What the Harvard Loneliness Estimate Does and Does Not Sample

The figure has traveled fast. Harvard’s Making Caring Common project reported in February that 36% of Americans experience serious loneliness, with the number rising to 61% among young adults aged 18 to 25 and 51% among mothers with young children. It has since been cited in policy discussions, opinion pieces, and public health commentary as evidence of a pandemic-driven loneliness crisis. The underlying finding deserves scrutiny before it is treated as a settled national estimate, because the survey that produced it was not built the way a national prevalence estimate usually is.

What the survey actually sampled

The Making Caring Common report does not describe a probability sample of the U.S. population drawn through the kind of address-based or random-digit-dial methods that underlie, for instance, Census Bureau or Pew estimates. The report presents findings from an online survey fielded to a national panel. Panel-based online surveys can produce useful and interesting data, but they carry a structural bias toward people who are already online, already willing to opt into survey panels, and disproportionately reachable through the recruitment channels the panel vendor uses. Whether the panel was weighted to Census benchmarks on age, income, race, education, and region, and how closely the achieved sample matched those benchmarks, are the first questions a reader needs answered before treating 36% as a population parameter rather than a sample statistic.

This is not a minor technical quibble. The gap between “36% of this panel report serious loneliness” and “36% of Americans experience serious loneliness” is exactly the gap that determines whether a policymaker should treat the number as an input to resource allocation or as a directional signal.

Compare the instrument, not just the number

The other place scrutiny belongs is the loneliness measure itself. The report describes “serious loneliness” without specifying whether respondents were classified using a validated multi-item scale, such as the UCLA Loneliness Scale, or a single self-report item asking directly whether they feel lonely. This distinction matters more than it sounds.

The AARP Foundation’s 2018 survey of adults 45 and older is useful precisely because it used the full 20-item UCLA Loneliness Scale, the same instrument academic researchers use, which is why its finding — one in three adults 45 and older report being lonely, from a sample of 3,020 — sits comfortably alongside the broader meta-analytic literature. Julianne Holt-Lunstad’s 2015 review in Perspectives on Psychological Science, synthesizing dozens of studies that mostly used validated scales, found loneliness associated with a 26% increase in odds of early mortality. That figure is legible against the AARP number because both rest on comparable measurement.

Cigna’s 2020 workplace survey took a different approach again, reporting that 61% of U.S. adults sometimes or always feel lonely — a figure that sounds close to the Making Caring Common number but was generated by yet another instrument, fielded to yet another sample (an online panel oversampling working adults), asking a differently worded question. That the two land near each other is not evidence they are measuring the same thing; it may be coincidence, or it may reflect that broad single-item loneliness questions tend to produce higher and more volatile estimates than validated multi-item scales regardless of the population sampled.

None of this means the Making Caring Common figure is wrong. It means the figure cannot yet be placed on the same axis as the AARP or Holt-Lunstad numbers without knowing more about how “serious loneliness” was operationalized and how the panel was weighted. A reader who wants to know whether loneliness increased during the pandemic needs a measure that was fielded consistently before and after, on a comparable sample, not two different instruments compared across two different survey vehicles.

The subgroup numbers deserve the same treatment

The subgroup findings — 61% loneliness among 18-to-25-year-olds, 51% among mothers of young children, and the claim that 43% of young adults reported increased loneliness since the pandemic began — are the parts of the report most likely to be quoted in isolation, and they are also the parts most sensitive to sampling error and panel composition. Online panels are known to have different reach into young adults depending on recruitment channel: a panel recruited through certain digital platforms will systematically overrepresent young people who are heavy internet users and underrepresent those who are not, which is a nontrivial confound when the topic under study is social isolation. If the young-adult subsample within the panel is itself unrepresentative of young adults generally — smaller in absolute number, drawn disproportionately from certain platforms — then a subgroup estimate built on it carries wider uncertainty than the topline figure, and the report as summarized publicly does not give a reader the tools to assess that uncertainty because it does not report subgroup sample sizes or confidence intervals.

The qualitative-adjacent finding that about half of lonely young adults said no one had asked how they were doing in a way that felt genuine in recent weeks is a striking data point, and it is the kind of finding online panel surveys are reasonably good at capturing, because it asks about a specific recalled experience rather than a diffuse trait. That finding is more trustworthy, on its face, than the topline percentage, precisely because it asks something narrower and more concrete.

Why the instrument gap keeps recurring

This is not unique to the Making Caring Common survey. Every large loneliness estimate in circulation right now was produced by a different combination of sampling frame and instrument, and the field has not converged on reporting them side by side with equivalent methodological detail. The National Academies’ 2020 consensus report on isolation and loneliness in older adults, drawing on the peer-reviewed literature rather than commissioning new survey data, put isolation prevalence among adults 65 and older at roughly one quarter — a figure it treats cautiously, because it is itself synthesizing studies with different sampling frames and different isolation measures, not producing a single new estimate. Even a careful synthesis inherits the comparability problems of its inputs.

The pattern to watch for is this: a survey organization fields an instrument to a convenience or panel sample, reports a striking topline percentage, and that percentage then circulates without the sampling frame or instrument attached, so it is compared as though it were commensurable with the peer-reviewed prevalence literature. It usually is not.

What would settle the question

A stronger version of the Making Caring Common finding would report the panel’s demographic composition against Census benchmarks, disclose whether weighting was applied and how, use a validated scale such as UCLA’s rather than an unspecified “serious loneliness” classification, and report subgroup sample sizes alongside subgroup percentages. Best of all would be a probability-based sample — something closer to the design the Census Bureau or a federal statistical agency would use — fielded with the same instrument at multiple time points bracketing the onset of the pandemic, so that “increased loneliness since the pandemic began” could be measured as an actual before-and-after comparison rather than inferred from a single retrospective self-report question, which is subject to its own recall and framing effects.

Until then, the 36% figure is worth treating as a real and useful signal that a substantial share of a large panel of U.S. adults reported serious loneliness in early 2021, and as considerably weaker evidence for the claim that exactly 36% of the U.S. population meets that description. The distinction sounds pedantic. It is the difference between a number that supports a resource-allocation decision and one that supports a general sense of direction.