Metacognitive efficiency and metacognitive bias
The interoceptive-taxonomy records that the field has three names for one confidence–accuracy construct: Garfinkel’s interoceptive awareness, Farb’s coherence, Khalsa’s insight. This page records something the taxonomies do not: that construct has two components that dissociate, and the interoception literature has often measured the wrong one.
The distinction
Metacognitive bias — mean confidence. A disposition to feel sure. Someone with high bias rates every trial confidently whether or not they are right. It is measurable from confidence ratings alone and requires no accuracy data.
Metacognitive efficiency — how well confidence discriminates correct from incorrect responses, corrected for how well the person actually performed. Operationalized as M-Ratio = meta-d’/d’ (Maniscalco & Lau 2012; Fleming’s HMeta-d), the ratio of the sensitivity an ideal observer would need to produce the observed confidence pattern, to the sensitivity actually observed. M-Ratio = 1 means confidence extracts all available evidence; below 1 means information is lost between deciding and rating.
The correction matters here more than in most fields. Better performers have more to be confident about, so an uncorrected confidence–accuracy correlation partly measures task performance. Banellis et al. make the point directly against the earlier interoceptive literature: “metacognitive measures that do not fully adjust for actual performance can misrepresent these relationships.”
The dissociation, and what it costs this wiki
Across cardiac, respiratory and auditory tasks in 241 participants:
| cardiac ↔ respiratory | respiratory ↔ auditory | cardiac ↔ auditory | |
|---|---|---|---|
| bias (mean confidence) | r = 0.510 | r = 0.642 | r = 0.606 |
| efficiency (M-Ratio) | ns (BF01 = 5.04) | ns | ns |
Bias travels everywhere, including to a task with no body in it. Efficiency travels nowhere. This mirrors the exteroceptive metacognition literature exactly (Ais et al. 2016; Mazancieux et al. 2020), which is itself informative — whatever the domain-general confidence factor is, it is not interoceptive.
Three consequences the wiki has to absorb:
Questionnaire “interoceptive awareness” sits on the general side. The maia, the Body Perception Questionnaire, and confidence ratings folded into awareness scores are all closer to bias than to efficiency. That reframes a finding this wiki has recorded repeatedly and explained piecemeal: sensibility measures keep failing to predict accuracy (interoceptive-sensitivity), meditators report more bodily contact without detecting better (does-mindfulness-enhance-interoceptive-accuracy), and attachment-anxious participants score high on noticing while worrying more. If sensibility instruments load on a domain-general confidence disposition, none of these are anomalies. They are one dissociation, met four times.
The Garfinkel cross-modal result splits in half. Garfinkel et al. (2016) reported cardiac–respiratory metacognitive correlation, which the field read as evidence for a general interoceptive awareness capacity. Banellis et al. reproduce the confidence half and, with four times the sample, do not reproduce the insight half. So the domain-general finding survives in the component that is not specifically interoceptive, and fails in the component that is. See is-interoception-domain-general.
“More interoceptive awareness” is at least two questions. is-more-interoceptive-awareness-better collides findings measured with sensibility scales against findings measured with task performance. On this page’s distinction those are not two measures of one thing; they are a person-level disposition and an organ-level competence.
What is not settled
M-Ratio is not clean. It corrects for first-order performance but retains residual dependence on confidence criterion and decision accuracy (Rahnev & Fleming 2019; Xue et al. 2021 — the shape of metacognitive noise confounds efficiency with bias). Banellis et al. use it because a systematic comparison finds it outperforms the alternatives (Rahnev 2025), while calling for a generative model of metacognition built for interoception tasks. So the bias/efficiency separation this page rests on is better than the alternatives and not airtight.
Its estimates are unstable where the wiki needs them most. Cardiac M-Ratio required excluding 96 of 241 participants for negative or implausible values, against 10 respiratory. M-Ratio is known to misbehave when first-order performance is near chance — and near chance is precisely where cardiac interoception lives (under 10% of adults meet a counting accuracy criterion; one forensic sample averaged d’ = 0.00). A measure that becomes unestimable at low performance is being asked to work in a literature whose participants mostly cannot do the task. That is a structural problem for interoceptive metacognition, not a technicality.
And the domain-general confidence factor is unexplained. Whether it is a genuine trait, a scale-use habit, or a mixture of both is undetermined by anything the wiki holds — no study here decomposes cross-task confidence correlation into shared disposition and shared method variance.