Contribution
Earlier reviews looked at safety behaviours inside exposure therapy: whether instructions to drop or keep them during exposure change its outcome (Meulders et al. 2016 found small effects either way). This is the first meta-analysis of studies that manipulate safety behaviours on their own, with no exposure: treatments that fade them, and experiments that tell people to cut them down or step them up, sometimes with no rationale given. Its inclusion criteria require studies that “directly target SB outside of the context of exposure” (Methods, full-text review). Both directions moved symptoms. Cutting safety behaviours down lowered symptoms, stepping them up raised them, and the bigger the drop in safety behaviour, the bigger the drop in symptoms.
Key insight
Safety behaviours are treated here as a manipulable cause, not only a symptom. If dropping them helps and adding them harms, in experiments where participants are not told why, they do more than accompany pathology. The literature that shows this, though, is young, small and concentrated on appearance and social anxiety in students.
Assumptions
- Safety behaviour. “overt actions and mental strategies used to prevent or reduce anxiety, avoid a potential threat, or prevent a negative outcome” (Introduction, after Salkovskis 1991). Mirror checking, body comparison, avoiding eye contact, searching symptoms online, checking and trauma-cue avoidance all count.
- Between-group effects only. Within-group effects are reported but not meta-analysed, after Cuijpers et al. 2017.
- Cohen’s d from standardized mean differences, random effects, REML meta-regression, outlier screening (Viechtbauer & Cheung 2010), funnel plots, trim-and-fill and Egger’s test (Methods).
- Search. PsycINFO, Academic Search Complete and PubMed, June 2025, updated September 2025; adults only; English; PROSPERO CRD420251063442 (Methods).
Key results
- The studies (Characteristics of Studies). 26 studies, 2,048 participants, mean age 24.90, 26.63% male, 67.30% white; 20 from the United States. Fifteen were non-treatment experiments and 11 were treatment trials. Nineteen were RCTs. By focus: eight social anxiety, seven appearance concerns, three eating disorders, three health anxiety, two broad anxiety, two obsessive-compulsive, one PTSD. Twelve recruited from universities. In nine experiments “deception” was used, and participants were not told why safety behaviours were being changed. Thirteen manipulations were daily digital reminders. “Interestingly, only five studies involved a control condition that included contact with a clinician.”
- Within-group (not meta-analysed). Reduction: d = 1.08 on safety behaviours and d = 0.99 on symptoms. Increase: d = 2.81 on safety behaviours and only d = 0.35 on symptoms.
- Reduction → safety behaviours. d = 0.75 (k = 16, I² = 86.7%), and d = 0.52 (95% CI 0.36–0.68, I² = 38.1%) with one outlier (Wilver et al. 2020) removed. Study type moderated it: 0.76 for non-treatment experiments, 0.31 for treatments. Trim-and-fill gives 0.46, and Egger’s test is not significant.
- Reduction → symptoms. d = 0.43 (95% CI 0.27–0.59, k = 16, I² = 45.8%). “In observing moderators, no variable moderated the pooled effect.” The size of the safety-behaviour effect did moderate it (B = 0.23, R² analogue 0.85), “suggesting that for every 1-point increase in SB effect size, the effect size for clinical symptoms increased by 0.23”. Egger’s test is significant (p = .012), and trim-and-fill imputes four missing studies and gives d = 0.31 (95% CI 0.12–0.51).
- Increase → safety behaviours. d = 2.96 (five effects, N = 368), heterogeneous, with no bias tests possible.
- Increase → symptoms. d = 0.65 (95% CI 0.46–0.83; five studies, seven effects, I² = 0%), with no bias tests possible.
- The authors’ cautions (Discussion). The increase studies “were examined in healthy college samples rather than symptom-elevated or clinical samples that were included in the SB decrease studies”. On expectancy: “It is possible that the effects seen in these studies may be in part due to expectancy effects”, since all 11 treatment trials gave a rationale. Demand effects and weak no-instruction controls may inflate the experimental effects. The verdict: “Taken together, the current findings allow for cautious optimism in the state of the standalone SB treatment literature.”
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Standalone reduction of safety behaviours lowers symptoms relative to control | moderate: d = 0.43, 0.31 after trim-and-fill; publication bias likely | Meta-analysis 2; k = 16 |
| C2 | Instructed increases in safety behaviours raise symptoms in non-clinical samples | moderate: d = 0.65 with no heterogeneity, but five studies, no bias test | Meta-analysis 4 |
| C3 | The larger the reduction in safety behaviour, the larger the symptom reduction | moderate: between-study meta-regression, not mediation | Meta-analysis 2 moderator |
| C4 | Experiments reduce safety behaviours more than treatments | weak: one moderator, possibly demand or weaker controls | Meta-analysis 1 moderator |
| C5 | Safety behaviours maintain psychopathology | weak as a general claim: the authors say the findings “align” with the theory | Discussion |
Concepts
- Standalone manipulation: a change to safety behaviours made outside exposure, either as a treatment with a rationale or as an experiment, sometimes deceptive (Introduction; Methods).
- Safety-behaviour fading: systematically reducing and eliminating safety behaviours, here often by daily text or app reminders (Introduction; Cougle et al. 2020).
Connections
The cognitive-behavioural tradition makes safety behaviours a maintaining factor (Salkovskis 1991; Ehlers & Clark 2000; Wells et al. 1995). The review sets itself beside Meulders et al. 2016, which meta-analysed safety-behaviour instructions within exposure, and Blakey and Abramowitz 2016, which reviewed safety behaviours from the inhibitory-learning view of Craske et al. (LIT-1500). Most of the included studies come from the senior author’s group (Cougle; Patel & Cougle; Stentz & Cougle; Summers & Cougle; Wilver et al.; Zech et al.).
Bearing on the record
- CLAIM-446 (governance-maintained anxiety). The exchange cited
it, unnamed, for the effect of reducing safety behaviours on its own,
and gave the counts and d = 0.43 correctly. Three corrections. The
bias-adjusted estimate is d = 0.31. Bias is shown for the symptom
outcome specifically (Egger p = .012), not for the review as a whole.
And none of the planned moderators moderated the symptom effect; only
the size of the safety-behaviour change did. The reading bears on the
claim as follows.
- For the claim, at its base. The claim needs protective behaviour to do causal work in keeping symptoms, not only to accompany them. The increase experiments are the best evidence the record holds that it does: in non-clinical people told to check, compare or seek reassurance more, symptoms rose (d = 0.65). The dose-response moderator points the same way for reductions.
- Silent on what is distinctive. The claim says the maintaining link is defensive authority over action and evidence, not belief. The review cannot tell those apart. No study measured threat estimates, exploration, or what evidence the participants gathered once the safety behaviour stopped. Symptoms could fall because the reduction acts as unprotected exposure that corrects belief, and that is the belief-maintained reading. It could also be because the defence loses its grip, and that is the claim’s. The claim’s four-intervention test is not approached.
- A narrowing of what “safety behaviour” covers. Most of the evidence is appearance-related checking and comparison and social safety behaviours in students. Avoidance-maintained anxiety disorders, the claim’s case, are thinly represented: one PTSD study, two checking studies and no panic or agoraphobia.
- THEORY-053 (efficacy is not evidence for a theory of change). The moderator is the kind of evidence THEORY-053 asks for, that moving the process moves the outcome, but it is between studies. No included study is reported as testing whether the safety-behaviour change precedes and mediates the symptom change. The authors flag expectancy effects as an alternative.
- THEORY-058 (self-reinforcing loops). The bidirectional result (adding raises symptoms, removing lowers them) fits a loop through behaviour. It does not show the loop closes, which would need symptoms to raise safety behaviour in turn.
- Filed from this reading: THEORY-903.
- No instruction for machine-learning practice; nothing for the anthology.
Limitations
- Small, early literature. The meta-analyses hold 16 and 5 studies. The authors call them “preliminary”.
- Publication bias and small-study effects. “Taken together, publication bias and small-study effects cannot be ruled out in the present findings” (both reduction analyses). For the increase analyses bias could not be tested.
- Samples. “most of the participants included in this review and meta-analysis were young women”. Increase studies used healthy college samples. Clinical diagnoses were rare.
- Heterogeneous outcomes. “primary symptoms” range from social anxiety to loneliness, distress tolerance, authenticity, disgust propensity and body satisfaction (Table 1), all pooled as one outcome.
- Risk of bias. 13 of 19 RCTs had “some concerns”, and all seven non-randomized studies were at moderate risk.
- One group’s work. Many included studies are from the authors’ own laboratory. The review does not discuss this.
- Reporting slips in the manuscript. The within-group symptom effect of reduction is d = 0.99 in Results and d = 1.03 in the Discussion. The increase-on-symptoms effect is printed “d = 65”. Table 1 labels van Uijen & Toffolo 2015 “decrease checking SB” but describes a checking increase. “Eger’s” appears for Egger’s. The version of record may correct these.
Open questions
- Do standalone reductions change threat estimates, and do they bring people into the situations they had avoided? Measuring both would show whether reductions work as exposure (belief correction) or by another route.
- Does the effect hold in diagnosed panic, agoraphobia and obsessive-compulsive disorder, with active controls and masked rationale?
- Within a study, does the safety-behaviour change precede and mediate the symptom change?