Social connection and mortality: what 90 cohorts and 2.2 million people show

Social isolation carries a pooled relative risk of 1.32 for all-cause mortality across 90 cohorts; loneliness, 1.14. A 2010 meta-analysis put stronger social relationships at an odds ratio of 1.50 for survival. Every one of these is observational, and not one of them has been through a trial.

mixedStudies disagree, or the mechanism is contested.Revised 23 Oct 2026

Two wooden chairs facing each other beside a window on a pale wall, with a third chair turned slightly away and a mug left on the floor

A relative risk of 1.32 over 2.2 million people is not a small finding. Whether it is a cause of death or a marker of it is a different question, and it is still open.

Two and a quarter million people, followed across ninety cohorts, and the pooled relative risk of dying is 1.32 if you are socially isolated. That is the shape of the finding. What it is not is a result of an experiment, and the distance between those two statements is where most of the writing on this subject goes wrong.

Three meta-analyses, published thirteen years apart, get to the same place. None of them has a control group that was assigned to be lonely.

The largest of them: 90 cohorts, 2,205,199 people

The 2023 review in Nature Human Behaviour registered its protocol (PROSPERO CRD42022299959) and pooled 90 prospective cohort studies with 2,205,199 individuals aged 18 and over. It asked two separate questions, because “loneliness” and “social isolation” are measured with different instruments and are not the same exposure: loneliness is a subjective state, isolation is an objective count of contacts.

In the general population, both were significantly associated with all-cause mortality. Social isolation: pooled effect size 1.32 (95% CI 1.26 to 1.39). Loneliness: 1.14 (95% CI 1.08 to 1.20).

The cause-specific numbers follow the same shape. Cancer mortality: isolation 1.24 (1.19 to 1.28), loneliness 1.09 (1.01 to 1.17). Cardiovascular mortality: isolation 1.34 (1.25 to 1.44). And in people who already had the diagnosis, the association is not softer: 1.28 (1.10 to 1.48) for all-cause mortality among those with cardiovascular disease, and 1.51 (1.34 to 1.70) among those with breast cancer, whose cancer-specific mortality ran at 1.33 (1.02 to 1.75).

Note the arithmetic. 1.32 is a 32% higher rate, not a 32% higher chance of dying — the risk it multiplies depends on the baseline, which differs by age, country and decade. Read the interval, not the point.

The one everybody quotes, and the one nobody should

The 2010 meta-analysis in PLoS Medicine is the origin of the modern conversation: 148 studies, 308,849 participants, random-effects weighted average odds ratio 1.50 (95% CI 1.42 to 1.59) for stronger social relationships, and the authors’ own conclusion that the influence of social relationships on mortality is comparable with well-established risk factors.

It is also the paper that produced the most common misreading in preventive medicine, because it contains a decomposition that usually does not travel. The association was strongest for complex measures of social integration — 1.91 (1.63 to 2.23). It was lowest, and not statistically significant, for the binary indicator of living alone versus living with others: 1.19 (0.99 to 1.44). The confidence interval crosses one.

So “living alone kills you” is not what the meta-analysis found. What it found is that the richest measures of social integration track mortality more strongly than the crudest one, and that the crudest one barely clears — or fails to clear — the line.

In older adults, the numbers hold and the noise gets loud

The 2025 review narrowed the population to older adults and was more thorough about method: 86 studies out of 11,964 identified, quality assessed on the Newcastle-Ottawa scale, random effects with restricted maximum likelihood, plus subgroup and meta-regression analyses.

Loneliness: HR 1.14 (1.10 to 1.18). Social isolation: HR 1.35 (1.27 to 1.43). Living alone: HR 1.21 (1.13 to 1.30). Living alone reaches significance here, with a confidence interval that does not cross one.

The heterogeneity is the number to keep in view: for loneliness, I² = 84.0%. That means the studies disagree with each other more than they agree — the pooled estimate describes the centre of a wide, uneven spread. The meta-regression then names what predicts a stronger association: longer follow-up, female sex, validated social network indices rather than ad-hoc ones, adjustment for cognitive function, and higher study quality. In other words, the better the study, the more the estimate depends on how carefully isolation was actually measured.

Who ends up isolated is not a random draw

A separate 2023 review looked at the risk factors for social isolation in adults aged 60 and over: 18 observational studies, 11 databases to August 2022, modified Newcastle-Ottawa. Age 80 and over carried an odds ratio of 1.68 (1.44 to 1.97), smoking 1.43 (1.18 to 1.73).

That direction of causation matters. The path that produces isolation in later life runs through health, mobility, bereavement and income. The same meta-analyses that link isolation to mortality are, in part, measuring the mortality of the health conditions that produced the isolation. Adjusting for baseline health status — which the 2010 review did across subgroups and found did not erase the association — helps, but observational adjustment is not the same as a controlled comparison.

What the data do not show

No randomized trial has assigned people to be socially isolated or to have more contact, and then counted deaths. There is no version of this literature in which the exposure was manipulated. Every number above is observational, which means residual confounding travels with each of them.

Nothing here establishes a mechanism. Inflammation, hypothalamic-pituitary-adrenal axis activity, blood pressure, sleep, alcohol, physical activity — each is proposed, and each is also a plausible consequence of isolation rather than a cause of death. The cohorts measure the outcome; they do not measure the road to it.

Living alone is not the same as being lonely, and neither is the same as being isolated. The 2010 review’s own decomposition shows the crude indicator carrying the weakest association. Treating them as synonyms in the conversation is the single most common error in the popular version of this research.

A pooled relative risk says nothing about an individual. The 1.32 comes from pooling 90 studies with different instruments, different follow-up periods and different populations. It is an average of averages, not a personal forecast.

There is no established threshold. No study in these reviews identifies the number of weekly contacts, the number of meals shared, or the hours of conversation at which the curve changes. “Three social contacts a week” is not a finding; it is a plausible-sounding number with no source.

The effect is not confined to mortality. The same exposures are associated with depression, cognitive decline and disability in the same literature. If you are lonely and healthy, the honest reading of these studies is that the observable consequences closest to you — mood, sleep, activity, adherence to treatment — are where the evidence is thickest, not your lifespan.

The useful reading

If you take one thing from ninety cohorts, take this: the association is large, consistent and graded across the whole literature, and it is attached to the quality and multiplicity of social contact rather than to the binary of living alone. The strongest estimate in the entire field is for complex social integration, 1.91.

That points at a measurement problem as much as a health one. If the crude proxy is weak and the rich measure is strong, then how much connection you have matters more than who you live with.

And the honest ceiling: this is one of the largest pooled associations in preventive medicine, and it still does not tell you that joining a choir will add years. It tells you the size of the pattern. Everything past that sentence is somebody’s opinion wearing a number.