NHANES is an ongoing national cross–sectional survey administered by the National Center for Health Statistics (NCHS) to assess the health and nutritional status of the civilian population in the United States since 1999 (9). Research Ethics Review Board approved the underlying protocol and written informed consent was obtained from all participants. Details on NHANES study design, study protocol, and data collection have been described elsewhere (10). The questions on sleep duration were added to NHANES from 2005. We included data on adults (≥ 20 years) during the 5 cycles of the National Health and Nutrition Examination Survey (NHANES) from 2005 to 2014 by linkage to the National Death Index (until December 31, 2015) in a prospective design (Supplementary Figure S1). In addition, we further estimated the population attributable fraction for the 10–year risk of developing cardiovascular events due to short/long sleep duration from the most recent cycle of NHANES (2017–2018), which could provide the newest estimates (Supplementary Figure S1).

During the 5 cycles of NHANES from 2005 to 2014, sleep duration was determined using the question, “How much sleep do you usually get at night on weekdays or workdays?” The response was recorded as the number of hours of sleep obtained on a typical night, which was assessed as a continuous variable. In the cycle of NHANES (2017–2018), two questions were asked, “What time do you usually fall asleep on weekdays or workdays?”, and “What time do you usually wake up on weekdays or workdays”. Hours of sleep on weekdays or workdays were calculated from the reported usual sleep time and wake time during main sleeping period.

All–cause and CVD mortality were determined by linking to the National Death Index until December 31, 2015. The anonymized data of serial NHANES 2005–2014 were linked to mortality data with the assigned sequence number. Death from CVD was defined as codes I00–I09, I11, I13, I20–I51, and I60–I69 using the International Statistical Classification of Diseases and Related Health Problems, Tenth Revision (ICD−10). The follow–up time was identified as the period between the date of interview and the date of mortality or the end of follow–up (December 31, 2015), whichever occurred first.

The study covariates included demographic characteristics and lifestyle factors. Age, sex, race/ethnicity, educational level, smoking status, physical activity, alcohol consumption, body weight, and height were collected during interviews. Smokers were defined as individuals who reported smoking at least 100 cigarettes during their lifetime. Drinkers were defined as participants who drank at least 12 drinks of alcohol in any given year. Participants with more than 2.5 h of exercise per week were classified as physically active. Diet quality was assessed by Healthy Eating Index 2015 (HEI−2015) and the methods have been described previously (11). Comorbidity conditions were determined to be diagnosed and/or to take prescribed medications for the disease. Family history of chronic disease (heart disease and diabetes) were defined if participants reported that their close relatives had such diseases.

Sleep duration was categorized into five categories (≤ 5, 6, 7, 8, and ≥ 9 h). Cox proportional hazards models were used to estimate hazard ratios (HRs) and 95% CIs with 7 h regarded as reference. Model 1 was adjusted for age and sex. Model 2 was further adjusted for race/ethnicity, body mass index (BMI), educational level, physical activity, alcohol consumption, smoking status, and diet quality based on model 1. Model 3 was further adjusted for family history of chronic disease and comorbidity conditions based on model 2. To describe the potential non–linear association between sleep duration and all–cause mortality, cox proportional hazard models were used with sleep duration treated as continuous variables and fitted on restricted cubic splines with 3 knots. Likelihood ratio tests were utilized to examine the non–linearity of the exposures. We further applied stratification analysis for associations of sleep duration with all–cause and cardiovascular mortality according to several potential confounding factors such as age (younger or older than 65 years), sex, race/ethnicity (non–Hispanic white or other), education (less than college or college or above), BMI (< 30 or ≥ 30), smoking (never smoker or ever smoker), alcohol drinking (never drinker or ever drinker), physical activity (physically inactive or physically active). The survey–weighted Wald F statistic was used to evaluate the interaction effect between sleep duration and subgroup variables, and the P–value for multiple testing was adjusted using the Bonferroni correction with statistical significance set at P < 0.005 (0.05/8 [subgroups]). Several sensitivity analyses were performed by (1) excluding the participants with heart disease or cancer; (2) excluding the participants died in the first year of follow–up.

Under an assumption of causality, we estimated the proportion of CVD events attributable to short/long sleep duration in the US. Sleep duration and risk factors for CVD were basing on the data from the NHANES 2017–2018 cycle (12). Pregnant women and subjects with self–reported CVD were excluded from the analysis (13). The American Heart Association released a report that presented updated risk equations, the Pooled Cohort Equations (PCE), for CVD (13). The PCE provide estimates of the 10–year risk of CVD for men and women 40 to 79 years of age. The variables that statistically merit inclusion in the risk assessment equations are age, total cholesterol, high–density lipoprotein cholesterol, systolic blood pressure (including treated or untreated status), diabetes mellitus, and current smoking status. The 10–year risk estimate for CVD was calculated through a series of steps. The natural log of age, total cholesterol, high–density lipoprotein cholesterol, and systolic blood pressure were first calculated with systolic blood pressure being either a treated or untreated value. Any appropriate interaction terms were then calculated. These values were then multiplied by the coefficients from the equation for the specific race–sex group of the individual. The sum of the “Coefficient × Individual Value” was then calculated for the individual. The estimated 10–year risk of a first CVD event (R_p) is formally calculated as 1 minus the survival rate at 10 years, raised to the power of the exponent of the “Coefficient × Value” sum minus the race– and sex–specific overall mean “Coefficient × Value” sum. The risk equation has been validated and showed to be well calibrated in the US population (14, 15). The altered predicted 10–year risk for each person (R_q) by adjusting the sleep duration to 7 h was evaluated. Furthermore, the risk attributable to short/long sleep duration was calculated with the value of (R_p−R_q). If one adult slept x hours longer/shorter than 7 hours, the risk for the individual (RR_x) could be calculated as follows: RR_x = e^lnHR*x, where the HR of per hour increment (>7 h per day, HR 1.12 95% CI 1.06 to 1.16) and per hour decrement (<7 h per day, HR 1.06 95% CI 1.03 to 1.08) from a most comprehensive meta–analysis (5). If the participant's sleep duration became 7 h, the altered predicted 10–year risk (R_q) would be R_p/RR_x (16). We estimated Σ(R_p−R_q) × population size as a number of CVD events attributable to sleep duration in the US. In addition, population attributable fraction was derived as Σ (R_p−R_q)/ Σ (R_p) (17). Detail description of the analysis has been described in our previous publication (18).

All analyses incorporated sample weights, stratification, and clustering of the complex sampling design to ensure nationally representative estimates. SAS version 9.4 (SAS Institute Inc) and Stata version 14 (StataCorp LP, TX, USA) were used for the analysis. Statistical significance was set at a 2–tailed P < 0.05.

Of the 25 481 US adults, 12 388 (48.6%) were men (mean age, 46.3 years) and 13 093 (51.4%) were women (mean age, 47.8 years). During 146 484 person–years of follow–up, we documented 2033 deaths due to all causes, including 378 deaths due to heart disease and 461 deaths due to cancer. Table 1 showed the baseline characteristics of participants according to sleep duration. Compared with the subjects with sleep duration of 7 h/day, slept more than 9 h/day were more often older, female, non-smoker, non-drinker, physically inactive, were less educated (Table 1). Participants with an estimated sleep duration of ≤ 5 h were more likely to be minorities, non-smoker, non-drinker, physically inactive, less educated, and had higher BMI, and poor diet quality. There were more participants with extreme sleep duration reporting family history of chronic disease and having morbidity conditions, such as diabetes, heart disease, and cancer.

As shown in Table 2, after adjustment for age and sex, sleep < 5 h and more than 9 h per day were associated with significant elevated risks of all–cause mortality, and the HRs across categories were 1.79 (95% CI, 1.47–2.18), 1.20 (95% CI, 0.99–1.45), 1 (reference), 1.44 (95% CI, 1.18–1.75), and 2.06 (95% CI, 1.67–2.53). In the multivariable models, the association was slightly attenuated but a significant U–shaped relationship of sleep duration with all–cause mortality was still observed, and the multivariable–adjusted HRs across categories were 1.40 (95% CI, 1.14–1.71), 1.12 (95% CI, 0.91–1.38), 1 (reference), 1.35 (95% CI, 1.12–1.63), and 1.74 (95% CI, 1.42–2.12). In the restrict cubic spline, significant nonlinear association was observed between sleep duration and all–cause mortality (P < 0.001) (Figure 1).

As shown in Table 2, after adjustment for age and sex, sleep < 5 h and more than 9 h per day were associated with significant elevated risk of cardiovascular mortality, and the HRs across categories were 2.09 (95% CI, 1.28–3.39), 1.22 (95% CI, 0.83–1.81), 1 (reference), 1.62 (95% CI, 1.10–2.39), and 2.11 (95% CI, 1.28–3.45). In the fully adjusted models, the associations were largely attenuated, and the multivariable–adjusted HRs across categories were 1.66 (95% CI, 1.02–2.72), 1.15 (95% CI, 0.77–1.73), 1 (reference), 1.55 (95% CI, 1.05–2.29), and 1.81 (95% CI, 1.09–3.02). Significant non-linear association was observed between sleep duration and cardiovascular mortality (P < 0.01) (Figure 2).

In subgroup analysis, the results remained consistent in most subgroups and there was no significant interaction observed in each subgroup (Supplementary Tables S1, S2 in the Supplementary). In sensitivity analyses, the results remained similar after excluding deaths during the first year (Supplementary Table S3 in the Supplementary), and excluding subjects with history of heart disease or cancer at baseline (Supplementary Table S4 in the Supplementary).

We analyzed 1805 subjects aged 40 to 79 years in the National Health and Nutrition Examination Survey, 2017–2018 cycle. After accounting for the nationally representative, complex, stratified, multistage design and the sample weight, approximately 112.8 million adults were represented. The absolute event rates were estimated to be 9.2% (10.3 million CVD events) in the US from 2018 to 2028. Under an assumption of causality, 187 000 CVD events over 10 years (population attributable fraction 1.8%, 0.9% to 2.3%) were estimated to be attributable to short sleep duration and 947 000 CVD events (population attributable fraction 9.2%, 6.4% to 11.6%) were estimated to be attributable to long sleep duration (Table 3). Adults younger than 60 years, men, and the non–Hispanic black would have a greater proportion of CVD events related to short sleep duration, while older adults (60–79 years), women, and other races would have a greater proportion of CVD events related to long sleep duration.