Does sleeping too little accelerate aging—and is too much sleep harmful too?

01What we found

A nonlinear association between reported sleep and mortality remains. In the complete-case main model, hazard ratios relative to seven hours were 1.22, 1.17, 1.21 and 1.32 for ≤5, 6, 8 and 9+ hours, respectively; the spline minimum was approximately seven hours. The two ends behave somewhat differently under restrictions, but these analyses do not identify causality. At a three-year landmark, the short-sleep interval includes one while the long-sleep interval remains above one. In our baseline-healthy subgroup, both intervals are wide and the long-sleep point estimate is close to one. The long-sleep association is relatively strong for heart-disease mortality, without establishing that heart disease accounts for almost all of the all-cause association. Simulation illustrates how adjustment can absorb a stipulated reporting-bias mechanism; it does not exclude reporting error in real people. Seven hours is this model's association minimum, not an individual sleep prescription.

Correction, September 19, 2026: We corrected the activity variable's “any yes” coding, functional-questionnaire special codes, age ceilings and survey standard errors. Covariate models now use the same complete-case sample. The original overall spline test was mislabeled as a nonlinearity test; a genuine nonlinearity test has now been added. Two literature links were corrected. We withdrew claims that heart disease drives the long-sleep excess, that observed data rule out reporting bias through measured health, and that differences in statistical significance establish differences in causal strength.

02What “hours of sleep” means here

Exposure is SLD010H: usual nighttime sleep on weekdays or workdays. Wording is consistent across the five cycles. A response of 12 means 12 hours or more; 77 and 99 represent refusal and unknown responses. Seven and nine are valid durations. SLQ documentation

Questionnaires, examinations and mortality files were linked by SEQN. Eligible participants were at least 20, had valid MEC weights, were eligible for mortality linkage and had an observed sleep response. Follow-up uses months since examination in the NCHS 2019 public files, with zero months treated as half a month. Selected follow-up times and causes are perturbed for confidentiality; vital status is not changed by that process. These are not untouched copies of complete death-register records. NCHS

Covariates include demographics, smoking, BMI, alcohol history, activity, self-rated health, counted chronic conditions, depression and functional limitation. ALQ101 asks whether a person has ever had at least twelve drinks in any one year, not whether they currently drink or drank in the previous year. It is a coarse measure of alcohol history. The 2005–2006 activity questionnaire differs from later work/leisure questions; we therefore also exclude the first cycle in a sensitivity analysis. “Do not do this activity” is not the same as inability, and functional-item skips follow verified screening rules. Public age values are harmonized to 80+, leaving actual age differences within that group unobserved. Covariate dictionaries

Because examination variables are included, we use WTMEC2YR/5 and estimate variance with SDMVSTRA and SDMVPSU. All HR confidence intervals and P values use the same survey-design covariance. Analytic guidance In the original eligible sample, alcohol history and depression are missing for approximately 13.2% and 14.1%; income and BMI also have missingness. Complete cases leave 22,120 participants and 2,829 deaths, losing approximately 22% of participants and 31% of events. We did not perform multiple imputation or identify the missingness mechanism.

Reported duration is not interchangeable with objectively measured sleep. MESA comparisons of self-report, actigraphy and polysomnography found systematic differences and limited agreement. Jackson 2018 An average error in another cohort cannot serve as a constant correction that shifts our curve's minimum one hour to the left.

03The curve remains, but adjustment is not causal decomposition

The original eligible sample contains 4,410, 6,660, 7,453, 7,594 and 2,238 people in the ≤5, 6, 7, 8 and 9+ hour groups. Weighted crude mortality rates are approximately 13.5, 7.7 and 22.5 per 1,000 person-years in the ≤5, seven-hour and 9+ groups. These rates mix sleep with differences in age and health composition.

M1–M3 use the same 22,120 complete cases, preventing changes in sample membership from masquerading as attenuation after adding covariates:

Point estimates decline with health adjustment. That is compatible with health status contributing to the observed pattern, but the HR difference is not the fraction of risk explained by health. Covariates may be confounders or lie on pathways from sleep to health, and hazard ratios are noncollapsible. This analysis does not identify mediation or a causal decomposition.

The natural spline has internal knots at six, seven and eight hours. Its grid minimum is approximately seven hours. The four-coefficient overall association test gives p approximately 1.6 × 10⁻⁵; the three independent contrasts testing departure from a linear direction give nonlinearity p approximately 5.6 × 10⁻⁵. These test different hypotheses. The minimum has no confidence interval and depends on reporting, knots and adjustment. It does not establish an optimal duration.

Adding reported sleep trouble and diagnosed sleep disorder leaves a long-sleep HR of approximately 1.31. An exploratory sleep-by-follow-up-time interaction at five years gives p=0.82, without detecting that particular time variation. It does not prove every proportional-hazards assumption. The original unadapted Schoenfeld test is no longer used as survey-valid evidence.

04What the sensitivity analyses can address

Two- and three-year landmark analyses include only people still at risk at the landmark and restart follow-up there. They address the influence of early deaths while changing the target population; they cannot eliminate all reverse causation.

The healthy subgroup requires self-rated health of at least good, none of the counted chronic conditions and no functional limitation as defined here. It contains 12,527 participants and 560 deaths. Self-rated health, which still varies within this group, remains adjusted for. The lower long-sleep point estimate makes baseline health and reverse causation plausible explanations worth considering, but precision also falls. A significant result in one subgroup and a nonsignificant result in another do not establish a significant difference between them. These analyses neither identify risk as concentrated in people with disease nor establish a causal effect at either end.

In the cycle-interaction model, long-sleep point estimates exceed one in all five cycles, but only 2005–2006 and 2011–2012 have intervals excluding one. These contrasts share covariate coefficients; they are not five independent replications. Wider intervals in later, shorter-followed cycles also do not identify a biological latency period.

05A clearer heart-disease signal does not account for the whole association

The main model includes 726 heart-disease deaths, 694 cancer deaths, 162 stroke deaths and 1,247 deaths from all remaining causes. Relative to seven hours, cause-specific HRs for 9+ hours are:

The heart-disease association is clearer among these comparisons, but other causes remain uncertain, and we did not estimate their contributions to all-cause excess risk. For ≤5 hours and cancer mortality, HR is 1.32 (0.90–1.95), p=0.16. The original narrower interval and its accompanying P value were calculated with different standard errors and should not have been reported together.

These Cox models remove competing deaths from the risk set when they occur. They estimate cause-specific hazards, not the cumulative probability of dying from that cause by a fixed time. This is a defined estimand, not merely an approximation to marginal mortality risk. Cumulative probabilities require a separate competing-risk analysis. Multiple cause and sleep-category comparisons are exploratory; selecting the most significant cell cannot establish a mechanism.

06A reporting-bias simulation is an example, not a real-world bound

The simulation preserves observed covariates, follow-up and outcomes, draws a random sleep duration, then adds a reporting shift determined by measured health indicators plus random noise. Across shifts γ from zero to three hours, with 40 repetitions per value, adjustment for those same health indicators leaves median simulated long-sleep HRs near one. The shaded band spans the 10th–90th percentiles of simulation results, not a confidence interval for real-world bias. For computational efficiency, it uses unweighted Cox models and compares them with the observed association fitted on the same unweighted basis.

The exercise shows that adjustment can absorb an association generated through the selected indicators under this setup. It does not establish that disease severity is measured accurately in real life or that the fitted functional form is correct. Unmeasured disease, medication, sleep-disordered breathing, selection and more complicated measurement processes remain possible explanations. Nondifferential error also cannot universally be summarized as attenuating associations without ever changing their shape in nonlinear models. This simulation neither estimates a causal share nor rules out reporting-error explanations.

07How this fits existing research

Earlier NHANES analyses also reported similar curves, but participants and mortality files overlap with ours; these are not additional independent replications. Gu 2024 2023 analysis 2022 analysis Our contribution is transparent coding, common-sample adjustment, estimates under restrictions, coverage of causes and explicit limits on simulation-based inference.

Cappuccio and colleagues' 2010 prospective-cohort meta-analysis reported RRs of 1.12 for short and 1.30 for long sleep. Jike and colleagues' 2018 review focused on long sleep. Itani and colleagues' 2017 review concerned short sleep, with mortality RR 1.12; it should not be cited as an analysis of both ends. Cappuccio Jike Itani These are observational syntheses of different populations, reference groups and adjustments. They do not causally validate our estimates.

This study cannot say how many years an individual would gain by changing sleep duration or infer aging speed from mortality associations. The clearest conclusion is that reported sleep duration jointly reflects sleep, health and living conditions. Single reports, complete-case selection, coarse covariates, top-coding, limited subgroup events and unmeasured confounding constrain interpretation. The observed minimum is not a personal target. Objective and repeated measurements, alongside stronger causal designs, are needed to distinguish these explanations further.