The Mortality Numbers Are Old, and the Field Keeps Building on Them Anyway

Julianne Holt-Lunstad’s 2010 meta-analysis in PLoS Medicine remains the single most cited number in this field: pooling 148 studies and 308,849 participants, it found that stronger social relationships were associated with a 50% increased likelihood of survival, an effect size comparable to quitting smoking. Her 2015 follow-up in Perspectives on Psychological Science refined the picture, separating social isolation (odds ratio 1.29), loneliness (1.26), and living alone (1.32) as independently predictive of early mortality. These two papers anchor almost every subsequent claim that loneliness is a health risk on par with obesity or physical inactivity.

They are also, by now, old data. The underlying studies in both meta-analyses were conducted predominantly before 2010, many considerably earlier. The National Academies’ 2020 consensus report on social isolation in older adults draws heavily on this same evidence base, as does the clinical commentary that followed it in the American Journal of Geriatric Psychiatry. A decade of reviews has restated the same effect sizes rather than generating new ones drawn from more recent, larger, or more diverse cohorts. This is worth pausing on, because it is a genuine and persistent gap in the literature, not a minor technical footnote.

What the underlying cohorts actually looked like

The 148 studies pooled in the 2010 meta-analysis were assembled from decades of existing prospective research, much of it drawn from populations enrolled in the 1970s through the 1990s. Social contact was often measured with instruments built for other purposes: single items about marital status, frequency of visits with relatives, or participation in organizations. Cacioppo’s 2008 work on the physiology of loneliness, published around the same period, helped establish why isolation might causally affect health — through cortisol dysregulation, disrupted sleep, and altered immune function — but the epidemiological evidence for the size of the effect still rests on cohorts that predate widespread mobile phone use, social media, and the specific demographic and economic conditions of the past decade.

This matters for a simple reason: the composition of social networks has changed. A person’s isolation profile in 1985 — measured by proximity to family, club membership, church attendance — is not necessarily comparable to a person’s isolation profile in 2021, where contact can be sustained at a distance through channels that did not previously exist, and where in-person associational life has continued the decline that Robert Putnam documented in Bowling Alone. Whether the same odds ratios apply to a population whose isolation looks structurally different is an open question. Nobody has run the equivalent large-scale prospective meta-analysis on cohorts assembled in the past ten years.

The reviews keep citing each other instead

This is the pattern worth naming explicitly. The National Academies’ 2020 report, produced by a standing consensus panel with access to the full academic literature, still leans on the 2010 and 2015 Holt-Lunstad figures as its central quantitative anchors. Its own contribution is largely to call for something the underlying evidence does not yet supply: that the health care system routinely assess isolation and loneliness in clinical settings. But routine assessment presumes a stable, current estimate of risk. If the odds ratios were established on cohorts recruited before 2000, a clinician screening a patient in 2021 is applying a risk model built on a different social world.

The 2020 journal commentary on the National Academies report makes a similar move: it argues for the clinical operationalization of isolation screening while treating the epidemiological base rates as settled. Nowhere across these documents does a genuinely new large-N mortality study, using contemporary cohorts and contemporary measures of network structure, appear. The citation trail terminates, repeatedly, at the same two Holt-Lunstad meta-analyses.

Why this hasn’t been fixed

Part of the explanation is structural. Mortality studies require long follow-up periods; a cohort enrolled in 2015 cannot yet generate the fifteen- or twenty-year survival data that the earlier meta-analyses were built on. Prospective epidemiology of this kind is slow by design, and there is no methodological shortcut around it. A meta-analysis published in 2021 using only genuinely recent cohorts would necessarily have shorter follow-up windows and correspondingly less statistical power to detect mortality effects, which tend to accumulate over decades rather than years.

But this only explains part of the pattern. It does not explain why cross-sectional and shorter-term studies of loneliness prevalence — of which there are many, and which are far cheaper to run — have not been better integrated with mortality-relevant biomarkers to produce interim estimates. The AARP Foundation’s 2018 national survey of adults 45 and older, using the 20-item UCLA Loneliness Scale, found that one in three respondents were lonely, and that the gap between socially connected and isolated respondents was stark: 33% of those who had spoken with neighbors were lonely, against 61% of those who had never done so. That is a useful, current, well-instrumented prevalence estimate. It says nothing directly about mortality risk, because it was not designed to. The Harvard Graduate School of Education’s Making Caring Common survey, published in February 2021, adds a contemporary snapshot showing 36% of Americans reporting serious loneliness and 61% of young adults aged 18 to 25 reporting the same — but again, this is prevalence, not a linkage to mortality outcomes.

The field has, in effect, two separate literatures that have not been joined: recent, well-measured prevalence data, and older, well-powered mortality data. Nobody has yet connected a 2021-era prevalence instrument to a 2021-era mortality outcome, because the mortality outcome does not exist yet — it cannot, given the follow-up time required. What the field has instead done is treat the old mortality coefficients as fixed constants and apply them to new prevalence numbers, which is a reasonable stopgap but not the same thing as updated evidence.

What would actually close the gap

A study that would meaningfully address this would need three features the current literature lacks in combination. First, a cohort enrolled within the past decade, so that the network structure and communication technology available to participants resembles the present rather than the 1980s. Second, a mortality or serious-morbidity endpoint with enough follow-up to generate a defensible hazard ratio — which, given the constraints above, likely means such a study could not report definitive results before the mid-2030s at the earliest. Third, measurement that distinguishes objective network structure from subjective loneliness at baseline, rather than folding both into a single isolation index, since the 2015 meta-analysis already showed these operate as separable risk factors with different odds ratios.

Until such a cohort matures, every claim about the magnitude of loneliness’s mortality effect in the present decade is an extrapolation from data collected under different social conditions. That extrapolation may well hold. The physiological mechanisms Cacioppo described are not obviously time-bound, and there is no strong reason to expect the relationship between isolation and mortality to have reversed. But the honest position is that this is currently an assumption carried forward by citation rather than a finding re-established by new data, and ten years of reviews restating the same two numbers have not changed that.