Published on 24 September 2026
Anxiety disorders, mortality, and suicide: what does a synthesis of 165 studies show?
The essentials
A systematic review with meta-analysis and meta-regression of 165 observational studies, published in World Psychiatry, examines mortality in people with an anxiety or stress-related disorder. Compared with the general population, these individuals show a pooled suicide mortality risk ratio of 2.88 (95% CI 2.13 to 3.89, 39 studies) and an all-cause mortality risk ratio of 1.54 (95% CI 1.14 to 2.08, 42 studies). Natural-cause mortality is also increased, though to a markedly smaller degree, at 1.25 (95% CI 1.09 to 1.44, 19 studies). These figures share a single weakness, and it is a major one: between-study heterogeneity reaches 98% to 100% on the main outcomes, meaning that almost all the dispersion in results reflects genuine differences between studies rather than sampling chance. A pooled average computed under these conditions poorly summarizes what it is meant to summarize, and every pooled estimate must be read with this caveat attached. A second limitation: publication bias is identified, and once accounted for, the association with all-cause mortality, natural-cause mortality, and cardiovascular mortality becomes non-significant. A third limitation, the most instructive of the three: when anxious individuals are matched to controls on comorbidities, the excess in all-cause mortality is no longer significant for most diagnostic subgroups, with the exception of generalized anxiety disorder and panic disorder. This disappearance does not demonstrate that anxiety alone carries no risk: it shows that comorbidities account for a large share of the observed association. The design remains observational throughout, and no causal inference is possible. What this leaves for the consulting room is a signal calling for vigilance, suicide first, somatic health second, not a mechanism.
Context
In private psychiatric practice, anxiety disorders occupy a place their reputation does not reflect. They are common, they are chronic, and they are often regarded as the least serious part of outpatient psychiatry, the part that does not raise alarm. The question of mortality is rarely asked here, even though it is asked spontaneously in front of a major depressive episode or a bipolar disorder. It is precisely this asymmetry that the synthesis interrogates.
What the literature already documented was the link between depression and suicide, and between severe mental disorders and reduced life expectancy, largely through somatic causes. For anxiety disorders, data existed but were scattered across national registries, clinical cohorts, and primary-care follow-up studies, with widely differing diagnostic definitions and follow-up durations. The authors note that earlier meta-analyses diverged, some reporting significantly elevated risk ratios and others not, depending on whether they relied on established diagnoses or on mere anxiety symptoms. A synthesis was therefore warranted. It runs, however, into a structural difficulty: pooling heterogeneous studies does not make them comparable, it only makes their incomparability measurable.
What this kind of work cannot produce must also be stated up front. All included studies are observational, randomized trials having been explicitly excluded. They describe people carrying an anxiety diagnosis and compare their mortality to that of other groups. No design of this kind establishes that an anxiety disorder causes premature death. Anxious individuals differ from others in far more than their anxiety: tobacco and alcohol use, comorbid depression, socioeconomic precarity, access to care, somatic illness already present at the time of diagnosis. The association is the result, causality remains beyond the reach of the design.
The methodological point
What a pooled estimate permits when heterogeneity reaches 99%
The I² statistic expresses the share of variability in results attributable to differences between studies rather than to sampling error. The values reported here, 98% to 100% on the main outcomes, do not describe strong heterogeneity in the usual sense. They describe a situation in which the studies are not estimating the same quantity. The pooled risk ratio then keeps a descriptive status: it indicates where the center of gravity of a highly dispersed set of results lies, it does not provide the estimate expected for a given patient or a given care system.
The practical consequence is direct. A confidence interval describes the precision of the mean, it does not describe the range of plausible values in a new population. With an I² of this magnitude, the prediction interval, which would answer that second question, would likely be far wider than the confidence interval reported, and could well include values showing no excess risk. This prediction interval is not reported by the authors, who rely on the random-effects model and the confidence interval alone. Citing 2.88 without mentioning the heterogeneity lends this value a stability it does not have.
This heterogeneity has identifiable sources: the definition of anxiety disorder, ranging from a structured clinical diagnosis to an administrative reimbursement code, the source of mortality data, follow-up duration, historical period, with observation windows spanning from 1947 to 2023, the adjustments applied, and above all the comorbidity composition of the groups compared. It is this last source that the synthesis addresses head-on through its matched analyses, and it is what makes those analyses more informative than the headline estimate.
Two further tools complete the picture. Egger’s test explores funnel-plot asymmetry. The authors set a suspicion threshold at p below 0.1: the test comes out at 0.054 for all-cause mortality, at 0.0021 for natural-cause mortality, at 0.0057 for cardiovascular mortality, and at 0.042 for cancer mortality. Asymmetry alone does not prove that negative studies were withheld. The fail-safe N then quantifies the fragility of the result: it stands at 42 for all-cause mortality, at 23 for natural-cause mortality, and at 10 for cardiovascular mortality, and in all three cases the association becomes non-significant once publication bias is accounted for. No trim-and-fill-adjusted risk ratio is reported in the text of the article.
The study at a glance
| Question (PICO) | |
|---|---|
| Population | |
| People with an anxiety disorder or a stress-related disorder, defined by DSM or ICD criteria through diagnostic interview or medical record: generalized anxiety disorder, panic disorder, post-traumatic stress disorder, phobias, and a mixed anxiety or stress-related disorder category used when studies did not distinguish subtypes. Studies identifying anxiety solely through rating scales were excluded. | |
| Exposure | |
| Presence of an anxiety diagnosis, treated as the exposure. No intervention is evaluated, no treatment is compared. | |
| Comparators | |
| Three comparisons, conducted separately: the general population, controls without an anxiety disorder matched on somatic or psychiatric comorbidities, and people with other mental disorders. | |
| Outcomes | |
| Co-primary outcomes: all-cause mortality and suicide mortality. Secondary outcomes: natural-cause mortality, non-natural-cause mortality excluding suicide, cancer mortality, cardiovascular mortality, suicide attempt. | |
| Corpus | |
| 165 studies, 7,395,722 people with an anxiety or stress-related disorder and 135,059,023 controls, across 27 countries. 133 cohort studies (80.6%) and 32 case-control studies (19.4%). Median follow-up 7.5 years (interquartile range 3 to 11). Mean age 49.9 years, 35.7% women. Eighty-nine of 165 studies (53.9%) report all-cause mortality. | |
| Design | |
| Systematic review, random-effects meta-analysis, and meta-regression. Protocol publicly registered on OSF before the review was conducted, reporting compliant with PRISMA 2020, search across three databases (PubMed via Ovid MEDLINE, PsycINFO, EMBASE) through August 15, 2024, supplemented by a manual search of references in prior systematic reviews. Quality of included studies assessed with the National Institutes of Health rating tool, adapted to the design. |
Quality control
| Point checked | Status |
|---|---|
| Protocol registration | Registered before the review |
| FindingProtocol made public on OSF before the review was conducted, which limits the risk of outcome selection after the fact. | |
| Conduct of the review | Complete |
| FindingReporting compliant with PRISMA 2020, three bibliographic databases searched through August 15, 2024, supplementary manual search, selection and extraction by five independent reviewers using Covidence. | |
| Volume of data | 165 studies |
| FindingA considerable corpus: 7,395,722 exposed individuals, 135,059,023 controls, 27 countries. It ensures broad literature coverage, it does not in any way offset the heterogeneity between the studies pooled. | |
| Quality of included studies | Mostly fair |
| FindingUsing the National Institutes of Health tool, 96 studies are rated fair quality, 62 good quality, and 7 poor quality. Excluding them does not change the results for the co-primary outcomes. | |
| Funding and conflicts of interest | Not reported |
| FindingThe version consulted contains neither a funding statement nor a conflict-of-interest statement. No conclusion can therefore be drawn in either direction. The acknowledgments mention only a methodological contribution to patient and public involvement. The work concerns no health product, which mechanically limits the promotional stakes, without amounting to a declaration. | |
| Heterogeneity | Major problem |
| FindingI² of 98% to 100% on the main outcomes. The legitimacy of statistical pooling itself is questionable, and every pooled estimate must be cited with this caveat. | |
| Publication bias | Identified |
| FindingEgger’s test suggestive for all-cause mortality (p = 0.054, threshold set by the authors at 0.1), for natural-cause mortality (p = 0.0021), cardiovascular mortality (p = 0.0057), and cancer mortality (p = 0.042). Once accounted for, these associations become non-significant. No publication bias emerges, by contrast, for suicide mortality or for suicide attempt. | |
Results
The results trace a coherent gradient: the excess risk observed is clear-cut for suicide deaths and suicide attempts, modest and fragile for all-cause and natural-cause mortality, and largely dissipates once comorbidities are accounted for through matching.
| Outcome | Estimate reported |
|---|---|
| All-cause mortality, versus general population | RR 1.54 (95% CI 1.14 to 2.08), 42 studies, I² = 99%, p = 0.005 |
| PEB readingA modest signal whose lower bound comes close to no effect at all, against a backdrop of extreme heterogeneity and identified publication bias. This is the least robust result in the series: it does not survive accounting for publication bias, with a fail-safe N of 42. | |
| Suicide mortality, versus general population | RR 2.88 (95% CI 2.13 to 3.89), 39 studies, I² = 100%, p < 0.001 |
| PEB readingThe most pronounced association, with an interval entirely above 1 and no publication bias detected. The direction of the signal is convincing, its magnitude is not, heterogeneity being at its maximum. Retain an order of magnitude, not a precise value. | |
| Suicide attempt, versus general population | RR 2.47 (95% CI 1.78 to 3.42), 15 studies, I² = 100%, p < 0.001 |
| PEB readingSame direction as suicide mortality, with an interval that stays clearly above 1. Subgroup values range from 6.33 (95% CI 4.08 to 9.82, 5 studies) in panic disorder to 2.74 (95% CI 1.72 to 4.35, 5 studies) in phobias, on small study numbers that call for caution. | |
| Somatic mortality, versus general population | Natural-cause RR 1.25 (95% CI 1.09 to 1.44), 19 studies; cardiovascular RR 1.26 (95% CI 1.06 to 1.50), 16 studies; non-natural cause excluding suicide RR 1.94 (95% CI 1.47 to 2.55), 10 studies; cancer not significant, 10 studies |
| PEB readingThese excesses are real, they are small, and they are fragile: once publication bias is accounted for, the association with natural-cause mortality and with cardiovascular mortality no longer holds. Cancer mortality is not increased in the overall analysis (p = 0.81), the only positive signal coming from a single study in generalized anxiety disorder. | |
| After matching on comorbidities | All-cause mortality not significant for most diagnostic subgroups, except generalized anxiety disorder (RR 1.46, 95% CI 1.23 to 1.73, 7 studies) and panic disorder (RR 1.12, 95% CI 1.04 to 1.21, 2 studies) |
| PEB readingThe most useful result of the whole synthesis. It does not prove that anxiety alone carries no risk: absence of a significant difference is not proof of absence of effect, and these analyses necessarily have lower power, the panic disorder exception resting on only two studies. It shows that comorbidities account for a large share of the observed association with overall mortality. The risk of suicide attempt, meanwhile, remains significantly elevated after matching (RR 2.43, 95% CI 1.16 to 5.11, 8 studies). | |
| Versus other mental disorders | No difference on co-primary outcomes beyond two studies |
| PEB readingA comparison very sparse in data. The rare significant results rest on one or two studies: lower all-cause mortality than other mental disorders for PTSD and for the mixed category, higher suicide attempt risk for phobias and for PTSD, each time based on a single study. Nothing can be concluded from this for practice. | |
The ranking among anxiety diagnoses depends closely on the outcome considered, which rules out any general hierarchy. For all-cause mortality compared with the general population, only generalized anxiety disorder (RR 1.48, 95% CI 1.23 to 1.78, 9 studies) and PTSD together with other stress-related disorders (RR 1.39, 95% CI 1.15 to 1.67, 21 studies) reach significance; panic disorder, phobias, and the mixed category do not. For suicide mortality, the order partly reverses: panic disorder carries the highest estimate (RR 3.58, 95% CI 1.39 to 9.25, 3 studies), ahead of PTSD (RR 3.13), the mixed category (RR 2.77), and generalized anxiety disorder, which comes last (RR 1.93, 95% CI 1.17 to 3.17, 3 studies). The authors state no diagnostic hierarchy, and the number of studies per subgroup, sometimes as few as three, would not support one.
The meta-regressions add a final element. The excess in all-cause mortality decreases significantly as the period of data collection becomes more recent (beta = -0.039, 95% CI -0.070 to -0.008, p = 0.015). A higher frequency of schizophrenia-spectrum comorbidity (beta = -0.148, p < 0.001) or bipolar comorbidity (beta = -0.122, p = 0.024) is associated with lower suicide mortality, which the authors cautiously interpret as a possible effect of the treatment received. Comorbid depression, for its part, has no significant effect on all-cause mortality (p = 0.55) or on suicide mortality (p = 0.59). These results rest on variables often poorly reported in the primary studies, as the authors themselves note.
Critical appraisal
| Domain | Judgment |
|---|---|
| Nature of the data | Observational |
| PEB readingNo randomized study, no intervention evaluated. The work describes associations between a diagnosis and mortality. Any wording suggesting that anxiety causes these deaths goes beyond what the design allows. | |
| Heterogeneity | Major weakness |
| PEB readingWith an I² of 98% to 100%, the pooled risk ratios mask extreme variability between studies. They keep value as a direction and lose almost all value as a magnitude. | |
| Residual confounding | Not controlled |
| PEB readingComorbid depression, substance use disorders, somatic illness, smoking, and psychological interventions: all factors difficult to measure uniformly across 165 studies, and which the authors themselves describe as under-reported in the primary studies. The direction of the distortion is not one-way: most of these factors pull the crude association upward, but the meta-regressions also show comorbidities associated with lower suicide mortality. | |
| Publication bias | Present |
| PEB readingIdentified for all-cause, natural-cause, cardiovascular, and cancer mortality, with loss of significance once accounted for the first three. It affects the broadest outcomes, the ones most often repeated in communication, and it spares the suicide-related outcomes. | |
| Comorbidity-matched analyses | Real contribution |
| PEB readingA demanding approach, rarely carried out at this scale, that usefully shifts the reading. It points to where much of the association resides, without allowing a conclusion that anxiety alone is harmless. The authors themselves note that the comorbidities considered were highly heterogeneous and the analyses sparsely populated. | |
| External validity | Limited |
| PEB readingThe studies come mainly from high-income countries in North America and Western Europe, the United States alone contributing 71 studies. The proportion of women, 35.7%, is markedly below the expected female prevalence of anxiety disorders, which the authors attribute to the weight of US veteran cohorts. | |
| Match between claim and evidence | Respected |
| PEB readingThe conclusions reported remain nuanced, the limitations are spelled out, the subgroup analysis is presented as such. The risk of overstatement lies in how the headline figure gets reused out of context, not in the source text. | |
Level of evidence
PEB appraisal: confidence is high regarding the conduct of the review, protocol registered before the search, reporting compliant with PRISMA 2020, three databases searched, selection by five independent reviewers, and quality rating of included studies with the National Institutes of Health tool, and also regarding the direction of the signal for suicide mortality, found across a large number of studies and with no publication bias detected. It is low regarding the magnitude of every pooled estimate, owing to a heterogeneity of 98% to 100% that makes the average poorly representative of the studies composing it. It is also low for all-cause mortality and for somatic mortality, which do not survive accounting for publication bias. It is nil for any causal reading: this work establishes an association, it says nothing about what produces it or what would modify it. What is demonstrated is that people followed for an anxiety disorder die by suicide more often than the general population in the cohorts pooled here, and that they make suicide attempts more often. What is suggested is the excess in all-cause and somatic mortality, smaller and sensitive to publication bias, along with the role of comorbidities, which the matched analyses point to without quantifying. What amounts to opinion is any stable hierarchy between anxiety diagnoses: it changes direction depending on the outcome chosen.
The colleague test
What an experienced colleague might say about this study in two minutes, between two consultations.
“I’ll take the message, I won’t take the number. That suicide mortality runs about three times that of the general population in these cohorts, I’m happy to believe it, it matches what I see. But with an I² of 100%, I’m not going to announce 2.88 as if it were a lab result. And what interests me most is that the excess in overall mortality fades once you match on comorbidities: that doesn’t tell me anxiety is benign, it tells me where to look.”
Translated for practice: the finding that changes something is not the pooled estimate, it is what that estimate becomes once you account for what accompanies the anxiety disorder. In an anxious patient, assessing suicide risk and screening for comorbidities are the same clinical gesture.
What you can do with this
- Ask explicitly about suicidal ideation in patients followed for an anxiety disorder, with the same regularity as in the follow-up of a depressive episode, and document it in the record.
- Revisit the inventory of comorbidities in your long-term anxious patients: comorbid depression, alcohol or benzodiazepine use, active somatic illness, social isolation. This is where most of the described association concentrates.
- Do not carry the figure of 2.88 into a consultation or a teaching session without mentioning the 98% to 100% heterogeneity that accompanies it. An order of magnitude can be cited, a precise value cannot be defended.
- Treat the excess in somatic mortality with the same caution: it exists in the raw data, it is small, and it does not survive accounting for publication bias. Somatic follow-up in these patients keeps its own justifications and does not need this one.
- Do not convert this data into an individual prognosis. It describes cohorts, it predicts nothing for a given patient, and an observed association says nothing about what produced it.
- Guard against the reverse mistake: the disappearance of excess all-cause mortality after matching does not justify relaxing vigilance in an anxious patient without apparent comorbidity, particularly since generalized anxiety disorder and panic disorder are exceptions, and the risk of suicide attempt remains elevated after matching.
- If you pass this work on to a non-psychiatrist colleague, pass on the nuance along with the number. This is exactly the kind of figure that travels fast and travels badly.
- The course of action is set out in the NICE decision tree for generalised anxiety and panic disorder.
Frequently asked questions
Can a patient be told that their anxiety disorder shortens life expectancy?
No. This data does not establish that, and the observational design would not allow it even with more robust figures. It describes higher mortality in groups of diagnosed individuals, without identifying what produces it. Turning this into an individual prognosis would be both methodologically wrong and clinically harmful, particularly in patients whose disorder centers precisely on anticipating the worst.
Does an I² of 99% make the meta-analysis unusable?
It does not cancel it, it changes what can be drawn from it. Dozens of studies converging on an excess in suicide mortality remains an argument for direction, especially since the confidence interval sits entirely above 1. The average magnitude, however, loses its practical meaning: it does not represent the studies included, still less any particular population. It is a qualitative finding reached by quantitative means.
If the excess mortality disappears after matching on comorbidities, is anxiety alone harmless?
That conclusion would be a misreading. A non-significant difference is not proof of no difference, and these matched analyses rest on few studies, with comorbidities that vary widely from one study to the next. What matching shows is that comorbidities account for a large share of the observed association with all-cause mortality. Two subgroups escape this pattern, generalized anxiety disorder and panic disorder, and the risk of suicide attempt remains significantly elevated.
Does this call for particular somatic surveillance in anxious patients?
The synthesis does report an excess in natural-cause mortality (RR 1.25) and in cardiovascular mortality (RR 1.26), and the authors draw from this an explicit call not to neglect general somatic care in these patients. Two caveats accompany this finding: the magnitude is small, and both associations lose significance once publication bias is accounted for. In other words, this data alone does not justify a specific surveillance protocol. Somatic follow-up of psychiatric patients keeps its own justifications, sturdier than this one.
Does the finding hold equally for post-traumatic stress disorder and for generalized anxiety disorder?
The population includes stress-related disorders, and these two categories are the only ones showing a significant excess in all-cause mortality against the general population. But the order changes with the outcome: for suicide mortality, it is panic disorder that carries the highest estimate and generalized anxiety disorder the lowest. The authors state no hierarchy between diagnoses, and the number of studies per subgroup, sometimes as few as three, would not support one. As things stand, a principle of uniform vigilance is safer to retain than a ranking.
What does this work change for patient management?
Nothing on the therapeutic side: no intervention is evaluated, and nothing here supports the claim that any treatment would change this mortality. What can change is clinical attention, meaning the weight given to suicide-risk assessment and comorbidity screening in a follow-up that, without it, tends to become routine.
Annotated bibliography
Source study. Wagner E, Mortazavi M, Poddighe L, Baldwin DS, Masdrakis V, Castle DJ, Serretti A, Oliva V, Fanelli G, Fornaro M, Shin JI, Colman I, Semchishen SN, Anderson KK, Wang JL, Brietzke E, Sabé M, Cortese S, Domschke K, Hasan A, Chang WC, Myran DT, Correll CU, Connor Gorber S, Højlund M, Solmi M. Risk of all-cause and cause-specific mortality, and suicide attempt in people with anxiety and stress-related disorders: a systematic review, meta-analysis and meta-regression analysis of 165 studies. World Psychiatry 2026; 25(2): 307-320. DOI 10.1002/wps.70072. PMID 42136520. Systematic review with meta-analysis and meta-regression of 165 observational studies, covering 7,395,722 people and 135,059,023 controls across 27 countries. Protocol registered on OSF, reporting compliant with PRISMA 2020, quality of included studies assessed with the National Institutes of Health tool. What it contributes: a pooled estimate of the excess mortality, suicide-related mortality above all, associated with anxiety and stress-related disorders, and above all comorbidity-matched analyses that shift the interpretation. Its limitations: heterogeneity of 98% to 100% on the main outcomes, which strips the pooled estimates of value as a magnitude, publication bias that erases the significance of all-cause, natural-cause, and cardiovascular mortality, under-representation of low- and middle-income countries, and residual confounding that is unavoidable in a body of work that is exclusively observational. The version consulted contains neither a funding statement nor a conflict-of-interest statement.
No context reference is cited in this version. Mortality and suicide data in other mental disorders, often invoked for comparison, are not reproduced here for lack of having been verified one by one, and an approximate reference is worth less than no reference at all.
What was consulted. Verification carried out on August 12, 2026 on the full text of the published version, 40 pages, including figures, tables, acknowledgments, and reference list. The published title, the complete author list, the volume, the pagination, the DOI, and the PMID were recorded directly from it. The supplementary material, deposited by the authors on an external platform, was not provided to us and was not consulted: the analyses that appear only there, notably the detail of the matched comorbidities and the search strategy, were therefore not verified. The version consulted contains neither a funding statement nor a conflict-of-interest statement, and neither could be established. This analysis underwent an independent double reading.
Editorial collections
Tags
Verified on August 12, 2026 against the full text of the publication and its supplementary material where available. This analysis underwent an independent double reading. The English version was checked for conformity on September 23, 2026, against the figures of the French version and against the source. How we verify what we publish
This analysis is intended for healthcare professionals. It does not constitute a prescribing recommendation and does not replace individual clinical judgment.
Analysis from Psychiatry Evidence Base, evidence-based psychiatry, explained with rigor.