Short sleep and mortality: what the cohorts actually associate

A per-hour meta-analysis of prospective cohorts put the risk at 1.06 per hour of sleep lost below seven hours. A 2025 meta-analysis of 79 cohorts put short sleep at 1.14 and long sleep at 1.34. Both are cohort associations, and both say so.

goodWell supported, with the usual caveats about effect size.Revised 20 Oct 2026

An empty bed with plain white linen and a folded pale blue blanket in a dim room with a bedside table and a blank-faced clock

The curve is a U and both arms of it are elevated. That is the finding, and it is not a reason to start sleeping more than you need.

The mortality literature on sleep is the widest dataset in this blog and also the one most often quoted as if it were an experiment. It is not. Everything in this article is observational, and the two meta-analyses below are explicit about that in their own design sections.

The per-hour estimate

Yin and colleagues, in the Journal of the American Heart Association 2017, ran a systematic review and per-hour meta-analysis of prospective cohort studies published before 1 December 2016, in generally healthy populations, examining sleep duration against all-cause mortality, total cardiovascular disease, coronary heart disease and stroke.

The associations were U-shaped for all four outcomes, with the lowest risk at roughly seven hours of sleep per night, and that optimum varied little by sex. For all-cause mortality: pooled relative risk 1.06 (95% CI 1.04 to 1.07) per one-hour reduction below seven hours, and 1.13 (95% CI 1.11 to 1.15) per one-hour increment above seven hours.

Read those two numbers side by side, because they are almost always quoted one at a time. Sleep that is one hour short of seven carries about six per cent more risk; sleep that is one hour long carries about thirteen per cent. The long arm of the U is steeper in these data, which is counter to the intuition that oversleeping is harmless.

The other three outcomes follow the same shape with the same asymmetry. Total cardiovascular disease: 1.06 (1.03 to 1.08) and 1.12 (1.08 to 1.16). Coronary heart disease: 1.07 (1.03 to 1.12) and 1.05 (1.00 to 1.10). Stroke: 1.05 (1.01 to 1.09) and 1.18 (1.14 to 1.21).

The category estimate, eight years later

The 2025 Geroscience meta-analysis from Ungvari and colleagues is larger and categorises rather than computing per-hour gradients. Systematic search to October 2024 across PubMed, Cochrane Central and Web of Science, retrospective and prospective cohort studies with at least one year of follow-up, hazard ratios pooled with a random-effects model, subgroup analyses by sex. Seventy-nine cohort studies.

Short sleep, under seven hours per night, against a seven-to-eight-hour reference: pooled hazard ratio 1.14 (95% CI 1.10 to 1.18). Long sleep, nine hours or more: hazard ratio 1.34 (95% CI 1.26 to 1.42). Both associations were present in men and women, with the long-sleep effect more pronounced in women.

The title of that paper — “imbalanced sleep increases mortality risk by 14–34%” — is the whole asymmetry in two numbers, and it is the same asymmetry the per-hour estimates show. Two meta-analyses, different methods, eight years apart, same shape.

The other arm, measured against incidence

Jike and colleagues’ 2018 Sleep Medicine Reviews analysis is the companion piece, with 137 prospective cohorts and 5,134,036 participants. Long sleep was significantly associated with mortality (RR 1.39, 95% CI 1.31 to 1.47), incident diabetes mellitus (1.26, 1.11 to 1.43), cardiovascular disease (1.25, 1.14 to 1.37), stroke (1.46, 1.26 to 1.69), coronary heart disease (1.24, 1.13 to 1.37) and obesity (1.08, 1.02 to 1.15). Not significantly associated with incident hypertension (1.01, 0.95 to 1.07), with insufficient data for depression and dyslipidaemia.

Their conclusion is the sentence to keep next to every number on this page: future studies should address whether the relationship is causal and modifiable. They are not claiming it is causal. They are saying it is not known.

The randomised data point that does exist

This blog’s entry on sleep and glucose metabolism covers it in detail, and the figure belongs here because it is the only randomised evidence in the set. In Zuraikat and colleagues’ 2024 crossover, 38 women went through two six-week phases; curtailing sleep to 6.2 hours per night for six weeks impaired insulin sensitivity, with fasting insulin up 6.8 ± 2.8 pmol/L and HOMA-IR up 0.30 ± 0.12, both P = 0.016.

Six weeks. Insulin sensitivity is not mortality. The mechanism is plausible and measured; the endpoint is six weeks long.

The cohort genre has a direct parallel in what the human data on NAD and NMN show — one exposure, one endpoint, and the same ceiling on what an association can carry: the whole field there rests on one randomised trial, of one precursor, at one amount.

What the data do not show

They do not show that sleeping more extends life. The long arm of the U is almost certainly doing something other than causing harm directly: illness causes both fatigue and early death, and a cohort cannot separate those. They do not show a threshold you can act on with confidence — “approximately seven hours” is an average across populations, not an individual target, and it varied little by sex in the per-hour analysis without being tested for individual variation. And the strongest number available here, 1.14, is a hazard ratio against a reference band, from observational data, in a literature where measurement error in self-reported sleep is known to be substantial and is not adjustable for in the pooled estimates.

This article summarises cohort meta-analyses and one randomized trial. It is not medical advice and it is not a recommendation for any individual.