Lower confidence and increased error sensitivity in OCD patients while learning under volatility

The study investigates how obsessive-compulsive disorder (OCD) affects the relationship between action updating and confidence when learning in volatile environments. Prior work suggests OCD is associated with lower confidence and potentially abnormal action–confidence coupling, but findings have been mixed: some report decoupling between action and confidence in OCD, others do not. Moreover, general-population studies using compulsivity analogues sometimes show increased confidence and decoupling, contrasting with clinical OCD. The authors aim to clarify whether OCD patients (without comorbid diagnoses and off medication) differ from healthy controls (HC) in action, confidence, and their coupling during a predictive inference task, and to compare these patterns to high versus low compulsive individuals from the general population to assess specificity and translational relevance.

Confidence shapes belief updating and behavior under uncertainty; higher confidence reduces the impact of new information, while lower confidence promotes evidence gathering. OCD has been linked to lower confidence and potential abnormalities in employing confidence to guide behavior. Prior predictive inference studies reported: (a) decoupling between action and confidence in OCD driven by excessive action to small prediction errors, consistent with clinical error sensitivity; (b) in general-population samples, compulsivity has been associated with inflated confidence and decoupling, although replication is mixed. The literature also highlights error sensitivity as an OCD endophenotype. The authors note discrepancies across clinical and analogue samples and propose that different mechanisms may underlie similar compulsive symptoms across populations.

Design: Case–control comparison of OCD patients versus healthy controls (HC) using a predictive inference task, with a secondary matched comparison between high-compulsive (HComp) and low-compulsive (LComp) general-population participants from an existing dataset. Analyses included model-free measures (action, confidence, learning rate) and model-based regressions using latent variables from a quasi-optimal Bayesian observer model (absolute prediction error [PE], change-point probability [CPP], and model confidence [MC]) plus trial feedback (hit/miss). Participants: Initially 43 OCD patients (ages 18–65) recruited via a university medical center and community websites; OCD diagnosis confirmed via structured interview. Exclusions: current MDD, hypomania, anxiety, substance use, psychotic disorders, and psychiatric medication use at inclusion. 45 HCs were recruited via advertisements and matched to OCD by age, gender, and education. General-population comparison groups (HComp/LComp) were selected from Seow & Gillan (Mechanical Turk sample), matched on demographics and OCD-symptom distributions to approximate the clinical sample numbers. After pre-registered task-based exclusions, final samples: OCD n=38, HC n=37; HComp n=76, LComp n=73. Questionnaires: MINI for psychiatric screening (HC/OCD). OC symptoms via OCI-R in all; OCD severity via Y-BOCS in patients (patients with Y-BOCS < 12 excluded). Anxiety/depression via DASS (OCD/HC) and STAI/Zung (HComp/LComp). Task: Web-based predictive inference task. On each trial, participants positioned a “bucket” on a circle to catch a particle whose landing positions were sampled from a Gaussian distribution with occasional change-points (abrupt shifts of the mean). After placing the bucket (action), participants rated confidence (1–100). OCD/HC completed 300 trials; HComp/LComp completed 150 trials. Behavioral measures: Action was operationalized as absolute change in bucket position from trial t to t+1. Prediction error (PE) δ_t = x_t − b_t (landing minus bucket). Learning rate (LR) α_t = |b_t − b_{t+1}| / |δ_t|. The action update used in some analyses was LR × |PE|. Exclusions: Pre-registered task-based exclusions similar to prior work (e.g., low use of confidence scale, insufficient correlation between confidence and feedback). Trial-level exclusions removed extreme learning rates (>95th percentile within-group), trials with very small PE susceptible to measurement noise, and trials without usable action. Statistical analysis: Linear mixed-effects models (lme4) tested group effects on action and confidence (random intercepts). Action–confidence coupling was tested with trial-wise action as dependent variable, predictors: z-scored confidence, group, and their interaction; random intercepts and random slopes for confidence. Correlations assessed relationships between coupling and symptom severity (OCI-R) within groups. Model-based regressions predicted action or confidence from z-scored PE, CPP, (1−CPP)×MC, and hit/miss, with group interactions and by-subject random intercepts and slopes.

- Confidence: OCD patients showed significantly lower confidence than healthy controls (β = −18.9, SE = 4.9, t = −3.83, p < 0.001). High-compulsive individuals showed higher confidence than low-compulsive individuals (β = −7.72, SE = 3.5, t = −2.19, p = 0.030); confidence correlated positively with OCI-R in HComp (r = 0.52, p < 0.001), but not in OCD. - Action: No group differences in mean action (bucket updates) for OCD vs HC, nor for HComp vs LComp. - Action–confidence coupling: Across groups, higher confidence predicted smaller action (negative coupling). No significant difference in coupling between OCD and HC. HComp showed weaker (more decoupled) action–confidence coupling than LComp (interaction β ≈ −2.27, t ≈ −2.16, p ≈ 0.033); coupling correlated with OCI-R in HComp (r ≈ 0.50, p < 0.001) but not in OCD. - Learning rates: Mean LR did not differ between groups. All groups increased LR with larger PEs. OCD patients exhibited greater error sensitivity at low PE magnitudes than HC (i.e., higher LR for small errors). In the HComp vs LComp comparison, an interaction indicated a less steep LR–PE relationship in HComp, but post hoc bin-wise differences were not significant. - Model-based: PE, CPP, model confidence-related term, and feedback significantly predicted action and confidence, consistent with quasi-optimal Bayesian updating. No significant group interactions for these model predictors (i.e., no group differences in perceived hazard or use of model variables).

The findings clarify that in clinically diagnosed, medication-free OCD patients without comorbidities, confidence is reduced relative to healthy controls, while average action updating and action–confidence coupling are largely intact. However, OCD patients show increased sensitivity to small errors, aligning with theories of hyperreactive error signaling in OCD and clinical observations of excessive checking or overcorrection despite limited utility. In contrast, high-compulsive individuals from the general population demonstrate increased confidence and reduced coupling between confidence and action, with both confidence and decoupling scaling with symptom severity. This divergence suggests that similar levels of obsessive-compulsive symptoms in clinical versus analogue samples can arise from distinct cognitive mechanisms: patients may exhibit low confidence and heightened error sensitivity, whereas high-compulsive non-clinical individuals may show inflated confidence and decoupling. These differences support models proposing that OCD involves difficulties integrating prior beliefs with new evidence under volatility, leading to low confidence and reliance on immediate feedback. The translational implication is that results from analogue samples may not generalize straightforwardly to clinical OCD, underscoring the need to consider sample context when interpreting metacognitive findings.

This study shows that OCD patients have lower confidence and heightened sensitivity to small prediction errors during volatile learning, without evidence for altered action–confidence coupling relative to healthy controls. By contrast, high-compulsive individuals in the general population exhibit higher confidence and decoupling between confidence and action, both relating to symptom severity. These results delineate distinct behavioral and metacognitive profiles across clinical and analogue compulsivity, cautioning against direct generalization. Future research should refine tasks to enhance ecological validity and cost of excessive action, improve reliability of model-based parameters, and examine transdiagnostic mechanisms across clinical and subclinical populations to better tailor interventions.

- All groups were tested online; OCD/HC were not recruited via specialized platforms, whereas HComp/LComp were, potentially affecting task familiarity and volatility estimation. - Small prediction errors are susceptible to motor/measurement noise; although the top 5% LRs were excluded, residual noise may inflate LR at low PE. - Model-based parameters may have poorer reliability and predictive validity compared to model-free measures; between-subject inferences from model parameters should be cautious. - Ecological validity is limited; tasks without penalties for excessive action may not fully capture real-world compulsive behavior; future paradigms could include costs for over-adjustment or symptom-relevant contexts. - Data exclusion decisions (e.g., LR thresholds, PE=0 handling) may influence results despite preregistration; sensitivity to exclusion criteria warrants caution in interpretation. - Potential inconsistencies in reported statistics from secondary analyses indicate some reporting noise in the general-population comparison; primary OCD vs HC conclusions are robust.