in (b) and (c) are expected to differ as a function of the different mechanisms described in (d). e Prediction for model D1, distortion in confidence learning – higher anxious-depression symptoms (blue lines) are associated with a blunted influence of higher local confidence on global SPEs compared to lower anxious-depression symptoms (gold lines). f Prediction for model D2, distortion in feedback learning—compared to the absence of feedback, higher anxious-depression symptoms are associated with a greater sensitivity to incorrect vs. correct feedback, and vice-versa for lower anxious-depression symptoms. g Prediction for D3, negative response bias – under a simple response bias model, we expect an overall reduction in global SPEs when comparing high vs. low anxious-depression symptoms in the absence of the nonlinearities depicted in (e) and (f).
Fig. 1 A schematic depiction of the hypotheses. a Outline of a canonical task in which individual trials are followed by local confidence reports and occasional (veridical) performance feedback. After completing the block, participants provide a global self-performance estimate (SPE) for that block. b, c Schematic illustration of different influences on global confidence within computational accounts of confidence formation. b Global SPEs are expected to increase proportionally with local confidence. c More frequent feedback on correct (respectively incorrect) trials is expected to increase (decrease) global SPEs. d Three candidate computational mechanisms for how the formation of global SPEs differs as a function of anxious-depression symptoms. e–g Visualisation of how the base patterns of global confidence formation depicted in (b) and (c) are expected to differ as a function of the different mechanisms described in (d). e Prediction for model D1, distortion in confidence learning – higher anxious-depression symptoms (blue lines) are associated with a blunted influence of higher local confidence on global SPEs compared to lower anxious-depression symptoms (gold lines). f Prediction for model D2, distortion in feedback learning—compared to the absence of feedback, higher anxious-depression symptoms are associated with a greater sensitivity to incorrect vs. correct feedback, and vice-versa for lower anxious-depression symptoms. g Prediction for D3, negative response bias – under a simple response bias model, we expect an overall reduction in global SPEs when comparing high vs. low anxious-depression symptoms in the absence of the nonlinearities depicted in (e) and (f).CC BY · PubMed Central
Fig. 2 Study design and stimuli. a The eight groups of participants and the order in which they performed the perception (P) and memory (M) tasks in Exp 1. Participants within a group performed one or both tasks. A task sequence began with two baseline blocks followed by two sets of intervention blocks (with feedback), interleaved with test blocks (without feedback). For groups that performed both tasks, their order during baseline was randomised across participants. Unfilled boxes indicate the absence of feedback. Green-filled boxes indicate positive feedback blocks (more feedback on correct than incorrect trials) and red blocks indicate negative feedback blocks (more feedback on incorrect than correct trials). The colour of the subsequent test block indicates whether feedback delivered on the preceding (intervention) block was positive (green) or negative (red). b Single trials of the perception (top) and memory (bottom) tasks depicted using screenshots of stimuli shown in the experiments. For the perception task, participants attended the green bush in the middle where 121 stimuli of two different kinds of berries (red raspberries and dark purple blackberries) appeared at non-overlapping random locations for 1000 ms. Participants were asked to respond (without time limit) as to whether there were more raspberries or blackberries, and then use a slider to report their confidence. After reporting their local confidence, participants pressed a key to proceed. At the end of each block of trials, participants were asked to report their global self-performance estimate on that block. Feedback for correct or incorrect trials appeared with a defined frequency depending on the intervention block (positive or negative). For the memory task, a set of fruits appeared on the bush for 1500 ms. Participants were tasked to memorise the fruits and then decide which of two fruits was present in the set. Confidence reporting and feedback for the memory task were similar to the perception task.
Fig. 2 Study design and stimuli. a The eight groups of participants and the order in which they performed the perception (P) and memory (M) tasks in Exp 1. Participants within a group performed one or both tasks. A task sequence began with two baseline blocks followed by two sets of intervention blocks (with feedback), interleaved with test blocks (without feedback). For groups that performed both tasks, their order during baseline was randomised across participants. Unfilled boxes indicate the absence of feedback. Green-filled boxes indicate positive feedback blocks (more feedback on correct than incorrect trials) and red blocks indicate negative feedback blocks (more feedback on incorrect than correct trials). The colour of the subsequent test block indicates whether feedback delivered on the preceding (intervention) block was positive (green) or negative (red). b Single trials of the perception (top) and memory (bottom) tasks depicted using screenshots of stimuli shown in the experiments. For the perception task, participants attended the green bush in the middle where 121 stimuli of two different kinds of berries (red raspberries and dark purple blackberries) appeared at non-overlapping random locations for 1000 ms. Participants were asked to respond (without time limit) as to whether there were more raspberries or blackberries, and then use a slider to report their confidence. After reporting their local confidence, participants pressed a key to proceed. At the end of each block of trials, participants were asked to report their global self-performance estimate on that block. Feedback for correct or incorrect trials appeared with a defined frequency depending on the intervention block (positive or negative). For the memory task, a set of fruits appeared on the bush for 1500 ms. Participants were tasked to memorise the fruits and then decide which of two fruits was present in the set. Confidence reporting and feedback for the memory task were similar to the perception task.CC BY · PubMed Central
Fig. 3 Manipulation of global confidence by feedback and transdiagnostic symptom axes. a Global confidence measured as self-performance estimates (global SPEs), (b) performance (mean accuracy), and (c) mean local confidence on intervention blocks, with more frequent positive (green) or negative feedback (red) plotted separately for the two tasks in Exp 1. For global SPEs and local confidence, baseline values were subtracted out for each individual. Each dot is an individual participant (N = 230). Error bars are centred on means and show bootstrapped 95% confidence intervals. d Regression coefficients depicting the relationship in Exp 2 (N = 287) between transdiagnostic symptom axes and mean local confidence, global SPEs measured retrospectively, global SPEs measured prospectively, and accuracy. Error bars show bootstrapped 95% confidence intervals. In all panels, statistical comparisons were performed using mixed regressions, and p-values were obtained through likelihood ratio tests with χ² distributions. For (a)−(c), p-values were adjusted for multiple comparisons using FDR correction. ****p <0.001, **p = 0.020, *p = 0.046.
Fig. 3 Manipulation of global confidence by feedback and transdiagnostic symptom axes. a Global confidence measured as self-performance estimates (global SPEs), (b) performance (mean accuracy), and (c) mean local confidence on intervention blocks, with more frequent positive (green) or negative feedback (red) plotted separately for the two tasks in Exp 1. For global SPEs and local confidence, baseline values were subtracted out for each individual. Each dot is an individual participant (N = 230). Error bars are centred on means and show bootstrapped 95% confidence intervals. d Regression coefficients depicting the relationship in Exp 2 (N = 287) between transdiagnostic symptom axes and mean local confidence, global SPEs measured retrospectively, global SPEs measured prospectively, and accuracy. Error bars show bootstrapped 95% confidence intervals. In all panels, statistical comparisons were performed using mixed regressions, and p-values were obtained through likelihood ratio tests with χ² distributions. For (a)−(c), p-values were adjusted for multiple comparisons using FDR correction. ****p <0.001, **p = 0.020, *p = 0.046.CC BY · PubMed Central
Fig. 4 Data and model fits show distorted learning from local metacognition. Results from Exp 1 (left column (a, c, e); N = 230) and Exp 2 (right column (b, d, f); N = 287). Mean global self-performance estimates (SPEs) and 95% confidence intervals depicted in (a–d) are obtained from linear regression models after regressing out random effects of participant, block number, and trial number from observed data (filled circles) and model fits (unfilled triangles). a, b Global SPEs plotted against local confidence (z-scored and split into 4 quantiles per individual) separately for individuals belonging to the lowest (gold) and highest (blue) tertiles of anxious-depression symptom scores (obtained using the GAD-7 questionnaire in Exp 1 and transdiagnostically in Exp 2). Depicted statistical comparisons are for slopes of higher confidence values between higher and lower anxious-depression scores. c, d Global SPEs plotted for blocks with more frequent negative feedback, no feedback, and more frequent positive feedback, separately for individuals with low and high anxious-depression scores. e, f Deviation Information Criteria (DIC) scores for the seven possible models of how different distortions in global SPE formation relate to AD symptom scores: D0—No distortion, D1—Confidence distortion only, D2—Feedback distortion only, D3—Response bias only, D4—Confidence and feedback distortions, D5—Confidence distortion and response bias, D6—Feedback distortion and response bias. Posterior distributions of regression slope parameters relating anxious-depression symptoms to (g) confidence and (h) feedback distortions for model D4 in Exp 1 and Exp 2. Regression coefficient values from Exp 2 are divided by 5 (corresponding to the difference in the range of GAD and Anxious-Depression axis scores) to allow direct comparison of coefficients.
Fig. 4 Data and model fits show distorted learning from local metacognition. Results from Exp 1 (left column (a, c, e); N = 230) and Exp 2 (right column (b, d, f); N = 287). Mean global self-performance estimates (SPEs) and 95% confidence intervals depicted in (a–d) are obtained from linear regression models after regressing out random effects of participant, block number, and trial number from observed data (filled circles) and model fits (unfilled triangles). a, b Global SPEs plotted against local confidence (z-scored and split into 4 quantiles per individual) separately for individuals belonging to the lowest (gold) and highest (blue) tertiles of anxious-depression symptom scores (obtained using the GAD-7 questionnaire in Exp 1 and transdiagnostically in Exp 2). Depicted statistical comparisons are for slopes of higher confidence values between higher and lower anxious-depression scores. c, d Global SPEs plotted for blocks with more frequent negative feedback, no feedback, and more frequent positive feedback, separately for individuals with low and high anxious-depression scores. e, f Deviation Information Criteria (DIC) scores for the seven possible models of how different distortions in global SPE formation relate to AD symptom scores: D0—No distortion, D1—Confidence distortion only, D2—Feedback distortion only, D3—Response bias only, D4—Confidence and feedback distortions, D5—Confidence distortion and response bias, D6—Feedback distortion and response bias. Posterior distributions of regression slope parameters relating anxious-depression symptoms to (g) confidence and (h) feedback distortions for model D4 in Exp 1 and Exp 2. Regression coefficient values from Exp 2 are divided by 5 (corresponding to the difference in the range of GAD and Anxious-Depression axis scores) to allow direct comparison of coefficients.CC BY · PubMed Central
Fig. 5 Metacognition and its relation to affective self-evaluation. The interaction of word valence (positive or negative) with (a) Anxious-Depression transdiagnostic scores (p < 0.001), and (b) mean local confidence (p < 0.001) during baseline runs when regressed upon self-endorsements at the first timepoint (T1) of the self-referential encoding task in Exp 2 (plotted as marginal effects from a logistic regression predicting self-endorsement). c Change in self-endorsement from T1 to T2 (i.e. post-feedback) separated by word valence and the type of feedback delivered during intervention blocks. While there was no difference in self-endorsements following positive feedback, negative words were endorsed more than positive words following negative feedback. For all panels, error bars show 95% bootstrapped confidence intervals, statistical comparisons were performed using mixed regressions, and p-values were obtained through likelihood ratio tests with χ² distributions. For (c), error bars are centred at the respective means and p values were adjusted using FDR.
Fig. 5 Metacognition and its relation to affective self-evaluation. The interaction of word valence (positive or negative) with (a) Anxious-Depression transdiagnostic scores (p < 0.001), and (b) mean local confidence (p < 0.001) during baseline runs when regressed upon self-endorsements at the first timepoint (T1) of the self-referential encoding task in Exp 2 (plotted as marginal effects from a logistic regression predicting self-endorsement). c Change in self-endorsement from T1 to T2 (i.e. post-feedback) separated by word valence and the type of feedback delivered during intervention blocks. While there was no difference in self-endorsements following positive feedback, negative words were endorsed more than positive words following negative feedback. For all panels, error bars show 95% bootstrapped confidence intervals, statistical comparisons were performed using mixed regressions, and p-values were obtained through likelihood ratio tests with χ² distributions. For (c), error bars are centred at the respective means and p values were adjusted using FDR.CC BY · PubMed Central

Verification

Claims the paper makes

0 Valid · 0 Not valid · 27 Not checked · 97 Own results and interpretations

61 more claims

Create a free account to see all 124 claims and to verify the 66 references still open.

See all claims — free account
  • Probably not a substantial claim. 42
  • The quote only partly states this claim. 1
  • It may state two separate claims as one. 25

Create a free account to see these claims.

Statistics recomputed

Every reported test result, recomputed from its test statistic and degrees of freedom. A result shows as not adding up only once an independent check has confirmed it.

  • Consistentt(226) = 2.19reported p = .03 · recomputed p = .030

    “t(226) = 2.19, p = .030”

  • Consistentt(275) = 3.12reported p = .002 · recomputed p = .002

    “t(275) = 3.12, p = .002”

  • Consistentt(593) = 4.21reported p < .001 · recomputed p < .001

    “t(593) = 4.21, p < .001”

  • Consistentt(593) = 0.65reported p = .516 · recomputed p = .516

    “t(593) = 0.65, p = .516”

4 of 4

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

3 of 3

Citations

Cites

Show all 61 references

Cited by

See all 47

Metadata

Authors
Katyal, S., Huys, Q. J., Dolan, R. J., & Fleming, S. M.
Affiliations
  • University of Copenhagen (DK)
  • UCL Queen Square Institute of Neurology (GB)
Corresponding author
  • Katyal, Sucharit
  • Fleming, Stephen M.
Venue
Nature Communications
Published
2025-02-21
Volume
16
Issue
1
Pages
1854-1854
ISSN
2041-1723 (ISSN-L)
OpenAlex
W4407837907
Cited by
47
References
61
Citations by year
2026: 22 · 2025: 15
FWCI
20.49
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Communications (publishedVersion, cc-by)
  • PsyArXiv (OSF Preprints) (submittedVersion, cc-by)
  • PubMed (publishedVersion)
  • PubMed Central (publishedVersion, cc-by)
  • Research at the University of Copenhagen (University of Copenhagen) (publishedVersion, cc-by)
  • RePEc: Research Papers in Economics (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
  • MPG.PuRe (Max Planck Society) (submittedVersion)
  • PsyArXiv (OSF Preprints) (submittedVersion)
Funders
Koa Health
Topics
  • Social Sciences › Psychology › Experimental and Cognitive Psychology › Anxiety, Depression, Psychometrics, Treatment, Cognitive Processes
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Memory Processes and Influences
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Neural and Behavioral Psychology Studies
Keywords
metacognition, underconfidence, transdiagnostic symptoms, anxiety, depression, self-confidence, feedback valence, external feedback, subclinical symptoms, self-belief
MeSH terms
Metacognition, Adolescent, Adult, Anxiety, Depression, Female, Humans, Learning, Male, Middle Aged, Self Concept, Young Adult

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-11-18, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-11-18 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Distorted learning from local metacognition supports transdiagnostic underconfidence. https://app.cochat.ai/papers/10.1038/s41467-025-57040-0
The paper
Katyal, S., Huys, Q. J., Dolan, R. J., & Fleming, S. M. (2025). Distorted learning from local metacognition supports transdiagnostic underconfidence. Nature Communications. https://doi.org/10.1038/s41467-025-57040-0

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.