Study design and participants
The CLHLS is a nationwide community-based survey with one of the largest samples of people in the oldest-old age group (≥80 years) in the world. The first wave was carried out in 1998, and six follow-ups with participant replacement accounting for attrition were conducted in 2000, 2002, 2005, 2008, 2011, and 2014. The sampling strategy included an attempt to interview all centenarians in selected communities, and also adopted a targeted random-sample design to ensure representativeness with approximately equal numbers of male and female nonagenarians, octogenarians, and young-old (aged 65–79 years). More details regarding study design and data quality have been described elsewhere [21, 22].
The current study draws on data from the last four waves 2005–2014. In the 2005 wave, 10,400 people aged 80–105 were interviewed, and 7520 and 918 were added in the 2008 and 2011 waves, respectively. We first excluded 427 participants who had incorrect death dates (n = 114), lacked sleep duration information (n = 97), or provided implausible sleep duration of < 3 or > 16 h per day (n = 216). This left a sample of 18,411 participants, of whom 11,582 died during follow-up. All participants are eligible for inclusion in the analysis of sleep duration and mortality. Three thousand three hundred sixty-three individuals who dropped out before a follow-up observation were however excluded, resulting in a sample size of 15,048 for this analysis. Those that died during the study period are eligible for the analysis of sleep duration and QOD. Nine hundred thirty-eight of these decedents were omitted due to the absence of information on bedridden days or suffering state, leaving a final sample of 10,644 deaths for this analysis (Fig. 1).
Outcomes measures
Ascertainment of mortality- Date of death in the CLHLS study was obtained in multiple ways. The preferred method was obtaining dates from death certificates whenever such was available. Otherwise, dates were obtained through reports provided by next of kin or neighborhood committees. Mortality data in the CLHLS has been shown in previous analyses to be of high quality and accurate [23].
Assessment of QOD- Duration of being bedridden and degree of pain before death, as evaluated by next of kin, were adopted as components of QOD. Bedridden days before death were dichotomized into less than 30 days and 30 or more days before death, and painfulness of death was classified into “no suffering” and “suffering” [24]. Finally, we constructed two end-of-life categories by combining the two components, which were defined as “good” if a participant was bedridden for fewer than 30 days and experienced no suffering before death and as “poor” if a participant was bedridden for 30 or more days or experienced suffering before death.
Measurement of sleep duration
The 2005 survey was the first wave to assess sleep duration. Average sleep duration per day of a participant was obtained at baseline using the question “how many hours on average do you sleep every day?” with respondents giving an integer number.
Assessment of potential confounders
Covariates were assessed across several domains including sociodemographic characteristics (i.e., age at baseline, age at death, sex, region of residence, years of education, marital status, primary lifetime occupation, and economic condition), lifestyle (i.e., regular exercise, current smoking, and current drinking), and health factors (i.e., cognitive impairment, functional limitation, depression and chronic conditions).
Region of residence was classified as urban and rural. Marital status was dichotomized as “in marriage” if a participant was currently married with spouse present and “not in marriage” if divorced, widowed, separated or single. Primary lifetime occupation was defined according to the longest-held job during lifetime and categorized into white collar versus others. Economic condition was classified as good, fair, and poor by the question “Compared with other local people, how do you rate your economic position?”
Cognitive impairment was defined as a score lower than 18 (maximum score of 30) as assessed by the mini-mental status examination (MMSE). Functional limitation was defined as needing assistance in one or more activities of daily living (ADL) including bathing, dressing, eating, indoor transferring, toileting, and continence. Depression was defined by survey items that indicate always feeling fearful/anxious, lonely/isolated, or useless. Chronic conditions included obesity, hypertension, diabetes mellitus, cardiovascular disease, stroke, respiratory disease, and cancer. Obesity was defined as body mass index (BMI) ≥30 kg/m2, and the others were recorded via self-reported doctor diagnosis.
Health status was assessed at baseline and dichotomized into ‘good’ if individuals had no cognitive impairment, functional limitations, depression and reported no chronic conditions and ‘poor’ if individuals had at least one of these characteristics.
Statistical analyses
Baseline characteristics of participants were summarized by survivorship status, and by good versus poor QOD. Cox and logistic regression models with penalized splines were used to explore the shape of the association between sleep duration and all-cause mortality and poor QOD and to identify the sleep duration with minimum risk of mortality and poor QOD after adjustment [25]. We used penalized partial likelihood to estimate parameters with sleep duration as smoothed terms. According to the corrected Akaike information criterion for smooth regression functions, a degree of freedom of 3 for sleep duration was selected [26]. 95% confidence intervals were estimated using a bootstrap method with 10,000 replicates [27]. Finally, we grouped the participants into three categories-short, recommended, and long sleep duration-based on the intersection of optimal intervals of sleep duration which had the lowest risk of all-cause mortality and poor QOD.
To quantify the associations with all-cause mortality, a Cox proportional hazards model was used to calculate hazard ratios for sleep duration as a categorical variable. In minimally adjusted models, we controlled for age at baseline and sex. Fully adjusted models included age at baseline, sex, region of residence, years of education, marital status, primary lifetime occupation, economic condition, regular exercise, current smoking, current drinking, cognitive impairment, functional limitation, depression, cardiovascular disease, stroke, respiratory disease, and cancer. The relationship between the categorical measurements for duration of sleep and risk of poor QOD was examined using a binary logistic model with similar procedures as was applied to the analyses of all-cause mortality, but adjusted for age at death instead of age at baseline. Diabetes mellitus, hypertension, and obesity might be induced by short or long sleep duration and in the causal pathway linking sleep duration with death or QOD, and therefore, were not adjusted in the primary analyses.
The proportion of missing data was 2.4% for BMI, and less than 0.5% for other covariates. We applied multivariate imputation by chained equations generating five complete datasets to deal with missing data [28]. Variables included in the primary analyses were used in imputation models. The following built-in imputation models were conducted in our analyses: for continuous variables, predictive mean matching; for binary variables, logistic regression; and for ordered categorical variables, proportional odds [29]. Final statistical inferences were obtained by pooling the separate estimates from imputed datasets according to Rubin’s rule [30].
Sensitivity and subgroup analyses
Several sensitivity analyses were performed: (a) To examine the relation of sleep duration with mortality and QOD in all participants, we included individuals who reported the duration of sleep less than 3 h or more than 16 h per day; (b) To investigate the independent association of sleep duration with mortality and QOD, we further adjusted for diabetes mellitus, hypertension, and obesity at baseline; (c) To address the possible reverse causation between sleep duration and mortality, we omitted deaths that occurred in the first 2 years after entry; (d) To test the cut-off points of bedridden days for QOD, we used 44/45 days.
Some socioeconomic characteristics and health status might modify the association of sleep duration with mortality or QOD. Therefore, we evaluated interactions of these factors with sleep duration and conducted subgroup analyses among octogenarians, nonagenarians, and centenarians at baseline (applicable for all-cause mortality) and at death (applicable for QOD), men and women, urban and rural residents, participants in and not in marriage, participants with good, fair, and poor economic condition, and participants with good health status and poor health status.
All statistical analyses were conducted using R version 3.4.2 (R foundation for Statistical Computing).