Dynamic influences on static measures of metacognition

The paper addresses how to quantify metacognitive accuracy—the ability to judge the correctness of one’s decisions via confidence—without confounding it with primary task performance or decision strategy. Traditional static measures (e.g., meta-d′ and M-ratio = meta-d′/d′) are widely used but ignore decision dynamics, particularly the speed-accuracy tradeoff governed by response caution (decision boundary). Because focusing on speed induces more premature, easier-to-detect errors, static measures may conflate metacognitive ability with response caution. The authors propose using a dynamic evidence accumulation framework (drift diffusion model, DDM) in which confidence arises from post-decisional evidence accumulation. Within this framework, they define v-ratio as the ratio of post-decision drift rate to pre-decision drift rate as a dynamic measure of metacognition. The core hypothesis is that M-ratio will vary with response caution, whereas v-ratio will capture metacognitive accuracy independently of caution.

Several approaches exist to quantify metacognitive accuracy: correlations between confidence and accuracy are confounded by task difficulty; alternatives include logistic mixed models, type 2 ROC curves, and meta-d′ within signal detection theory. The meta-d′ framework quantifies how well confidence discriminates correct from incorrect responses while controlling for d′, enabling computation of M-ratio as metacognitive efficiency. This framework has been used to study domain-generality of metacognition, neural correlates, bilingualism effects, and links with psychiatric dimensions. However, all these static approaches disregard the dynamics of decision making and the speed-accuracy tradeoff. Evidence accumulation models (e.g., DDM) conceptualize choices as noisy accumulation to decision boundaries, with boundary height reflecting response caution. Prior work suggests confidence can be computed from post-decisional accumulation, aligning with typical findings that stronger evidence yields higher confidence. The authors build on this literature to propose v-ratio as a dynamic metacognitive measure that should be insensitive to boundary setting.

The study combines model simulations, a within-participant experimental manipulation of speed-accuracy emphasis (Experiment 1), and re-analyses of four datasets without explicit speed instructions (Experiments 2A–D).

• Simulations: 100 simulated agents with 1000 trials each using a DDM with post-decisional accumulation. Parameters varied across agents: drift rate, decision boundary, non-decision time, and post-decision drift (encoded via v-ratio). Confidence was computed as evidence integrated after a post-decisional interval sampled to mirror empirical confidence RT distributions. M-ratio was computed from confidence ratings binned into four levels and meta-d′/d′ (Maniscalco & Lau method), and v-ratio as post-decision drift divided by drift rate. Correlations among parameters and measures were assessed.
• Experiment 1 (N=32): Random dot motion direction discrimination (20% coherence), with block-wise instructions to emphasize accuracy versus speed. After each decision, confidence was reported on a 6-point scale. Behavioral measures: accuracy, RTs, confidence, and confidence RTs. Data were fit with an extended DDM featuring parameters for drift rate (v), decision boundary (a), non-decision time (Ter), v-ratio (post-decision drift/drift), and two mapping parameters (M, SD) linking integrated evidence to confidence scale. The duration of post-decisional accumulation per trial was set from the empirical confidence RT distribution. Fitting used quantile optimization across RT and confidence distributions for correct and error trials. Statistical tests compared parameters and measures across instruction conditions.
• Experiments 2A–D (total N=430): Four perceptual discrimination datasets with confidence reports and without explicit speed instructions: 2A dot numerosity comparison with continuous confidence (N=63); 2B similar task with 6-point confidence (N=96); 2C temporal two-interval Gabor contrast discrimination with continuous confidence (N=204); 2D average color decision with 6-point confidence across four datasets (N=67). The same DDM with post-decisional accumulation was fit per participant. Hierarchical mixed-effects models assessed relations among M-ratio, v-ratio, and estimated decision boundary across datasets, including confidence response mode as a factor where relevant. Additional analyses related d′ and meta-d′ to boundary settings. Parameter recovery analyses validated the modeling approach.
• Data and code: Data and analysis code are publicly available; some datasets sourced from prior publications and the Confidence Database.

• Simulations: M-ratio correlated positively with v-ratio (r=0.436, p<0.001) and negatively with decision boundary (r=-0.552, p<0.001), indicating dependence of M-ratio on response caution. By design, v-ratio was unrelated to boundary (r=-0.044, p=0.670). Full parameter correlations showed M-ratio also negatively related to drift rate (r=-0.24, p=0.017).
• Experiment 1 (speed vs accuracy instructions, N=32): Behavioral manipulation succeeded: faster RTs under speed (650 ms) vs accuracy (832 ms), t(31)=5.67, p<0.001; lower accuracy under speed (76.2%) vs accuracy (78.8%), t(31)=2.20, p=0.035; lower confidence under speed, t(31)=2.41, p=0.022. Model fits captured RT and confidence distributions. Estimated decision boundary was lower in speed (1.29) vs accuracy (1.63), t(31)=5.59, p<0.001; drift rate did not differ (p=0.478); non-decision time modestly shorter in speed, t(31)=2.13, p=0.041. Critically, M-ratio was higher under speed (0.74) vs accuracy (0.51), t(31)=2.29, p=0.029, whereas v-ratio did not differ (p=0.647).
• Experiments 2A–D (N=430, no explicit speed instruction): Hierarchical models showed M-ratio negatively related to decision boundary (b=-0.200, t(424)=-2.43, p=0.015), independent of confidence response mode; v-ratio positively related to M-ratio (b=0.31, t(424)=5.548, p<0.001) and not related to decision boundary (p=0.233), with no effects or interactions of response mode. Reduced models confirmed that v-ratio’s independence from boundary did not depend on including M-ratio.
• Per-experiment summaries: 2A: M-ratio vs v-ratio r=0.21 (p=0.090), M-ratio vs boundary r=-0.41 (p<0.001), v-ratio vs boundary r=-0.03 (p=0.797). 2B: M-ratio vs v-ratio r=0.63 (p<0.001), M-ratio vs boundary r=0.11 (p=0.282), v-ratio vs boundary r=0.22 (p=0.029). 2C: M-ratio vs v-ratio r=0.277 (p<0.001), M-ratio vs boundary r=-0.16 (p=0.024), v-ratio vs boundary r=0.109 (p=0.120). 2D: M-ratio vs v-ratio r=-0.019 (p=0.878), M-ratio vs boundary r=-0.30 (p=0.015), v-ratio vs boundary r=-0.18 (p=0.139).
• Relations with d′ and meta-d′: Simulations: d′ positively correlated with boundary (r=0.683, p<0.001), meta-d′ not significantly related (r=-0.141, p=0.160). Experiment 1: no significant differences in d′ or meta-d′ across instruction conditions (ps=0.083). Experiments 2A–D combined: d′ unrelated to boundary (b=-0.045, p=0.444); meta-d′ negatively related to boundary (b=-0.358, t(427)=-3.527, p<0.001).

Findings demonstrate that the commonly used static metacognitive efficiency measure M-ratio is influenced by response caution (decision boundary), conflating metacognitive ability with speed-accuracy strategy. This effect appears in simulations, under explicit instruction manipulations, and across multiple datasets without speed emphasis. In contrast, v-ratio, derived from a dynamic evidence accumulation framework and capturing the strength of post-decisional accumulation relative to pre-decisional drift, correlates with M-ratio (shared variance in metacognition) but is independent of decision boundary, aligning with the conceptual requirement that metacognitive accuracy should be strategy-invariant. The results caution interpretations of domain-generality of metacognition, neural correlates, and group differences when using static measures, as observed effects might reflect boundary settings rather than metacognitive computations. The dynamic framework offers a principled alternative that integrates reaction times and confidence timing. The authors note that while M-ratio’s dependence can arise via effects on both d′ and meta-d′, patterns vary across contexts, underscoring complexity in static measures. They also discuss the need to consider potential changes in drift rate under speed stress, which could influence v-ratio, and emphasize ruling out response caution confounds in future work on metacognition’s neural and cognitive bases.

Across simulations and human data, M-ratio is confounded by response caution, whereas a dynamic measure, v-ratio (post-decision drift rate relative to drift rate), is unaffected by speed-accuracy tradeoff and thus better isolates metacognitive accuracy. The work advances metacognition measurement by embedding confidence within evidence accumulation dynamics, leveraging both decision and confidence timing. Future research should: (1) collect precise confidence RTs to estimate v-ratio reliably; (2) model and test mechanisms of post-decisional processing and its termination; (3) ensure that observed differences in metacognition reflect post-decision evidence rather than global drift changes; and (4) re-examine reported domain-general, neural, or group differences in metacognition controlling for response caution.

• Experiments 2A–D are correlational; boundary-metacognition associations cannot establish causality and may reflect strategic choices by participants.
• Some datasets (including Experiment 1 and 2B) were collected online, limiting control over hardware and timing precision for between-participant comparisons.
• Accurate estimation of v-ratio requires precise measurement of confidence reaction times; many paradigms may lack sufficient temporal resolution (e.g., cursor-based scales).
• The modeling assumes confidence arises via post-decisional evidence accumulation; the exact source of post-decision evidence (sensory buffer, memory resampling, etc.) and stopping rules are not fully specified.
• If speed instructions reduce drift rate (not only boundary), v-ratio can be affected because it is a ratio involving drift; careful experimental checks are needed.