Evidence review

Rejection: mechanism, evidence, and what follows

What the research actually establishes about social rejection — including which famous findings were reversed, corrected, or shrank to nothing.

31 July 2026  ·  ~10,000 words  ·  every claim labelled HIGH MEDIUM LOW UNVERIFIED

1. The mechanism

The finding everyone knows, and the state it is now in

Eisenberger, Lieberman and Williams reported in Science in 2003 that being excluded from a ball-tossing computer game activated dorsal anterior cingulate cortex, and that dACC activity correlated with self-reported distress at r = 0.88. The paper has roughly 6,900 citations and is the origin of the claim that social pain and physical pain share neural machinery.

The design facts are in the paper itself: thirteen participants, an uncorrected threshold of p = 0.005 with a ten-voxel extent, condition order never counterbalanced, one exclusion episode. The r = 0.88 was computed at voxels selected by a whole-brain search for that very relationship, which is the double-dipping problem; with N = 13 a brain–behaviour correlation of 0.88 exceeds what the joint reliability of fMRI and self-report plausibly allows. Anterior insula also activated and showed no relationship to distress — a detail that vanished from the downstream literature, which routinely cites "dACC and anterior insula" as the social-pain pair. HIGH on the design facts, read from the primary paper; MEDIUM on the inflation inference.

What has since happened to it, in order:

Iannetti and Mouraux argued the "pain matrix" is a multimodal salience network that responds to any behaviourally significant stimulus, so overlap proves nothing. Lieberman and Eisenberger escalated in 2015 with a PNAS paper titled "The dorsal anterior cingulate cortex is selective for pain," using Neurosynth. Wager and nine co-authors (including Poldrack, Botvinick, Iannetti and Yarkoni, who built Neurosynth) replied in the same journal: they hand-classified the first 240 of 647 studies activating within 8 mm of the claimed pain-selective coordinate, and the probability that such a study was about pain was 12%. Fifty per cent were cognition studies. They also noted that the selectivity paper had never computed posterior probabilities — it compared Z-scores across keywords, which is a different operation. The rebuttal has roughly 220 citations to the reply's 24, both published in the same journal in the same month. HIGH

Two empirical results are more damaging than the methodological ones:

The inferential gap, stated plainly. "The regions overlap" is a claim about spatial coincidence at ~3 mm resolution, averaging over hundreds of millions of neurons. "Social pain is physical pain" is a claim about shared representational content. The second was tested directly and failed. Coarse co-activation is exactly what a region implementing a domain-general function — salience, self-relevance, aversive monitoring — would produce.

Acetaminophen. DeWall et al. 2010 (N = 62 behavioural, N = 25 fMRI, not preregistered) reported that three weeks of paracetamol reduced hurt feelings. No adequately powered preregistered direct replication exists. The nearest test, Hofman, Wieser and van der Veen 2021 (N = 72), is titled for its result: acetaminophen did not affect cardiac or brain responses to social rejection. What the literature grew into instead is valence-general affective blunting — reduced empathy for others' pain (Mischkowski 2016), reduced empathy for others' pleasure and reduced personal pleasure (Mischkowski 2019), blunted evaluative processing regardless of valence (Durso 2015), increased risk-taking on the BART across three double-blind studies, N = 545 (Keaveney 2020). A drug that flattens everything is a poor fit for a pain-specific mechanism. MEDIUM-HIGH on the blunting; MEDIUM on the null verdict for the rejection claim, since absence of replication attempts is weaker evidence than a failed one.

What is solid

The behavioural effect is enormous and unusually invariant. Hartgerink, van Beest, Wicherts and Williams 2015 (PLoS ONE, 120 Cyberball studies, N = 11,869, preregistered — and note Williams himself is an author, which makes the negative findings more credible, not less):

HIGH

It survives conditions designed to break it. Participants told they were playing a computer were still distressed (Zadro, Williams & Richardson 2004). Being excluded while it was financially profitable to be excluded still hurt (van Beest & Williams 2006). Being excluded by people identified as the Australian KKK still hurt — and that one replicated directly (Gonsalkorale & Williams 2007; Fayant, Muller, Hartgerink & Lantian 2014). HIGH

One boundary condition matters and is under-cited: Law & Williams 2006 found that stripping human characteristics from the game and giving no instruction to visualise produced no distress at all. The pain requires an inferred mind on the other side. MEDIUM

Cortisol is the most methodologically sound item in the whole package. Dickerson & Kemeny 2004 (Psychological Bulletin, 208 laboratory studies): overall cortisol effect d = 0.31, but partitioned — tasks with social-evaluative threat d = 0.67 [0.50, 0.84] versus d = 0.15 [0.02, 0.27] without; Trier-type designs d = 0.87. Emotion induction produced nothing (d = −0.13). The active ingredient is being evaluated by others combined with uncontrollability, not aversiveness in general. HIGH

Opioids implicate a different circuit than the fMRI story. Hsu et al. 2013 (μ-opioid PET, n = 18) found rejection-induced opioid release in ventral striatum, amygdala, midline thalamus and periaqueductal grey — notably not dACC or anterior insula. Unreplicated. The strongest leg here is old and comparative: Panksepp & Herman 1978 showed low-dose morphine sharply reduces separation-distress vocalisations in puppies, guinea pigs and chicks, with naloxone increasing them. That supports "mammals have an opioid-modulated separation-distress system" — not "human social pain is repurposed physical pain." MEDIUM on the human PET; HIGH on the animal phenomenon.

Inflammation is weaker than its citation count implies. Slavich, Way, Eisenberger & Cole 2010 (PNAS, N = 124) found sTNFαRII and IL-6 increases to social-evaluative stress; but the brain–immune coupling comes from an fMRI subsample of n = 31 and holds for sTNFαRII and not IL-6. [Inflammatory response to social-evaluative stress: HIGH. dACC–inflammation coupling: LOW-MEDIUM.]

The stage model is not supported in the form it is usually cited

Williams' temporal need-threat model posits reflexive → reflective → resignation stages, with the reflexive stage predicted to be impervious to moderation. Hartgerink et al. found the opposite pattern on both halves: moderation on the first measure was significant (Δd = −0.46, CI [−0.64, −0.28], k = 52), on the last measure it was not (Δd = −0.20, p = .052). The authors' own words: "support for this proposition is tenuous." HIGH

The resignation stage — the one that describes actual chronic rejection — rests on Zadro's interviews with 28 self-selected long-term targets of the silent treatment, plus two theoretical borrowings. Williams attaches the disqualifying caveat himself: they "sought to be part of the study… Cause and effect are impossible to determine." Anyone invoking learned helplessness and unworthiness as an established consequence of repeated rejection is citing a hypothesis dressed as a stage. HIGH

Two distinctions the literature usually collapses

Rejected versus ignored. Molden, Lucas, Gardner, Dean & Knowles 2009 (JPSP, four studies): explicit, active rejection produces a sense of social loss and prevention-focused responses — withdrawal, thoughts about what one should not have done, agitation. Being ignored produces a sense of failed social gain and promotion-focused responses — re-engagement, thoughts about what one should have done, dejection. Different motivational signatures, not just different intensities. MEDIUM single paper, 339 citations, no replication located.

Silence versus a delivered no. Telari, Pancani & Riva 2025 (Computers in Human Behavior) ran multi-day diary experiments (N = 46 over six days; N = 90 over nine) where a confederate ghosted, explicitly ended the interaction, or continued. Both hurt; ghosting produced a slower and more prolonged negative response, attributed to uncertainty blocking closure. Pancani et al. 2022 (N = 176, recall-based) found ghosting produced worse outcomes than rejection across most measures. Small samples, but the experimental design is better than most of this literature. MEDIUM

2. Individual differences

The most-cited claim in this area needs correcting before use. The Hartgerink meta-analysis is routinely invoked to show the initial sting is universal. It cannot bear that weight. The moderators it tested were experimentally manipulated situational variables, not personality traits — continuous individual-difference variables that primary authors had dichotomised were deliberately collapsed away. And the moderation analyses were catastrophically underpowered: detecting the interaction at 80% power would require N = 2,186; mean sample size in included factorial designs was 110, giving mean power of .08. A literature with 8% power to detect moderation will report that effects are unmoderated whether or not they are. The authors' own reading is more careful than the abstract: "differences in recovery from ostracism… occur somewhere between initial pain and final recovery." HIGH

Rejection sensitivity is a real construct with a broken measure. Downey & Feldman 1996 (~3,880 citations) defined RS as a processing disposition to anxiously expect, readily perceive, and overreact to rejection. Prospective prediction is genuine: couples containing a high-RS partner are roughly three times more likely to break up within a year, with the mechanism running through the high-RS person's own behaviour — a self-fulfilling prophecy, not just biased perception (Downey et al. 1998). Meta-analytically (Gao et al. 2017, 75 studies): r = .33 with depression, .41 anxiety, .39 loneliness, .41 BPD, similar in clinical and community samples — but the RS–depression association is negatively moderated by follow-up length, decaying as the horizon lengthens. HIGH

The measurement problem is serious. Lord et al. 2022 (Psychological Assessment, N = 346 and N = 540) found the A-RSQ's expectancy × concern product score does not hold together: a two-factor solution fit better in both samples, and the factors behaved differently — concern tracked negative affect, expectancy tracked diminished positive affect. Two decades of findings rest on a composite the data do not support. HIGH

"Rejection Sensitive Dysphoria" is a folk construct, not a finding. Popularised by psychiatrist William Dodson through lectures and the consumer site ADDitude rather than peer review. Total quantitative evidence base: none. The careful public accounting (Kahn 2026) counts five studies, samples of 4 to 43 participants, all qualitative — including a four-patient case series and a study in which participants defined RSD inconsistently within individuals across the study. There is no validated measure and no demonstrated discriminant validity against rejection sensitivity, social anxiety, atypical depression, BPD or general emotional dysregulation. HIGH

One correction worth preserving, because it explains why clinicians don't dismiss the phenomenon while rejecting the label: "a long-standing pattern of interpersonal rejection sensitivity… resulting in significant functional impairment" is an actual criterion of the DSM's atypical-features specifier for depressive and bipolar disorders. Rejection sensitivity is recognised; the ADHD-specific packaging is what lacks standing. HIGH

Where individual differences actually live: trajectory, not amplitude. The convergence here is unusually consistent across otherwise unrelated literatures.

Recovery speed has resisted measurement, which is the honest limit. The two serious attempts to make it a trait have failed reliability audits. Frontal alpha asymmetry is 40–60% occasion-specific and carries almost no signal for depression (van der Vinne 2017: grand mean d = −0.007 across 16 studies; Kołodziej 2021 multiverse: 13 of 270 analysis paths significant, below chance). Emotional inertia is meta-analytically associated with poorer wellbeing at ρ = −.151, and Wendt et al. 2020 (N = 1,192, 51,278 occasions) found split-half reliability of residualised negative-affect autoregression of .09–.27 — reliability that low caps any true correlation at its square root. Shao & Ong 2026 add that inertia estimates partly reflect sampling interval. Whether recovery speed is a real trait badly measured, or an artifact, is currently open. HIGH

One unresolved contradiction to keep in view. Gerber & Wheeler 2009 (88 studies) report rejection lowers mood (d = −0.50) and self-esteem (d = −0.70). Blackhart et al. 2009 (192 studies, same year, overlapping literature) report rejection shifts affect toward neutrality rather than distress, and that self-esteem among rejected participants did not differ from neutral controls. Two meta-analyses cannot agree on whether rejection lowers self-esteem. HIGH that the contradiction exists; LOW on the answer.

3. Interpretation: how much is the story?

Less than its cultural standing suggests, and the direction of the evidence is against the applied version of the thesis.

The foundational paradigm was reversed by its own author. Maier & Seligman 2016, Psychological Review, "Learned helplessness at fifty," ~1,300 citations. Verbatim: "the original theory got it backward. Passivity in response to shock is not learned. It is the default, unlearned response to prolonged aversive events." Intense aversive stimulation sensitises dorsal raphe 5-HT neurons, which suppresses active escape; what is actually detected — by a prelimbic vmPFC → dorsal medial striatum circuit — is the presence of control, and a separate prelimbic population then inhibits the DRN. There is no "detection of uncontrollability" circuit. Uncontrollability is a non-event neurally.

This matters because learned helplessness was the empirical foundation for the claim that a construal of uncontrollability produces the damage. On the current account the damage is a brainstem response requiring no construal, and the psychologically interesting variable — the thing that can be learned and later deployed — is prior experience of actual control. That relocates the practical implication from reframing to mastery experience. HIGH

The immediate hit is close to attribution-proof. This is where the interpretation thesis takes its hardest empirical hit. An external, maximally derogation-friendly attribution — the people rejecting me are contemptible — does not spare you (the KKK study, replicated). Whatever room interpretation has is in recovery, and the meta-analytic evidence that this room is large is weak. HIGH

The correlational half of the thesis holds; the causal half does not. Cognitive style prospectively predicts depression onset with large odds ratios (Temple-Wisconsin CVD Project: high-risk participants 3.5–6.8× greater odds of major, minor and hopelessness depression) — but in a selected-extremes, non-randomised design, and the same project found negative cognitive style predicts the occurrence of negative life events, so the stressor is partly endogenous to the person. HIGH on the finding; MEDIUM-HIGH on the caveats.

Meanwhile every applied programme built on changing interpretation shrinks toward zero, and — this is the part worth dwelling on — fails precisely at the mediator it was built around:

The middle row of that table is the finding. When you restrict to studies that verified the intervention changed the purported mediator, the effect disappears. That is a manipulation-check failure at the level of an entire literature, and it is the same shape as PRP's explanatory-style null.

Adjacent casualties. Depressive realism has essentially collapsed: Moore & Fresco 2012 (75 studies) put the effect at d = −0.07, with both dysphoric (d = .14) and non-dysphoric (d = .29) groups showing positive bias; Dev, Moore, Fresco et al. 2022 ran a preregistered replication and found none — "despite its popular acceptance, depressive realism is not replicable." Note Moore is an author on both, revising his own conclusion. HIGH

One useful synthesis about attribution. Crocker et al. 1991 found attributing negative feedback to a specific, clearly-flagged prejudiced evaluator buffered affect — and the same ambiguity that lets you discount criticism also lets you discount praise, so the mechanism is double-edged. Against that, Schmitt et al. 2014 (328 effect sizes, N ≈ 144,246) found perceived discrimination correlates with wellbeing at r = −.23. The reconciliation most consistent with the data: a situational, bounded external attribution protects; a stable, global external attribution does not. Which is the internal/stable/global structure from 1978 reappearing with the locus flipped — stability and globality do the work; internality does much less. MEDIUM my synthesis across two HIGH-confidence sources.

4. Rumination versus processing: where the boundary actually is

The best-supported answer is level of construal, not amount of thinking, not valence, not duration, and not "why versus what."

It is not amount. Mor & Winquist's meta-analysis (226 effect sizes) found self-focused attention correlates with negative affect at r ≈ .51 overall, but the relationship depends on the type of self-focus, not its quantity. HIGH

It is not "why." This is the most common popular misreading. Kross, Ayduk & Mischel 2005 ran a 2×2 (immersed/distanced × what/why) and found distanced-why was the only good cell — lower implicit anger (d = 0.45), explicit anger (d = 0.69) and negative affect (d = 0.59) than the average of the other three. "Why" from an immersed perspective was as bad as anything else. Roughly 38 per cell, so underpowered by current standards. MEDIUM

It is construal level. Watkins 2008 (Psychological Bulletin, 3,800+ citations) explicitly tested three candidate principles — content valence, situational context, level of construal — and concluded construal does the explanatory work. The causal evidence is unusually clean: Watkins, Moberly & Moulds 2008 ran three experiments training participants into abstract versus concrete construal on emotionally balanced scenarios (so valence could not explain the result), then exposed them to a standardised anagram failure. Concrete-trained participants showed less emotional reactivity in all three, including one using implicit induction rather than explicit instruction. The training itself produced no mood effect; only the response to subsequent stress differed. Construal level does not change how you feel now; it changes how much the next setback costs. HIGH

Abstract-evaluative processing — what does this mean about me, why does this keep happening — generalises a single event into a property of the self. Concrete-specific processing — what exactly happened, in what order, what specifically would I do differently — keeps the event bounded to its circumstances.

The counterfactual literature converges on the same line from a different direction. "If only I had sent that email on Tuesday" is concrete, names a controllable antecedent, and generates an action. "If only I were the kind of person who doesn't screw things up" is abstract, names an uncontrollable trait, and generates nothing. Same grammatical form; opposite construal level. LOW my synthesis, but it is the strongest argument that the boundary is real.

Honest limits on the therapeutic side. Concreteness Training in a Phase II RCT (N = 121, primary-care major depression) beat treatment-as-usual (HAMD difference 4.28, CI 1.29–7.26) but did not beat relaxation training (1.98, CI −1.14 to 5.11) — though it did outperform relaxation on rumination and overgeneralisation specifically. Rumination-focused CBT showed d = 0.94–1.10 in Watkins' own Phase II trial but only d = 0.38 in a larger head-to-head against group CBT (N = 131), with no difference in rumination and no difference at six months. A trial that beats an active comparator on symptoms while failing to move its own proposed mediator is a mechanism problem. HIGH

Brooding versus reflection is descriptively useful and prescriptively weak. The two-factor structure replicates widely; the claim that reflection is adaptive is shakier (Treynor found reflection predicted lower depression at one year while being positively correlated with it cross-sectionally). Olatunji et al. 2013 (179 studies) found associations strongest for brooding — strongest, not absent, for reflection. Construal cuts through the reflection subscale rather than aligning with it, which is why it is the better boundary. Nolen-Hoeksema's own 2008 review is unusually candid: distraction did not consistently correlate with lower depression as predicted, and "rumination appears to more consistently predict the onset of depression rather than the duration" — the reverse of the original claim. HIGH

Two secondary variables with real support. Arousal at the time of thinking: Kjærvik & Bushman's 2024 meta-analysis (154 studies, 184 samples, N = 10,189) found arousal-decreasing activities reduced anger and aggression at g = −0.63, while arousal-increasing activities were flatly null (g = −0.02). Venting does not work; neither does jogging it off. The functional variable is arousal, not expression. And immediate structured processing of a fresh trauma can harm: the Cochrane review of single-session psychological debriefing found no prevention of PTSD, with one trial showing increased risk at one year (OR ≈ 2.5). HIGH for the meta-analyses; MEDIUM for "may be harmful," which rests on a single trial.

There is no established "wait N days" rule — Pennebaker concedes no parametric studies exist. The evidence supports arousal state, not elapsed time, as the gate. High arousal converts analysis into rehearsal. MEDIUM inference from the above, not a directly tested claim.

No validated physiological discriminator exists. Perseverative cognition is associated with higher blood pressure (g = .45/.51 — and these are the two effects the authors report as publication-biased), heart rate (g = .28/.20), cortisol (g = .36/.32) and lower HRV (g = .15/.27 — the smallest in the set, and the one most often sold as a rumination biomarker). No study validates HRV or cortisol as a within-person classifier of the two states. HIGH on the meta-analysis; UNVERIFIED on the classifier claim.

5. What actually helps, ranked

A framing note first: almost nothing here was designed to answer "what helps after rejection." Self-affirmation was built for defensive processing of health messages; expressive writing for trauma disclosure and immune function; reappraisal effect sizes come from film clips and IAPS images. Much of this is transported evidence, and the transport is often the weakest link.

A rule that governs the whole table: the effect size you are quoted depends almost entirely on the comparison condition. Exposure for social anxiety drops from d = 0.86 to 0.35, self-compassion drops from significant to null, and meditation drops to no-better-than-exercise, all at the moment an active control is introduced. An effect size cited without its comparison condition is not information.

Tier 1 — supported, with the real number attached

Exposure-based CBT for clinical social anxiety. Powers et al. 2008 (32 RCTs, N = 1,479): versus waitlist d = 0.86, but versus psychological placebo d = 0.34 and pill placebo d = 0.36. Acarturk et al. 2008 (29 RCTs, N = 1,628): mean d = 0.70, with waitlist-controlled studies producing significantly larger effects and smaller effects in samples meeting full DSM criteria. Carpenter et al. 2018 (41 placebo-controlled RCTs): g = 0.56 overall. Cite the placebo number. HIGH

Important mechanistic caveat: exposure does not work by habituation, and the field established this decades ago. Craske et al. 2008/2014: "neither fear reduction nor ending fear levels predict long-term outcome from extinction or exposure." It works by inhibitory learning — building a competing expectancy that violates prediction. Repeated exposure without expectancy violation is just repeated exposure. HIGH This is the single most important thing to know before assuming that volume of rejection builds tolerance.

Time, with a corrected expectation. Not an intervention, but the most robust finding in the set, and it is not what people think. Lench et al. 2019 (JPSP, two field studies — students receiving midterm grades, N = 643; US citizens after the 2016 election, N = 706 — plus three experiments) decomposed the impact bias: people were accurate at forecasting emotional intensity, and inaccurate at forecasting emotional frequency and general mood, with the mood error growing with time since the event. Eastwick et al. 2008, measuring through actual breakups, found the forecasting error was entirely an initial intensity bias with no decay bias — people are accurate about how fast they recover. And Levine et al. 2012 showed a large share of the classic intensity bias is a procedural artifact of asking people to forecast feelings "about the event" and later report feelings "in general." HIGH

The consolation people offer — "it won't hurt as much as you think" — is roughly the wrong one. The pain per occurrence is about what you fear. What collapses is the number of occurrences. That is a more precise claim, better supported, and the only thing in this report that requires nothing of the person.

Tier 2 — real but small

Self-distancing (observer-perspective reflection, written or spoken). Guo 2022 (48 studies, 102 effects): g = −0.26 [−0.36, −0.15]. Murdoch et al. 2022, preregistered (25 experiments, N = 2,397): g = 0.19 [0.05, 0.33], with the authors' own risk-of-bias assessment flagging "uncertainty regarding the benefit." The one reliable moderator: it works when externalised through writing or talking, not through bare pronoun-swapping. HIGH on the range

Three things stop this being the recommendation it is usually made into. There is no Registered Replication Report and no many-labs replication. A preregistered RCT (Riddell et al. 2023, N = 257) came back null on wellbeing and stress. And there is a twice-run harm finding: Giovanetti et al. 2019 randomised participants to third-person self-distanced daily writing over two weeks in two studies (n = 104, n = 80) and found greater depressive symptoms than first-person expressive writing and than no writing, driven by participants high in cognitive vulnerability; their conclusion was that it "should not be used for the prevention of depressive symptoms." MEDIUM It is a nudge, not a mechanism.

There is also an unresolved theoretical tension nobody should paper over: construal-level theory says psychological distance produces abstract construal, and Watkins says abstract construal is the pathogenic ingredient. Kross's intervention and Watkins' intervention appear to push the same dial in opposite directions. LOW my read: distancing changes whose event it is; concreteness changes how far it generalises. Nobody has cleanly reconciled them.

Situational reappraisal and perspective-taking. Webb, Miles & Sheeran 2012 (306 experimental comparisons) is the definitive taxonomy, and it embarrasses the folk version. Cognitive change overall d+ = 0.36. Within it: reappraising the situation d+ = 0.36 and perspective-taking d+ = 0.45 both beat reappraising the response ("it's okay that I feel this") at d+ = 0.23 — which is the version most self-help teaches. Distraction, routinely dismissed as avoidance, works at 0.27; concentrating on the feeling is actively counterproductive at −0.26. And suppression splits: suppressing expression works (0.32) while suppressing experience (−0.04) and suppressing thoughts of the event (−0.12) do not. "Suppression bad, reappraisal good" collapses a distinction the experimental data insist on. HIGH

One boundary: Troy, Shallcross & Mauss 2013 (N = 170) found a crossover — high reappraisal ability predicted lower depression when the stressor was uncontrollable and higher depression when it was controllable. Reappraising a situation you could have changed appears to substitute for changing it. Good news for rejection specifically, which is mostly uncontrollable; a caution for rejections that are actually a fixable conflict. MEDIUM well-cited single study, no replication located.

Belonging reframing, where the environment can confirm it. Walton & Cohen 2011 (N = 92) found three-year GPA and health effects; Walton et al. 2023 (Science, N = 26,911, 22 institutions, preregistered) found increased persistence concentrated in historically lower-progressing groups — and only where institutions provided genuine opportunities to belong. That boundary condition is the whole lesson. HIGH

Compassion-based interventions. Ferrari et al. 2019 (27 RCTs): self-compassion g = 0.75, stress 0.67, depression 0.66 — with the authors noting "publication bias may be present." Kirby et al. 2017 (21 RCTs, N = 1,285): d = 0.70 self-compassion, 0.64 depression. But Wilson et al. 2019 found that restricted to trials with active controls, change scores were not significantly different from control on any outcome; Wakelin et al. 2021 found greater reductions against passive than active controls; Han & Kim 2023 (56 RCTs) reported high overall risk of bias and few active comparisons. Three of five meta-analyses converge on the effect shrinking or vanishing against anything that is also a plausible treatment. HIGH that this is unresolved

There is also a measurement problem. Muris & Otgaar argue the Self-Compassion Scale total fuses compassionate self-responding with uncompassionate self-responding (self-judgment, isolation, over-identification), and the latter subscales approximate reverse-scored neuroticism — so a total score correlated with depression is partly a construct correlating with itself. A parallel exchange in European Journal of Personality ended with the critics restating that "our initial critique of the self-compassion scale holds." HIGH that the critique is live

The two problems compound in a way I did not see stated anywhere: Wilson et al. note the interventions improved the negative subscales more than the positive ones. If the negative subscales are the neuroticism-flavoured ones, the literature may be substantially measuring reductions in self-reported negative affect, on a scale whose negative half is a negative-affect measure, in unblinded trials against waitlists. MEDIUM my synthesis from two HIGH sources.

Tier 3 — small, fragile, or aimed elsewhere

Mindfulness meditation. Goyal et al. 2014 (JAMA Internal Medicine, 47 trials, N = 3,515, insisting on active controls): anxiety SMD 0.38, depression 0.30, pain 0.33, and no evidence of superiority over any active treatment — drugs, exercise, or other behavioural therapies. HIGH

Self-affirmation. Epton et al. 2015 (144 tests): message acceptance d+ = .17, intentions .14, behaviour .32. Sweeney & Moyer 2015 found effect sizes on intentions did not predict effect sizes on behaviour within studies, undermining the assumed causal chain. Hanselman et al. ran a well-powered replication in the same setting as a prior positive large-scale trial and got a null with precision sufficient to rule out effects larger than 0.10. And the transport problem is severe: self-affirmation was designed to reduce defensive resistance to threatening information. Rejection is not a message you are resisting; it already landed. UNVERIFIED as a rejection intervention — untested, not disproven.

Tier 4 — folk wisdom with a citation veneer

Expressive writing. Track the number: Smyth 1998 d = 0.47 (this built the reputation) → Frattaroli 2006 (146 randomised studies) r = .075, d ≈ 0.15 → Mogk et al. 2006 (30 RCTs) null on both somatic and psychological health → Reinhold, Bürkner & Holling 2018 (39 RCTs) post-test g ≈ 0.09, follow-up 0.03, both non-significant → a 2023 meta-analysis reporting 0.33 raw falling to 0.16 after publication-bias correction. The true effect is around a tenth of a standard deviation. It is free and harmless; it is not a treatment, and the popular account overstates it roughly fourfold. HIGH

"Reach out and reconnect." This has the worst evidence-to-confidence ratio in the set. Quarmley et al. 2022 (Aggressive Behavior, three meta-analyses, N = 3,864) found rejection increases aggression (d = 0.41, k = 19) and decreases overt prosocial behaviour (d = 0.59, k = 7; d = 0.71 when participants could freely choose either), concluding that the results "cast doubt on the theory that rejection triggers prosocial behaviour." No trial shows that deliberately reconnecting repairs mood. HIGH

Meaning-making and benefit-finding. Helgeson, Reynolds & Tomich 2006 is the meta-analysis, and its design is the finding: 87 cross-sectional studies. Benefit finding related to less depression and more positive wellbeing — and also to more intrusive and avoidant thoughts about the stressor — and was unrelated to anxiety, global distress, quality of life, and subjective physical health. No causal claim is supported; people coping well may narrate growth rather than growth producing coping. HIGH

Rejection therapy. No controlled evidence exists. Searches return transplant immunology and 1970s desensitisation. Jia Jiang's 100 Days is an unblinded, unmeasured, single-subject demonstration by someone whose subsequent career — TED talk, book, speaking business — depends on its success. The mechanism it borrows (graduated in-vivo exposure with disconfirmation) is the one carrying the d ≈ 0.35 above, which makes it plausible; but it differs from clinical exposure in ways that could break it — no hierarchy, no therapist, no response prevention, and stunts calibrated for video rather than for the specific fear. HIGH that no trial exists

6. Repeated and structural rejection

The literature is sharply lopsided. Single-episode rejection has 120 randomised experiments and a preregistered meta-analysis. Rejection as a recurring condition has almost nothing, and Hartgerink et al. state plainly that the lab version is "not feasible (and even unethical)" to run.

Habituation, what little is known. The only genuine repeated-exposure test I found: Davidson et al. 2019, 30 adolescents, Cyberball twice one month apart — distress and anxiety significantly attenuated at retest, but most participants continued to experience distress. Partial habituation, not extinction. Across hundreds of episodes over a year, there is no experimental evidence at all. MEDIUM for the study; HIGH that the year-scale evidence is absent.

Unemployment is the strongest natural experiment, and it says people do not fully recover. Lucas, Clark, Georgellis & Diener 2004 (Psychological Science, German Socio-Economic Panel, 15-year longitudinal with baseline estimated from pre-event years): people reacted strongly and shifted back toward baseline, but "did not completely return to their former levels of life satisfaction, even after they became re-employed." Reactions were similar regardless of sex, age, income, or length of time unemployed — duration did not predict severity. Reverse causation is partly ruled out: the eventually-unemployed had high pre-event life satisfaction. HIGH on direction; the specific SD figures in circulation are inconsistent and I would not quote one.

Magnitudes: Paul & Moser 2009 report d = 0.51 for distress, unemployed versus employed — but the longitudinal subset, people moving into unemployment, gives only d = 0.19. That gap is the selection effect made visible: roughly two-thirds of the cross-sectional difference is not the transition. HIGH

Job search: intention holds, behaviour collapses. Wanberg et al. 2005 (JAP, 903 Minnesota UI recipients, 10 waves over 18 weeks, 5,371 observations). Search intensity slope −3.07 (p < .01), attitude −0.093 (p < .01), self-efficacy −0.025 (p < .05) — all declining significantly. Intention slope −0.0149, not significant. People keep meaning to look and stop looking. The quadratic on intensity is positive, so the collapse decelerates rather than going to zero. HIGH

One caveat that runs opposite to the usual survivorship worry: the sample falls from 903 to 370, and a substantial share of the exits are re-employment, which search intensity predicts. So the highest-intensity searchers are systematically leaving the panel and a meaningful share of that −3.07 is composition change, not within-person demoralisation. HIGH that the attrition is outcome-correlated; MEDIUM on the magnitude of the inference.

And the daily timescale runs the other way: Wanberg, Zhu & van Hooft 2010 found lower perceived progress on any given day predicted more effort the next day. Over months, effort declines; day to day, falling behind increases it. Both are true, and neither generalises to the other's timescale. HIGH

The best test of "what doesn't kill you," and why it cannot be read the way it usually is. Wang, Jones & Wang 2019 (Nature Communications), regression discontinuity at the NIH R01 funding threshold for junior scientists:

Credit where due: the authors tested screening by removing the worst ex post performers from the narrow-win group until attrition rates matched, and near-misses still outperformed. I found no published critique or failed replication. HIGH on the findings

But the estimand is the effect on those who persevered — the paper says so — and the 12.6% who vanish contribute no performance data whatsoever. Their outcome is measured as absence, not as harm; there is no wellbeing, income, or career-elsewhere data. The clean statement is: 6.1 points of upside, purchased at a 12.6-point risk of career death, with the upside measured only on the survivors. In practice only the first half travels. LOW-MEDIUM as an inference, flagged as such; the authors' screening argument has not been independently re-analysed.

Sales. "Call reluctance," the field's famous construct, is a commercial product: Dudley and Goodson are principals of the company that owns, publishes and sells SPQ*GOLD, and the owner and its own distributors publish three different factor counts (16, 12, and 6). No independent peer-reviewed psychometric validation surfaced. The peer-reviewed replacement is Verbeke & Bagozzi 2000's Sales Call Anxiety (n = 189), independently replicated in South Africa (n = 112), with coping tactics that must be matched to symptom channel. HIGH

Whether tolerance builds with tenure is unanswerable from existing data: no cohort is followed from entry with repeated measurement and modelled dropout. And the selection filter is enormous — roughly 68% of life-insurance agents are gone within two years, with four-year retention around 11–17%. If 85% of entrants are gone by year four, the observed "veterans handle rejection better" gradient is exactly what pure selection produces with zero individual adaptation. HIGH on direction; MEDIUM on any single retention figure.

Discrimination as chronic rejection. Pascoe & Smart Richman 2009 (134 samples): mental health r = −.20 (−.16 trim-and-fill), physical health −.13. Paradies et al. 2015 (293 studies): mental −.23 (−.18 adjusted), physical −.09, and blood pressure/hypertension r = 0.00, p = .814, k = 24. That last figure is the tell — the association is strongest where shared method variance is total and exactly zero for the one outcome measured by an instrument rather than a questionnaire. Paradies also found the effect halving over time: cross-sectional r = −.22 (k = 197) versus long-term longitudinal −.11 (k = 7), with only 9% of articles longitudinal. HIGH

The causal evidence is real and roughly half the size: Emmer et al. 2024 (Psychological Bulletin, 73 RCTs, 12,097 participants, preregistered) found g = −0.30 overall — and, directly relevant, pervasive-discrimination manipulations g = −0.55 versus single-event g = −0.25. Framing rejection as recurring and inescapable roughly doubles the acute effect. Two flags: no publication-bias correction was performed, and the manipulation definition was broad enough to include generic social exclusion. HIGH on the numbers; MEDIUM on interpretation.

Two high-profile causal claims in this area did not survive reanalysis and belong in any careful treatment: Hatzenbuehler et al.'s "12 years shorter life expectancy" was corrected to null after a coding error in the survival model, and Raifman et al.'s finding that marriage equality reduced adolescent suicide attempts became statistically indistinguishable from zero once standard errors were clustered at the state level, the level at which treatment varies. HIGH

Dating. Base rates are brutal and well measured (Bruch & Newman 2018, several million users: >80% of first messages from men; women's average reply rate under 20%). The wellbeing link is modest and concentrated: a 2026 Communications Psychology meta-analysis (27 studies, N = 21,263) found behavioural dysregulation g = 0.44 and body-related outcomes g = 0.32, but general wellbeing small and non-significant. And the most-quoted claim in the space is dead: Strübel & Petrie's 2017 "Tinder: swiping self-esteem?" rested on 31 male users versus 203 non-users; the same authors' 2022 follow-up (187 men, 547 women) found Tinder use "not related to psychological well-being," η² = .004. Nine years of coverage still cites the first one. HIGH

7. The strongest counterargument

Grit is a jangle fallacy with a small correlation attached. Credé, Tynan & Harms 2017 (JPSP, 88 samples, ~67,000 people): grit correlates with performance at ρ ≈ .18 — under 4% of variance — and with conscientiousness at ρ ≈ .84, close to the reliability ceiling. The higher-order two-facet structure was not confirmed; perseverance of effort carries essentially all the validity, consistency of interests almost none. The authors' own conclusion: "interventions designed to enhance grit may only have weak effects." A nationwide cluster-RCT in North Macedonia (~33,000 students) raised perseverance of effort while reducing consistency of interests, with no significant effects on conscientiousness, self-regulation, frustration tolerance or locus of control. HIGH

Resilience is common as an outcome and unpredictable as a trait. Bonanno is simultaneously the strongest evidence that most people are fine (Galatzer-Levy, Huang & Bonanno 2018: 54 trajectory studies, resilience 65.7%, chronic 10.6%) and the strongest evidence that resilience training rests on a measurement failure. His "resilience paradox" argues that although resilience is modal, who will be resilient is essentially unpredictable: resilience scales predict post-trauma mental health poorly, different scales disagree on content, and machine-learning models do not rescue prediction. Training meta-analysis (Vanhove et al. 2016, 42 samples, n = 16,348): d = 0.21, decaying from 0.26 proximal to 0.07 distal, with one-on-one coaching working best and computer-based and train-the-trainer formats — the ways these actually scale — working worst. HIGH

Honest caveat against the contrarian's own tool: Infurna & Luthar 2016 showed that relaxing the growth-mixture-model assumptions dissolves the distinct resilient class. So "most people are fine" and "most people are damaged" are both underdetermined by the trajectory literature. You can say the alarmist number is unsupported; you cannot cite 65.7% as bedrock. HIGH

Post-traumatic growth is the strongest single result in the contrarian case. Frazier, Tennen, Gavian, Park, Tomich & Tashiro 2009 (Psychological Science) is the rare prospective design: participants completed measures of PTG-relevant domains at T1; 122 experienced a traumatic event before T2 two months later; the researchers compared actual measured change against retrospective Posttraumatic Growth Inventory scores. Verbatim: "PTGI scores generally were unrelated to actual growth in PTG-related domains." Worse for the growth narrative: "perceived growth was associated with increased distress from pre- to posttrauma, whereas actual growth was related to decreased distress." The reported growth is a coping narrative, and the people telling it loudest are on average doing worse on measured outcomes. The field's own methodologists agree (Infurna & Jayawickreme 2019, "Fixing the Growth Illusion"; Jayawickreme et al. 2021). HIGH

Limits: Frazier's sample is undergraduates over two months with a broad trauma definition. It shows the instrument is invalid as a measure of change; it does not prove nobody ever changes for the better.

The replication base rate in this exact neighbourhood is poor. Ego depletion collapsed: Hagger et al. 2016, 23 labs, N = 2,141, d = 0.04 [−0.07, 0.15] — with the manipulation working (participants rated it more effortful and frustrating) and producing no downstream decrement. The canonical rejection-impairs-self-control finding (Twenge, Ciarocco, DeWall & Baumeister 2005) is built on that framework and shares authors, labs and measures; I found no direct multi-lab replication of it — UNVERIFIED, and the inference that it inherits the fragility is LOW, though natural. The broader base rate comes from an unlikely source: Baumeister, Tice & Bushman 2023 reviewed 36 multisite replication projects in social psychology and counted 27 (75%) failures. They argue the replications are themselves flawed; the count is theirs. HIGH

Survivorship, formally. Denrell 2003 (Organization Science) is the strongest statement: under undersampling of failure, risky and resource-concentrating practices appear positively associated with performance in the observed sample even when they are not in the population. "Twelve publishers turned down Rowling" is drawn from the survivor sample by construction; there is no reachable comparison group of writers rejected twelve times who stopped. Lists of famous rejections contain zero information about the expected value of persisting. HIGH on Denrell; LOW/definitional on the application, since it is a logical point.

Leaving it alone can beat intervening. The Cochrane review of single-session psychological debriefing (11 trials) found no reduction in PTSD incidence or severity at any follow-up, no benefit for depression or anxiety, and no superiority over education, with one trial finding significantly increased PTSD risk at about a year. The authors advise against compulsory debriefing. Scope limit: this licenses scepticism about mandatory, immediate, single-session, universal intervention. It says nothing against treating the minority who develop persistent symptoms. HIGH for "no benefit"; MEDIUM for "may harm."

The loneliness-as-killer frame took a serious hit in 2024. Holt-Lunstad's 2010 meta-analysis (148 studies, N = 308,849) gives OR = 1.50 for survival with stronger social relationships; her own better-adjusted 2015 figures are much smaller — isolation 1.29, loneliness 1.26. The "equivalent to 15 cigarettes a day" line is genuinely traceable to her and she defends the wording, but three things are wrong with how it is repeated: it has been narrowed to loneliness when the original was an aggregate; it benchmarks against 1.50, the largest number in the literature, and nobody re-benchmarks against 1.26; and it smuggles in causality the design cannot support. Then Liang et al. 2024 (Nature Human Behaviour) compared observational against Mendelian-randomisation estimates in UK Biobank: observationally loneliness predicted elevated risk in 13 of 14 disease categories; under MR there was little evidence of causal effect for 20 of 26 diseases, concluding loneliness functions largely as a surrogate marker rather than a cause. HIGH on the numbers; MEDIUM-HIGH on the MR result, which is the most important disconfirming evidence in the health literature.

Where the contrarian case is itself weak, and should not be pushed. Cyberball's d = −1.36 is real and its decay is not cleanly demonstrated, so "rejection barely hurts" is false. Gerber & Wheeler's mood (−0.50) and self-esteem (−0.70) effects are moderate, not trivial. The 65.7% resilience base rate is model-dependent. And Gilbert's impact bias and Levine's artifact critique cannot both be deployed — pick one; the version surviving both is narrow (forecasts of post-event duration and frequency for consequential setbacks are inflated).

8. What the people living it say, and where it contradicts the lab

Sourced from Reddit (via the Arctic Shift archive, whose comment scores are early snapshots and unreliable as quality signals) and Hacker News (which does carry karma). Read as texture, not evidence. Four divergences are worth carrying.

The dose variable is probably wrong. The literature counts rejections. Operators consistently report that automated, impersonal rejection barely registers while invested rejection is devastating. A job seeker on Reddit: "I have played this game so many times that this is like water off of a duck's back… It takes me a few months to apply to 1000 roles." Against, in the same thread (21 pts): "the rejection after 5 or 6 rounds of interviews hit harder than those auto mailers. Feels like I almost had it." Nobody describes an unreturned swipe as painful.

This maps directly onto Law & Williams 2006: distress requires an inferred agent. A form rejection from an applicant tracking system supplies no mind to be rejected by. The literature's implicit dose model — rejection count — is probably measuring the wrong thing. HIGH as a convergent pattern

The academic thread is the sharpest natural test, because output volume is stated. One commenter: "8000 citations from 100 papers over 22 years… maybe 200 rejections total, completely numb by now." Another, same rejection count, opposite outcome: "I have 150 citations, 15 papers published, and over 200 rejections over 12 years. It still sucks." The difference is not volume; it is hit rate.

Ambiguity may be the primary harm. In a hiring thread, the two commenters who prefer ghosting are both hiring managers, and one names the institutional mechanism: companies withhold the explicit rejection because an explicit rejection produces a real state that makes the candidate unavailable for a second-choice offer. The cost of that suppression is transferred to the candidate as open-ended ambiguity. Several people describe self-manufactured closure as an invented technology — "I assume that if they haven't contacted me after X days I wasn't selected… it's healthier to give it closure than to keep waiting" — converting a non-event into a dated, closed one. That maps onto nothing in the standard coping taxonomies, and it is exactly what the Telari ghosting experiment would predict. HIGH as a pattern

What practitioners report as helpful is exposure control, not reappraisal. The most-upvoted protocol anywhere in this material is entirely structural: "I only apply to jobs for a maximum of three hours a day… I only check my email three times a day so I don't constantly see rejection emails… I don't think about jobs until I get the phone screen even if the job is something I would love to do." Another: "If I get a rejection email, immediately delete it" — paired with tracking phone screens in a spreadsheet, which is a process-metric substitution. Self-compassion and cognitive reappraisal, the two most-studied interventions, are essentially absent from what people spontaneously report doing. HIGH

Salespeople largely deny that fear habituates: "The fear never really goes away, which is good" (35 pts); "It's not that you get less fearful, it's that you get more courageous." What they attribute improvement to is procedural mastery and, strikingly, eliminating the inter-trial interval: "don't put the phone down. If you put it down you have to mentally prep for the next call." Practitioners have independently concluded that appraisal and rumination need a gap, and the intervention is to remove the gap — something fixed-interval lab paradigms would never detect. This population is also survivorship-selected by its own admission: "You get comfortable and good at it within a few months or you don't survive." MEDIUM-HIGH , badly confounded

Deliberate rejection-seeking, where people tried it, works by base-rate correction rather than desensitisation — and reverses when the yes is genuinely wanted. The most detailed account: "I think it doesn't help with rejection, rejection still hurts. But it teaches you that chances of rejection aren't as high as you think they are." The boundary condition, from someone it damaged: "Lots of rejection only hurt my confidence more… I thought I could push through nos until I got to yes and it just destroyed my confidence." That is the investment moderator arriving from a third direction, and it contradicts the practice's own stated theory. MEDIUM

One last observation, because it is the counter-case to attribution theory as usually taught: an accurate, external, stable attribution can be as corrosive as a distorted internal one. A job seeker who correctly identified his résumé gap as the cause: "It started feeling like a crime to have a gap on resume." The external cause was also uncontrollable and permanent, which removed agency without removing blame. MEDIUM

9. What is still unknown

The year-scale question is empty. Nobody has measured what happens to a person subjected to hundreds of rejections over a year with outcomes tracked on the people who quit as well as those who don't. The lab version is infeasible and, per Hartgerink et al., unethical. Every field study that could answer it loses the failures. The confident-sounding literature is about minutes; the question people actually have is about years.

Recovery speed may or may not be a trait. The two best attempts to measure it have failed their reliability audits. The dissociation between initial hit and recovery has been shown four times in samples of 47 to 91.

Whether interpretation can be changed, as opposed to whether it correlates. The correlational half is solid. Every applied programme built on the causal half fails at its own mediator.

Whether social pain is a distinct kind at all. The overlap interpretation is dead; the replacement is unclear. Woo et al. show distinct representations, but what rejection is neurally — a socially-tuned threat system, a salience response, an aversive learning signal — is unsettled.

The base rate of resilience. Genuinely underdetermined by the modelling.

10. Practical account, ordered by evidence

  1. Wait, with the right expectation. The pain per occurrence is roughly what you fear. The frequency of occurrence is what collapses, and people badly mispredict that. This is the best-supported item and the only one that asks nothing of you.
  2. If it is a clinical-grade fear of rejection, exposure-based CBT works — d ≈ 0.35 against a real placebo — and it works by expectancy violation, not by wearing the fear out. Repetition without disconfirmation is just repetition.
  3. Keep the thought concrete. The boundary between processing and rumination is construal level. "I handled that meeting badly and should have prepared the second slide" is processing. "I always handle these badly, what's wrong with me" is rumination. Same event, same duration, same valence, same "why" — different construal, and the difference causally changes how much the next setback costs. The practical test: if every attempt to make the thought specific slides back into a general statement about what you are, more thinking will not open it. (This test is my inference from the construal evidence, not a validated instrument.)
  4. Reappraise the situation, not your reaction to it (d+ = 0.36 versus 0.23), and know that distraction is not cheating (0.27) while concentrating on the feeling is actively harmful (−0.26).
  5. Lower arousal before analysing. Venting does not work (g = −0.02); reducing arousal does (g = −0.63).
  6. Self-distancing, in writing or aloud, is worth trying at g ≈ 0.19–0.26 — with the caveat that sustained daily distanced writing produced worse depressive symptoms in two studies of vulnerable people.
  7. Reduce investment per attempt and keep a parallel pipeline. This is not in the intervention literature; it is what every high-volume population independently converges on, and it is consistent with the finding that distress scales with what a rejection destroys rather than with rejection count.
  8. Things to stop recommending: expressive writing as a treatment, "reach out and reconnect" (the meta-analytic finding is that rejection reduces prosocial behaviour), benefit-finding, and rejection therapy as a protocol.

Sources

Full citation lists with URLs are held in the nine agent reports underlying this document. Key primary sources by section:

Mechanism. Eisenberger, Lieberman & Williams 2003 Science 302:290–292. Woo, Koban, Kross, Lindquist & Wager 2014 Nat Commun 5:5380. Cacioppo et al. 2013 Sci Rep 3:2027. Lieberman & Eisenberger 2015 PNAS 112:15250; Wager et al. 2016 PNAS 113:E2474. Iannetti & Mouraux 2010 Exp Brain Res 205:1–12. Hartgerink, van Beest, Wicherts & Williams 2015 PLoS ONE 10:e0127002. Gonsalkorale & Williams 2007 EJSP 37:1176; Fayant et al. 2014 Social Psychology 45:489. Dickerson & Kemeny 2004 Psych Bull 130:355–391. Hsu et al. 2013 Mol Psychiatry 18:1211. Panksepp & Herman 1978. Slavich, Way, Eisenberger & Cole 2010 PNAS 107:14817. DeWall et al. 2010 Psych Sci 21:931; Hofman, Wieser & van der Veen 2021 Social Neuroscience 16:362. Molden et al. 2009 JPSP. Telari, Pancani & Riva 2025 Comput Human Behav 177:108756; Pancani et al. 2022 Cyberpsychology 16(2).

Individual differences. Downey & Feldman 1996 JPSP; Downey et al. 1998 JPSP. Gao, Assink, Cipriani & Lin 2017 Clin Psych Rev. Lord et al. 2022 Psych Assessment 34:1062. Mendoza-Denton et al. 2002 JPSP 83:896. Wesselmann et al. 2012 PAID. Zadro, Boland & Richardson 2006 JESP 42:692. Berenson et al. 2016 Cogn Ther Res. Wendt et al. 2020 Eur J Pers; Dejonckheere et al. 2019 Nat Hum Behav. Blackhart et al. 2009 PSPR; Gerber & Wheeler 2009 Perspectives.

Interpretation. Maier & Seligman 2016 Psych Review 123:349–367. Alloy, Abramson et al. 2006 J Abnorm Psych. Brunwasser, Gillham & Kim 2009 JCCP; Bastounis et al. 2016 J Adolescence. Sisk et al. 2018 Psych Sci 29:549; Yeager et al. 2019 Nature 573; Macnamara & Burgoyne 2023 Psych Bull 149. Moore & Fresco 2012 Clin Psych Rev; Dev et al. 2022 Collabra 8:38529. Schmitt et al. 2014 Psych Bull 140:921.

Rumination. Watkins 2008 Psych Bull; Watkins, Moberly & Moulds 2008. Kross, Ayduk & Mischel 2005 Psych Sci. Guo 2022 Cognition & Emotion; Murdoch et al. 2022 Stress & Health; Riddell et al. 2023; Giovanetti et al. 2019. Kjærvik & Bushman 2024 Clin Psych Rev. Rose, Bisson, Churchill & Wessely, Cochrane CD000560. Ottaviani et al. 2016 Psych Bull.

Interventions. Webb, Miles & Sheeran 2012 Psych Bull. Ferrari et al. 2019 Mindfulness; Kirby et al. 2017 Behav Ther; Wilson et al. 2019 Mindfulness; Muris & Otgaar 2020 Mindfulness. Frattaroli 2006 Psych Bull; Reinhold, Bürkner & Holling 2018. Quarmley et al. 2022 Aggressive Behavior. Walton & Cohen 2011 Science; Walton et al. 2023 Science. Powers et al. 2008; Acarturk et al. 2008 Psych Med; Carpenter et al. 2018 Depress Anxiety; Craske et al. 2008/2014 BRAT. Helgeson, Reynolds & Tomich 2006 JCCP. Gilbert et al. 1998 JPSP; Eastwick et al. 2008 JESP; Levine et al. 2012 JPSP; Lench et al. 2019 JPSP. Goyal et al. 2014 JAMA Intern Med.

Repeated/structural. Lucas, Clark, Georgellis & Diener 2004 Psych Sci; Paul & Moser 2009 JVB. Wanberg et al. 2005 JAP; Wanberg, Zhu & van Hooft 2010 AMJ. Wang, Jones & Wang 2019 Nat Commun 10:4331. Verbeke & Bagozzi 2000 J Marketing. Bruch & Newman 2018 Sci Adv; Strübel & Petrie 2017/2022. Pascoe & Smart Richman 2009 Psych Bull; Paradies et al. 2015 PLoS ONE; Emmer et al. 2024 Psych Bull. Davidson et al. 2019 JPBA.

Contrarian. Credé, Tynan & Harms 2017 JPSP. Vanhove et al. 2016 JOOP; Bonanno 2021 EJPT; Galatzer-Levy, Huang & Bonanno 2018; Infurna & Luthar 2016 Perspectives. Frazier et al. 2009 Psych Sci; Jayawickreme et al. 2021 J Personality. Hagger et al. 2016 Perspectives; Baumeister, Tice & Bushman 2023 Perspectives. Denrell 2003 Org Sci. Holt-Lunstad et al. 2010 PLoS Med / 2015 Perspectives; Liang et al. 2024 Nat Hum Behav.

Clinical. Slavich et al. 2009 J Soc Clin Psych 28:223–243 (N = 27 — note this is frequently miscited to Archives of General Psychiatry). Chu et al. 2017 Psych Bull; Franklin et al. 2017 Psych Bull. Takizawa, Maughan & Arseneault 2014 Am J Psychiatry. Masi, Chen, Hawkley & Cacioppo 2011 PSPR.