The 50% Figure That Split Into Three Smaller Ones

The most-cited number in the social connection literature is the claim that weak social relationships carry a mortality risk comparable to smoking. It traces to a single figure: a 50% increased likelihood of survival for people with stronger social relationships, from Julianne Holt-Lunstad’s 2010 meta-analysis in PLoS Medicine, pooling 148 studies and 308,849 participants. That figure still circulates in policy documents, op-eds, and conference slides, usually attached to the word “loneliness.”

It should not be. The 2010 meta-analysis did not measure loneliness. It measured “social relationships” — a composite that pooled together studies using wildly different instruments: some assessed marital status, some assessed the size of a person’s network, some assessed the frequency of contact, and a smaller number assessed subjective loneliness. The 50% figure describes an omnibus construct, not a specific one. That distinction went largely unnoticed for five years, in part because the paper’s abstract described the effect as comparable to established risk factors like smoking, and that comparison was what circulated.

In 2015, Holt-Lunstad revisited the question in Perspectives on Psychological Science and did something the field had not done at scale: split the composite apart. The follow-up meta-analysis reported separate odds ratios for social isolation (1.29), loneliness (1.26), and living alone (1.32), each for early mortality. All three remained statistically significant after adjusting for health status. But none of them, individually, comes close to reproducing a 50% figure. An odds ratio of 1.29 describes a 29% increase in the odds of an outcome — smaller, and importantly, a different statistical quantity than the “likelihood of survival” language used in 2010.

Why the number shrank

The gap between the two papers is not evidence that the underlying risk got smaller between 2010 and 2015. It reflects what happens when a heterogeneous composite gets decomposed into its parts.

A composite measure of “social relationships” that includes marital status, network size, contact frequency, and subjective loneliness will tend to produce a larger aggregate effect than any single component, partly because the studies feeding into it varied enormously in what they measured and how, and partly because a person who scores badly on one dimension often scores badly on several — isolated people are more likely to also be lonely and to also live alone. Pooling those correlated measures into one estimate inflates the apparent size of any single one of them.

The 2015 paper’s contribution was to show that isolation, loneliness, and living alone are not simply three names for the same underlying deficit. They behave differently across subgroups. The paper found that social deficits were more predictive of death in samples with an average age under 65 than in older samples — a pattern that a single pooled estimate would have obscured entirely, since it flattens age structure along with everything else.

This matters for anyone using the 2010 figure now. If it is being cited to argue that loneliness specifically carries smoking-level mortality risk, that is not what the underlying study measured. If it is being cited to argue that the general absence of adequate social relationships — isolation, loneliness, and living alone together — carries that level of risk, the citation is closer to accurate, but even then the 2015 breakdown suggests the pooled figure overstates any one component taken alone.

The definitional split shows up downstream, too

Once the field treated isolation and loneliness as distinct constructs rather than interchangeable proxies, later work built the distinction into its design rather than discovering it as an afterthought.

The National Academies’ 2020 consensus report on older adults defines social isolation as a structural property — a person having few relationships or infrequent contact — and estimates that roughly a quarter of U.S. adults aged 65 and older meet that description. That is a network-structure measure. It is not asking anyone how they feel.

Compare that to the AARP Foundation’s 2018 survey of adults 45 and older, which found that one in three respondents were lonely, using the 20-item UCLA Loneliness Scale. That is a subjective measure. It is not counting anyone’s contacts.

These two figures — roughly a quarter structurally isolated, roughly a third subjectively lonely — are frequently reported side by side as though they describe the same population, or as though one confirms the other. They do not. A person can be embedded in a large, frequent-contact network and still score high on the UCLA scale; a person with almost no social contact can score low on it. The AARP survey’s own finding illustrates the gap concretely: 33% of respondents who had spoken with a neighbor were nonetheless lonely, against 61% of those who had never spoken to a neighbor. Contact reduces the odds of loneliness. It does not eliminate the category difference between measuring contact and measuring the feeling.

The American Heart Association’s 2022 scientific statement pushed the distinction further still, building it directly into its terminology: “objective” versus “perceived” social isolation, evaluated separately for cardiovascular and brain health outcomes. The statement reports isolation and loneliness associated with roughly a 30% increased risk of heart attack, stroke, or death from either — 29% for heart attack and death from heart disease, 32% for stroke — figures in the same rough range as the 2015 mortality odds ratios, and notably smaller than the 2010 composite figure that still dominates public conversation. The AHA statement is explicit that this is a scientific statement, not a set of clinical guidelines, and it names the absence of intervention trial evidence as the central gap: the association is established: what reduces it is not.

What the revision cost, and bought

The methodological lesson generalizes past this one pair of papers. A single pooled effect size, built from studies that operationalize a construct in incompatible ways, will almost always look larger and cleaner than the effect sizes that emerge once the construct is split along its natural joints. That is not a reason to distrust pooled estimates outright, but it is a reason to ask, before citing one, what exactly was pooled.

The 2010 figure was not wrong; it answered a different question than the one it is usually invoked to answer now. The 2015 revision cost the field a simple, quotable number and bought it three smaller, more defensible ones, each attached to a specific and measurable construct. A study that would improve on both would need to track isolation and loneliness longitudinally, with consistent instruments across cohorts, rather than reconstructing the comparison after the fact from studies that never intended to be pooled together.