The study team revisited a previously reported anxiety-related EEG finding using two open probabilistic-learning datasets. In the primary cohort — described as 46 participants with elevated depressive symptoms — investigators reported that greater trait anxiety was associated with a larger negative-minus-positive FCz theta feedback response after adjustment for depression severity. The cohort showed heterogeneity: the trait-anxiety relationship differed between the sampled depression-symptom groups, suggesting the association is contingent on symptom sampling rather than uniform across all participants.
This analysis was run on feedback-locked EEG responses. The authors emphasized that these scalp measures are often interpreted through the lens of computational variables such as prediction error, but the data here forced a reappraisal of how well those computational attributions hold when predictors are tightly coupled.
Within-subject predictor correlations emerged as a central barrier. Data indicated that feedback valence and signed prediction error were highly correlated inside individual participants (reported r = −0.964). That degree of collinearity implies that, for these task designs, attempts to assign variance in the EEG signal uniquely to signed PE will be severely constrained.
The authors tested this separability empirically. Participant-specific learning-model refits recovered overall signed PE more reliably than they recovered the smaller within-valence variation required to distinguish PE from raw valence signals. In other words, model refitting can capture a global signed-PE regressor but struggles to isolate the nuance that would make PE distinct from valence on a per-subject basis.
Investigators evaluated reliability and frequency specificity of the feedback contrast. The negative-minus-positive FCz feedback contrast showed useful within-session reliability, meaning the contrast produced consistent signal patterns within the same recording session. Parallel analyses that matched delta-band activity and a P3-minus-N2 ERP contrast did not confirm theta specificity. The matched delta and P3-minus-N2 analyses failed to establish that the effect is exclusive to the theta band, so the theta label should be treated with caution when interpreting the signal’s mechanistic origin.
The study team ran exact-design simulations to probe recovery properties. Simulated outcomes showed a separation between population-level and participant-level inference: population-level model discrimination could succeed even when individual neural slopes were poorly recovered. This asymmetry matters clinically. Group-level inferences about whether a model explains variance can be valid, while attributing a specific computational slope to an individual’s EEG response may not be supported by the data or design.
These simulations further revealed that the commonly used fitting procedures and the experimental design limited participant-level identifiability of the computational predictors. The authors therefore urged restraint in claiming single-subject computational mappings from these types of feedback-task EEG datasets.
An independent dataset reproduced the FCz feedback contrast. The replication held in untreated and placebo cohorts, providing evidence that the feedback contrast replicates across samples and pharmacologic contexts. That replication dataset, however, lacked trait-anxiety scores, so the original symptom-linked association could not be tested there. The absence of anxiety measures in the replication constrains conclusions about generalizability of the symptom-related effect beyond the primary cohort.
Model dependence emerged as a further constraint. When investigators added signed PE to feedback valence in predictive models, they observed a small held-out improvement under one reinforcement-learning specification. That improvement was limited: it showed little consistency across sessions and did not replicate under a dual-rate learning model or in the placebo arm. The data therefore imply that modest predictive gains from including computational regressors are model-dependent and fragile across analytic choices.
The authors explicitly noted several limits from their reported analyses. The preprint status means these results have not been certified by peer review. Participant-level collinearity between valence and signed PE (r = −0.964) restricts the ability to make definitive computational attributions in individual EEG recordings. The independent replication reproduced the feedback contrast but lacked trait-anxiety scores, so symptom-level generalization remains untested in that sample. Simulations indicated group-level discrimination can coexist with poor individual recovery, leaving open the question of when, if ever, single-subject computational mapping from feedback EEG is defensible under similar task designs.
These are the concrete constraints documented by the study team; they frame what the data support (a symptom-linked feedback EEG contrast at FCz) and what remains uncertain (theta-band specificity, participant-level computational recovery, and robustness across reinforcement-learning specifications).