The Panel Nobody Answers: Sampling Behind the Household Pulse Numbers

A 2026 analysis in Community Mental Health Journal used the 2024 Household Pulse Survey to break down loneliness and social isolation prevalence by demographic and socioeconomic group in the United States. It is a useful dataset: large, frequent, and administered by the Census Bureau, which lends it an authority that a market-research panel does not have. It is also worth asking, before citing the headline breakdown, who actually answered the questions.

What the Household Pulse Survey is, mechanically

The Pulse survey was built during the pandemic to produce near-real-time estimates faster than the Bureau’s flagship surveys allow. Households are sampled from the Census Bureau’s Master Address File — in principle a full population frame, which is the strongest kind of sampling frame available. But contact happens by email and text message, using addresses matched to phone numbers and email addresses from commercial and administrative databases. That matching is imperfect and it is not random with respect to who gets found. People who move often, share phone plans, or simply are not in the commercial databases that get matched to census addresses are less likely to be reachable at all, independent of whether they respond.

Then there is response. Pulse survey response rates have run in the single digits to low double digits across its various waves, a fact the Bureau has disclosed but that gets dropped when the resulting percentage travels into a policy brief. A phone number does not answer a loneliness question. A person does, and the smaller the fraction of contacted households that becomes a completed survey, the more weight the analysis is asking post-stratification weighting to carry.

Weighting can correct for known imbalances — age, sex, race, region, income band — if the survey has enough respondents in each cell and if nonresponse within a cell is unrelated to the outcome being measured. That second condition is the one that should worry anyone using this data specifically for loneliness. There is no strong reason to think that within, say, “men aged 45-54 in the South,” the people who respond to an unsolicited text from the Census Bureau are equally likely to be lonely as the people who do not. If socially withdrawn people are systematically less reachable — no stable address, no smartphone, screening unknown numbers — a demographic weight cannot fix that, because it does not know isolation status ahead of time. It can only balance on the traits it has measured.

What a different frame does to the same question

The comparison worth making is with the Health Information National Trends Survey, HINTS-6, whose 2025 analysis in the American Journal of Preventive Medicine reported that roughly 1 in 7 U.S. adults experienced severe loneliness and closer to 1 in 5 moderate loneliness, with over 37% falling into the combined moderate-to-severe category. HINTS uses a mail-based, address-based sample with a paper option throughout, not a digital-first contact strategy. That single design choice changes who is reachable: people without reliable internet access, older adults who do not check email, and people wary of unfamiliar text messages are not excluded by construction the way they can be in a survey that leads with email and SMS.

Neither design is free of bias. Mail surveys skew toward people who open mail and complete forms, which correlates with education and stability of housing in its own way. But the two instruments are not measuring loneliness through the same funnel, and a reader comparing their prevalence figures side by side is comparing two different populations who happened to answer, not two estimates of the same fixed number with sampling noise layered on top.

The CDC’s 2024 MMWR surveillance report on loneliness adds a third design entirely: it draws on data collected in 2022 through established federal health surveillance instruments rather than a purpose-built loneliness panel, and its publication date sits two years after its data date. That gap matters less for the sampling question than for a different one — whether the prevalence figure being cited describes 2026 conditions or a snapshot from a period that included lingering pandemic disruption to social contact. Citing it as current risks conflating both problems, sampling and staleness, in one number.

Why this is not a purely academic complaint

The Foundation for Art & Healing aside, most organisations working on loneliness policy need a number to argue with, and the temptation is to reach for whichever figure is newest or largest. The 2023 BMC Public Health review of the state of loneliness and social isolation research named inconsistent measurement as a structural barrier to comparing findings across the field, and sampling frame is a large, underdiscussed part of that inconsistency — bigger, in practice, than the difference between a ten-item and a twenty-item loneliness scale, which gets far more scrutiny.

The AARP Foundation’s 2018 survey of adults 45 and older offers a useful contrast precisely because it did the opposite of what Pulse does: it used a phone- and mail-recruited panel calibrated against the U.S. Census, achieved a substantial completed sample of 3,020 respondents, and administered the 20-item UCLA Loneliness Scale rather than a bespoke question. The 2025 follow-up, AARP’s Disconnected report, kept that same instrument, which is what makes the two comparable over time in a way that switching survey vendors would not allow. That comparability is worth more than a marginally larger sample size collected through a leakier funnel.

None of this means the Household Pulse estimate is wrong. Large surveys with low response rates can still produce usable estimates if the weighting model is well specified and the outcome is not strongly correlated with reachability. But loneliness is exactly the kind of outcome where that assumption deserves scrutiny rather than a pass, because withdrawal from contact is both the condition being measured and, plausibly, a cause of not answering the phone.

What would settle this

A study designed to test the bias directly would need to compare loneliness prevalence among Pulse respondents against a benchmark sample recruited through a method with a much higher response rate — in-person interviewing, ideally — for the same population and the same time window. Short of that, the more modest fix is disclosure: reporting the completion rate alongside the topline figure, and treating cross-survey prevalence differences as evidence about instruments and frames before treating them as evidence about the world.