Oral health and longevity: what four million people in a meta-analysis do and do not show

Thirty-nine cohorts, 4,389,263 people: periodontal disease carries a relative risk of 1.31 for all-cause mortality. Mendelian randomization in the same field finds no evidence of genetic causality. Both are in the same literature and both are correct.

mixedStudies disagree, or the mechanism is contested.Revised 14 Oct 2026

A white ceramic mug and a polished steel dental mirror on a folded sage-green cloth on a windowsill, a blurred green view outside

A relative risk of 1.31 sounds small until you multiply it by the population in the meta-analysis. It is also the kind of number that a genetic instrument failed to reproduce.

Four million three hundred and eighty-nine thousand two hundred and sixty-three people, pooled across 39 prospective cohorts, carry a relative risk of 1.31 for all-cause mortality if they have periodontal disease.

That is one of the largest pooled associations in preventive medicine. It is also, by the standards of causal inference, an unusually soft one, and the same literature contains a Mendelian randomization analysis that failed to reproduce any genetic evidence of causality. Both findings are published. Almost nobody reports both.

The meta-analysis, in full

The 2023 paper in PLoS One is the reference, and the numbers deserve to be read together rather than selectively:

major adverse cardiovascular events: RR 1.24 (1.15 to 1.34)
coronary heart disease: RR 1.20 (1.12 to 1.29)
myocardial infarction: RR 1.14 (1.06 to 1.22)
stroke: RR 1.26 (1.15 to 1.37)
cardiac death: RR 1.42 (1.10 to 1.84)
all-cause mortality: RR 1.31 (1.07 to 1.61)

Thirty-nine cohort studies, 4,389,263 individuals, random-effects model, and the authors report that sensitivity analyses indicated the pooled conclusions for cardiovascular outcomes and all-cause mortality were robust.

They also state, in the same paragraph, that the associations could be affected by region, study design, periodontal definition, follow-up duration and study quality. That is five different ways the number could be an artefact, listed by the authors themselves.

Every one of these is a relative risk. A 1.31 is not “31% more people die”; it is the ratio of rates, and the absolute difference in a population with a low baseline mortality rate is a great deal smaller than the headline suggests. The all-cause mortality interval, 1.07 to 1.61, is a threefold spread in the point estimate across the included studies, which is itself a signal.

The same pattern appears in the population where the association should be strongest. A 2017 meta-analysis in BMC Nephrology covered 8 cohort studies with 5,477 people with chronic kidney disease. Periodontal disease was associated with all-cause death at RR 1.254 (1.046 to 1.503) with moderate heterogeneity, but showed no evident association with cardiovascular mortality at RR 1.30 (0.82 to 2.06) and substantial heterogeneity. So in people with already-advanced kidney disease — where a chronic inflammatory driver ought to have the most room to act — the cardiovascular association disappears.

Tooth loss and cognition, the same pattern

The 2023 meta-analysis in Frontiers in Neurology pooled 18 cohort studies with 356,297 participants and a mean follow-up of 8.6 years, ranging from 2 to 20.

dementia: RR 1.15 (1.10 to 1.20)
cognitive decline: RR 1.20 (1.14 to 1.26)
Alzheimer’s disease: RR 1.12 (1.02 to 1.23)
vascular dementia: RR 1.25 (1.06 to 1.47)

The interpretation of 1.15 is the thing people skip. In a population with, say, a 15% ten-year dementia risk, a relative risk of 1.15 raises it to about 17%. That is a real difference. It is not a cause, and it is not a mechanism.

The paper’s own subgroup results are the most informative part: pooled estimates varied by geographic location, sex, use of dentures, number of teeth or edentulous status, dental assessment, and follow-up duration. A pooled estimate that moves when you stratify by which teeth are missing, and by where the study was done, is measuring something other than the disease.

The 2021 critical examination in the Journal of Dental Research says this more directly than the meta-analyses do. Three mechanisms have been proposed — tooth loss leading to compromised nutrition, tooth loss resulting in fewer interocclusal contacts and so less somatosensory feedback, and chronic periodontitis causing tooth loss only after the inflammation has already affected the central nervous system. The reviewer’s finding: none of these is supported by compelling empirical evidence.

And then the alternative explanation, which is the one that fits the data. Using a life course approach, the authors argue the association arises from cognitive function much earlier in life: people with better childhood cognitive function go on to keep more teeth, and this is not a periodontal pathway at all. Longitudinal evidence from longstanding cohort studies supports the view that the association between missing teeth and poorer cognitive function in older people originates decades earlier.

If that is right, the mouth is a marker of a childhood, not a cause of a dementia.

What the biology says, and what the instrument says

The 2024 review in Molecular Cells is the most useful single paper in this field, because it takes the causal question seriously and answers it honestly.

Three strands:

Randomized controlled trials of periodontal interventions have demonstrated reductions in surrogate markers of cardiovascular risk, but remain inconclusive regarding their direct effects on cardiovascular events.
Preclinical animal models suggest a causal relationship, but are limited in their ability to definitively prove causality.
The positive associations in epidemiological studies may be influenced by biases such as confounding and collider stratification — and a systematic review of Mendelian randomization studies on the causal relationship between periodontitis and cardiovascular disease found no evidence of genetic causality.

Mendelian randomization is the tool that exists precisely to solve this problem. It uses genetic variants that shift exposure but are unrelated to confounding factors, so a genuine causal path should show up in the instrument test. It did not.

Collider stratification deserves a sentence for non-statisticians, because it is the specific way a real association gets inverted. Conditioning on a common effect of two causes — say, on being admitted to hospital for a cardiovascular event, in a population where both periodontitis and cardiovascular disease raise the risk — can manufacture an association between two things that are not causally linked. In a literature where people with severe periodontitis see dentists and cardiologists, and people with neither do not, this is not hypothetical.

The intermediate pathway has been measured directly, though, and it is real. A 2025 population study in Inflammation Research analysed 9,202 adults free of cardiovascular disease in the 1999 to 2004 NHANES survey. Mean age 40.65. Both mild and moderate-to-severe periodontitis were associated with elevated high-sensitivity cardiac troponin and NT-proBNP — markers of subclinical myocardial injury and impaired cardiac structure and function.

So the chain plausibly runs: periodontal inflammation → subclinical myocardial injury → more cardiovascular events. Every link in that chain except the last has been measured in humans. The last link is where the Mendelian randomization came back empty.

The microbiome result that nobody can interpret yet

The 2026 paper in the Journal of Periodontology is the newest and the strangest. 5,037 adults from the 2009-2010 and 2011-2012 NHANES cycles, mean age 48, salivary 16S rRNA sequencing, full periodontal examination, mortality ascertained through 2019. Median follow-up 8.8 years, during which 81 cardiometabolic and 267 all-cause deaths occurred.

After multivariable adjustment, the microbial indicator of periodontitis — the log-ratio of Treponema, linked with periodontal disease, to Corynebacterium, linked with periodontal health — was associated with an increased risk of cardiometabolic mortality, hazard ratio 2.10 (1.30 to 3.38) per standard deviation.

Neither microbial diversity nor any of the clinical periodontitis measures were associated with cardiometabolic mortality.

Read that again. A ratio of two bacterial taxa doubled the hazard. The pocket depths, the attachment loss, the bleeding, the actual clinical disease that a dentist looks at — none of it predicted. The authors are explicit that longitudinal studies are needed to contextualize the oral microbiota, and they describe this as a prediction, not a finding about causation.

It is possible this is the discovery. It is equally possible that a single time-point ratio in saliva is standing in for diet, smoking, socioeconomic position, diabetes, and dental attendance — all of which predict both what lives in your mouth and how you die. With 81 events, the study is not powered to separate any of that.

The one study that shows the intervention works

The 2004 paper in the Journal of Clinical Periodontology is thirty years old and structurally different from everything else here: it is an intervention study, not an observational one.

More than 550 adults were recruited in 1971 and 1972. 375 formed a test group; 180 formed a control group, which was discontinued after six years. The test group was maintained in a preventive programme — individual case presentation, education in self-diagnosis, and additional education on plaque control and interdental cleaning tailored to need, with visits ranging from every two months in the first two years to every three to twelve months across years 3 to 30 — and re-examined after thirty years.

That is a real demonstration that a well-run oral hygiene programme can be maintained for three decades in a cohort of adults who started middle-aged, and the study measured plaque, caries, probing pocket depth, attachment loss and the community periodontal index of treatment need at baseline and after 3, 6, 15 and 30 years.

What it does not have is a mortality endpoint, or a control group after year six. So it cannot answer the longevity question. What it establishes is that thirty years of structured prevention is achievable, which is a more practical finding than any relative risk.

What these cohorts do not show

No study shows that treating periodontal disease lengthens a life. Not one. The randomized trials, as the Molecular Cells review states plainly, reduce surrogate markers and remain inconclusive for events. Everything on the mortality side is observational.

The Mendelian randomization found no genetic causality, and in a field where a causal claim would change clinical practice, that is the single most decisive negative result available.

The two meta-analyses both report that their pooled estimates move when you stratify by region, by design, by disease definition, by follow-up duration, by tooth count, by denture use, by sex and by dental assessment. A pooled estimate that unstable across subgroups is a summary of a literature, not a measurement of a fact.

The relative risks are relative. 1.31 for all-cause mortality across 4,389,263 people is a rate ratio. The absolute effect depends entirely on the baseline mortality of the population, and neither paper reports an absolute number.

The confounding candidates are not hypothetical. Smoking, diabetes, obesity, socioeconomic position, education, healthcare access and dental attendance all predict both periodontitis and mortality, and the largest cohort in the field — NHANES — is a survey explicitly designed to sample a general population, with all the selection that implies.

No trial has an endpoint of any consequence. Not a cardiovascular event, not a stroke, not a death, not a year of healthy life. The surrogate markers are troponin, NT-proBNP, pocket depth and bacterial ratios.

The tooth-loss literature may not be about the mouth at all. The critical examination’s life course account — childhood cognitive function predicts adult tooth retention — would explain the entire association with no periodontal pathway at all, and it is consistent with the subgroup variation the meta-analysis reports.

Nobody has shown a graded relationship in the direction causality requires. If periodontitis drives cardiovascular risk, more disease should mean more events after adjustment. The mechanism paper shows the biomarkers scale with severity, but the mortality meta-analysis does not test that, and the NHANES mortality study found the clinical measures did not predict at all.

The defensible summary is two sentences. Periodontal disease is associated with cardiovascular events and all-cause mortality across a very large body of cohorts, with relative risks in the 1.1 to 1.4 range, and it tracks with genuinely measurable myocardial injury in the middle of the pathway. And when anyone has tried to test that association with the tools designed to remove confounding — genetics — it has not held. That does not make the mouth irrelevant to longevity. It makes it a well-marked, well-measured, poorly understood correlation, and the only thing anyone has demonstrated you can change about it is plaque.