Feedback Valence, Prediction Error, and the Limits of Computational Attribution in Symptom-Related EEG

Abstract

Feedback-locked electroencephalography (EEG) responses are often interpreted using model-derived prediction errors (PEs), although signed PE can be strongly coupled to feedback valence in binary-outcome tasks. We revisited a published anxiety-related EEG association using two open probabilistic-learning datasets to examine measurement reliability, predictor separability, and participant-level recovery. In the primary dataset, greater trait anxiety was associated with a larger negative-minus-positive FCz theta response among 46 participants with elevated depressive symptoms after adjustment for depression severity. The association also differed between the sampled depression-symptom groups. The feedback contrast showed useful within-session reliability, but matched delta and P3-minus-N2 analyses did not establish theta specificity. Within participants, feedback valence and signed PE were highly correlated (r = -0.964). Exact-design simulations showed that population-level model discrimination could succeed despite poor recovery of individual neural slopes. Participant-specific learning-model refits recovered overall signed PE more consistently than the within-valence variation needed to distinguish it from feedback valence. An independent dataset reproduced the FCz feedback contrast in untreated and placebo cohorts but lacked trait-anxiety scores. Adding signed PE to valence yielded a small held-out predictive improvement under one reinforcement-learning specification, with limited support across sessions and no comparable support under the dual-rate model or in placebo. Together, these findings support a symptom-related feedback EEG association while constraining its theta-specific and participant-level computational interpretation.