CoChat

Predicting emotional valence in autism: a preregistered study in the Bayesian Brain framework

Open access
Plank, Irene Sophia, Pior, Alexandra, Yurova, Anna, Nowak, Julia, Shi, Zhuanghua, Falter-Wagner, Christine M. · 2026 · Molecular Autism

The DOI matches its Crossref record; 47 of the 47 references it cites were found in a registry or an index; 35 of the 46 sources its claims cite were checked for support.

Abstract

BACKGROUND: Autism spectrum disorder (ASD) affects social interaction, communication and behavioural flexibility. Recent theories grounded in the Bayesian Brain Framework propose that autistic individuals process environmental uncertainty differently, overweighting volatility when updating predictions about the world. However, the robustness of these predictive processing differences and their relevance to core symptoms remains unclear. METHODS: Lawson and colleagues [1] reported a tendency to overestimate environmental volatility in autistic adults based on belief states extracted using a Hierarchical Gaussian Filter. We extended their paradigm to a domain central to the autistic experience: rather than distinguishing houses from faces, participants detected valence in emotional facial expressions. Preregistered hypotheses were evaluated using Bayesian linear mixed models based on data of 22 autistic and 22 non-autistic participants. RESULTS: Despite using an identical computational model to extract belief states, we did not find robust differences between autistic and non-autistic adults, though autistic participants showed a non-credible trend towards increased processing of phasic volatility. Largely comparable probabilistic associative learning was also reflected in response times and pupil sizes. LIMITATIONS: While participants reacted faster to expected than unexpected trials in the beginning, this effect decreased over the course of the experiment; thus, possibly indicating that the individual blocks of stable cue-outcome associations were too short. CONCLUSIONS: We were unable to find credible differences in environmental volatility and learning rate updates using a task designed to probe a key autistic difficulty, facial emotion recognition. The current results call into question the generalisability and clinical relevance of prior findings.

Overview

CoChat Summary

## TL;DR

This preregistered study tested whether previously reported Bayesian Brain differences in autism would replicate in a socio-emotional emotion-recognition task with 22 autistic and 22 non-autistic adults. The authors found no credible group differences in tonic volatility, reaction times, or pupil sizes, with only suggestive but inconclusive evidence for higher phasic volatility in autistic adults.

## Key findings

  • In the preregistered HGF analyses, there was no credible difference between autistic and comparison adults in phasic volatility: estimate = -0.1 [-0.22, 0.01], posterior probability = 92.54%.
  • There was no evidence for higher environmental tonic volatility in autistic adults: estimate = 0.03 [-0.24, 0.31], posterior probability = 41.61%.
  • For reaction times, there was no credible evidence for group differences in Expectancy or transition effects; however, across groups, expected trials were faster than unexpected trials (estimate = -0.07 [-0.1, -0.05]) and easy trials were faster than hard trials (estimate = -0.07 [-0.09, -0.04]), both with posterior probability = 100%.
  • Accuracy was high overall (90.8%); exploratory analysis found lower accuracy in the autistic group (88.7 ± 14.2%) than the comparison group (92.9 ± 11.1%), and in unexpected versus expected trials (89.3 ± 15.7% vs 92.3 ± 9.2%).

## Methods

  • Design: Preregistered Bayesian study extending Lawson et al.’s probabilistic associative learning paradigm to facial emotion recognition; participants judged whether emotional faces were positive (happy) or negative (fearful) while cues varied in volatility and expectancy.
  • Sample/data: Final analyzed sample was 22 autistic and 22 comparison participants; pupil-size analyses used 18 per group due to missing/technical issues. The task used 336 trials, with analysis of reaction times, accuracies, and pupil sizes.
  • Main analysis: Reaction times were modeled with a three-level Hierarchical Gaussian Filter (oHGF/eHGF) and compared with a Rescorla-Wagner model; hypotheses were tested using Bayesian linear/mixed models in brms, with posterior predictive checks, simulation-based calibration, and ROPE/HDI-based inference.

## Limitations

  • The paper states that the block structure may have been too short, because the expected-vs-unexpected effect diminished over the course of the experiment; this may mean participants did not form strong enough stable cue-outcome associations.
  • The authors note that although the sample reached the preregistered target size, the analysis differed from the power analysis, so the group sizes may have been insufficient to detect subtle differences.
  • They also state that the valence decision was very simple, producing a ceiling effect in accuracy that limits interpretation. Evident from the text: pupil data were only available for 18 per group and the HGF model showed suboptimal fit for some effects, so conclusions about pupillometry and model-specific parameters should be cautious.

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

Integrity signals

Registry

Crossref · 10.1186/s13229-026-00730-3

Cites

47 of 47 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Example stimuli based on the KDEF identity F05. For each identity, six versions were created, each three difficulty levels of the positive (happy) and negative (fearful) emotion
Fig. 1 Example stimuli based on the KDEF identity F05. For each identity, six versions were created, each three difficulty levels of the positive (happy) and negative (fearful) emotionCC BY · PubMed Central
Fig. 2 Trial order for the experiment. All participants first encountered a stable, prevolatile phase in which 83.3% of the high tones were followed by an emotionally positive facial expression and thus expected, while 16.7% were unexpectedly followed by an emotionally negative facial expression. In the middle phase, expected associations changed multiple times (volatile phase) and short, neutral blocks had equal associations between the emotions and tones. In the postvolatile phase, high tones were associated with emotionally negative expressions with a stable association for the duration of the phase; thus, this phase was complementary to the prevolatile phase with switched associations
Fig. 2 Trial order for the experiment. All participants first encountered a stable, prevolatile phase in which 83.3% of the high tones were followed by an emotionally positive facial expression and thus expected, while 16.7% were unexpectedly followed by an emotionally negative facial expression. In the middle phase, expected associations changed multiple times (volatile phase) and short, neutral blocks had equal associations between the emotions and tones. In the postvolatile phase, high tones were associated with emotionally negative expressions with a stable association for the duration of the phase; thus, this phase was complementary to the prevolatile phase with switched associationsCC BY · PubMed Central
Fig. 3 Conceptual visualisation of the perceptual and the response model forming the Hierarchical Gaussian Filter (HGF) and their link to the processing and the state of the world. Predictive processing relies on belief states that estimate the true hidden states of the world based on the sensory input. These belief states govern behavioural responses which in turn influence the world. The perceptual model of the HGF formalises how belief states are formed on the three levels of uncertainty about the generation of the stimuli. The response model links this perception to the behavioural responses. The perceptual and response model are fitted in an iterative process to determine the most probable parameter values given sensory inputs and observed responses
Fig. 3 Conceptual visualisation of the perceptual and the response model forming the Hierarchical Gaussian Filter (HGF) and their link to the processing and the state of the world. Predictive processing relies on belief states that estimate the true hidden states of the world based on the sensory input. These belief states govern behavioural responses which in turn influence the world. The perceptual model of the HGF formalises how belief states are formed on the three levels of uncertainty about the generation of the stimuli. The response model links this perception to the behavioural responses. The perceptual and response model are fitted in an iterative process to determine the most probable parameter values given sensory inputs and observed responsesCC BY · PubMed Central
Fig. 4 Difference in reaction times between expected and unexpected trials across the phases. While during the prevolatile and the volatile phase most participants reacted faster on expected trials, this effect seems to be lost by the postvolatile phase
Fig. 4 Difference in reaction times between expected and unexpected trials across the phases. While during the prevolatile and the volatile phase most participants reacted faster on expected trials, this effect seems to be lost by the postvolatile phaseCC BY · PubMed Central
Fig. 5 Posterior predictive checks comparing the reaction times of the participants to those simulated based on the oHGF. While the expectancy effect observed in the participant data is largely reproduced in the simulated reaction times, the difficulty effect capturing the difference in responses to hard and easy stimuli is not reproduced by the model. Error bars show standard deviations
Fig. 5 Posterior predictive checks comparing the reaction times of the participants to those simulated based on the oHGF. While the expectancy effect observed in the participant data is largely reproduced in the simulated reaction times, the difficulty effect capturing the difference in responses to hard and easy stimuli is not reproduced by the model. Error bars show standard deviationsCC BY · PubMed Central
Fig. 6 Participant-specific HGF parameters extracted from the perceptual (2nd and 3rd level tonic volatility) and the response model. The dots show individual participant parameters, while the boxplot and distribution graphs summarise all participants in a group. Boxplot hinges show the first and third quartiles with the bold line denoting the median. Whiskers extend ± 1.5 times the interquartile range. While tonic volatility parameters show little difference between groups, phasic volatility estimates are higher in the autistic than the comparison group
Fig. 6 Participant-specific HGF parameters extracted from the perceptual (2nd and 3rd level tonic volatility) and the response model. The dots show individual participant parameters, while the boxplot and distribution graphs summarise all participants in a group. Boxplot hinges show the first and third quartiles with the bold line denoting the median. Whiskers extend ± 1.5 times the interquartile range. While tonic volatility parameters show little difference between groups, phasic volatility estimates are higher in the autistic than the comparison groupCC BY · PubMed Central
Fig. 7 Learning rate updates on the second and the third level, capturing the absolute change from prevolatile to volatile and volatile to postvolatile phase. Learning rate updates were higher for the switch from prevolatile to volatile phase than for the switch from volatile to postvolatile as well as for learning about cue-outcome associations, captured by the learning rate on the second level, than learning about the environment, captured by the learning rate on the third level. Black, horizontal lines show the learning rate updates based on the Bayes optimal perceptual parameters. Participants in both groups showed stronger updates from the prevolatile to the volatile phase on the environmental level, while only autistic participants additionally showed increased updates on the environmental level moving from the volatile to the postvolatile phase
Fig. 7 Learning rate updates on the second and the third level, capturing the absolute change from prevolatile to volatile and volatile to postvolatile phase. Learning rate updates were higher for the switch from prevolatile to volatile phase than for the switch from volatile to postvolatile as well as for learning about cue-outcome associations, captured by the learning rate on the second level, than learning about the environment, captured by the learning rate on the third level. Black, horizontal lines show the learning rate updates based on the Bayes optimal perceptual parameters. Participants in both groups showed stronger updates from the prevolatile to the volatile phase on the environmental level, while only autistic participants additionally showed increased updates on the environmental level moving from the volatile to the postvolatile phaseCC BY · PubMed Central

Verification

Claims the paper makes

16 Valid · 11 Not valid · 4 Not checked · 74 Own results and interpretations

42 more claims

Create a free account to see all 105 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 40
  • Probably not a substantial claim. 30
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Citations

Cites

Show all 47 references

Research works printed but not linked: 1

  1. Hautzinger M, Bailer M, Worall H, Keller F. Beck-depressions-inventar (BDI). Bern: Huber; 1994.

Cited by

See all 2

Metadata

Authors
Plank, I. S., Pior, A., Yurova, A., Nowak, J., Shi, Z., & Falter-Wagner, C. M.
Affiliations
  • LMU Klinikum (DE)
  • University of Groningen (NL)
  • Ludwig-Maximilians-Universität München (DE)
Corresponding author
Plank, Irene Sophia
Venue
Molecular Autism
Published
2026-07-21
Volume
17
Issue
1
ISSN
2040-2392 (ISSN-L)
OpenAlex
W7169849249
Cited by
2
References
44
Citations by year
2026: 1
FWCI
4.11
Type
journal article
Language
en
Open access
Licence
Locations
  • Molecular Autism (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • PubMed Central (publishedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
  • PsyArXiv (OSF Preprints) (submittedVersion, cc-by)
Funders
Klinikum der Universität München
Topics
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Autism Spectrum Disorder Research
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Face Recognition and Perception
  • Social Sciences › Psychology › Social Psychology › Action Observation and Synchronization
Keywords
autism spectrum disorder, predictive processing, emotional valence, facial emotion recognition, environmental volatility, belief states, hierarchical Gaussian filter, pupil size, learning rate
MeSH terms
Autism Spectrum Disorder, Adult, Autistic Disorder, Bayes Theorem, Brain, Emotions, Facial Expression, Female, Humans, Male, Reaction Time, 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.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-11, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-11 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

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: Predicting emotional valence in autism: a preregistered study in the Bayesian Brain framework. https://app.cochat.ai/papers/10.1186/s13229-026-00730-3
The paper
Plank, I. S., Pior, A., Yurova, A., Nowak, J., Shi, Z., & Falter-Wagner, C. M. (2026). Predicting emotional valence in autism: a preregistered study in the Bayesian Brain framework. Molecular Autism. https://doi.org/10.1186/s13229-026-00730-3

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.