Testing the independence of self-reported interoceptive accuracy and attention

Study at a glance

• Authors: Jennifer Murphy, Rebecca Brewer, David Plans, Sahib S. Khalsa, Caroline Catmur, Geoffrey Bird
• Year: 2020
• Journal: Quarterly Journal of Experimental Psychology
• DOI: 10.1177/1747021819879826
• N: 451 / 117 / 76 / 59 / 35 across five samples; CFA on 375 pooled IAS and 350 pooled BPQ records
• Design: Six-study scale development and validation: PCA (n=451), 30-day test–retest against ICQ and BPQ (n=117), convergent validity against alexithymia and self-esteem (n=76), and two studies pairing the new scale with the heartbeat counting task (n=59; n=35), plus a confirmatory factor analysis (n=375)

Key findings

• Develops the Interoceptive Accuracy Scale (IAS) — 21 items of the form ‘I can always accurately perceive when I am hungry / breathing fast / going to sneeze’, rated 1–5 — as a trait self-report of believed accuracy across interoceptive domains, deliberately excluding attention items. Internal consistency α = .88–.90; 30-day test–retest r = .754, statistically indistinguishable from the ICQ (.814) and BPQ (.684).
• The independence result, replicated across three samples. Self-reported accuracy measures correlate strongly with each other (IAS ↔ ICQ r = −.631 and −.712 across two sessions, and −.574 in Study 5) while self-reported attention (Porges BPQ) correlates with neither (IAS r = .066, .022, .040, .091; ICQ r = .013, .091, .291 ns). Every IAS↔ICQ correlation was significantly larger than the corresponding correlation with the BPQ.
• Objective heartbeat-counting accuracy correlated with the accuracy self-report and only with it. Study 4: r(52) = .271, p = .047 (r(57) = .266 under the overestimator-tolerant scoring). Study 5: r(33) = .336, p = .049 — while the BPQ, ICQ and TAS-20 were unrelated to HCT performance (all ps > .13). The IAS↔HCT relationship survived controls for age group, gender, BMI, time-estimation performance, heart-rate variability, resting heart rate and (Study 5) participants’ beliefs about their own resting heart rate.
• State confidence behaves the same way. Post-trial HCT confidence — moment-to-moment perceived accuracy — correlated with the IAS (r(33) = .409, p = .015) but not the BPQ, and with HCT performance itself (r = .507, .454, .806 across scoring schemes and studies). Study 4’s IAS↔confidence correlation was null (r = −.026), which the authors attribute to ambiguous confidence instructions later corrected in Study 5.
• Alexithymia tracks the accuracy dimension, not the attention dimension. TAS-20 correlated with the IAS (r = −.430 in Study 3, −.572 in Study 5) and the ICQ (r = .648) but not the BPQ (r = .079, .067). The TAS-20↔IAS relationship survived partialling out self-esteem (partial r = −.255, p = .027) and, in regression, held independently of depression and anxiety. In the same regressions the BPQ was predicted only by anxiety (b = .588, p = .022) — and even that vanished under robust regression.
• Interoceptive insight is independent of trait believed accuracy. Confidence–accuracy discrepancy on the HCT was unrelated to the IAS (r = .090 in Study 4; r = .000 in Study 5), confirming the scale asks how accurate people are, not how well they know how accurate they are.
• Factor structure is unresolved. PCA suggested two factors (19% and 18.6% of variance) splitting signals perceivable from interoception alone (hunger, thirst, heartbeat) from signals with external correlates or social sanction (itch, bruise, wind, blood sugar), but confirmatory analysis rated the two-factor solution acceptable-but-imperfect (RMSEA = .072, CFI = .863) and only modestly better than one factor. The authors recommend using the total score, which is what every result above rests on.

Limitations

• The objective arm is small: n = 52–59 and n = 35 after exclusions, with correlations of r = .27–.34 sitting at p = .042–.049. The headline accuracy↔self-report finding is two underpowered samples, not a demonstration.
• The objective measure is the heartbeat counting task, whose validity is contested throughout this wiki. The authors survey the criticisms at length and proceed anyway, on the grounds that HCT is the task most of the interoception→cognition literature was built on — so the finding inherits every objection on Is the heartbeat-counting task a valid measure of cardiac perception?.
• A non-replication inside the paper, and it cuts against the controls. Contrary to previous reports — including the authors’ own (Murphy, Brewer et al. 2018) and Ring et al. (2015) — HCT accuracy was not related to beliefs about resting heart rate or to time-estimation ability in Study 5. So the claim that the IAS↔HCT relationship ‘survives’ those controls is weaker than it sounds: in this sample the controls had nothing to remove.
• Cardiac only. The IAS asks about 21 interoceptive domains and is validated against one, in a literature where channels do not predict each other.
• Studies 1–4 were run online (Study 4 entirely via a smartphone PPG application on MTurk), without the full control battery, with HCT and time-estimation intervals unmatched and uncounterbalanced. Study 5’s fingertip pulse oximeter ‘may have increased sensation of the pulse.’
• Discriminant validity against a general disposition to endorse positive self-descriptions rests on one self-esteem measure in one sample (n = 76).
• Study 3’s alexithymia correlations and Study 5’s HCT correlations come from partly overlapping analytic decisions (overestimator exclusion, two scoring schemes), and the alexithymia/heartbeat data in Study 5 were previously reported in Murphy, Brewer, Hobson, Catmur & Bird (2018).

Connections

• Methods: Interoceptive Accuracy Scale (IAS), Heartbeat detection task, Body Perception Questionnaire (BPQ)
• Concepts: Interoceptive taxonomy (Farb et al. Box 1), Interoceptive sensitivity, Alexithymia, State vs trait interoception, Interoceptive propensity, Metacognitive efficiency and metacognitive bias
• Debates: Is the heartbeat-counting task a valid measure of cardiac perception?, Is interoception domain-general or organ-specific?, Is more interoceptive awareness better?
• Related studies: Ventura-Bort et al. (2021), Murphy (2022), Banellis et al. (2026), Desmedt et al. (2022), Wallman-Jones et al. (2023), Harrison et al. (2021), Dunn et al. (2010)

Six studies developing and validating the Interoceptive Accuracy Scale, from the Bird/Catmur lab with Khalsa as co-author. The stated aim is modest — build a self-report measure of believed interoceptive accuracy, because one barely exists — but the paper is doing something more consequential, and the wiki should read it for that: it is the first study here to test the sensibility split against an objective task.

The model being tested: a 2×2, not a list

Interoceptive taxonomy (Farb et al. Box 1) tracks five partitions of interoception, all of them flat lists of constructs. Murphy, Catmur & Bird (2018, 2019) propose a different shape, and this paper is its empirical test:

	what is measured: accuracy	what is measured: attention
how: performance	heartbeat counting / detection	experience sampling of bodily attention
how: self-report belief	confidence ratings, IAS, ICQ	BPQ, MAIA Noticing

Two crossed factors: what is assessed (can you perceive accurately vs is the body the object of attention) and how (objective performance vs belief). The point of the crossing is a prediction the flat lists cannot make. Measures should relate when they share the what, whichever the how — so a self-report of believed accuracy should track objective accuracy, and a self-report of attention should not.

The quadrant with almost nothing in it was self-reported accuracy: existing questionnaires either measure attention (BPQ, MAIA) or bundle the two together (MAIA, Body Awareness Questionnaire). The IAS was built to fill it, which is what makes the test possible at all.

What they found

Self-report splits exactly along the what axis. In three independent samples, the two accuracy-belief scales (IAS and the reverse-scored ICQ) correlate at r = .57–.71 while the attention scale (BPQ) correlates with neither. This is Ventura-Bort et al.’s (2021) believed-accuracy/deployed-attention factor split arriving from a different lab, a different country and a different method — correlation structure rather than forced PCA — with the same seam.

And the how axis is crossed successfully. Heartbeat counting accuracy correlated with the IAS (r = .271, .336) and with nothing else administered: not the BPQ, not the ICQ, not the TAS-20, not time-estimation performance. Post-task confidence — the state version of believed accuracy — behaved identically, correlating with the IAS and with HCT performance but not with the BPQ.

That is the 2×2’s central prediction confirmed on both axes, in the paper built to test it. Its value to the wiki is not the scale.

Why this matters more than the paper claims: the wiki’s most durable null gets a boundary

Across Interoceptive sensitivity, Interoceptive taxonomy (Farb et al. Box 1), Multidimensional Assessment of Interoceptive Awareness (MAIA) and Heartbeat detection task the wiki records the same anomaly over and over: self-reported interoception does not correlate with objective interoceptive accuracy. Meditators attend without detecting. Attachment-anxious participants notice without perceiving. Harrison et al. find sensibility outranking accuracy against anxiety with a sparse cross-block correlation matrix. Banellis et al. find the generalizing component of confidence is not interoceptive at all. Wallman-Jones et al. produced the wiki’s first positive association and it was filed as friction.

This paper supplies the boundary condition those nulls were missing, and it is not subtle. Every one of those nulls used an attention-flavoured self-report — MAIA Noticing, the BPQ, generic body-awareness scales — against an accuracy-flavoured task. Match the what, and the correlation appears; mismatch it, and it does not, in the same participants, in the same session, twice.

Ventura-Bort et al.’s page closes by naming the discriminating study: “run these questionnaires alongside an objective interoceptive measure and a physiological one.” Half of that study is this one, and it was published first. The wiki’s sensibility↔accuracy dissociation is therefore, on current evidence, an attention↔accuracy dissociation — which is not the same claim, and is far less damaging to the taxonomies than the version the wiki has been carrying.

Two brakes worth applying before that reframing hardens.

The effect is small and the samples are tiny. r = .27 and r = .34 at n = 52 and n = 33, p = .047 and p = .049. Roughly 7–11% shared variance, in studies with the statistical power to detect little else. Nothing here rules out the reframing being partly a chance separation of two weak correlations, one of which happened to clear .05.

The objective anchor is the contested one. If the HCT is largely guessing from a believed heart rate, then a correlation between the HCT and a questionnaire asking “can you accurately perceive your heartbeat?” is not obviously a perception result. The paper’s defence against this is the control battery, and see below for why that defence is weaker in this dataset than in the literature it cites.

What it contributes to the heartbeat-task debate — honestly scored

This is the wiki’s first pro-validity empirical datum on the HCT in some time, and it should be recorded as one without being oversold.

For validity. Under a pure guessing account, HCT performance is driven by beliefs about resting heart rate plus time-estimation skill. Neither of those is a plausible cause of the selective pattern found here: time-estimation task performance correlated with no questionnaire, the HCT correlated with the accuracy self-report and not the attention self-report, and the relationship held with estimated resting heart rate partialled out. A guessing account has to explain why a belief-driven score tracks one belief questionnaire and not the other.

Against. The paper’s own control-variable analyses failed to replicate: in Study 5, HCT accuracy was unrelated to heart-rate beliefs and to time estimation, contrary to Ring et al. (2015) and to Murphy, Brewer et al. (2018) — the authors’ own paper, the one that established the control battery. Controls that show no effect cannot demonstrate that a relationship survives them. And the enormous HCT-confidence↔HCT-accuracy correlation in Study 5 (r = .806) is exactly as consistent with both quantities flowing from one shared belief about one’s own heart as with genuine metacognitive insight.

Net: a real point for the defenders, at n = 35, with its own controls non-replicating. See Is the heartbeat-counting task a valid measure of cardiac perception?.

Alexithymia lands on the accuracy side

Alexithymia on this wiki has an awkward split: the TAS-20 factors with bodily belief scales (Ventura-Bort) while the objective-task version of the claim is dead flat (Desmedt et al. 2022, r = −.01). This paper sharpens the belief half into something more specific than shared method variance: the TAS-20 tracks the accuracy self-reports (r = −.43, −.57 with the IAS; .65 with the ICQ) and not the attention self-report (r ≈ .07 with the BPQ), holds after partialling self-esteem, and holds independently of depression and anxiety in regression — while the BPQ, in the same regressions, is predicted by anxiety and nothing else.

So “alexithymia is a general deficit of interoception” is, at the level of beliefs, a claim about believed accuracy, not about how much attention goes to the body. Which is a cleaner claim, and still a claim about beliefs: the objective-task null stands untouched, and this paper’s own HCT data show no TAS-20↔HCT relationship either (all ps > .13), consistent with Desmedt.

What the BPQ turns out to be

A byproduct worth keeping, because the BPQ is used across this wiki as a general interoceptive-sensibility scale. Here it is: excellent internal consistency (α = .98), good test–retest (r = .684, though poorer than the ICQ’s .814, Z = 2.282, p = .022), an imperfect one-factor CFA fit (RMSEA = .085, CFI = .877), uncorrelated with objective cardiac accuracy, uncorrelated with the two accuracy self-reports, uncorrelated with alexithymia — and predicted by anxiety alone, an effect that did not survive robust regression. Reliable, and measuring something narrow: how much of the body is attended, with a distress component.

Placement in the wiki

Creates Interoceptive Accuracy Scale (IAS) as a method page — the instrument Ventura-Bort et al. used and the wiki had named without hosting. It is the second Murphy source in raw/ and the first with data: her 2022 commentary argues that ability is not use, and this paper is the measurement work that argument stands on. It refines rather than refutes the accuracy/sensibility dissociation on Interoceptive sensitivity and Interoceptive taxonomy (Farb et al. Box 1), adds a defender’s datum to Is the heartbeat-counting task a valid measure of cardiac perception?, and narrows Alexithymia’s belief-level claim. Held at its actual weight: a careful psychometric paper whose most important result rests on two correlations of r ≈ .3 in samples of 35 and 52.