What a Repeated Measure Would Show That a Single Survey Cannot

Almost every widely cited loneliness statistic comes from a single survey administered once, to a sample of people who were not surveyed before and will not be surveyed again. The AARP Foundation’s finding that one in three U.S. adults over 45 are lonely, Cigna’s 61%, Gallup’s 24% worldwide — all of these are cross-sectional. They describe a population at a moment, not a person over time. That distinction sounds like a technicality. It is not. It determines what the number can and cannot be used to argue.

The problem a snapshot cannot solve

A cross-sectional survey can establish that lonelier people also report worse health, fewer friends, or lower income. It cannot establish which came first. If the AARP survey finds that lonely respondents are less likely to have spoken with a neighbor — 61% lonely among those who never have, versus 33% among those who have — two explanations fit the data equally well. Isolation may cause loneliness, as the intuitive reading suggests. Or loneliness may cause withdrawal: people who feel lonely may stop initiating the small exchanges that would otherwise reduce it. A single measurement occasion cannot distinguish these, because both produce the same correlation.

This is not a flaw in any particular study’s design; it is a structural limit of the design type. Holt-Lunstad’s 2015 meta-analysis in Perspectives on Psychological Science, pooling isolation and loneliness data across dozens of studies, found isolation carried an odds ratio of 1.29 for early mortality and loneliness 1.26 — genuinely large effects, on a par with other recognized mortality risk factors. But most of the underlying studies measured isolation and loneliness once and then tracked who died, which supports a claim about prediction, not about mechanism. Whether reducing isolation would reduce mortality, as opposed to isolation simply flagging who was already frailer, is a separate question the design cannot answer.

What longitudinal data adds, concretely

A longitudinal design measures the same people repeatedly over time. That structure buys three things a cross-sectional survey cannot offer, even a very large one.

The first is temporal ordering. If a cohort is measured for isolation at time one and depression at time two, a researcher can at least establish which changed first within the same individuals, which narrows — though does not eliminate — the space of causal explanations. The JACSIS study of Japanese adults, tracking the same national sample across 2020 and 2021, is useful for exactly this reason: it can describe how isolation and loneliness moved for specific people across the pandemic period, rather than comparing two different snapshots of Japan and inferring a trend from the difference. A repeated cross-sectional design — a different sample surveyed each year — cannot separate a genuine shift in the population from a shift in who happened to answer that year’s poll.

The second is within-person change versus between-person difference. Cross-sectional data can only compare different people to each other: lonelier people versus less lonely people, at one point in time. This conflates two very different questions — why some people are lonelier than others, and what makes a given person’s loneliness rise or fall. A national friendship survey from the Survey Center on American Life at AEI found that the share of American men reporting no close friends rose from 3% in 1990 to 15% in 2021, and the share with six or more close friends fell from 55% to 27%. That is a striking comparison, but the 1990 and 2021 figures come from different men. It tells us the aggregate shifted; it cannot tell us whether individual men lost friends as they aged, whether successive cohorts of men simply started adulthood with fewer friends and never caught up, or some mixture of both. Untangling age effects from cohort effects requires following actual cohorts across time, which none of the widely cited friendship or loneliness trend figures do.

The third is the ability to test an intervention against a real counterfactual. The 2023 difference-in-differences evaluation of the UK’s Campaign to End Loneliness is one of the few studies in this literature that does this properly: it compares changes in loneliness and mental health outcomes in areas exposed to the campaign against comparable areas that were not, over time, rather than simply measuring outcomes once after the campaign ran. Most social prescribing evaluations and awareness-campaign assessments in this field lack that structure entirely — they report outcomes among participants after the fact, with no comparison group and no baseline, which cannot distinguish a program’s effect from the fact that people who enroll in a program are already inclined to improve.

What this changes about the mortality claim

The most consequential number in this literature is Holt-Lunstad’s earlier 2010 finding, in PLoS Medicine, that stronger social relationships carried a 50% increased likelihood of survival across 308,849 participants and 148 studies — an effect size comparable to smoking cessation or maintaining a healthy weight. That figure is now cited constantly, including as backing for the National Academies’ 2020 recommendation that health systems routinely screen older adults for isolation. But the studies feeding that meta-analysis are, overwhelmingly, ones that measured social integration once and then followed survival — a design that does establish temporal order (loneliness measured before death, obviously) but still cannot establish that changing someone’s social integration changes their mortality risk. It establishes that social integration predicts survival, which is a genuinely strong and replicated finding. It does not establish that a clinical intervention targeting isolation would move the outcome, because no study in that pool actually intervened and then tracked survival against a control group. The National Academies’ 2020 report is candid about this gap: it recommends screening as sound clinical practice while stopping short of claiming that screening, on its own, changes health outcomes — the evidence to support that stronger claim would need to come from trials, not from observational cohorts, however large.

What repeated national surveys can and cannot substitute for

Gallup’s 2023 survey of loneliness across 142 countries, finding 24% of adults worldwide report feeling very or fairly lonely, is valuable precisely because of its scale and its consistent instrument across countries. If Gallup repeats this survey with the same questions in the same countries in future years, the resulting series would start to approximate a genuine trend — not because any one wave becomes longitudinal, but because a stable repeated cross-section, unlike an ad hoc comparison between two differently designed surveys, at least holds the measurement constant. That is a lesser thing than following individuals, but it is far more informative than comparing, say, Cigna’s 61% to Harvard’s 36% and treating the eleven-point gap as movement, when the two surveys used different instruments on different populations in different years. The Harvard Making Caring Common report, which found 36% of Americans reporting serious loneliness and 61% among young adults specifically, cannot be read against the Cigna figure as a trend line at all; they are not measurements of the same thing taken twice.

What a better study would look like

The improvement this field needs most is not a bigger sample or a more elaborate instrument. It is the same cohort, measured with the same tool, at intervals long enough to observe change, with a comparison group for any intervention tested against it. A study that measured isolation, loneliness, and health outcomes in the same 10,000 people every two years for two decades, with a randomized subset receiving a social-connection intervention and a subset receiving none, would settle questions the current literature can only gesture at: whether isolation precedes decline or follows it, whether the relationship differs by age cohort rather than age itself, and whether an intervention that reduces measured isolation actually moves mortality or is simply correlated with people who were going to do better regardless. No study of that design exists yet in the published record. Until one does, the strongest claims in this field will remain claims about association, dressed, understandably but not accurately, in the language of cause.