On the relationship between interoceptive awareness, emotional experience, and brain processes (Pollatos et al. 2005)
The wiki’s first EEG study, its first look at the pre-Dunn interoception-and-emotion literature on its own terms, and — unexpectedly — the counter-evidence to the claim four wiki pages are currently built on.
Forty-four people counted their heartbeats, then looked at 60 IAPS pictures while 61 electrodes recorded their brains, rating each picture for valence and arousal. Good heartbeat perceivers said the emotional pictures were more arousing, and had bigger P300s and slow waves. The authors read this as support for James.
The conflict with Dunn et al. (2010)
This is the reason the ingest matters, so it goes first.
Dunn et al. (2010) Study 1 and this study run nearly the same experiment. Both show affective pictures for 6 s. Both collect felt arousal and felt valence per picture. Both measure interoceptive accuracy with Schandry heartbeat counting. Both regress one on the other in a healthy adult sample of roughly the same size (58 vs 44).
They disagree:
| Pollatos et al. (2005), n = 44 | Dunn et al. (2010) Study 1, n = 58 | |
|---|---|---|
| interoceptive accuracy → arousal ratings | F(1, 39) = 5.90, P < .05; r = 0.34 | ns |
| interoceptive accuracy → valence ratings | ns (F(1, 39) = 0.14) | ns |
| interoceptive accuracy → cardiac response | not measured | ns |
| bodily response x accuracy → arousal | not testable | delta-R-squared = .12, P < .01 |
The wiki has been stating Dunn’s null in strong terms — “interoceptive accuracy predicts nothing on its own,” on interoceptive-sensitivity, heartbeat-detection-task, is-more-interoceptive-awareness-better and dunn-2010-listening-to-your-heart itself. That claim now has a five-years-earlier study reporting the opposite in the same paradigm, and this page exists partly to stop those four pages from asserting the null as settled.
Four things keep it from being a simple contradiction, and they are worth separating.
1. The samples are not comparable, and this one is stacked. Pollatos et al. selected 22 good perceivers from a screen of ~140 and matched 22 comparisons to them. That is an extreme-groups design. It has more power to detect a group difference than Dunn’s unselected sample does, and it inflates any correlation computed across it. So “Pollatos found r = 0.34, Dunn found r = .08” is not two estimates of one quantity — it is one selected-sample coefficient and one unselected-sample coefficient, and the first is expected to exceed the second even if the truth is identical. The disagreement is smaller than the numbers suggest, and neither paper’s r should be quoted as the population value.
1b. But the selection may be doing something defensible, and the wiki did not see this until the Van der Does et al. (2000) ingest. Van der Does et al. pooled 709 participants and found that only ~8% of healthy adults meet an accuracy criterion while over 95% produce a count — and argue the task is valid for that minority and measures something else entirely in the rest. If so, taking the top sixth of a screen of ~140 is approximately selecting the subpopulation the task works on, and Dunn’s unselected correlation is computed across a mixture where roughly four in five participants contribute noise. Selection as purification, not only as inflation. The two accounts are compatible — the selection can isolate a real subgroup and inflate the coefficient within it — and the wiki now carries both rather than only the sceptical one. See is-the-heartbeat-counting-task-valid.
1c. And the price it charges this paper is steeper than the one above. Two things follow from Van der Does et al. that this design does not survive comfortably. The bimodality in Fig. 2 is not evidence of a natural kind — the gap between 0.82 and 0.87 is manufactured by the selection procedure, so it cannot be cited as independent support for a two-population reading of heartbeat perception, and this page should not do so. And “good perceiver” may not be a stable category: Van der Does et al. report fewer than half of accurate patients still accurate at a second session, and exercise manufacturing accurate perceivers above ~100 bpm who vanish by ~95. Pollatos et al. screened in one session and tested in another, treating the classification as a trait. If it is partly state, some of their “good perceivers” were not good perceivers on the day that mattered — which attenuates their effect rather than inflating it, and is the rare consideration on this page that runs in the paper’s favour.
2. They fit different models, and Pollatos et al. cannot fit Dunn’s. Dunn’s claim is not that accuracy is unrelated to arousal; it is that accuracy is a gain term on the bodily signal — arousal is predicted by (bodily response x accuracy), not by accuracy. Testing that needs a measured bodily response. Pollatos et al. recorded ECG and never analyzed it against the pictures, so the moderator has nothing to moderate. This study tests a model Dunn et al. argue is the wrong one, using a design that cannot evaluate the alternative.
3. Which means a main effect here is not necessarily inconsistent with a null there — arithmetic, recorded as this wiki’s reading and not either paper’s. If feeling really is b1*(bodily response) + b3*(bodily response x accuracy) with no accuracy main effect, then averaging over bodily responses leaves a marginal accuracy slope of b3 x (mean bodily response). Dunn’s images all produced HR deceleration, so the mean bodily response in an affective-picture study is not zero — and a marginal main effect of accuracy is exactly what an interaction predicts under those conditions. Pollatos et al., who measured only the margin, would see it. The account is incomplete and should be flagged as such: Dunn’s own marginal was null, and this reading owes an explanation of why. Sample selection (point 1) supplies a candidate. Neither paper is in a position to adjudicate.
4. They agree exactly where it counts most. Both find nothing for valence. That convergence is not weakened by any of the above — it survives the design difference, the selection difference, and the model difference, and it is the finding this wiki should take from the pair.
The honest summary: on whether interoceptive accuracy has a main effect on felt arousal, the wiki now holds two studies pointing opposite ways, neither decisive, differing in sampling and in what they measured. That is a real open question and it is recorded as one on is-more-interoceptive-awareness-better. What the wiki should stop doing is citing Dunn’s null as though it were unopposed.
The arousal/valence dissociation now has three independent legs
The more durable contribution, and it strengthens rather than disturbs existing pages.
Interoceptive accuracy raised arousal ratings for pleasant and unpleasant pictures, and did nothing to valence ratings — no main effect, no interaction. Same result, same direction, as Dunn et al. (2010), in a different sample, five years earlier, with a different analysis. And Ferré et al. (2024) report the split a third time from an entirely different direction, in the lexicon: “feeling” ratings correlate −.043 (ns) with valence across 1,286 words.
Three methods — ERP/ratings, cardiac moderation, semantic norms — spanning 2005 to 2024, all saying the bodily dimension of emotion and the hedonic dimension come apart. core-affect carries this, and the leg added here is the cleanest of the three in one respect: it is a straightforward group difference requiring no product term, no partialling, and no theory to interpret.
The dissociation also lands where Dunn et al. said it should, on hybrid two-factor accounts over strong Jamesian ones. Strong James predicts bodily perception should gate all emotional experience, valence included. It does not. Which makes the paper’s own framing — “as would be postulated for example by the James–Lange theory of emotions,” in the abstract’s last line — an overstatement of its own result: the data pick out Schachter and Singer’s arousal term specifically and leave the quality of the emotion untouched, which is closer to two-factor theory than to James. See dror-2017-two-factors for the irony that the two-factor lineage reached this prediction without thinking about bodies at all.
What the ERPs do and do not show
The paper’s title promises brain processes, and this is where its reach exceeds its grasp.
The P300 effect is real and is not about emotion. Good perceivers had larger P300s than poor perceivers — F(1, 42) = 5.17 — averaged across all 60 slides, “independent of the affective content” in the authors’ own words. Pictures of household objects included. A group difference that does not care whether the stimulus is a mutilation or a hairdryer is a fact about these people’s P300s, not about emotional processing, and the abstract’s “strong relationship between the perception of cardiac signals and the cortical processing of emotional stimuli” is not what this effect shows.
It also sits badly with the paper’s own mediation story. The discussion proposes that higher arousal in good perceivers drives the VEP differences. But the arousal difference is emotion-specific (nothing for neutral slides, F = 0.09) and the P300 difference is not. The proposed mediator has a specificity profile the effect it supposedly mediates does not share, and no mediation was tested.
The slow wave is the affect-specific result, and it is the paper’s best one. Between 550 and 900 ms, the good/poor difference appears for pleasant and unpleasant slides and not for neutral ones, lateralized to the left at anterior and medial sites, with the paper’s largest correlations (r = 0.46 antero-inferior left). If interoceptive accuracy touches emotional processing anywhere in this dataset, it is here.
The anatomy is borrowed. Every structural claim — insula, ACC, medial prefrontal, somatosensory cortices — is inferred from scalp topography plus other people’s imaging. The paper is candid that P300 generator studies disagree (frontal per Anderer et al. 1998 and Yamazaki et al. 2001; posterior/parietal per Keil et al. 2002) and then reasons from that disagreement to a conclusion anyway. The N100 null grounds the one inference that is properly its own: interoception does not touch the earliest response, so whatever it does, it does late. On Damasio’s (1999) staging that points to first- and second-order structures rather than trigger sites — insular-cortex, anterior-cingulate-cortex, somatic-marker-hypothesis — which is a reasonable inference from a null in an EEG study, and no more than that.
What this paper is, in the wiki’s terms
It is a good example of the literature Dunn et al. were pushing against, and reading it makes their contribution legible in a way the Dunn ingest alone did not.
Pollatos et al.’s own introduction states the prior consensus: “Most of the studies addressing this question found a positive relationship between heartbeat perception and emotional experience,” citing Schandry (1981), Wiens et al. (2000), Critchley et al. (2004), Ferguson & Katkin (1996), Montoya et al. (1993) — with Blascovich et al. (1992) the lone dissent. This paper is a member of that consensus, and the wiki previously held that consensus only through Dunn’s rebuttal of it. Now it holds one instance first-hand, and can see what the instance actually does: it measures the perceiver and never the body, correlates the perceiver with the report, and concludes the body causes the report.
That is precisely the gap Dunn identified — a bodily theory of emotion whose evidence never includes a bodily measurement — and Pollatos et al. illustrate it while claiming to support James. The paper’s Jamesian conclusion depends on a chain (heartbeat counting → visceral sensitivity generally → more perceived bodily change → more felt arousal) in which only the first and last links were measured. Whitehead & Drescher (1980) is cited to justify the middle step; it is a 1980 gastric-motility study of different organs in different people, and it is carrying the paper’s central inference.
None of which makes the finding wrong. It makes it a finding about a heartbeat score and a rating scale — and, given heartbeat-detection-task’s cardiodynamic confound, plausibly a finding about sympathetic tone. Good perceivers rated everything emotional as more arousing and had bigger P300s to everything full stop. “These are people whose hearts are louder and whose cortices are more responsive” fits the whole result set without any perception in it, and the design cannot exclude it.