The Age-Reversal Buried in the Mortality Meta-Analyses

Julianne Holt-Lunstad’s 2015 meta-analysis in Perspectives on Psychological Science is cited constantly as evidence that isolation kills older people. It is usually cited correctly for the headline numbers: social isolation carried an odds ratio of 1.29 for early mortality, loneliness 1.26, living alone 1.32. What gets dropped in translation is a secondary finding in the same paper, one that sits uneasily next to almost everything written about loneliness and aging since: the association between social deficits and death was stronger, not weaker, in samples with an average age under 65.

This is worth sitting with, because it runs against the shape of the entire policy conversation that followed. The National Academies’ 2020 consensus report frames social isolation as a problem of old age, reporting that roughly one quarter of adults 65 and older are socially isolated and recommending that health systems screen for it in that population specifically. The UK’s 2018 national loneliness strategy, the first of its kind, was not exclusively about older people, but much of its public framing and subsequent media coverage treated isolated elders as the paradigmatic case. AARP’s 2018 survey of adults 45 and older, finding that one in three report loneliness on the UCLA scale, reinforced the same association in the public mind: loneliness is what happens to people as they age out of work, marriage, and mobility.

None of that is wrong as description. Older adults do face structural risk factors for isolation – retirement, bereavement, declining mobility, smaller networks – that are less common at 35 than at 75. The question the 2015 finding raises is narrower and more technical: given that someone is isolated or lonely, does that fact predict death more strongly in youth and midlife, or in old age? Holt-Lunstad’s data say the former.

Why relative risk and absolute risk tell different stories

The resolution is not that isolation is harmless for elders. It is that mortality risk in older populations is already high and driven by so many concurrent factors – cardiovascular disease, cancer, frailty, multimorbidity – that isolation’s marginal statistical contribution gets crowded out. In a 40-year-old, death from any cause is rare enough that the ones that do occur are more likely to trace back to a small number of dominant risk factors, and social isolation appears to be one of them with disproportionate weight. In an 80-year-old, so many pathways to mortality are already active that isolation’s relative contribution, measured as an odds ratio, shrinks even if its absolute contribution — deaths attributable to it in real numbers — stays large or grows, because the base rate of death is so much higher to begin with.

This is a standard problem in survival analysis often described loosely as competing risks, and it means a relative-risk figure and a public-health-burden figure can point in different directions without either being wrong. The 2015 meta-analysis is reporting the former. The National Academies report, when it cites isolation as consequential for older adults, is implicitly making a claim closer to the latter — that even an attenuated relative risk, applied to a population with such elevated baseline mortality, produces large numbers of excess deaths. Both statements can be true. They are not the same statement, and confusing them is exactly the kind of error a methods note exists to flag.

A second complication: isolation and loneliness are not the same instrument at every age

There is a further wrinkle. Holt-Lunstad’s earlier 2010 meta-analysis in PLoS Medicine, covering 148 studies and over 308,000 participants, established the baseline claim that stronger social relationships predict a 50% greater likelihood of survival, an effect size comparable to well-established risk factors like smoking. That paper pooled measures of social integration broadly. The 2015 paper separated isolation (a structural property — how many people a person has contact with, and how often) from loneliness (a subjective state — how connected a person feels regardless of contact frequency) and from living alone (a household fact that correlates with both but is neither). These three constructs do not track each other cleanly, and they are not measured the same way across the studies that feed into a meta-analysis, or across the surveys that inform policy.

AARP’s 2018 survey measured loneliness in adults 45 and older using the 20-item UCLA Loneliness Scale, the same instrument used across much of the academic literature, which is one reason its 33% figure is comparable to other academic estimates rather than an artifact of a bespoke question. But AARP’s survey also found something that complicates a purely structural account of isolation: 61% of respondents who had never spoken to a neighbor reported loneliness, against 33% of those who had. That gap tracks a specific, modest behavior — a hallway conversation — rather than network size in the aggregate. It suggests the isolation construct used in most large studies, typically built from contact frequency and network size, may be missing something that a single low-effort interaction captures better than a count of total ties.

Cacioppo’s 2008 account of loneliness as an aversive biological signal, akin to hunger, offers one reason this might be so: the signal is triggered by a perceived shortfall relative to a person’s own baseline expectation of connection, not by an absolute number of contacts. A retired person with three regular contacts may feel adequately connected if that matches their expectation; a person of any age with the same three contacts, but a different reference point, may not. This is a subjective calibration that a structural isolation index, by design, cannot see.

What this changes, and what it does not

None of this undermines the case for taking isolation and loneliness in older adults seriously as a health issue. The absolute burden is large, the National Academies’ clinical recommendations are reasonable on their own terms, and the UK’s decision to fund social prescribing at a national scale does not depend on which age band shows the largest odds ratio. What the age-reversal in the 2015 data should change is the reasoning offered for that attention. If the justification is “isolation predicts death most strongly among older adults,” the meta-analytic evidence does not support it. If the justification is “older adults carry the largest absolute health burden from isolation because their baseline mortality risk is already elevated,” it does.

A study designed to settle this properly would report age-stratified relative risks alongside age-stratified attributable fractions in the same population, using a single consistent instrument for isolation and a separate one for loneliness across the full adult age range, rather than pooling studies that used different instruments in different age bands and then interpreting the pooled odds ratio as if it meant the same thing at 30 and at 75. Until that exists, the honest statement is that older adults face more isolation, in raw prevalence terms, but the mortality signal attached to that isolation may be doing more damage, relatively speaking, to people decades younger.