What the WHO's '871,000 Deaths' Figure Actually Measures

The figure that travelled fastest out of the World Health Organization’s June 2025 report, From Loneliness to Social Connection, was this one: loneliness is linked to an estimated 871,000 deaths annually, or roughly 100 an hour. It is a striking number, deliberately built for repetition, and it has already been repeated widely since the report’s launch. It is worth asking what it actually measures before repeating it further.

The short answer is that it is a modeled estimate of attributable mortality risk, built by applying relative-risk figures from the epidemiological literature to population-level prevalence data on loneliness and isolation. It is not a count of deaths that a coroner or a health system recorded as loneliness-caused. No such category exists. Understanding the difference matters for anyone using the figure to argue for a specific intervention, because the number answers a narrower question than the headline suggests.

What “attributable” means here

Attributable-risk estimates work by combining two things: how much a given exposure raises the relative risk of death, and how many people in a population have that exposure. The WHO Commission’s figure rests on the body of relative-risk work assembled over the past fifteen years, most prominently Julianne Holt-Lunstad’s two meta-analyses. Her 2010 review, spanning 148 studies and 308,849 participants, found that stronger social relationships were associated with a 50% increased likelihood of survival, an effect size the authors compared to the risk reduction from quitting smoking. Her 2015 follow-up disaggregated the construct further, finding an odds ratio of 1.29 for social isolation, 1.26 for loneliness, and 1.32 for living alone, each estimated separately and each surviving adjustment for baseline health status.

Those are real, peer-reviewed effect sizes from a substantial evidence base. But they are association measures from largely observational studies, not causal estimates from trials. The Commission’s global mortality figure inherits that limitation. It takes an odds ratio derived mostly from cohort studies in high-income countries and extrapolates it against loneliness prevalence estimated from a different, much larger dataset: the 2023 Gallup-Meta survey of roughly 142 countries, which found 24% of respondents worldwide reporting loneliness “very” or “fairly” often. Multiplying a relative risk estimated in one set of populations against a prevalence figure measured in a much broader and more heterogeneous set of populations is a defensible modeling choice, and it is the standard method for producing burden-of-disease estimates in public health. It is also a choice that compounds every assumption embedded in each input.

The instrument problem underneath the model

The prevalence side of the equation is where instrument mismatch becomes hardest to ignore. Gallup’s 24% figure, drawn from its partnership with Meta surveying roughly 1,000 people aged 15 and older per country, used a single-item measure of loneliness frequency. The American Heart Association’s 2022 scientific statement on cardiovascular and brain health, by contrast, worked from studies using instruments such as the UCLA Loneliness Scale, a validated multi-item measure that captures loneliness as a graded psychological state rather than a yes/no frequency report. The U.S. Surgeon General’s 2023 advisory cited a figure of roughly half of American adults experiencing loneliness, again drawn from a different set of instruments and time windows than either the Gallup or AARP figures.

These are not competing claims about the same fact. They are different questions producing, unsurprisingly, different numbers. A global mortality estimate built by feeding one country’s or one survey’s prevalence rate into a relative risk estimated from a separate literature is only as sound as the assumption that “loneliness” means the same measured thing across all of them. The BMC Public Health review of the state of loneliness and social isolation research, published in 2023, flagged inconsistent measurement as one of the central obstacles to comparing findings across the field. The 871,000 figure does not escape that obstacle. It sits on top of it.

Isolation and loneliness, collapsed into one number

There is a second, more subtle aggregation problem. Isolation is a structural property of a person’s network, typically measured by network size or contact frequency. Loneliness is the subjective experience of feeling disconnected, and it is entirely possible to have one without the other: a person with few social ties who does not feel lonely, or a person embedded in a large network who feels acutely isolated regardless. Holt-Lunstad’s 2015 meta-analysis treated these as distinct exposures with distinct odds ratios precisely because they behave differently, and because social deficits of either kind were found to be more predictive of death in samples averaging under 65 than in older cohorts. The AHA’s 2022 statement, similarly, separates “objective” from “perceived” isolation in its title, and reports risk increases in the range of 29% for heart attack or death from heart disease and 32% for stroke as distinct estimates rather than a single blended figure.

The WHO’s 871,000 figure, and the “1 in 6” prevalence estimate that accompanies it, necessarily compress this distinction to produce a single global headline. That is a reasonable choice for a report trying to establish that disconnection deserves public health priority status, which is the stated purpose of the Commission’s work since its 2023 launch under the co-chairmanship of Vivek Murthy and Chido Mpemba. It is a less reasonable input for anyone trying to design or evaluate a specific intervention, because an intervention aimed at expanding someone’s network will not necessarily touch their subjective loneliness, and vice versa.

What the figure is useful for, and what it is not

None of this means the underlying concern is exaggerated. The consistency of the relative-risk estimates across multiple independent meta-analyses and the AHA’s scientific statement is one of the stronger patterns in this literature, and the AHA statement itself is unusually candid about the limitation that matters most: it explicitly identifies the absence of intervention trial evidence as the central research gap. Knowing that social disconnection correlates with elevated mortality risk is not the same as knowing which intervention reduces that risk, at what dose, for whom.

The 871,000 figure is useful as an advocacy tool and as a rough order-of-magnitude signal that a large global health authority takes this seriously enough to model it at the scale of other major risk factors. It is not useful as a precise accounting of cause and effect, and it should not be treated as evidence that a specific policy lever, whether social prescribing, third-place investment, or a national loneliness strategy, would avert a specific number of deaths. The WHO’s own Director-General framed the report’s purpose as establishing social connection as a determinant of health, comparable in stature to other established public health priorities, not as producing a number precise enough to plug into a cost-benefit calculation for a particular program.

What would tighten this

A stronger version of this estimate would separate isolation and loneliness throughout, rather than aggregating them into one prevalence figure before applying risk. It would also use a single harmonized measurement instrument across the countries feeding the prevalence estimate, rather than combining Gallup’s single-item measure with relative risks drawn from cohorts using the UCLA scale or comparable multi-item instruments. And it would report a plausible range around 871,000 rather than a point estimate, since every step in the calculation, from country-level prevalence to the pooled odds ratio itself, carries its own confidence interval that the headline number obscures. Until that kind of estimate exists, 871,000 deserves to be cited as what it is: a serious, method-consistent extrapolation from a real body of evidence, not a body count.