Contribution
Before this study, two studies had tested avoidance learning in anxiety patients (Gillan et al. 2014 on OCD; Pittig et al. 2021, a mixed sample). Both looked at whether patients go on avoiding once avoidance is pointless or costly. This one looks at how avoidance is first learned. It tracks three things trial by trial: when patients avoid, whether they come to expect that their avoidance works, and how much relief they report when the aversive picture does not come. Patients and controls did not differ in avoiding the danger cues. They differed in avoiding the safety cue, in residual threat expectancy after effective avoidance, and in relief.
Key insight
How often someone avoids does not show what they learned from avoiding. Patients and controls pressed the button equally often to the avoidable danger cue, but patients (OCD and PTSD) still expected the picture more after pressing it. On the authors’ reading, they learned to avoid without learning as well that avoiding makes them safe. Each omission then stays surprising and relieving. In the authors’ words, the relief “reinforces the avoidance action further, thereby leading to higher levels of avoidance, but apparently not by strengthening the action → safety association (which is weakened in the patients)” (p. 7).
Assumptions
- Expectancy as action-safety learning. A high rated expectancy of the picture after an avoidance response is read as a weak action → safety association, and a low one as an acquired association (p. 2).
- Relief as surprise. Relief ratings after omissions are taken to gauge the prediction-error (surprise) signal that computational models make the reinforcer of avoidance (p. 2, after Moutoussis et al. 2008 and Vervliet et al. 2017).
- The task. Three lamp colours in an office picture. Two (CS+) were followed by one of 12 mildly aversive IAPS pictures, the third (CS−) never. Pressing a red button within 2 s cancelled the picture after one CS+ (CS+avoidable) but not the other (CS+unavoidable). The Pavlovian phase had 2 trials per CS; the avoidance phase had 12 per CS in two blocks, with two unannounced trials between the blocks in which the button did not work (p. 3). Participants were told which colours could be followed by the picture “in order to speed up learning” (p. 3).
- No cost to avoiding. Pressing the button cost nothing (p. 8).
- Sample. Patients were recruited at an outpatient clinic before intensive exposure treatment, with diagnoses made by the M.I.N.I. Comorbidity overlapped: 19 of the 66 PTSD patients also had PDA, OCD or both (p. 3). “The specific size per group was a convenience size”, and “The study was not pre-registered” (p. 3). Controls were tested at home by a research assistant, patients at the clinic, after two other tasks (p. 3).
- Analysis. Repeated-measures ANOVAs with age, gender and education as covariates, repeated with rated picture unpleasantness added, and Bonferroni-corrected pairwise comparisons (p. 4).
Key results
- Pavlovian phase (p. 5). All groups expected the picture more after the CS+‘s than after the CS−, with no group difference and no Group × CS interaction. But relief to the CS−, before any avoidance was available, already differed by group (F(3, 177) = 15.95, eta = 0.21): PDA and PTSD above controls, and PTSD above PDA and OCD (p. 5; Fig. 2B).
- Avoidance (p. 5; Fig. 3A). A Group × CS interaction (F(4.15, 244.72) = 4.90, p = .001, eta = 0.08), driven by OCD (p = .004) and PTSD (p = .028) patients avoiding the CS− more than controls. “No significant differences between groups were found in reaction to the CS+ ’s” (p. 5). PDA patients did not differ from controls on any CS. The interaction held with rated unpleasantness added.
- Expectancy (p. 6; Fig. 3B). Group × CS interaction (F(5.43, 320.53) = 3.50, p = .003): OCD and PTSD patients expected the picture more than controls after the CS− and the CS+avoidable, but not after the CS+unavoidable. On trials where the button was pressed during the CS+avoidable, OCD (p = .045) and PTSD (p < .001) patients expected the picture more than controls. Both effects held with unpleasantness added, which also had a main effect of its own.
- Relief (p. 7; Fig. 3C). A main effect of group (F(3, 141) = 28.71, eta = 0.38) with no Group × CS interaction: PDA and PTSD patients above controls, PTSD above PDA and OCD, “especially during the first half of the task” (Group × Block, p = .007). OCD patients did not differ from controls. Unpleasantness had a large main effect (eta = 0.16) and left the group effect intact (eta = 0.28).
- Aversiveness (Table 1). Patients rated the pictures as more unpleasant (χ²(3) = 28.07, p < .001); PTSD patients rated them higher than controls and OCD patients.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | OCD and PTSD patients avoid a safety cue more than healthy controls in this task | moderate: one cross-sectional sample, post-hoc contrasts | p. 5; Fig. 3A |
| C2 | Patients and controls avoid the danger cues equally, avoidable and unavoidable alike | moderate: a null, with no cost to avoiding | p. 5 |
| C3 | OCD and PTSD patients keep a higher expectancy of the aversive outcome after effective avoidance | moderate | p. 6; Fig. 3B |
| C4 | PDA and PTSD patients report more relief at omitted aversive outcomes, controlling for rated aversiveness | moderate: relief was already higher before avoidance was possible | pp. 5, 7 |
| C5 | Low confidence in avoidance makes omissions surprising, and the extra relief reinforces avoidance without strengthening action → safety learning | weak: the authors’ interpretation, not tested | p. 7 |
| C6 | Elevated relief may build a habitual CS → action association | speculative: called “untested” by the authors | p. 7 |
| C7 | Avoiding the safety cue reflects behaviour-as-information, the act of avoiding giving the cue a threat meaning | speculative: offered as “A remaining possibility” | p. 7 |
Concepts
- Action → safety association: learning that an executed avoidance action is followed by the omission of the aversive outcome, indexed by low expectancy after the action (p. 2).
- Relief: rated pleasant feeling at an omitted expected aversive outcome, taken to track surprise (p. 2).
- Unnecessary avoidance: avoidance to the CS−, which no outcome follows anyway. Ineffective avoidance: avoidance to the CS+unavoidable, which the action cannot prevent (p. 2).
Connections
The task is Vervliet and Indekeu’s (2015; Vervliet, Lange & Milad 2017). Relief as a reinforcer comes from the same group (Vervliet et al. 2017; Willems & Vervliet 2021). The computational frame is Moutoussis et al.’s temporal-difference account. The habit proposal follows LeDoux and Daw (2018) and LeDoux et al. (2017). The study cites Craske et al. 2014 (LIT-1500) for the translation of fear learning into exposure, and Gangemi, Mancini and van den Hout 2012 (LIT-1498) for behaviour-as-information. Its co-author Vervliet is also an author of Craske et al.
Bearing on the record
- CLAIM-446 (governance-maintained anxiety). The exchange cited the
study for patients avoiding even during safety cues, with higher threat
expectancy and more relief. That is accurate, with two corrections:
each effect held for two of the three patient groups, not all three, and
the relief difference was there before any avoidance was learned. The
study bears on the claim’s clauses separately.
- “Relief reinforces the defense without being evidence of danger.” This is the closest the record has to data for the clause, and it is an interpretation. The authors propose that relief reinforces avoidance “but apparently not by strengthening the action → safety association”. Relief was rated, and reinforcement was not measured. Relief was also elevated in the Pavlovian phase, to a cue that only ever predicted nothing, so excess relief is not specific to avoidance. It goes with the person, the diagnosis or the aversiveness of the pictures. Supports the clause weakly.
- “Under avoidance … repeated safety need not correct the threat estimate.” This complicates the clause, because the study finds a different gap. The claim’s agent learns correctly that avoidance prevents harm and learns nothing about the world without it. Patients here had less confidence than controls that the action prevents harm. They went on avoiding without trusting it. Nor does the design test the claim’s identifiability point: no condition let participants find out what happens to the CS+avoidable without the button.
- “The regime can outlast a corrected belief.” Mixed. Patients expected the picture no more than controls after the CS− in the Pavlovian phase, and the authors read this as “suggesting that patients did acquire explicit knowledge that the CS- was safe” (p. 7). Yet OCD and PTSD patients avoided it more later. That looks like defence outrunning belief. But during the avoidance phase their expectancy after the CS− was also higher, so the belief moved with the avoidance. The authors’ candidate explanation is that the button, or pressing it, “conferred a threat meaning to the otherwise safe CS-” (p. 7). That is the claim’s self-evidencing loop (“if I felt compelled to protect myself, the danger must have been serious”), but offered as a possibility and not tested. On this evidence governance-maintained and belief-maintained readings cannot be told apart.
- The defeater. The study does not touch it. Nothing is intervened on, authority is not measured, and the design is cross-sectional (p. 8).
- QUESTION-045 (is a maintained disorder a protective lock-in?). A cross-sectional difference in learning, measured before treatment. It cannot say whether the learning profile maintains the disorder or follows from it, and the authors say so (p. 8).
- THEORY-285. That model’s agent stops sampling and never gets corrected. This study’s patients did sample: they made the same number of avoidance responses and saw the same outcomes. What differed is what they took from the outcomes. The two failures are different.
- THEORY-058 (disorders maintained by self-reinforcing loops). The proposed relief → avoidance loop is a candidate link of that kind. It is proposed, not shown.
- Filed from this reading: THEORY-912 (the measured group differences, without the interpretation).
- No instruction for machine-learning practice; nothing for the anthology.
Limitations
- Cross-sectional, with convenience group sizes and no pre-registration (pp. 3, 8). The authors: “We used a cross-sectional design, which does not allow testing whether the observed effects and proposed mechanism are causal to the disorder” (p. 8).
- Aversiveness is a rival explanation, and the authors raise it. “This begs the question whether the differences between the groups are the result of differences in US aversiveness (threat appraisal) rather than learning processes per se” (p. 7). Adding the rating as a covariate does not settle it, and they propose raising aversiveness in healthy people as the test (p. 8).
- No cost to avoiding. “There was no cost attached to the avoidance action” (p. 8). This may be why nobody stopped pressing to the unavoidable cue. It also means the study says nothing about avoidance at a cost, where pathology is usually located.
- Uneven testing conditions. Controls were tested at home, patients at the clinic after two other tasks. The CSs were not counterbalanced, fear was not rated and there was no physiology (pp. 3, 8).
- Overlapping groups. Comorbidity crosses the diagnostic groups, so the per-diagnosis pattern (OCD and PTSD on expectancy, PDA and PTSD on relief) should not be read as diagnosis-specific.
- Reporting. “2 sonds” for 2 seconds (p. 3); section 2.4 precedes 2.3 in the layout. The results report Group × Block for relief without saying which block was higher beyond “especially during the first half”.
Open questions
- Does the profile predict, or follow, the response to the exposure treatment the patients were about to receive? The design allowed it, and the paper does not report it.
- With a cost to avoiding, does the safety-cue avoidance persist? And does it persist when the threat expectancy is brought down to control levels? That would separate defence outrunning belief from belief driving defence.
- Is relief to the CS− in the Pavlovian phase, before any avoidance, what predicts later avoidance? If it is, the excess relief is a trait that the person brings, not something avoidance produces.