The association between social ties and mortality

The finding has a long history. The Alameda County study in California, published by Lisa Berkman and Leonard Syme in 1979, and the synthesis by James House and his colleagues in Science in 1988 had already established the association in the epidemiological literature.

Julianne Holt-Lunstad, Timothy Smith and Bradley Layton published a meta-analysis covering more than one hundred and fifty follow-up studies, totalling over three hundred thousand participants.

It finds an association between indicators of social relationships and lower mortality, and the authors compare the size of that association with that of other known risk factors. The indicators pooled are highly heterogeneous, from marital status to participation in associations.

The comparison concerns statistical effect sizes, not mechanisms. It indicates that the measured association is of the same order, which is information about the data and not about causation.

The authors say so explicitly. The widely repeated formula comparing loneliness to smoking is a simplification of that point, and it has been criticised including by researchers sympathetic to the general finding. It also slides from isolation, which most of the pooled studies measured, to felt loneliness, which they measured far less.

Poor health reduces social participation. Studies excluding deaths occurring in the first years of follow-up, precisely to limit that bias, find reduced but persistent associations.

Adjusting for baseline health poses a symmetrical problem: if relationships act through health, adjusting for it removes part of the effect one is trying to measure. An analysis by Andrew Steptoe and his colleagues, published in 2013 on an English cohort, thus found that the association with isolation persisted after adjustment, whereas that with felt loneliness did not.

Three families are proposed: an effect on health behaviours, an effect on use of care, and a direct physiological effect through stress pathways. The first refers to concrete situations, such as having someone who notices a change or comes along to an appointment.

The evidence is uneven. The first two are documented through intermediate measures; the third rests mainly on short experimental work whose transposition remains disputed.

Trials aiming to reduce isolation measure effects on participation and sometimes on reported loneliness. None is sized to measure an effect on mortality.

That absence is structural: a trial able to detect such an effect would require sample sizes and durations beyond the ordinary funding of this field. The evidence therefore rests, and will no doubt long rest, on observational data, with the limits just described.