Contribution
The largest assessment of the mortality-salience (MS) literature, and the first to correct it for publication bias and p-hacking. It codes the literature twice: once as Burke et al. (2010) did, for simple MS-versus- control effects, and once for the focal test of each study, as p-curve requires. It runs five tools on the result and reports that they disagree. Its second contribution is methodological: no single tool should be trusted to judge a literature.
Key insight
Burke et al.’s moderate effect is reproduced exactly when the same uncorrected method is used, and shrinks by about half, or to nothing, when publication bias is modelled. The literature is also so heterogeneous that an average says little about any one paradigm: under the selection model, 95% of the true effects lie between g = −1.05 and 1.77 (r = −.47 to .66). The authors’ conclusion (Conclusion): the literature “contain[s] studies that appear to be testing nonzero effects, although the literature is highly heterogeneous and underpowered, rendering many individual effects to be likely spurious.”
Assumptions
- TMT in brief (Introduction). People manage the anxiety of knowing they will die by defending cultural worldviews “to achieve feelings of symbolic immortality”. Because a culture outlasts its members, they share its immortality by bolstering it or living up to its standards.
- Only the MS hypothesis is assessed, not the anxiety-buffer or death-thought-accessibility hypotheses. Defenses are distal (worldview defense, strongest after a delay) or proximal (such as resolving to live more healthily, when death is consciously in mind).
- Direction is taken from the authors. Effects in the predicted direction are coded positive, against it negative. The authors grant that HARKing can bias this.
- Published, English, peer-reviewed only. Dissertations were excluded so that p-curve, which uses only significant results, and the other tools work on comparable literatures.
- Not preregistered. An analysis plan was posted on OSF at the outset, but the analyses changed in revision (Statement of Limitations).
Key results
- Search (Methods). PsycINFO, prior reviews’ references and the TMT website gave 1,373 manuscripts. After exclusions, the remaining papers held 1,396 studies.
- Burke et al. as reported (Past Evaluations; Results of Traditional Meta-Analysis). The most extensive earlier meta-analysis: random effects over 277 studies, average r = .35, no bias correction. Yen and Cheng (2013) re-analysed it and found experimenter effects accounted for much of the variance. This review adapts Burke et al.’s search and selection rules so its uncorrected result can be compared directly.
- Traditional meta-analysis: 643 effects coded, 8 with g > 5 dropped
as likely reporting errors, leaving 635 tests from 557 studies.
- g = 0.805, 95% CI [0.757, 0.854], τ = .533; about r = .37. Heterogeneity Q(634) = 2972.7.
- The funnel plot is strongly asymmetric.
- Dropping the largest 10% (64 effects): g = 0.622, about r = .30.
- Selection model (three thresholds, JASP defaults). 430 significant,
69 marginal, 130 non-significant and 6 “surprising” effects.
- Non-significant effects are estimated published at .259 the rate of significant ones.
- Adjusted g = 0.357, 95% CI [0.168, 0.546], τ = 0.716, p < .001; about r = .18. Without the largest 10%: g = 0.457, about r = .22.
- PET-PEESE.
- PET gave a significant negative estimate: g = −0.114, 95% CI [−0.202, −0.026], t(633) = −2.527, p = .012. By the procedure’s rule, no positive PET effect means no evidence for MS, and PEESE is not read.
- The authors judge the negative estimate spurious: PET underestimates under heterogeneity, small samples and p-hacking.
- Without the largest 10%: PET g = 0.035, p = .322; still no effect.
- WAAP-WLS. Only 139 of 635 effects (21%) had 80% power. Their estimate: g = 0.385, 95% CI [0.340, 0.430]; about r = .19. Without the largest 10%: g = 0.354.
- p-curve (825 studies; one or two focal tests each).
- Significant right skew: half curve Z = −13.76, full curve Z = −14.23, both p < .001. So some studies test non-zero effects.
- Average power 25%, 90% CI [21%, 29%], below the 33% benchmark.
- Mean N was 114, about 57 per cell, enough to detect r = .18 with 27% power.
- Without the largest 10%: right skew holds (full Z = −4.08; half Z = −1.42, p = .0785), but estimated power falls to 8%.
- z-curve (the same 825 tests).
- Expected replication rate (ERR) .19, 95% CI [.13, .26]: an exact replication of a published significant effect would succeed about one time in five.
- Expected discovery rate (EDR) .08, 95% CI [.05, .17], close to the 5% a literature of only null effects would give.
- Without the largest 10%: ERR and EDR both .06.
- Preregistered replications. Of 9 found, 5 supported the MS hypothesis (Arnoult et al. 2023; Courtney et al. 2022; Dunn et al. 2020; Schindler et al. 2019; Vail et al. 2019) and 4 did not (Rodríguez-Ferreiro et al. 2019; Sætrevik and Sjåstad; Schindler et al. 2021 and 2022). Those studies were powered on published, likely inflated, effects.
- Many Labs 4 (Past Evaluations; General Discussion). Reported as 21 labs, 2,220 participants and g = 0.03. Chatard et al. (2020), the original authors’ critique, re-analysed only the sites that met the preregistered minimum and found g = 0.13, p = .045, for the author-advised sites under one exclusion rule, and nothing under the others.
- Two takes (What Can We Conclude).
- Liberal: p-curve, selection models and WAAP-WLS find an effect; five of nine preregistered studies replicated; studies after 2011, and those with longer delays or at least two distractor tasks, look better in the SOM.
- Conservative: z-curve’s EDR is near the null; PET-PEESE finds nothing in any subgroup; Many Labs 4 failed; conditional power is 19–25%.
- Reconciled: effects are real in some studies, highly heterogeneous and mostly underpowered. Replications should assume r = .18, follow expert protocols, and expect conceptual replications to fail more often.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Uncorrected, the MS literature yields a moderate-to-large average effect (about r = .37), matching Burke et al. 2010 | strong as a replication of the method; the authors reject the estimate as inflated | traditional meta-analysis |
| C2 | Corrected for publication bias, the average MS effect is small, about r = .18 | moderate: two tools agree; a third finds no effect; all three are known to misbehave under this literature’s heterogeneity and p-hacking | selection models, WAAP-WLS, PET-PEESE |
| C3 | Some published MS studies test non-zero effects | moderate to strong: p-curve right skew survives dropping the largest effects; p-curve’s null is that every effect is zero | p-curve, z-curve ERR |
| C4 | Most MS studies are underpowered and many published effects are likely spurious | moderate: 25% (p-curve) and 19% (z-curve) conditional power, 8% unconditional power, 21% of effects adequately powered | p-curve, z-curve, WAAP |
| C5 | MS effects vary so much across designs that an average is of limited use | strong: τ = .72; the 95% range of true effects spans negative to large positive | selection model |
| C6 | Replications fare better with expert guidance and strict protocols | weak: rests on Chatard et al.’s re-analysis under one exclusion rule and on SOM subgroups not read here | General Discussion |
Concepts
- mortality salience (MS) hypothesis: reminders of one’s death increase defense of one’s cultural worldview.
- symbolic immortality: in TMT, the sense of outliving death by belonging to, upholding or living up to a culture that will outlast one.
- distal and proximal defenses: worldview defense after death thoughts leave attention, against direct responses such as health resolutions while death is in mind.
- selection model, PET-PEESE, WAAP-WLS: corrections that model publication by p-value, regress effect on standard error, or keep only adequately powered studies.
- p-curve; ERR and EDR: the distribution of significant p-values; z-curve’s power of published significant studies (ERR) and of all studies run, file drawer included (EDR).
Connections
- Burke, Martens and Faucher 2010 (LIT-1687), the work this note stands in for. The review takes its search and selection rules, its pooled r = .35 over 277 studies, and its subgroup ideas (with Pyszczynski et al. 2015). It reproduces the uncorrected result and then corrects it.
- Many Labs 4 (Klein et al. 2022), read alongside. The review’s figures for it (21 labs, 2,220 participants, g = 0.03) do not match the accepted version the record read (2,281 collected, 1,550 analysed, g = 0.07). They may come from the first preprint.
- Lifton (LIT-1714) is not cited. The review traces symbolic immortality to Becker (1973).
- Castano et al. (LIT-1719) is not named. Its outcomes, in-group identification and bias, are among the many worldview-defense measures the review pools.
Bearing on the record
- The legacy question (QUESTION-301). The question cites
Burke et al. as the evidence for TMT, whose account is the record’s
route from death to an enlarged self: people defend the worldview and
group that will outlast them, for symbolic immortality. What this
review changes, as inference:
- The effect is smaller and less certain than the question assumes. The best-supported average is about r = .18, half Burke et al.’s; one correction finds none; the expected discovery rate is near the null. Whether a given study’s effect is real cannot be read from its publication.
- It is not gone. p-curve’s right skew says some MS studies test real effects. The review cannot say which. Subgroups with longer delays and post-2011 studies look better, by the authors’ report of an SOM the record has not read.
- The mechanism is untested. The review assesses whether death reminders change worldview defense, not whether the change is a search for symbolic immortality. That reading belongs to the theory. Even a robust effect would support the extended-self interpretation only as the theory’s gloss.
- The owner’s thesis (CLAIM-491). The thesis does not depend on TMT. TMT is a candidate mechanism for enlarging the self under threat, across a group and past death. It is now a weak one to cite: a small, highly variable effect whose core paradigm failed in Many Labs 4.
- The triviality objection (CLAIM-492). The MS literature’s outcome measures, worldview defense in many forms, record evaluative attachment to one’s group and its standards. Those are measures of caring, not of the self’s extent apart from caring. The review offers no criterion of the kind the objection asks for (an inference).
- A dated note should be added to QUESTION-301, jointly with Many Labs 4. The proposed text is in the Many Labs 4 draft; the part this review supplies: “Chen et al. (NOTE-1271) reproduce Burke et al.’s uncorrected effect (r = .37 over 635 effects, against .35 over 277) and then correct it for publication bias: selection models and WAAP-WLS give about r = .18, PET-PEESE no effect, and z-curve an expected discovery rate of 8%. Some mortality-salience effects are real, but the average is about half the published one and most studies could not detect it.” No note is needed on CLAIM-491.
- LIT-1687’s summary should be checked. “A moderate effect” is right for Burke et al. (r = .35). The record now holds a source for the figure but not for its moderators, such as delay.
- No THEORY filed.
- No ML instruction.
Limitations
- The figures and SOM were not read. The subgroup results behind the “liberal take” (delay, distractors, post-2011 studies, country) are reported only in summary.
- Quality was not checked study by study, by the authors’ own account. A few errors were found and corrected or dropped.
- Pooling across heterogeneous paradigms. The authors say an average over well- and badly-designed studies may misrepresent both.
- Inconsistencies in the copy.
- The search arithmetic does not close: 1,396 studies less 578 excluded is 818, not the 825 analysed.
- Mean per-cell N is given as about 57 (Results of P-Curve) and as about 28 (A Conservative Take).
- Table 1 gives PET as g = −0.14 and WAAP as “13% of effects”; the text gives −0.114 and 21% (139 of 635).
- Many Labs 4 is described as 21 labs and 2,220 participants with g = 0.03, against 2,281, 1,550 and 0.07 in its accepted version.
- Sætrevik and Sjåstad are dated 2022 in one place and 2019 in another.
- Typos. “Hedge’s g” throughout; “Stanley and Douliagos”; “Iaonnidis”; “Simmons, Nelson, & Simonsohn, 2014a” for Simonsohn et al.; “expected replication rate (EDR)” where ERR is meant; “worth nothing” for “worth noting”; “performed in in the R”; “Sladevokova” in the references; a garbled sentence in Results of Selection Models where text has dropped out (“nonsignificant effects (one tailed .05 ≤ pour universe of studies yielded”).
Open questions
- Which MS paradigms carry the real effects? The authors’ Shiny application and coded data allow subgroup p-curves; a subgroup of studies whose outcome is identification with a lasting group (the Castano et al. design) would bear most directly on the legacy question.
- Does any MS study measure symbolic immortality, the putative mediator, rather than inferring it from worldview defense?