Nighttime sleep duration, 24-hour sleep duration and risk of all-cause mortality among adults: a meta-analysis of prospective cohort studies

A dose-response meta-analysis was conducted to summarize evidence from prospective cohort studies about the association of nighttime sleep duration and 24-hour sleep duration with risk of all-cause mortality among adults. Pertinent studies were identified by a search of Embase and PubMed databases to March 2015. A two-stage random-effects dose–response meta-analysis was used to combine study-specific relative risks and 95% confidence intervals [RRs (95% CIs)]. Thirty-five articles were included. Compared with 7 hours/day, the RRs (95% CIs) of all-cause mortality were 1.07 (1.03–1.13), 1.04 (1.01–1.07), 1.01 (1.00–1.02), 1.07 (1.06–1.09), 1.21 (1.18–1.24), 1.37 (1.32–1.42) and 1.55 (1.47–1.63) for 4, 5, 6, 8, 9, 10 and 11 hours/day of nighttime sleep, respectively (146,830 death cases among 1,526,609 participants), and the risks were 1.09 (1.04–1.14), 1.05 (1.02–1.09), 1.02 (1.00–1.03), 1.08 (1.05–1.10), 1.27 (1.20–1.36), 1.53 (1.38–1.70) and 1.84 (1.59–2.13) for 4, 5, 6, 8, 9, 10 and 11 hours/day of 24-hour sleep, respectively (101,641 death cases among 903,727 participants). The above relationships were also found in subjects without cardiovascular diseases and cancer at baseline, and other covariates did not influence the relationships substantially. The results suggested that 7 hours/day of sleep duration should be recommended to prevent premature death among adults.


Discussion
Results from this dose-response meta-analysis showed that, compared with 7 hours/day, sleep duration was associated with the risk of all-cause mortality in a U-shaped manner among adults, regardless of nighttime sleep or 24-hour sleep. Overall, study design, population characteristics and other covariates did not influence the relationship substantially.
Five theoretical pathways for the relationship between short sleep duration and mortality have been put forth 32 : (I) short sleep directly causes mortality itself; (II) short sleep may result from variety of social, environmental, and physiological changes that lead to increased mortality risk; (III) short sleep itself causes physiological and social outcomes that may lead to increased mortality; (IV) short sleep is associated with other characteristics causally linked to mortality, such as age; and (V) the possibility of reverse causality is also of concern. Proposed mechanisms for mortality associated with long sleep include 33 : (I) long sleep is linked to increased sleep fragmentation that is associated with a number of negative health outcomes; (II) long sleep is associated with feelings of fatigue and lethargy that may decrease resistance to stress and disease; (III) changes in cytokine levels associated with long sleep increase mortality risk; (IV) long sleepers experience a shorter photoperiod that could increase the risk of death in mammalian species; (V) a lack of physiological challenge with long sleep decrease longevity; (VI) underlying disease processes mediate the relationship between long sleep and mortality. In addition, this meta-analysis suggested that the risk estimations of all-cause mortality associated with longer duration (10 hours and 11 hours) of 24-hour sleep were larger than those of nighttime sleep. The excess risk might be partly attributable to daytime napping that is positively associated with risk of all-cause mortality 34 .
One major concern with the observed associations is that short and long sleep duration might be just markers of poor health status rather than independent predictors. Among the studies included, a U-shaped association was found among subjects with one or more chronic diseases but no association was found among participants without chronic diseases in the 45 and Up Study 22 . However, the U-shaped association was found in both subjects with one or more chronic diseases and subjects without chronic diseases in another study 25 . In addition, no statistical differences in association between sleep duration and all-cause mortality were found in stratified analysis by self-perceived health 13,19 , and excluding subjects with prevalent chronic diseases made little or no difference to the overall results 35,36 . Reverse causality is of another concern; however, the included studies found that the associations did not changed substantially after excluding early death occurring within the first 1-5 years of follow-up 11,13,14,16,24,27,28,30,31,35,[37][38][39] . In addition, in this meta-analysis, the U-shaped relationship was also found among subjects without cardiovascular diseases and cancer at baseline. These findings indicated that preexisting chronic diseases might influence but not fully explain the observed associations. Furthermore, the observed associations remained after adjustment for other covariates. Therefore, these results suggested that both long and short sleep duration should be independent predictors of all-cause mortality.
Other factors influencing the observed associations should also be considered. Although no significant interactions between sleep duration and moderate-to-vigorous physical activity and body mass index on all-cause mortality were found in National Institutes of Health -AARP Diet and Health Study 28 , the U-shaped relationship was mainly observed among subjects with lower levels of physical activity in other studies 20,24 , and the interaction with body mass index was also found in another study 20 . In addition, the association of long duration of sleep with all-cause mortality was only observed among subjects taking napping 15 , and the U-shaped relationship was also more pronounced among subjects who napped daily 40 . However, the limited data precluded a more detailed analysis in this meta-analysis.
An earlier review found that both short and long duration of sleep are significant predictors of death 10 . Since the review was published, 19 population-based prospective cohort studies [11][12][13][14][15][16][17][18][19][20][21][22][23][24][25][26][27][28][29] were also included in this meta-analysis, making it possible to describe the dose-response relationship of sleep duration with risk of all-cause mortality, calculate the risk of all-cause mortality associated with specific duration of sleep, and assess whether other factors influence the association of sleep duration with all-cause mortality.
Other limitations should also be of noteworthy. First, sleep duration was self-reported and, therefore, might be subject to error and misclassification, and participants might also change their sleep pattern during follow-up. Second, while we extracted RRs that reflected the greatest degree of control for potential confounders, the extent to which they were adjusted for and potential residual confounding by other unmeasured factors [41][42][43][44][45] should also be considered. In addition, adjustment for potential factors that may be on the biological pathway between sleep duration and mortality risk would result in over adjustment. However, this meta-analysis showed that covariates adjusted for in the original studies did not influence the observed associations substantially. Third, publication bias could be of concern because small studies with null results tend not to be published; however, no publication bias was detected. In spite of the limitations, results from prospective cohort studies are still the best evidence available to assess the longitudinal effect of sleep duration on mortality.
In summary, results from this meta-analysis showed that both short and long sleep duration were independent predictors of all-cause mortality, and 7 hours/day of sleep should be recommended to prevent premature death among adults.

Materials and Methods
Literature search and selection. We performed a literature search up to March 2015 using the databases of Pubmed and Embase, using the following search terms ((((((prospective) OR cohort) OR longitudinal) OR follow-up)) AND ((death) OR mortality)) AND sleep), without restrictions. Moreover, we reviewed the reference lists from retrieved articles to search for further relevant studies. There is no protocol for this meta-analysis.
Two investigators (X.S. and Y.W.) independently reviewed all identified studies, and studies were included if they met the following criteria: (1) a prospective design; (2) the exposure of interest was sleep duration; (3) the outcome of interest was all-cause mortality; (4) relative risk (RR) with 95% confidence interval (CI) for 3 or more categories of sleep duration was provided; (5) the study was conducted among adults. If data were duplicated in more than one study, we included the study with the longest follow-up duration.  Data extraction. The following data were extracted from each study by two investigators (X.S. and Y.W.): publication year, the first author's last name, country where the study was performed, follow-up duration, total number of participants and death cases, health status of participants at baseline, sleep duration assessment method, variables adjusted for in the analysis, the numbers of death cases and participants (or person-years) and RR estimates with corresponding 95% CI for the each categories of sleep duration. Information was also gained with a request to the corresponding author of original studies 17,19 . We extracted RRs that reflected the greatest degree of control for potential confounders.
Statistical analysis. The trend from the correlated log RR estimates across levels of sleep duration was computed using a two-stage random-effects dose-response meta-analysis 46 , taking into account the between-study heterogeneity. In the first stage, the generalised least-square regression was used to estimate a restricted cubic spline model with three knots at the 25th, 50th and 75th percentiles of the levels of sleep duration. Then a multivariate random-effects meta-analysis was adopted to combine the study-specific estimates using the restricted maximum likelihood method. A P value for non-linearity was calculated by testing the null hypothesis that the coefficient of the second spline is equal to 0. This dose-response meta-analysis requires that the number of death cases and participants (or person-years) and RR estimates with corresponding 95% CI for the each categories of sleep duration should be available. The mean level of sleep duration in each category was assigned to the corresponding RR for every study. If the upper boundary of the highest category was not provided, we assumed that the boundary had the same amplitude as the adjacent category. For studies 15,20,25,26,39,40,47,48 that did not report the numbers of death cases for each categories of sleep duration, these numbers were inferred based on total numbers of cases and the reported crude risk estimates or risk estimates that were adjusted for least number of covariates 49 . Between-study heterogeneity was assessed with I2, and the values of 25%, 50% and 75% represent low, moderate and high heterogeneity 50 , respectively. Subgroup analysis and meta-regression were conducted to explore potential sources of heterogeneity and perform comparisons between groups, and a permutation test of 1000 was used to obtain P values from meta-regression to control spurious findings 51 . Publication bias was evaluated using Egger test. A sensitivity analysis was performed with one study removed at a time to assess whether the results could have been affected markedly by a single study. Study quality was assessed using the 9-star Newcastle-Ottawa scale (http://www.ohri.ca/programs/clinical_epidemiology/oxford.asp, accessed 12/21/2015). All statistical analyses were performed with STATA version 12.0 (Stata Corporation, College Station, TX, USA). All reported probabilities (P-values) were two-sided with P < 0.05 considered statistically significant.