What Loneliness Surveys Measure, and What They Cannot

The U.S. Surgeon General’s 2023 advisory put loneliness at roughly half of American adults. Gallup’s 2023 global survey, drawing on 142 countries, put it at 24%. The CDC’s 2024 surveillance report, using 2022 data, gives yet another figure. These are not competing estimates of the same fixed quantity that will converge with a larger sample. They are answers to different questions, asked in different ways, about a state that has no agreed operational definition. That is the methods problem underneath the headline numbers, and it is worth being precise about before citing any of them.

The instruments in circulation

Most self-report loneliness data comes from one of three approaches. The UCLA Loneliness Scale, a 20-item instrument developed decades ago, asks about subjective feelings of disconnection across multiple items and produces a continuous score. AARP Foundation’s 2018 survey of adults 45 and older used the full UCLA scale, which is part of why its “one in three” figure is directly comparable to the academic literature that also uses it.

Other surveys use a single direct question: “how often do you feel lonely?” Gallup’s global work and the Cigna index both work this way, and single-item measures tend to produce different distributions than multi-item scales, because a person may resist self-labeling as “lonely” even while endorsing several UCLA items describing exactly that state. Cigna’s 2020 report found 61% of U.S. adults sometimes or always lonely; Gallup’s 2023 survey found 24% worldwide feeling lonely “a lot.” Some of that gap is real variation across populations and time. Some of it is instrument.

A third family, exemplified by the CDC’s Behavioral Risk Factor Surveillance System item behind its 2024 MMWR report, asks about lack of social and emotional support rather than loneliness directly — a related but distinct construct. The BMC Public Health review of the loneliness research field, published in mid-2023, names this inconsistency as a structural barrier to comparing findings across studies, not a solvable footnote.

Self-report versus something outside the respondent’s head

All of the figures above share one property: they ask a person to characterize their own emotional state and take the answer as the data. This is defensible, because loneliness is by definition a subjective experience — nobody but the respondent can report it. But it means none of these surveys can be checked against an independent record of whether the respondent is, structurally, isolated. Self-reported loneliness and self-reported network size are both collected from the same person in the same interview, so any relationship between them is still a relationship between two things one person said, not between a feeling and an observed fact.

Genuinely observed or administrative measures of social isolation are rare in this literature, and that scarcity is itself informative. The National Academies’ 2020 consensus report on older adults distinguishes social isolation — an objectively measurable property of network size, contact frequency, and social participation — from loneliness, a subjective state, and stresses that the two are not interchangeable. But even the “objective” isolation measures in most studies are still self-reported: a respondent is asked how many people they see and how often, which is a self-report of a structural fact, not an independently verified one. Administrative alternatives — health system records of visits, geolocation data on movement patterns, network data drawn from actual contact logs rather than recall — appear occasionally in smaller studies but essentially never in the national surveys that produce the widely cited prevalence figures.

The Scientific Reports study published in December 2024 on isolation, age, and loneliness during the pandemic is useful precisely because it treats the two constructs as separately measured and separately variable by age, rather than collapsing them into a single “connection” index. It finds the relationship between isolation and loneliness itself shifts across the life course — which is a finding that would be invisible in any survey using a single blended measure. Gallup’s 2023 analysis of the world’s social connections makes a parallel point: connectedness and loneliness are not simple inverses of one another. A person can be embedded in a large network and still report loneliness; a person living alone can report none. Both facts undercut the common shorthand of treating isolation as loneliness’s objective proxy.

Where a closer-to-observed measure exists

The clearest instance of researchers pairing a subjective report with something closer to a behavioral record is the HEAL-HOA trial published in The Lancet Healthy Longevity in November 2024. It is a randomised controlled trial of volunteering as an intervention for loneliness among older adults in Hong Kong, and it matters here less for its result than for its design: participants were randomly assigned to a prosocial engagement condition, and outcomes were tracked against that assignment rather than relying solely on before-and-after self-report from an uncontrolled group. That structure — a defined intervention, a control condition, an outcome measured over time — is what is missing from nearly everything else in this literature, including the large national surveys. Most social prescribing evaluations, by contrast, are uncontrolled programme evaluations that report improvement without a comparison group, which the American Journal of Geriatric Psychiatry’s 2020 commentary on the National Academies report flags as a persistent evidentiary gap.

The Japanese JACSIS study, tracking the same respondents from 2020 to 2021, offers a different kind of improvement: it is longitudinal rather than cross-sectional, so it can describe change in isolation and loneliness within the same people over time rather than comparing two unrelated snapshots. That is a meaningfully different claim than what a single-wave survey can support, even though the underlying measure is still self-report.

Why this matters for reading the prevalence numbers

None of this means the self-report figures are wrong. Loneliness is a subjective state, and there is no more direct way to measure it than asking. The concern is what gets done with the number afterward. When a report cites “one in three older adults are lonely” or “61% of Americans feel lonely,” the natural next move is to treat that as equivalent to counting isolated people — and then to design interventions aimed at increasing contact, on the assumption that more contact will reduce the reported feeling. But if isolation and loneliness are only loosely coupled, as the Scientific Reports and Gallup analyses both suggest, an intervention that increases objective social contact without addressing the subjective experience of disconnection may show no effect on loneliness scores at all, and a programme evaluation using only self-report before and after would not be able to distinguish “the intervention didn’t work” from “the intervention worked on the wrong variable.”

A stronger evidence base on this specific question would look like the HEAL-HOA trial multiplied several times over: randomised designs, a defined and observable intervention, outcomes tracked longitudinally in the same people, and loneliness and isolation measured as genuinely separate variables rather than folded into a single index. It would also mean more studies pairing self-report against something the respondent did not generate themselves — attendance records, network data drawn from actual contact rather than recall, health system utilization — so that the correlation between feeling and structure could be estimated rather than assumed. Until then, the honest reading of the national prevalence figures is that they describe how people say they feel, measured in at least three incompatible ways, and say comparatively little about how isolated those same people actually are.