Why the Odds Ratio Matters More Than the P-Value Here

A meta-analysis of 148 studies covering 308,849 participants is not going to produce an ambiguous p-value. With a sample that size, almost any real association between social relationships and mortality will clear the conventional significance threshold, and so it did: Julianne Holt-Lunstad’s 2010 meta-analysis in PLoS Medicine found that stronger social relationships were associated with a 50% increased likelihood of survival, an effect the authors judged comparable in magnitude to well-established risk factors such as smoking. That comparison is doing more work in the literature than any significance test could.

This is the point worth sitting with. In a field this size, statistical significance stopped being the interesting question some time ago. The interesting question is how large the effect is, whether it holds up once other risk factors are accounted for, and whether it is large enough to justify a policy or clinical response. Those are effect-size questions, and they get answered with numbers like odds ratios, hazard ratios, and relative risk — not with “p < .05.”

What an odds ratio of 1.29 actually says

Holt-Lunstad’s 2015 follow-up in Perspectives on Psychological Science reported three separate figures: an odds ratio of 1.29 for social isolation, 1.26 for loneliness, and 1.32 for living alone, each in relation to early mortality. An odds ratio of 1.29 means the odds of the outcome — dying earlier than expected — were about 29% higher among the socially isolated group compared to the reference group, holding the comparison as specified in the pooled studies.

Two things about that figure deserve attention before anyone reaches for a policy conclusion. First, 1.29 is a modest effect by the standards of clinical epidemiology, well below the kind of odds ratio associated with, say, heavy smoking and lung cancer. Second, and this is the part that gets lost when the finding is summarized in a press release, the review reported that the effects remained after adjusting for health status, and that social deficits were more predictive of death in samples averaging under 65 than in older samples. That second detail complicates a simple isolation-is-bad-for-elders narrative: the relative predictive power was actually stronger among younger cohorts.

The three separate figures — isolation, loneliness, living alone — are also not interchangeable, and the paper does not treat them as such. They are distinct constructs with overlapping but non-identical effect sizes. A brief that reports “social disconnection raises mortality risk by 30%” without specifying which of the three is being cited has already lost precision the original analysis took care to preserve.

Why “comparable to smoking” is a claim about magnitude, not mechanism

The comparison to established mortality risk factors, repeated across Holt-Lunstad’s work including the 2021 review in the American Journal of Lifestyle Medicine, is an effect-size claim. It says nothing about mechanism — whether loneliness kills through immune suppression, sleep disruption, cardiovascular strain, or behavioral pathways like reduced medical adherence is a separate question the meta-analytic effect size cannot answer. The 2021 review positions social connection as a modifiable protective factor suitable for population-level prevention, in the register of diet, exercise, and smoking cessation. That framing borrows credibility from decades of causal and mechanistic research on those other risk factors. The loneliness literature has an effect size in the same range; it does not yet have the same density of mechanistic and interventional evidence.

This matters for how much weight the comparison should bear. A 50% increased likelihood of survival associated with stronger relationships is a real, well-replicated statistical association across many cohort studies. Whether strengthening a given person’s relationships would produce a 50% survival benefit for that person is a different, causal claim, and it is not what the meta-analysis measured. Most of the underlying studies are observational. Reverse causation — poor health leading to social withdrawal, rather than isolation causing poor health — is a standing concern the field has tried to address through longitudinal designs and adjustment for baseline health, but it is not eliminated.

Where the size of the effect changes what follows

The National Academies’ 2020 consensus report on social isolation and loneliness in older adults did not simply cite these effect sizes and stop. It used them to argue that the health care system should routinely assess isolation and loneliness, reasoning that an effect of this magnitude, replicated across a large literature, clears the bar for clinical relevance even though it is not enormous. The report estimated that roughly one quarter of adults aged 65 and older are socially isolated — a prevalence figure, not an effect size, but one that matters because it tells a clinician how often the risk factor will actually appear in a caseload. A modest odds ratio applied to a quarter of the older population is a different policy problem than the same odds ratio applied to a rare condition.

The clinician-facing commentary on that report, published in the American Journal of Geriatric Psychiatry, pressed on what routine assessment would actually require: an instrument, a workflow, a referral pathway. None of that follows automatically from an odds ratio. The commentary’s contribution was to note that having a statistically robust, moderately sized effect is necessary but not sufficient for building a clinical response — someone still has to decide what threshold on what instrument triggers action, and no consensus figure in the literature answers that.

The AARP Foundation’s 2018 survey of adults 45 and older offers a useful contrast in how effect-size language shows up outside the epidemiological literature. Rather than an odds ratio, it reported a straightforward proportion: 33% of respondents who had spoken to their neighbors were lonely, against 61% of those who never had. That is not a mortality effect size and should not be treated as one — it is a cross-sectional association measured with the 20-item UCLA Loneliness Scale, comparable to academic usage but silent on causation and silent on magnitude in the odds-ratio sense. The nearly 30-point gap is striking as a descriptive fact. It says nothing about how much loneliness would fall if a lonely person started talking to neighbors, because the survey was not designed to test that.

What would sharpen this

The strongest version of this literature would report effect sizes with explicit comparison bands — this odds ratio versus that one, for the same outcome, in the same cohort, adjusted the same way — rather than leaving readers to import comparisons to smoking or obesity from a single meta-analysis. It would also separate the “how large” question from the “compared to what” question, since both get compressed into single soundbite comparisons. A trial that randomly varied an isolation intervention and reported a hazard ratio with a confidence interval, rather than an observational odds ratio, would move the field closer to a causal claim it does not yet have grounds to make. Until then, the size of the association is well established. What produces it, and how much of it is reversible, is not.