Impaired sense of agency and associated confidence in psychosis

Psychosis involves disturbances in self-boundaries and experiences of diminished control over thoughts and actions. The Sense of Agency (SoA)—the feeling of controlling one’s own actions—arises from predictive sensorimotor processes and explicit, context-driven inferences. Prior work indicates impaired SoA in schizophrenia-spectrum disorders, potentially due to aberrant predictive coding. Most prior paradigms are non-embodied and rely on learned action–outcome contingencies, potentially missing strong priors from lifelong bodily control. It remains unclear whether patients possess metacognitive insight about their SoA deficits. This pre-registered study asked: (1) Are embodied SoA judgments (detection of sensorimotor conflicts) impaired in psychosis for temporal and spatial alterations? (2) Is metacognition of SoA (alignment between accuracy and confidence) diminished in psychosis? (3) Can SoA performance classify psychosis vs healthy controls, suggesting clinical utility?

Theoretical accounts propose SoA results from comparing efference-copy based predictions with sensory feedback; matching promotes self-attribution and sensory attenuation. In psychosis, evidence shows reduced sensory/neural attenuation, impaired prediction of action outcomes, and erroneous agency judgments, consistent with predictive coding abnormalities. Non-embodied tasks may underestimate embodied priors from lifelong motor control. Metacognitive deficits, including lack of insight, are common in psychosis, though some simple perceptual metacognitive abilities may be preserved when difficulty is controlled. The field lacks direct assessment of metacognition of embodied SoA in psychosis.

Design: Pre-registered case-control study using an embodied virtual reality (VR) virtual hand (VH) paradigm. Sensorimotor correspondence between real and virtual finger movements was manipulated in temporal (delays) or spatial (angular deviations) domains. Participants: 30 healthy controls (HC; mean age 24.4±3 years; 15 females) and 30 psychosis inpatients (mean age 30.9±8.3 years; all males) from Beer Yaakov-Ness Ziona Mental Health Center. Inclusion: right-handed, normal/corrected vision, informed consent. Ethics approvals obtained. Task: Participants’ real right hand was occluded and tracked by a Leap Motion controller. A realistic 3D VH mirrored the index-finger bending movement. Trial structure: fixation, VH presentation with single index finger bend, then two responses. In 25% of trials, VH matched the real movement; in 75% a sensorimotor alteration was introduced. Temporal delays: 100/200/300 ms; Spatial deviations: 6/10/14 degrees toward the thumb. Each magnitude per aspect appeared 30 times across five blocks (total 240 trials). Measures: (1) SoA judgment (Yes/No to “Was the VH movement identical to my movement?”). (2) Confidence rating on a continuous slider from “Not confident” to “Very confident.” After the task, psychosis patients were assessed with PANSS; HCs completed SPQ-B. Data preprocessing: Trials with no movement, tracking malfunction, or missing responses were removed (1.9% HC; 8.4% psychosis). Analyses in MATLAB and R. Statistical analysis: SoA (self-attribution) analyzed with logistic mixed-effects models (lme4), selecting models via BIC, with random intercepts and slopes for Alteration Magnitude where possible. Signal detection metrics computed across magnitudes: sensitivity (d′) and bias (c). Confidence analyzed with linear mixed-effects models including a quadratic term for Alteration Magnitude (observed hyperbolic effect). Metacognition quantified by Goodman–Kruskal’s gamma (γ) rank correlation between accuracy and confidence per participant. Correlations (Pearson) examined associations between performance and clinical scales (PANSS in patients; SPQ-B in HCs). Exploratory classifier: participant-level SoA linear-fit slopes (temporal and spatial) compared to group means using Euclidean distance; accuracy assessed with repeated subsampling across varying proportions of left-out participants and trials, including first-block-only subsets.

Impaired SoA: Mixed-effects modeling showed significant main effects and interactions. Alteration Magnitude: β=-1.18, p<0.0001, Z=18.2, 95% CI [-1.31, -1.05]; Group (psychosis vs HC): β=-0.61, p<0.0001, Z=5.9, 95% CI [-0.82, -0.41]; Alteration Magnitude × Group: β=-0.52, p<0.0001, Z=8.1, 95% CI [-0.65, -0.40]. Psychosis patients more often self-attributed altered movements and showed a shallower decrease in SoA with increasing alteration. Signal detection: HC d′=1.8 vs psychosis d′=0.76 (t56=7.39, Cohen’s d=1.9, p<0.0001); bias c: HC -0.43 vs psychosis -0.73 (t52=2.88, d=0.74, p<0.01). Impaired metacognition: Confidence mixed-model showed a significant three-way interaction Alteration Magnitude × Group × Accuracy: β=0.18, p<0.001, t=6.1, 95% CI [0.12, 0.24], driven by patients’ elevated confidence despite low accuracy, especially at large alterations. Main effects: Alteration Magnitude β=-0.18, p<0.001, t=7.47, 95% CI [-0.23, -0.13]; Accuracy β=-0.47, p<0.001, t=11.44, 95% CI [-0.55, -0.39]; Group ns (β=-0.18, p=0.07). Goodman–Kruskal’s γ: HC 0.26 (95% CI [0.17, 0.35]) vs psychosis -0.02 (95% CI [-0.12, 0.08]); between-group difference t57=4.27, p<0.0001, d=1.1. Within-group: HC γ>0 (t29=5.8, p<0.001, d=1.07); psychosis γ≈0 (t29=0.49, p=0.63). Clinical correlations: In psychosis, sensitivity d′ not correlated with PANSS total or subscales (e.g., r=-0.03 with total, p=0.86). Bias c not correlated with PANSS totals/subscales. Exploratory: metacognitive γ negatively correlated with PANSS Positive (r=-0.47, p<0.01, uncorrected). In HCs, d′ not correlated with SPQ-B total (r=-0.12, p=0.53); bias correlated with SPQ-B Disorganization (r=0.48, p<0.001, uncorrected). Classifier performance: Overall accuracy ~90% (significantly above chance; Kolmogorov–Smirnov D=0.95, p<0.001). Robustness: using half the trials and leaving out 80% of participants yielded 85% accuracy. Using only the first block (48 trials) with 20% left-out gave 81% accuracy; with 24 trials and 80% left-out, 73% accuracy.

Findings demonstrate that psychosis patients have markedly reduced embodied SoA for both temporal delays and spatial deviations, over-attributing altered movements to themselves. Their confidence failed to track accuracy, indicating impaired metacognitive insight into SoA judgments. These results support models of psychosis involving aberrant sensorimotor predictive mechanisms and an increased reliance on top-down priors, potentially explaining high-confidence misattributions of agency. Metacognitive impairment of SoA, rather than global confidence differences, may relate to positive symptom severity, as suggested by the negative correlation between γ and PANSS Positive. The high accuracy of a simple slope-based classifier underscores that individual SoA tuning curves are distinctive markers of psychosis and could be developed for clinical screening and monitoring, including remote applications.

Using an embodied VR paradigm, the study shows psychosis patients are impaired at discriminating self-generated from altered actions and lack metacognitive insight into these deficits. The work highlights disruptions across predictive sensorimotor processes and higher-order evaluative mechanisms underlying SoA in psychosis. The SoA-derived features enabled high-accuracy classification of patients versus controls, suggesting translational potential for digital assessment tools. Future research should: (1) test metacognition of SoA with tightly controlled first-order performance; (2) include broader psychiatric control groups and larger, diverse samples; (3) longitudinally track SoA and metacognition in relation to symptom fluctuations and treatment; and (4) integrate metacognitive metrics into classifiers to improve clinical utility.

Sample limitations include a modest psychosis cohort, all-male patient group, and diagnostic heterogeneity. The task did not equate first-order performance across groups, constraining some metacognitive inferences and limiting the use of newer metacognitive metrics. Absence of a non-psychotic psychiatric control group precludes claims of specificity to psychosis. Control and patient groups were not age-matched, though no associations with age were detected for SoA or confidence measures. Power was limited for detecting correlations between clinical symptoms and task metrics.