Introduction

Over the past 30 years, there has been an increasing interest in the association between sleep and health outcomes1. Sleep problems are common and are anticipated to increase with the rapid advent of the ‘24/7’ society involving round-the-clock activities and increasing night-time use of television and internet1,2. Previous studies showed that short duration of sleep is associated with increased risks of stroke, coronary heart disease, metabolic syndrome, hypertension, central adiposity, obesity and type 2 diabetes2,3,18,19,20,21,22,23,24,25,26,27,34.

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 Study22. However, the U-shaped association was found in both subjects with one or more chronic diseases and subjects without chronic diseases in another study25. In addition, no statistical differences in association between sleep duration and all-cause mortality were found in stratified analysis by self-perceived health13,19, and excluding subjects with prevalent chronic diseases made little or no difference to the overall results35,36. 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-up11,13,14,16,24,27,15, and the U-shaped relationship was also more pronounced among subjects who napped daily40. 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 death10. Since the review was published, 19 population-based prospective cohort studies11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29 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 factors41,42,43,44,45 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 studies17,19. 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-analysis46, 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 studies15,20,25,26,39,40,47,48 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 covariates49.

Between-study heterogeneity was assessed with I2, and the values of 25%, 50% and 75% represent low, moderate and high heterogeneity50, 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 findings51. 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.

Additional Information

How to cite this article: Shen, X. et al. Nighttime sleep duration, 24-hour sleep duration and risk of all-cause mortality among adults: a meta-analysis of prospective cohort studies. Sci. Rep. 6, 21480; doi: 10.1038/srep21480 (2016).