How Big Is the Effect, Actually

Julianne Holt-Lunstad’s 2010 meta-analysis is cited so often as “loneliness is as bad as smoking” that it is worth going back to what the number actually was. Across 148 studies covering 308,849 participants, stronger social relationships were associated with a 50% increased likelihood of survival over the follow-up period. That figure is an odds ratio pooled across studies with wildly different designs, follow-up windows, and measures of “social relationships” — some used marital status, some used network size, some used a composite index. The headline is real. What it means in practical terms is less settled than the citation count suggests.

Three numbers, one problem

The 2015 follow-up, also led by Holt-Lunstad, split the earlier composite into three separate exposures and reported them individually: social isolation carried an odds ratio of 1.29, loneliness 1.26, and living alone 1.32, all for early mortality. These are markedly smaller than the 2010 figure. That is not a contradiction — it reflects a different, more granular analytic strategy — but it is a case study in what happens when a pooled effect gets decomposed. The 50% figure from 2010 described a broad contrast between people with strong versus weak social relationships across many measurement approaches. The 2015 figures describe three specific, narrower exposures, each estimated on its own. Reporting on the 2015 paper very often kept the “50% increased survival” language from the 2010 paper while attaching it to the newer, narrower claims. The two papers are not measuring the same thing, and treating their numbers as interchangeable overstates the precision of what either one shows.

It is also worth sitting with how modest 1.29 and 1.26 are as odds ratios. In epidemiology, ratios in that range are not trivial, but they are not dramatic either. They sit in roughly the territory of moderate risk factors — noticeably above 1.0, well short of the 3s and 4s associated with heavy smoking or severe obesity. The comparison to smoking that circulates in press coverage of this literature derives more from the 2010 pooled estimate, and even there the analogy is a comparison of effect magnitude across very different types of exposure and outcome measurement, not a claim that the biological mechanisms are equivalent.

What “50% increased survival” is not

The 2010 statistic is frequently rendered in secondary coverage as “lonely people are 50% more likely to die.” That is not what the paper reports. The odds ratio describes the likelihood of survival associated with stronger relationships, pooled across all included studies and their varying follow-up periods, adjusting for the covariates each individual study happened to adjust for. It is an association across a heterogeneous body of research, not a hazard rate for any identifiable individual, and it should not be read as a prediction of what happens to a specific lonely person over a specific number of years. Holt-Lunstad’s later 2021 review, written for a lifestyle-medicine audience explicitly interested in prevention framing, is more careful on this point: it positions social connection as a modifiable risk factor comparable in category, not necessarily in magnitude, to diet, exercise, and smoking cessation, and frames the evidence as sufficient to justify screening and intervention rather than as a precise dose-response curve.

That distinction matters because effect size and clinical actionability are not the same question. A moderate odds ratio, replicated across 148 studies with over 300,000 participants, is a much stronger basis for population-level action than a large odds ratio from a handful of small studies. The 2010 meta-analysis’s strength is its consistency across a very large and heterogeneous evidence base, not the size of the number itself. The National Academies’ 2020 consensus report on isolation in older adults draws on this same logic: it does not lead with a single dramatic ratio but instead builds its case for routine clinical assessment on the breadth and consistency of association across many outcomes — cardiovascular disease, dementia, depression, and mortality among them — while stopping short of claiming a precise, quantified return on any specific intervention.

The confound problem effect sizes cannot settle

Every ratio discussed above comes from observational data. The 2015 analysis notes that effects remained after adjusting for health status, which addresses the most obvious form of reverse causation — that sick people become isolated rather than isolation causing illness — but adjustment is not the same as randomization. No meta-analysis in this literature can rule out unmeasured confounding: a lonely person may also, on average, sleep worse, drink more, or delay medical care for reasons no covariate in these studies captures. The National Academies’ consensus report and its accompanying clinical commentary are candid that the causal pathway between isolation and downstream health outcomes is not fully mapped, even as they treat the associational evidence as sufficient grounds for screening. Cacioppo’s physiological framing offers one account of mechanism — loneliness as an aversive signal that alters stress hormone regulation, sleep architecture, and immune response — but that is a proposed pathway, not a demonstration that the pathway accounts for the observed mortality odds ratios. The size of an odds ratio does not tell you how much of it is causal.

Why the emphasis on statistical significance obscures this

The three Holt-Lunstad-associated papers discussed here are all large enough that statistical significance is close to guaranteed; with 308,849 participants, a modest true effect will nearly always clear a p-value threshold. This is precisely why the significance framing common in press summaries is unhelpful: it tells the reader that an effect exists, not how large it is or what population-level consequence follows from it. A regulatory or clinical decision — whether to add isolation screening to a primary care visit, for instance, which is what the National Academies report ultimately recommends — depends on effect size, not on whether p was below 0.05. The 2021 clinical commentary on the National Academies report is explicit that translating “isolation is a risk factor” into a specific, resourced screening protocol requires knowing how much risk reduction a given intervention would plausibly produce, and that number is largely absent from the literature. Nobody has published an intervention trial with a comparable sample size and comparable rigor to the observational mortality meta-analyses.

What would settle this

A study that would meaningfully advance the question is not another observational cohort producing a fourth odds ratio to add to the pile, but an intervention trial — ideally cluster-randomized, given the difficulty of individually randomizing people into or out of social contact — that measures a specific, well-defined intervention against a specific, well-defined outcome over a long enough follow-up to capture mortality or major morbidity, not just self-reported wellbeing at three months. Until that exists, the honest summary of this literature is that social disconnection is consistently associated with worse survival across an unusually large and heterogeneous body of observational evidence, that the size of that association is moderate rather than extreme once decomposed into its component exposures, and that the causal share of that association remains unknown. All three of those statements can be true simultaneously, and none of them is well served by reducing the literature to a single borrowed comparison with cigarettes.