Heartbeat perception in panic disorder: a reanalysis (Van der Does et al. 2000)
The wiki has been building on a task for twenty pages. This is the paper that asks whether the task measures anything, in most of the people it is administered to.
Seven studies, 709 participants, contributed by the four authors who had spent the 1990s disagreeing in print about whether panic patients perceive their heartbeats better than anyone else. Rather than review each other, they pooled their raw data and re-scored it. That alone makes this the wiki’s most unusual source: a dispute settled by the disputants, using the disputants’ own participants.
The answer to the original question is yes-but-trivially. The answer to the question they discovered underneath it is the reason this page is long.
The finding: accurate perception is rare
| group | n | accurate | probably accurate | inaccurate |
|---|---|---|---|---|
| panic disorder | 275 | 17.1% | 14.5% | 68.4% |
| social phobia | 20 | 25.0% | 5.0% | 70.0% |
| GAD | 15 | 20.0% | 6.7% | 73.3% |
| specific phobia | 27 | 11.1% | 11.1% | 77.8% |
| normal controls | 191 | 7.9% | 13.1% | 79.1% |
| palpitations, no PD | 99 | 6.1% | 19.2% | 74.7% |
| infrequent panic | 50 | 6.0% | 12.0% | 82.0% |
| mood disorder | 32 | 0.0% | 9.4% | 90.6% |
χ²(14) = 29.4, p = 0.009. PD exceeds normal controls, palpitation patients, depressed patients, and (dichotomized) infrequent panickers. PD does not exceed social phobia, GAD, or specific phobia — the three groups too small to say much about, one of which is numerically ahead of it.
Read the right-hand column instead. In every group, most people cannot do this task. In the modal group — normal controls, the population nearly every study in this wiki recruits — the accurate rate is under 8%.
And 3.5% of the whole sample reported feeling no heartbeat at all (18.8% of depressed patients, 1.1% of PD patients; χ²(7) = 34.2, p < 0.0001). Everyone else produced a number.
The fork the paper is built on
This is the part that matters beyond panic research, and it is a disagreement about measurement, not about panic.
Everyone agrees on the data. The dispute is over what to do with a participant whose count is wrong by 30%.
- Ehlers’s position: the task is valid, so error is error — a continuous quantity, and a person 30% off is simply worse at a thing everyone does to some degree. Dividing people into accurate/inaccurate imposes an artificial boundary on a continuum. Every Ehlers paper after 1992 reports only % error.
- Van der Does’s position: the task is valid for the minority who genuinely perceive their heartbeats, and for the rest the count is unrelated to their actual heartbeats — they are doing something else entirely, and averaging them with real perceivers produces a number that is a mixture of two populations. Then the boundary is not artificial; it is the only thing in the data that means anything.
The evidence the authors marshal for the second reading is the strongest thing in the paper, and it is mostly self-report: participants do not say they lost track of a few beats. They say they felt a regular rhythm, somewhat slower than their actual heart rate. That is not what missing beats feels like. It is what feeling something-other-than-your-heart feels like.
This cuts directly against the standard validity argument — ‘almost everyone undercounts, which is what you would expect if people accurately perceive their heartbeats but miss a few.’ Van der Does et al. point out the argument has an equally good competitor: if people expect and count a rhythm slower than their real one, undercounting is also what you would expect, with no perception in it. And it predicts a nasty artefact. If PD patients feel more nervous during the test and therefore expect a faster rhythm, they will count faster, and — since nearly everyone undercounts — land closer to the truth. They would earn lower error scores by being anxious, not by perceiving.
See is-the-heartbeat-counting-task-valid, where this is tracked as a debate with named positions.
The exercise result, and why it is the most important thing here for this wiki
Schandry, Bestler & Montoya (1993) found that stroke volume correlates with heartbeat-counting performance. This wiki has carried that as the cardiodynamic confound on four pages — the single objection that keeps regenerating, and the one that fits every result on both sides of the Dunn/Pollatos dispute with no perceiving anywhere in it.
It has been a correlation. Here it is a manipulation.
Antony et al. (1995) exercised their participants and re-ran the task over seven trials while heart rate decayed back to baseline:
| trial | 1 | 2 | 3 | 4 | 5 | 6 | 7 |
|---|---|---|---|---|---|---|---|
| actual HR (bpm) | 130 | 104 | 95 | 93 | 90 | 91 | 92 |
| accurate perceivers | many | many | back to baseline | — | — | — | — |
Accuracy rose while the heart was loud and vanished when it quieted, at a threshold the authors put at about 100 bpm. Twenty-five of 60 participants showed the transient gain. Exactly one baseline-inaccurate participant became durably accurate. And the effect was identical in PD patients, social phobics and controls, who did not differ in actual HR at any trial.
Nobody in that experiment learned to perceive anything. The signal got louder and the score went up. Raise the amplitude of the thing to be detected and you manufacture accurate perceivers; let it fall and they dissolve. That is what a signal-strength account predicts and what a perceptual-skill account does not.
Two further teeth in it:
- The mean lied, and the categories caught it. Antony et al. had concluded exercise produced a general improvement in error scores. It did not: the post-exercise error scores of inaccurate perceivers were no better than baseline. The population mean moved because a subpopulation changed state. This is the paper’s categorical thesis demonstrated on the continuous score’s own home turf, and it is the best argument in it — better than the bimodal histograms, because it is a mechanism rather than a shape.
- It hands the clinical literature a problem. The authors immediately ask whether panic attacks reach the heart rates that manufacture accuracy, and answer, via McNally (1994), mostly no — HR increases during natural panic are ‘rather modest (if they occur at all)’ and largely confined to situational attacks. So the state in which patients actually panic is not a state in which the body is loud enough to be perceived. Which is precisely the setup for the schema account below.
What the other 80% are doing
If more than 95% of participants report perceiving their heart rate and 80% of them are wrong by ~30%, the interesting question is not who is accurate. It is what the majority are experiencing.
The paper’s candidate: vague bodily sensations misinterpreted as heartbeats, organized by cognitive schemata (Pennebaker 1982). Once a PD patient perceives a situation as threatening, an anxiety schema activates — attentional shift, selective perception, expected high HR, arousal symptoms — and symptom perception becomes ‘more guided by the schema (that is, by past information) than based upon present physiological status.’
Read that sentence with Seth (2013) in hand. Perception of the body’s state driven by a prior rather than by the afferent signal, with the prior supplied by past experience and activated by context, is interoceptive-inference. Van der Does et al. get there in 2000, from panic research and Pennebaker’s psychosomatics, with no Bayes and no generative model — and they propose the research programme that follows from it: work out ‘the conditions under which schema-guided versus physiology-based processing occur.’ That is the predictive-coding research programme, named thirteen years early by people who did not know they were in it. Recorded on feedforward-vs-predictive-interoception and schema-guided-symptom-perception.
The evidence for it here is thin — confidence ratings stay low (60.4% even under ‘count only those heartbeats about which you are sure’), which fits, and the somatosensory amplification scale failed to correlate with counted beats, which does not. The authors report the failure.
What this does to the wiki’s Dunn/Pollatos dispute
The wiki’s most active live disagreement is whether interoceptive accuracy predicts felt arousal: Pollatos et al. (2005) found the main effect (F(1, 39) = 5.90, r = 0.34), Dunn et al. (2010) found nothing (r = .08). Four pages carry it. The wiki’s explanation has been sampling: Pollatos used extreme groups, which inflates correlations and eases detection, so their r is not comparable to Dunn’s.
Van der Does et al. supply a second explanation, pointing the opposite way, and the wiki should hold both.
If the task is valid only for the ~8–17% who are genuinely accurate, and the rest supply counts unrelated to their hearts, then:
- Pollatos’s selection is not merely inflationary — it may be isolating the valid subpopulation. They screened ~140 people and took 22 good perceivers (≈16%), which is almost exactly the prevalence of genuine accuracy this paper reports. On Van der Does’s reading, that design does not distort the sample so much as purify it: it compares people the task works on against people it does not.
- Dunn’s unselected correlation runs across a sample in which roughly four in five participants contribute noise on the moderator. A correlation computed across a mixture of one valid subpopulation and one measuring-something-else subpopulation is attenuated toward zero — which is the shape of r = .08 and r = .07.
So the wiki’s tidy story (‘extreme groups inflate, therefore discount Pollatos’) has a mirror image (‘unselected samples dilute, therefore discount Dunn’s null’), and the same fact — 16% selection — is the evidence for both. Note that these are not exclusive: selection can isolate a real subgroup and inflate the within-selection correlation. Both are probably true.
This is the wiki’s inference, not the paper’s. Van der Does et al. wrote in 2000 and mention neither study (Dunn is a decade later); they are talking about panic, not about affective picture-viewing. What they supply is a prevalence estimate and an argument that the score is a mixture — and the mixture reading has consequences for every unselected heartbeat-counting correlation in this wiki, which is nearly all of them.
And note the price it charges Pollatos too. If accuracy is state-dependent and manufacturable by heart rate — as the exercise result says — then ‘good perceiver’ is not a stable category to select on. Pollatos’s 22 good perceivers were selected on a screening session and tested later; this paper says less than half of accurate PD patients were still accurate at a second session across treatment. Selecting on a labile state and then treating it as a trait moderator is a problem neither study addresses.
What it does to ‘interoceptive sensitivity’ as a trait
interoceptive-sensitivity defines its subject as ‘a characterological trait reflecting individual sensitivity to interoceptive signals.’ Two results here bear on the word trait:
- Accuracy is not stable. 8 of 17 accurate PD patients were still accurate after treatment; 4 became inaccurate. Ehlers & Breuer (1996) and Antony et al. (1994) had concluded from % error scores that good perception is a stable individual characteristic. The categorical re-scoring says it is not, and the authors say the earlier conclusion ‘may have to be reconsidered.’
- Inaccuracy is stable. Only 5 of 51 baseline inaccurate perceivers became accurate. Whatever is trait-like here is the inability — which is what you would expect if most people are not perceiving their hearts at all and no amount of treatment changes that, and also what you would expect if the trait is cardiodynamic and treatment does not change stroke volume.
Both readings are available. Neither is the one the construct’s name implies.
The literature’s habit of publishing against itself
Worth recording, because this is now the second instance on the same instrument.
Schandry built the heartbeat-counting task and then published the stroke-volume confound that undermines reading it as perceptual skill. Ehlers built the case that panic patients perceive their hearts better, defended the task’s validity, and then co-authored the reanalysis that reduces her group difference to a prevalence claim, reverses her lab’s conclusion about the stability of good perception, and shows her distraction manipulation to have a large effect she had reported as minimal. Antony’s exercise conclusion is corrected here with Antony as second author.
There is a related irony on the validity argument itself. Ehlers cites the stroke-volume correlation as evidence the task is valid — if counting performance tracks cardiodynamics, participants must be responding to real cardiac events. This wiki, following Oldroyd et al., carries the identical finding as the confound that undermines the task. Both readings are coherent and they are not compatible: the same correlation shows the task is measuring something cardiac (Ehlers) and that what it measures is the heart’s loudness rather than the perceiver’s skill (this wiki). See rainer-schandry, where the inversion is now recorded.
Why this paper is not on the wiki’s usual axis
Most sources here argue about what interoception does — for emotion, decision, wellbeing. This one argues about whether the number in the interoception column is a measurement. It is a clinical-psychology paper about panic disorder, and the panic findings are its least important contribution to this wiki.
What it leaves behind:
- A prevalence estimate for genuine cardiac accuracy — under 10% in healthy adults — that every unselected heartbeat-counting study in this wiki is implicitly averaging over.
- An experimental demonstration that the score moves with signal amplitude and not with skill.
- Evidence that the score is state-dependent, not trait-like, in the direction that matters (the ability is labile; the inability is not).
- A schema-driven account of what the majority are doing instead of perceiving, which is interoceptive-inference avant la lettre.
- A demonstration that the categorical and continuous scores disagree about the same manipulations, on the same data — which is the strongest form the measurement worry can take, because it means the choice of score changes the conclusion.
And the thing it does not leave behind: a test of its own central claim. The bimodality that would license the whole categorical enterprise is asserted from three histograms and never tested. Held here as the best-evidenced challenge to the wiki’s foundational instrument, and as a challenge that is itself one formal test away from being properly made.