Iowa Gambling Task

Method at a glance

• Modality: behavioural (decision-making under ambiguity), typically paired with electrodermal recording
• Measures: The capacity to shift choice toward long-term advantage under conditions where reward/punishment probabilities are never disclosed and cannot be learned explicitly — operationalized as the net score of advantageous minus disadvantageous deck selections across 100 trials.
• Procedure: Four decks (A, B, C, D) and a loan of play money. Subjects make 100 card selections, not told in advance how many, free to switch decks at will. Decks A and B yield high immediate gain but larger future loss (long-term loss); decks C and D yield lower immediate gain but smaller future loss (long-term gain). Schedules are pre-programmed and known to the examiner but not the subject. Net score = (C+D) − (A+B), usually reported in five blocks of 20. Commonly run with concurrent skin conductance recording, dividing the inter-selection interval into a 5 s reward/punishment window after each card and an anticipatory window before the next.
• Also called: IGT, gambling task, GT, ABCD task

Strengths

• Creates genuine ambiguity rather than risk: subjects never acquire the probabilities, even after reaching explicit knowledge that some decks are bad. This is the feature that distinguishes it from Rogers’ ‘betting’ task and from most economic paradigms, where odds are stated.
• Sensitive to the lesion it was built for, and to a specific double dissociation: VM patients generate reward/punishment SCRs but no anticipatory SCRs; amygdala patients generate neither. Both fail to avoid the bad decks.
• Performance corresponds to real-life decision impairment in a patient group whose deficits are invisible to conventional neuropsychological testing (normal IQ, language, working memory, attention).
• Pairing behaviour with concurrent SCR allows the covert bias to be observed before it reaches report — the pre-hunch effect that carries the hypothesis’s central claim. But see limitations: whether the pre-hunch window is real is exactly what Dunn et al. (2006) contest.
• Genuinely generative, whatever its problems. Dunn et al. (2006) — the task’s most thorough critics — catalogue ~35 clinical studies using it before July 2005 (schizophrenia, OCD, pathological gambling, substance dependence, psychopathy, anorexia, Huntington’s, MS, TBI, ADHD, chronic pain, suicide attempters) and credit it as ‘a sensitive, ecologically valid measure of decision-making impairment.’ Behavioural findings replicate outside Iowa ‘more often than not’; the psychophysiology rarely travels with them.
• Robust to several parameter changes: real vs. facsimile money (Bowman & Turnbull 2003), different inter-trial delays (Bowman et al. 2005), manual vs. computerized administration (Bechara et al. 2000a). Shows orderly lifespan effects (improves to adulthood, declines in older age) and some predictive validity — IGT performance predicts pharmacotherapy response in OCD (Cavedini et al. 2004b).
• Partial convergent validity — corrected. This page previously claimed the IGT ‘correlates with the gamble task and delay-discounting tasks (Monterosso et al. 2001), supporting a common VM-linked decision mechanism,’ taken from Bechara & Damasio’s framing. Read directly in Dunn et al. (2006), the study is narrower: 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 correlating.’ Dunn et al.’s own reading is that this tilts the IGT toward a difficulty considering future outcomes rather than risk-taking — and that it speaks only to this clinical group, not to healthy populations. The convergence is real, one-sided, and untested in controls.

Limitations

• Ambiguity is a strength and a confound at once. Rogers argued the unspecified pay-off structure makes the underlying deficit hard to characterize — a person may choose badly because they discount long-range interests or because they never learned the contingencies, and the task cannot separate these.
• The ambiguity may not be there. The design goal is a schedule no one can consciously grasp. Maia & McClelland (2004), reported in Dunn et al. (2006): with focused rather than open-ended probes, healthy subjects hold reportable knowledge sufficient to guide choice, and hold it more reliably than they act on it. Bowman et al. (2005) find above-chance deck-goodness ratings within the first 20 trials. Dunn et al. add two design reasons the task invites explicit reasoning: outcomes are presented as explicit numbers with unlimited deliberation time, and reward magnitude never varies (always 50 or 100), so only punishment must be tracked. If the schedule is penetrable, the paradigm’s central rationale — that arithmetic is unavailable, so something else must separate good from bad performers — is gone.
• Anticipatory SCR may index variance rather than value. Tomb et al. (2002) rebuilt the task so the advantageous decks carried the larger rewards and punishments; subjects still learned the good decks but now showed larger anticipatory SCRs before the good decks. In the original design the disadvantageous decks are also the high-variance ones, so the two accounts were confounded from the outset. Damasio et al. (2002) replied that markers record positive long-term consequences too, and that the effect may be a non-conscious danger signal overridden by conscious assessment.
• Anticipatory SCR is not necessary to learn the task. Crone et al. (2004), n = 96 split into good/moderate/poor performers: only the good group showed anticipatory SCR or heart-rate differentiation between decks. The moderate group acquired the task with no psychophysiological differentiation at all. The available defence — that unmeasured channels (e.g. facial EMG) carried the signal — is unfalsifiable as stated.
• The ‘anticipatory’ window may be post-decisional. Amiez et al. (2003) found, in rhesus monkeys on an equivalent task, that SCRs track anticipation of reward after a decision has been made rather than preceding the decision. SCR’s slow time course makes ‘prior to selection’ and ‘prior to feedback’ hard to separate at all; Dunn et al. propose higher-resolution acquisition with tonic/phasic decomposition (Lim et al. 1997), or a faster-responding measure.
• Control performance is far more variable than the headline implies. ~20% of normal adults perform disadvantageously (Bechara & Damasio 2002); in two Iowa studies 31% and 37% of controls scored in the VMPFC patients’ range. Poorly-performing normals with normal anticipatory SCRs were classified as ‘high-risk takers’ overriding their markers by deliberation. This makes null group differences uninterpretable (absent deficit, or a bad control sample?) and raises an ecological-validity problem: controls who fail badly presumably do not show gross real-life decision impairment. The ‘post hoc’ half of this criticism has been softened by the primary source — see ‘The high-risk-taker label’ below: the descriptor is self-report, in principle available before the task, and it comes with a claimed psychophysiological difference rather than being a bare relabelling. What it lacks is any means of checking either.
• The working-memory alternative was measured by the Iowa group itself, and it is a partial contributor rather than a substitute explanation. Bechara, Damasio & Damasio (2000) ran nine VM patients, ten dorsolateral patients and 21 controls on this task and on two delay tasks in the same session. Working memory was normal — sometimes superior — in severely impaired VM patients; but right-dorsolateral patients with severe working-memory impairment scored low-normal rather than normal here. So the two functions are ‘asymmetrically dependent’ in the authors’ own words. This weakens working memory as a rival account of the VM deficit and simultaneously concedes that the task is not a pure measure of deciding. Nobody has reported how much of the VM deficit survives once working memory is covaried out — nine patients, no statistics.
• The deficit has at least five other explanations (Dunn et al. 2006): working memory, reversal learning, risk-taking, feedback insensitivity, and apathy. The sharpest is reversal — Fellows & Farah (2005a) rearranged the opening trials so the bad decks had no early advantage, and VMPFC patients then performed identically to controls, implying the deficit is in reversing early learning rather than in marking futures. Note this is the objection D. Bishop put to Damasio in 1996 (see below), tested nine years later and confirmed.
• The insula has been lesioned and tested, and it does not tell you what the insula contributes. Added with Jones, Ward & Critchley (2010). Bar-On et al. (2003) ran this task on patients with lesions spanning vmPFC, amygdala and insular/somatosensory cortices against lesion controls outside limbic and viscerosensory regions: the experimental group chose disadvantageously and finished at a net loss, with no difference in general cognitive performance — but the lesions were non-focal and the groups too small for site-by-site comparison, so nothing is attributable to the insula. The two studies with focal insula lesions used other paradigms and disagree in direction (Clark et al. 2008: no risk adjustment, persistent high betting on the Cambridge Gambling Task; Weller et al. 2009: significantly reduced risk-taking even when advantageous), reconciled by Jones et al. as shared insensitivity to expected value. None of the three recorded a bodily measure. So the interoceptive cortex has now been subtracted from decision-making three times and the wiki still cannot say whether what was lost was the somatic signal, the perception of it, or an expected-value computation. See Does somatic feedback guide decisions?.
• Deficits are not specific to any condition, which limits the task’s diagnostic use: most patient groups tested show impairment, and replication within a condition is patchy — of six schizophrenia studies, one found no deficit, one a disadvantageous preference, one a medication-dependent effect, one a preference for infrequent-punishment decks, and two impairment only in subtypes.
• Design flaws Dunn et al. itemize: deck position is not counterbalanced (choices may be location bias); decks are scored 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 last block; the psychophysiological ANOVA has grossly unequal cell counts (many good-deck picks, few bad-deck picks), and elevated SCRs to rarely-chosen decks may be a novelty/habituation artefact; block-of-20 net scores are an insensitive index that may mask slower learning in patients (a trend analysis against chance, not only against controls, would be better).
• The certainty→as-if / ambiguity→body loop mapping rests on comparing the IGT against the betting task, which differ in payoff structure, instructions, and provenance as well as in epistemic character. The comparison is reported without statistics in Bechara & Damasio (2005).
• Anticipatory SCR is a single sympathetic index (see Skin conductance response (SCR)); it can register that a somatic state occurred, not which one, so the task cannot address emotion-specific patterning.
• The task’s headline dissociation is between-groups on small, heterogeneous lesion samples, and Bechara & Damasio (2005) report it as figures without n’s, effect sizes, or CIs — it is a review restating earlier primary papers. Damasio (1996) supplies the missing n’s for the original comparison and they are small and badly matched: 44 normal controls (21 women, 23 men; ages 20–79) against 6 ventromedial patients (2 women, 4 men; ages 43–84). The control range extends two decades younger, the patient group is seven times smaller, no age-matched subgroup analysis is reported, and no statistics accompany the result in either paper.
• The age mismatch is not a pedantic objection here. This wiki separately records (Age-related interoceptive decline) that felt bodily emotion dampens steadily across adulthood. An age-confounded comparison of somatic responsivity is therefore confounded on precisely the variable the task is measuring. The effect is large and the patients’ real-life impairment is independently documented, so this is an argument about what the finding is worth alone — roughly what a six-patient lesion study should be worth — not an argument that it is wrong.
• Interoception is rarely measured, and never as a moderator on this task. Whether individual differences in interoceptive sensitivity moderate anticipatory SCR or net score is the obvious question the Iowa programme did not ask. Dunn et al. (2006) name this explicitly as the study to run — ‘looking at how variance in interoceptive ability relates to IGT performance’ (citing Katkin et al. 2001) — and Dunn et al. (2010) then ran it and found the moderation is real and large enough to reverse the sign of the body’s effect on choice, but on the Intuitive reasoning task (IRT), not on this task. The one study that paired this task with a heartbeat measure — Werner et al. (2009), good vs poor counters on the IGT — found a main effect (good perceivers choose more advantageously, F(1,48)=5.32) but ran it as an extreme-groups main-effect design with no per-choice bodily term, so it cannot instantiate the moderation either. So the informative omission stands unrepaired for the IGT: the paradigm that generated the entire somatic marker literature has been paired with who can feel the signal exactly once, and never with a measure that could test whether feeling it changes what the signal does.
• Deck B is a known problem in the wider literature (high frequency of gains, rare large losses, so healthy subjects often prefer it) — not discussed in the Iowa sources. Now sourceable: Wilder et al. (1998) and Shurman et al. (2005), reported in Dunn et al. (2006), both found control groups preferring the infrequent-punishment decks (B and D) over the frequent-punishment ones (A and C), which Dunn et al. tie to Tomb et al.’s variance interpretation. The task offers no objective advantage either way, so the preference is a pure frequency/magnitude effect the net score silently absorbs.

Connections

• Key studies: Bechara & Damasio (2005), Damasio (1996), Bechara et al. (2000), Dunn et al. (2006), Dunn et al. (2010), Werner et al. (2009), Jones et al. (2010)
• Related methods: Skin conductance response (SCR), Pavlovian defense conditioning, Delay tasks (delayed response and delayed non-matching to sample), physiology congruence, Heartbeat detection task, Intuitive reasoning task (IRT), Lesion-symptom mapping

The paradigm on which the Somatic marker hypothesis was tested, and the source of nearly every empirical claim in Bechara & Damasio (2005). Described in full in Bechara et al. (2000b), which is not in raw/; the account here is drawn from the 2005 review, with the original samples and two early manipulations from Damasio (1996), and the critical literature from Dunn et al. (2006).

Read the limitations block in the frontmatter before the account below. This page was built from the Iowa sources, and describes the task as they present it — a paradigm whose ambiguity forces subjects onto a non-cognitive strategy. Dunn et al. (2006) reviewed every published use of the task to July 2005 and concluded it “is no longer sufficient to be a major source of evidence for the SMH.” Their case is that the ambiguity is not as complete as claimed, the anticipatory SCR admits at least three readings, and the deficit has at least five candidate explanations. None of that stops the IGT being a useful and generative measure of decision impairment — which Dunn et al. say plainly. It stops the IGT being evidence for a specific mechanism.

What the task is for

The design goal is ambiguity, in the behavioural-economics sense (Einhorn & Hogarth 1985; Ellsberg 1961) — as distinct from certainty and from risk. Subjects never acquire the probabilities of reward and punishment, even once they know perfectly well which decks are good and bad. The level of uncertainty “remains high throughout.”

This is why the task carries the argument. Under stated odds, a decision failure can always be attributed to bad arithmetic. Under ambiguity, arithmetic is unavailable to everyone, so what separates good from bad performers has to be something else — which is the space the somatic marker is proposed to fill.

The four periods

The task’s most important result comes from stopping the game every 10 cards and asking subjects to say what they know (Bechara et al. 1997). Four periods emerge:

period	subject’s knowledge	normal anticipatory SCR	normal behaviour
pre-punishment	sampling; no punishment yet	none	prefers high-paying A/B
pre-hunch	punishment encountered, “no clue”	substantial rise	a hint of shift away from bad decks
hunch	riskier decks suspected, not certain	sustained	shift more pronounced
conceptual	knows which decks are good and bad	sustained	advantageous

The anticipatory SCR rises in the pre-hunch period — before any conscious knowledge, and alongside the first behavioural shift. VM patients never reported a hunch, never developed anticipatory SCRs, and kept choosing A and B.

The dissociation that does the work

The result the hypothesis rests on is not that VM patients do badly. It is the crossing of knowledge and performance:

• 30% of controls never reached the conceptual period — and still performed advantageously.
• 50% of VM patients did reach the conceptual period — and still performed disadvantageously.

So explicit knowledge is neither necessary nor sufficient. VM patients “may ‘say’ the right thing, but they ‘do’ the wrong thing.” Bechara & Damasio extend the dissociation to addiction (knows the consequences, takes the drug anyway) and psychopathy — an extension asserted from surface similarity, with no data, and worth resisting on this evidence.

The lesion double dissociation

group	SCR to reward/punishment	anticipatory SCR	deck choice
normal	yes	yes — larger before risky A/B	avoids A/B
VM lesion	yes (slightly lower)	no	prefers A/B
amygdala lesion	no	no	prefers A/B

This is the pattern Primary and secondary inducers predicts: amygdala damage kills primary induction and therefore starves secondary induction of stored patterns; VM damage kills only the secondary trigger. Amygdala patients “can no longer register how painful it feels when one loses money,” which “misleads” the VM cortex about how painful a loss should feel.

Manipulations reported in the 2005 review

Emotional induction (Fig. 8). Ten healthy volunteers performed the task after recalling a neutral event (mowing the lawn) and after recalling an emotional one (the death of a loved one), order counterbalanced. Emotional induction reduced selections from advantageous decks. This is the entire empirical basis for the paper’s claim that emotion unrelated to the task at hand is disruptive — n = 10, no statistics reported, a bar chart with large error bars that visibly overlap.

Pharmacology. Dopamine and serotonin manipulation produce different effects on covert vs overt decision-making (Bechara et al. 2001) — DA biasing covertly via striatum, 5-HT overtly via ACC/SMA. The citation is a Society for Neuroscience abstract.

Three things only the 2000 source records

Added with the Bechara, Damasio & Damasio (2000) ingest — the Iowa laboratory’s own mid-point review, and the most patient-focused of the three primary statements this wiki holds.

The deficit is stable across four testings. Six VM patients and five controls were retested at one month, 24 hours later, and again at six months (Fig. 3). Controls improved significantly across repetitions; the VM group did not, and their net score declined. The authors read this as the laboratory counterpart of the patients’ real-life inability to learn from previous mistakes.

Worth holding at the same discount as everything else on this page — n = 6 and n = 5, no statistics, an odd retest spacing (1 month, then 24 h, then 6 months) that the paper does not explain. But it does bear on one live objection: if the VM deficit were a reversal-learning failure that extended practice could overcome, four exposures might have shown it. They did not. That is not an answer to Fellows & Farah (2005a), whose manipulation removes the reversal requirement rather than repeating the task, but it is a datum the reversal account has to accommodate.

The high-risk-taker label. The limitations block above records, via Dunn et al., that poorly-performing normals with intact anticipatory SCRs were classified “post hoc” as high-risk takers. The primary source makes the charge partly unfair and partly worse.

Partly unfair: these are described as “normal adults who describe themselves as high-risk takers in real life” — a self-report descriptor that is in principle available before the task, not a category invented to absorb the result. And the paper reports a psychophysiological distinction, not just a label: in these individuals anticipatory SCRs to the bad decks are slightly lower than to the good decks — the bias inverted — where advantageous performers show the normal direction, and where VM patients show no anticipatory SCRs at all. The claimed contrast is between an inverted-or-overridden bias and no bias, which is a claim about data.

Partly worse: no instrument is named for the self-description, no n is given, no statistics, no criterion for “slightly lower,” and no statement of whether the descriptor was collected before or after performance. So the correct charge is that the distinction is unverifiable, not that it is post hoc — and an unverifiable distinction load-bearing enough to explain away a third of the control sample is not obviously the better position.

Ageing, and its shape. Denburg et al. (1999): adults over 64 perform poorly relative to ages 26–56 — but dichotomously, some performing very well and some very poorly, rather than the group shifting as a whole. Recorded on Age-related interoceptive decline, where this wiki’s aging thread reaches the same structure from the interoceptive side.

Two things only the 1996 source records

The reversed-design control (Anderson et al. 1996). Damasio (1996) names three candidate accounts of why VM patients take high immediate reward with severe delayed punishment — hypersensitivity to reward, insensitivity to punishment, or general insensitivity to future consequences — and reports a task variant built to separate them: punishment placed up front, with unpredictable reward schedules as the unexpected variable. The patients behaved the same way, which cuts against the first two accounts and leaves the third. This is the most useful methodological control in the Iowa programme and the 2005 review does not mention it. (Citation is a Society for Neuroscience abstract, in press at the time.)

Damasio also considers and rejects defective response inhibition as the sole source of the deficit — not on data, but on the grounds that the task’s complexity, the knowledge available to players, and the length of time over which the result holds make it implausible as a complete account. He grants there is much in animal and human work supporting the inhibition idea in general.

The conditioning dissociation. In the transcribed discussion, D. Bishop presses exactly the right objection: subjects do not merely learn a mapping, they must change one, since everyone starts on A and B because those decks pay more — so is this a failure to inhibit an earlier association rather than a failure to form associations? Damasio concedes the inhibition reading is conceivable, then reports that most patients who fail the gambling task acquire classical conditioning normally. If that holds, the gambling deficit is not a conditioning deficit and “conditioning” is not one thing. Filed also on Pavlovian defense conditioning.

His actual reply to Bishop is the more revealing part: the intriguing thing is not whether inhibition failed but that such a failure is not compensated by the patients’ own realization that their strategy is losing — reasoning unaided by the marker does not prevail in guiding behaviour. That is the 1997 knowledge/performance dissociation stated as clinical intuition a year before Bechara et al. tested it.

Bishop was right, and someone eventually ran the experiment. Dunn et al. (2006) report Fellows & Farah (2005a), who did to the task exactly what Bishop’s objection implies: they rearranged the opening trials so the two disadvantageous decks no longer carried an early advantage, removing the reversal requirement while leaving everything else intact. VMPFC patients then performed the same as controls.

That is the strongest single result in the critical literature, and it lands on a specific claim. Damasio dismissed the inhibition account in 1996 not on data but on plausibility — the task’s complexity, the knowledge available to players, the duration over which the deficit holds. Nine years later the discriminating manipulation was run and the deficit disappeared when the reversal did. The conditioning dissociation above still stands (these patients do condition normally), so the honest statement is that “reversal” and “classical conditioning” are not the same capacity and the gambling deficit tracks the first.

Bechara et al. (2005) reply that the task contains more than contingency-reversal, and Turnbull et al. (in press, 2005) offer partial support — schizophrenic patients with negative symptoms acquired the standard task but failed added shift phases, suggesting the two are dissociable. Dunn et al. also note the framework can absorb reversal by claiming it needs an emotion-based “stop signal” — and then close that door: reversal learning survives amygdala lesions (Izquierdo et al. 2004), so it does not depend on emotion-processing structures.

Its replacement, built by its critics

Dunn et al. (2006) catalogued this task’s design flaws more thoroughly than anyone. Four years later Dunn et al. (2010) built the Intuitive reasoning task (IRT) — four decks, 100 trials, real money — that repairs them one by one: the reversal confound removed, deck position and card backs counterbalanced, magnitude crossed with profitability in a 2 × 2, the deck-exhaustion limit lifted, and the pre-decision window separated from the pre-feedback window.

Two results from that task bear directly on this page.

The first is a vindication. Crossing magnitude with profitability tests Tomb et al.’s (2002) variance interpretation — the objection in the limitations above that the anticipatory response may index reward variance rather than long-run goodness, since in this task the bad decks are also the high-variance ones. In the IRT they are not, and anticipatory bodily responses followed profitability with no magnitude effect and no interaction (Fs < 1). The body tracks goodness. That is the best available answer to Tomb et al., and it arrives from the critic’s side.

The second is a caution. 27% of a healthy sample still preferred the unprofitable decks (M = 25.74, SD = 40.71) — the same order as the 20–37% of IGT controls who perform disadvantageously, which this page flags as an ecological-validity problem. Fixing every design flaw did not fix that. Either a large minority of healthy adults really do decide badly under ambiguity, or four-deck card games measure something other than what their names claim, and no source in this wiki distinguishes the two.

What the IRT does not do is rescue this task’s data. Its clean answer to Tomb et al. is an answer about the IRT. In the IGT the magnitude/profitability confound remains exactly as Tomb et al. described it, and every published IGT result is downstream of it.

What the task cannot tell you

Two limits worth holding, both structural rather than fixable:

1. It measures that a somatic state occurred, not which one. SCR is one sympathetic channel. Every claim in the framework about positive versus negative somatic states — including the whole Background somatic states signal-to-noise model — is therefore unsupported by IGT data and imported from elsewhere. See Do basic emotions have distinct autonomic signatures?.
2. It rarely touches Interoception, and never as the wiki needs. The task shows a bodily signal precedes and predicts good choices. It does not show that anyone perceives it. Pairing a decision task with the Heartbeat detection task is what Dunn et al. (2010) did — and the answer matters: interoceptive accuracy has no relationship to decision quality (r = .08) but reverses the sign of the bodily signal’s effect depending on whether the body is right. A moderator that large, left unmeasured, sits in the error term of every result on this page. Werner et al. (2009) did pair this task with heartbeat counting and found good perceivers choose better — but as an extreme-groups main effect with the marker unmeasured, which is the Pollatos-style design the wiki discounts, not the moderation. So because Dunn measured the interaction against the Intuitive reasoning task (IRT) instead, and Werner measured only a main effect on the IGT, the moderator sits in this task’s error term still.