Interoception Wiki

← RADlab

The somatic marker hypothesis: A critical evaluation

25 min read

Study at a glance

  • Authors: Barnaby D. Dunn, Tim Dalgleish, Andrew D. Lawrence
  • Year: 2006
  • Journal: Neuroscience and Biobehavioral Reviews 30:239-271
  • DOI: 10.1016/j.neubiorev.2005.07.001
  • Design: critical review (no new data). Systematically evaluates the somatic marker hypothesis in three parts: the Iowa Gambling Task as its evidential base, the proposed neural substrate, and conceptual issues of novelty, specification and parsimony. Includes a table of every IGT clinical study published before July 2005 (~35 studies).

Key findings

  • The verdict is split, and the split is the paper’s main contribution. The psychological mechanism of the hypothesis requires substantial revision; the neural substrate is ‘reasonably well supported.’ Damasio ‘seems to have accurately identified some of the brain regions involved in decision-making, emotion, and body-state representation, but exactly how they interact at a psychological level is still somewhat unclear.’ The authors are explicit that none of their reservations falsify the framework — they show it needs evidence it does not have.
  • Identifies three assumptions the IGT must satisfy to support the hypothesis, and argues each fails: (1) the reward/punishment schedule is cognitively impenetrable, so early learning must be non-conscious; (2) the learning proceeds via anticipatory marker signals arising from the body; (3) the VMPFC deficit is a ‘myopia for the future.’
  • Assumption 1 fails (cognitive penetrability). Maia & McClelland (2004) re-ran the task with focused rather than open-ended knowledge probes in 20 healthy participants: advantageous performance was nearly always accompanied by reportable, reasonably accurate quantitative and qualitative knowledge sufficient to guide behaviour. Participants reported accurate knowledge more reliably than they chose advantageously. A 20-participant control given Bechara et al.’s original questions behaved identically, ruling out the objection that detailed questioning manufactured the insight. Bowman et al. (2005) independently found above-chance deck ‘goodness’ ratings as early as the first block of 20 trials.
  • The Iowa rebuttal is read as a retreat. Bechara et al. (2005, TICS) answer that transparency is not a problem because ‘the central feature of the SMH is not that non-conscious biases accomplish decisions in the absence of conscious knowledge, but rather that emotion-related signals assist cognitive processes even when they are non-conscious.’ Dunn et al.: this abandons the earlier emphasis on implicit markers (Tranel et al. 1999) ‘and makes the SMH hard to distinguish from other accounts of decision-making.’
  • Assumption 2 fails three ways (what the anticipatory SCR means). (a) Variance, not value — Tomb et al. (2002) reversed the magnitude structure so the advantageous decks carried larger rewards and punishments; anticipatory SCRs followed the high-magnitude decks, not the bad ones, consistent with SCR indexing variance rather than long-run goodness. (b) Not necessary — Crone et al. (2004), 96 students split into good/moderate/poor performers: only the good group showed anticipatory SCR (and HR) differentiation between decks; the moderate group acquired the task with no psychophysiological differentiation at all. (c) Possibly post-decisional — Amiez et al. (2003) found in rhesus monkeys on an equivalent task that SCRs track anticipation of reward after a decision is made rather than before, so the ‘anticipatory’ response may be an expectancy consequence rather than a cause of choice.
  • Assumption 3 fails (alternatives to ‘myopia for the future’). At least five other mechanisms can account for IGT impairment: working memory (Hinson et al. 2002; Jameson et al. 2004 — though Turnbull et al. 2005 found no dual-task effect on the original task), reversal learning (Fellows & Farah 2005a rearranged the opening trials so the bad decks had no early advantage — VMPFC patients then performed identically to controls), risk-taking (Sanfey et al. 2003), insensitivity to reward/punishment feedback (Suzuki et al. 2003 found feedback SCRs, not anticipatory ones, tracked learning), and apathy (Fellows & Farah 2005b). Naccache et al. (2005) report patient RMB, impaired on the IGT with a largely intact affective response system — a deficit without a marker deficit.
  • The causal evidence is the weakest link, and it mostly runs the wrong way. Patients with pure autonomic failure (peripheral denervation of autonomic neurons; Heims et al. 2004, n = 6) performed better than controls on the IGT — and Critchley et al. (2003) show long-standing PAF also reduces grey matter in ACC and insula, so both the body loop and the substrate of the as-if loop are compromised, and still no deficit appears. Patients with spinal cord section at C6 showed no IGT deficit (North & O’Carroll 2001). Vagus nerve stimulation produced no overall group difference (Martin et al. 2004). The one supportive study — peripheral neuropathy, mild impairment (Bechara et al. 1998) — is an unrefereed conference abstract.
  • The unfalsifiability charge, stated outright. ‘Nearly all of the negative findings can be explained away through some aspects of peripheral feedback remaining intact in the patient groups studied (for example, the somatic system in PAF, endocrine feedback following spinal injury) or through recourse to the as-if loop. It seems virtually impossible to test the theory in a scenario where all peripheral feedback routes are disturbed and the as-if loop cannot be utilised, therefore making it difficult to disprove.’
  • Control performance is far more variable than the headline result implies. Bechara & Damasio (2002) report ~20% of normal adults perform disadvantageously; in two of their own studies 31% and 37% of controls performed in the range of VMPFC patients. Within that poorly-performing subgroup, anticipatory SCR variance was high — some normal, some VMPFC-like. The badly-performing normals with normal SCRs were classified post hoc as ‘high-risk takers’ overriding their markers by conscious deliberation. Dunn et al. note this raises an ecological-validity problem: if controls can fail badly and presumably do not show gross real-life decision impairment, it is unclear what the task measures.
  • Task design problems, itemized: deck position is not counterbalanced (choices may reflect location bias); reward magnitude is fixed at 50 or 100 so only punishment varies — the task measures punishment sensitivity, not reinforcement generally (consistent with neuroticism positively correlating with IGT success; Carter & Smith-Pasqualini 2004); decks are classified by eventual rather than running net outcome, so early bad-deck sampling is rational exploration miscoded as impairment; the 80-card deck limit penalizes early learners in the final block; the psychophysiological ANOVA has grossly unequal cell counts (many advantageous selections, few disadvantageous), and elevated SCRs to rarely-chosen bad decks may be a novelty/habituation artefact.
  • The neural substrate holds up better, with two revisions. Lesions to VMPFC, amygdala, insula and somatosensory cortex do impair the IGT. But (a) the effect is right-lateralized — Tranel et al. (2002) found bilateral and right-only VM lesions impair while left-only performance was low-normal; Clark et al. (2003), controlling laterality, found deficit severity correlated with right- but not left-lesion size, implicating right middle and superior frontal gyri beyond Damasio’s VMPFC; and (b) DLPFC matters, which the framework can absorb (markers allocate attentional resources) at the cost of making its predictions hard to distinguish from non-somatic theories.
  • Lesion and imaging evidence against VMPFC centrality. Manes et al. (2002) found patients with discrete OFC lesions (including VMPFC) performed in the normal range; only large PFC lesions were impaired — though laterality confounds the comparison. Fukui et al. (2005), using event-related fMRI to isolate the risk-anticipation period, found risk anticipation activated the medial frontal gyrus and not orbitofrontal regions. Dunn et al. fairly note OFC signal dropout in MRI as a partial mitigation.
  • Psychopharmacology is mixed to negative. O’Carroll & Papps (2003): reboxetine at 4 mg and 8 mg vs. placebo (30 participants) produced no difference in IGT performance, against the framework’s claim that noradrenaline mediates somatic biasing. The supportive DA/5-HT dissociation (Bechara et al. 2001) is a Society for Neuroscience abstract, and its covert/overt interpretation depends on the very cognitive impenetrability the paper has just dismantled. Dunn et al. add that the dopamine result is equally consistent with DA mediating exploratory behaviour (Montague et al. 2004).
  • The novelty critique, and its best find: Nauta (1971). Dunn et al. argue the SMH’s core idea has a long pedigree — Mowrer’s (1947) two-factor learning theory, Marston (1928) and James (1884, note 4) for the as-if idea, Pribram (1970) on ‘feelings as monitors’, Dienstbier’s false-feedback work, and in parallel Loewenstein & Lerner’s affect-in-decision literature and Epstein’s CEST. The closest antecedent is Nauta (1971), who proposed that frontal damage produces an ‘interoceptive agnosia’ — an impairment in integrating information from the internal milieu with neocortical reports of the environment — and that interoceptive information establishes ‘a temporal sequence of affective reference points serving as navigational markers.’ Twenty-plus years before the SMH, and ‘remarkably similar conclusions.’
  • The parsimony critique (Rolls 1996): if a central appraisal must already ‘know’ a bad decision is coming in order to generate the SCR, routing the signal through the body (or a simulation of it) is a circuitous way to reach motor output. Reinforcement-learning algorithms solve n-armed bandit problems — of which the IGT is a variant — under uncertainty, are well specified anatomically, and need none of the somatic machinery. Pointedly: Oya et al. (2005), from the Iowa laboratory itself, model IGT data with an RL algorithm making no reference to the SMH.
  • Specification. How a high-dimensional body state (viscera, vascular bed, skeletomotor, endocrine) is reduced to a single good/bad marker is ‘unspecified in the SMH.’ Dunn et al. suggest the framework needs an explicit data-reduction algorithm (they float PCA, and Bar-Gad et al.’s 2003 model of basal ganglia dimensionality reduction as a template), and point to Craig (2002) as the emerging map of how body signals actually reach the brain.
  • What they credit. The SMH integrates a century of scattered claims into one framework; extends bodily states beyond visceral feedback to hormonal routes; specifies the as-if loop far more clearly than James or Marston; specifies a neural substrate where earlier models (ICS, SPAARS, Lambie & Marcel) did not; makes emotion-based biasing operate at both conscious and non-conscious levels; and — via the IGT — made the ideas empirically testable at all. The as-if loop specifically rescues the framework from the standard Cannon (1927) objections to Jamesian models.
  • Their recommended tests explicitly name this wiki’s question. Among proposed designs: causal manipulation in facial paralysis (Keillor et al. 2002); the peripherally-acting beta-blocker nadolol in healthy volunteers; combining causal methods (peripheral blockade in PAF patients) to close off residual feedback routes; false-feedback paradigms in patients with disturbed bodily feedback; IGT variants removing the reversal confound and using artificial-grammar structures to restore impenetrability; multi-channel peripheral recording (facial EMG, ECG) instead of SCR alone. And: ‘an individual differences approach could be adopted, for example looking at how variance in interoceptive ability relates to IGT performance’ (citing Katkin et al. 2001).

The source Does somatic feedback guide decisions? has been waiting for. That page has said, through three ingests, that the wiki held the proponents’ case first-hand and the critics’ case only secondhand — via Friedman’s (2010) one-line mention that Maia & McClelland and Dunn et al. contest the hypothesis. This is Dunn et al., read first-hand. See Barnaby D. Dunn.

It does not settle the debate. It does something more useful: it separates the debate into parts and reports a different verdict for each. The neural substrate survives. The psychological mechanism does not, on this evidence. And the evidential base — the Iowa Gambling Task — is judged ‘no longer sufficient to be a major source of evidence for the SMH.‘

What kind of paper this is

Not a refutation, and the authors are careful about it. The closing position: ‘none of these reservations falsify the SMH; they just suggest that other sources of evidence need to be gathered.’ They call the framework ‘intriguing,’ ‘elegant,’ and ‘a valuable contribution,’ and credit it with reintroducing emotion-as-benefit to neuroscience, supplying a plausible substrate, and generating a paradigm that produced the ~35-study clinical literature their Table 1 catalogues.

The argument is that the hypothesis is underdetermined by its own evidence, and the reason to read it here is that it says so with the evidence laid out.

The three assumptions

The paper’s spine. For the IGT to support the hypothesis, three things must be true. Dunn et al. take each in turn.

assumptionstatus after this paper
the reward/punishment schedule is cognitively impenetrable, so early learning is non-consciousfails — subjects know more, earlier, than the Iowa probes detected (Maia & McClelland 2004; Bowman et al. 2005)
the learning proceeds via anticipatory somatic markersfails three ways — SCR may index variance not value (Tomb et al. 2002); is absent in subjects who learn anyway (Crone et al. 2004); may be post-decisional (Amiez et al. 2003)
the deficit is ‘myopia for the future’fails — at least five alternatives account for it, and one (reversal learning) abolishes the group difference when controlled

The one that should change how this wiki reads the pre-hunch result

Iowa Gambling Task records the knowledge/performance dissociation as ‘the result the hypothesis rests on.’ Maia & McClelland’s attack is aimed precisely there, and its design deserves recording because the wiki has only ever described it at second hand.

The Iowa probe (Bechara et al. 1997) asks broad, open-ended questions every ten cards. Maia & McClelland argued this fails Shanks & St. John’s (1994) exhaustiveness criterion — it does not index all the conscious knowledge a subject holds. With focused questions after each block of 20, advantageous performance was ‘nearly always accompanied by verbal reports of reasonably accurate quantitative and qualitative knowledge about the outcomes of the decks that was sufficient to guide behaviour.’

The finding inside the finding is the sharper one: subjects reported accurate knowledge more reliably than they chose advantageously. That is the same pattern the Iowa lab reports for its VMPFC patients — knowing the right thing, doing the wrong thing — appearing in healthy controls. If knowing-without-doing is the normal condition, it stops being diagnostic of a missing marker.

And the obvious objection was pre-empted: a second group of 20 answered Bechara et al.’s original questions and behaved identically, so the detailed probing did not construct the insight it detected.

Bechara et al.’s reply, and why Dunn et al. call it a retreat. From the 2005 TICS exchange: ‘The central feature of the SMH is not that non-conscious biases accomplish decisions in the absence of conscious knowledge, but rather that emotion-related signals assist cognitive processes even when they are non-conscious.’ Compare Tranel et al. (1999): ‘without the help of such biases, overt knowledge may be insufficient to ensure advantageous behaviour… the autonomic responses detected in our experiment (especially those evident in the pre-hunch period) are evidence for a non-conscious signalling process.’

The retreat matters for Does somatic feedback guide decisions?, which now tracks three proponent positions rather than two, and they are not the same theory.

What the anticipatory SCR might actually be

Skin conductance response (SCR) already records that SCR indexes arousal from any source and cannot report valence. Tomb et al. (2002) turn that from a caveat into a result. They rebuilt the task so the advantageous decks carried the larger rewards and punishments. Subjects still learned to pick the good decks — and their anticipatory SCRs were now larger before the good decks. The SCR tracked magnitude and variance, not long-run goodness.

Damasio et al. (2002) replied that markers record long-term positive consequences too, and that the larger SCR to good decks may be a non-conscious danger signal overridden by conscious assessment. Dunn et al. do not need to refute this to make their point: if the SCR is larger before whichever decks are more variable, then the original finding — larger SCRs before the bad decks — is explained without any reference to somatic marking, because in the original design the bad decks were also the high-variance ones. The two accounts were confounded from the start and nobody noticed until 2002.

Crone et al. (2004) supply the necessity failure independently, and it is the cleanest single datum in the paper: of 96 students split by performance, the moderate performers acquired the task while showing no anticipatory SCR or heart-rate differentiation between decks at all. Markers are not necessary to learn the IGT. (The available defence — that unmeasured channels like facial EMG carried the signal — is the escape hatch again, and Dunn et al. name it as such.)

The causal section, which is the one this wiki should care about most

Everything above concerns whether the correlation means what Iowa says. This section asks the question Body loop and as-if body loop calls the framework’s most useful open empirical question, and reports that it has been asked, four times, and mostly answered no.

manipulationpredictionresult
peripheral neuropathy (Bechara et al. 1998, n = 20)impairedmildly impaired — but an unrefereed conference abstract
pure autonomic failure (Heims et al. 2004, n = 6)impairedbetter than controls
spinal cord section at C6 (North & O’Carroll 2001)impairedno deficit
vagus nerve stimulation (Martin et al. 2004)improvedno overall group difference

The PAF result is the load-bearing one. Pure autonomic failure denervates autonomic neurons peripherally — the body loop is cut. Patients did not merely fail to show a deficit; they outperformed controls, which is why Dunn et al. can rule out low power as the explanation. And Critchley et al.’s (2003) VBM finding that long-standing PAF also reduces grey matter in ACC and insula closes the obvious escape: these patients have a compromised body loop and a compromised substrate for the as-if loop, and still decide well.

This is Does somatic feedback guide decisions?’s ‘Everitt’s experiment’ item, and the wiki must now revise it. That page and Body loop and as-if body loop both record Everitt asking Damasio in 1996 whether manipulating the periphery would change gambling behaviour, and both say the experiment is ‘thirty years old and still unrun.’ It is not unrun. It was run — by the routes Damasio himself named in his reply, cord damage and peripheral neuropathy, plus a route he did not (PAF) — and the answer favours Damasio’s own 1996 prediction of relatively little impact over the 2005 indexing claim that ambiguity engages the body. The gambling task is the paradigm case of ambiguity; the 2005 framework says the body loop should be maximally engaged there; the patients without a body loop are unimpaired.

And then Dunn et al. state the wiki’s own escape-hatch analysis, in 2006

Does somatic feedback guide decisions? contains a table headed ‘The framework’s own escape hatches, listed,’ assembled from reading Bechara & Damasio and Damasio 1996 directly, and concludes that without the loop-indexing constraint ‘a bodily theory of decision that does not need the body has given the game away.’

Dunn et al. got there first, from the empirical side:

Nearly all of the negative findings can be explained away through some aspects of peripheral feedback remaining intact in the patient groups studied… or through recourse to the as-if loop. It seems virtually impossible to test the theory in a scenario where all peripheral feedback routes are disturbed and the as-if loop cannot be utilised, therefore making it difficult to disprove.

That is independent arrival at the same structural complaint, twenty years earlier, by authors who had read more of the primary literature than this wiki has. It is corroboration, and the wiki should record it as such rather than continuing to present the analysis as its own reading.

Nauta (1971), and a citation worth checking

The novelty section produces the paper’s best archival find, and it lands directly on this wiki’s subject.

Nauta (1971) argued that the frontal lobes integrate adaptive behaviour using feedback from the periphery, and that frontal damage produces — his phrase — an ‘interoceptive agnosia’: ‘an impairment of the subject’s ability to integrate certain informations from his internal milieu with the environmental reports provided by neocortical processing mechanisms.’ He further proposed that comparing strategic alternatives is ‘in the final analysis… one between the affective responses evoked by each of the various alternatives,’ and that interoceptive information establishes ‘a temporal sequence of affective reference points serving as navigational markers.’

Navigational markers. Interoceptive agnosia. In 1971, from anatomy alone.

The pointed part is one the wiki can verify against its own pages. Ventromedial prefrontal cortex (VM) records that Damasio (1996) cites Nauta (1971) as anatomical warrant — the VM cortices are ‘the only known frontal source of projections to central autonomic control structures.’ So Damasio cites Nauta for the wiring and not for the theory built on that wiring, which Dunn et al. show was already most of the somatic marker hypothesis. Whether that is an oversight or a judgement about priority is not something these sources settle, and the wiki should not adjudicate it. But the citation asymmetry is real and checkable, and it is on the record in both papers.

Dunn et al. are even-handed about what this proves: resonance with a century of theory ‘is both a strength and weakness; it gives the model strong concurrent validity… but at the same time it challenges the novelty of the framework.’ Their conclusion is that the SMH does make a novel contribution — by integrating, by specifying the neural substrate, and by building a task.

Where this leaves the neural substrate

The half that survives, with revisions the wiki should absorb into Ventromedial prefrontal cortex (VM):

  • Right-lateralized. Tranel et al. (2002): bilateral and right-only VM lesions impair everyday decision-making and the IGT; left-only performance is low-normal. Clark et al. (2003), controlling for the laterality confound: deficit severity correlates with right- but not left-lesion volume, and implicates right middle and superior frontal gyri — territory outside Damasio’s VMPFC.
  • DLPFC is involved, and the framework can absorb it (markers allocate working-memory resources) only by making its predictions harder to distinguish from non-somatic accounts. Dunn et al. name that cost explicitly.
  • Two findings cut against VMPFC centrality. Manes et al. (2002) found discrete OFC lesions performed in the normal range while only large PFC lesions were impaired. Fukui et al. (2005), using fMRI’s temporal resolution to isolate the risk-anticipation window, found medial frontal gyrus activation and, ‘surprisingly,’ no orbitofrontal difference. Dunn et al. fairly note OFC signal dropout as partial mitigation.

The wiki’s Ventromedial prefrontal cortex (VM) page separately flags that the region’s posterior→anterior gradients are ‘not dissociated’ and that ‘VM’, ‘VMPFC’ and ‘OFC’ do not pick out identical territory. Dunn et al. corroborate the naming problem from the lesion side: Damasio’s VMPFC spans BA 10, 11, 12, lower 24, 25 and 32; others use ‘OFC’ for overlapping-but-different territory; and Öngur & Price (2000) argue the region contains two distinct networks (an orbital sensory-integration system and a medial visceromotor one). Dunn et al. adopt Damasio’s term while noting ‘the lack of specificity of this term.’

That last point connects to something neither source makes: Öngur & Price’s medial visceromotor network is close kin to the Visceromotor areas (VMAs) Seth & Friston place atop the interoceptive hierarchy. The wiki’s conjecture on Ventromedial prefrontal cortex (VM) — that triggering a somatic state from a thought and issuing a descending interoceptive prediction may be one operation in two vocabularies — now has a third source using the same anatomical carve-up. Still a conjecture; better furnished.

Two corrections this paper forces on wiki pages

1. The Monterosso claim on Iowa Gambling Task is overstated. That page lists as a strength that the IGT ‘correlates with the “gamble” task and delay-discounting tasks (Monterosso et al. 2001), supporting a common VM-linked decision mechanism.’ The claim came from Bechara & Damasio’s framing. Dunn et al. report the study directly: in 32 cocaine-dependent patients, IGT performance correlated with delayed discounting, but the link to the Cambridge Gambling Task was ‘less clear-cut, with reaction times but not behavioural choices on the CGT correlating with IGT performance.’ So the convergent validity is partial and one-sided, and the sample is a clinical group whose relationships may not generalize — Dunn et al. say so. Corrected on that page.

2. Deck B is discussed after all — just not by the Iowa sources. Iowa Gambling Task flags the deck B problem as ‘not discussed in this source, flagged here.’ Dunn et al. supply the sourcing: Wilder et al. (1998) and Shurman et al. (2005) both found control groups preferring the infrequent-punishment decks (B and D) over the frequent-punishment ones (A and C), and Dunn et al. tie this to Tomb et al.’s variance interpretation. The wiki’s flag was correct and can now be cited.

Why this matters beyond the somatic marker literature

The paper’s proposed remedies read like a research programme this wiki has been assembling backwards. Dunn et al. want multi-channel peripheral recording instead of SCR alone; false-feedback designs; causal manipulations that actually close the peripheral routes (nadolol; facial paralysis; peripheral blockade within PAF); and — the sentence this wiki should note — ‘an individual differences approach… looking at how variance in interoceptive ability relates to IGT performance,’ citing Katkin et al. (2001).

That is Does somatic feedback guide decisions?’s question 3, named as an open problem by the critic, in 2006. The wiki’s next two queued sources on this thread — Werner et al. (2009) on cardiac perception and decision-making, and Dunn et al. (2010) on interoception and intuitive decision-making — are that programme being carried out, the second by this paper’s first author. Read in order, Bechara & Damasio (2005) → this → Dunn (2010) is one researcher identifying the gap in someone else’s framework and then going to measure it.

Graph View

  • Studies
  • Concepts
  • Methods
  • Debates
  • Gaps
  • Researchers, other
  • Studies
  • Concepts
  • Methods
  • Debates
  • Gaps
  • Researchers, other

Backlinks