Association between sleep duration and mortality risk among adults with type 2 diabetes: a prospective cohort study

The NHIS is a nationally representative and multistage stratified survey of non-institutionalised individuals in the USA. In the survey, one adult and one child are selected randomly from each household for a detailed interview on health status and lifestyle behaviours. Details of the survey design and methods are displayed at the website of NHIS from National Center for Health Statistics (NCHS) (www.​cdc.​gov/​nchs/​nhis/​about_​nhis.​htm; accessed 7 January 2018). NHIS was approved by the NCHS ethics review board. We could not influence the design of the prior studies upon which this work is based and cannot comment on individual ethics approval or consent. The data we used are publicly available and considered as exempt under the ethical board review of the corresponding author’s institution.

A total of 289,187 adults who participated in the NHIS from 2004 to 2013 were eligible to be linked to a mortality database over the follow-up period (up to 31 December 2015). After excluding missing data about sleep duration (n = 14,382), diabetes (n = 148), age at diagnosis of diabetes (n = 497) or use of oral glucose-lowering medication (n = 18) or insulin (n = 4), 274,138 remained. Among those with diabetes, we excluded individuals with possible type 1 diabetes (n = 1109), defined by use of insulin and age at diabetes onset <30 years (validated as accurate in 97% of cases [9]); the remaining 24,212 individuals were defined as having type 2 diabetes. The total study population of 273,029 adults also included 248,817 without diabetes. NHIS were reviewed and approved by the NCHS Research Ethics Review Board, and Institutional Review Board approval was not required because this study was based on secondary analyses of publicly available, de-identified data.

Mortality outcomes were determined by the National Death Index (NDI) records. Accuracy of the all-cause and cause-specific death information and consistency of the matching algorithm in the NDI records were verified [22]. The underlying causes of death were collected and classified according to the International Classification of Diseases (ICD-10) guidelines (http://​apps.​who.​int/​classifications/​icd10/​browse/​2016/​en). In addition to all-cause mortality, we were interested in the following cause-specific mortality outcomes as they are the eight leading causes of death in the general population: CVD, including heart disease and stroke; cancer; chronic lower respiratory disease (CLRD); Alzheimer’s disease; diabetes mellitus; influenza and pneumonia; and kidney disease. Electronic supplementary material (ESM) Table 1 shows specific codes for causes of death.

Sleep duration data were obtained from the following self-reported interview question: ‘On average, how many hours of sleep do you get in a 24 h period?’. The smallest unit of increments was 1 h, and sleep duration was subsequently categorised into six groups (≤5, 6, 7, 8, 9 and ≥10 h/day). Extreme sleep duration was defined as sleep duration ≤5 h/day or ≥10 h/day.

The following covariates were included as adjustment variables: age; sex; race (Hispanic, non-Hispanic White, non-Hispanic black, and others); education level (less than high school degree, high school degree, and more than high school degree); household income; lifestyle variables of BMI (normal weight or underweight [<25 kg/m²], pre-obese [25 kg/m² to ≤30 kg/m²], obese [>30 kg/m²]), physical activity (meeting or not the 2008 Physical Activity Guidelines for Americans that recommend at least 75 min of vigorous physical activity or 150 min of moderate physical activity in 1 week), smoking status (non-smoker, former smoker, current smoker), alcohol drinking status (lifetime abstainer, former drinker, current drinker); clinical variables with self-reported diagnoses of hypertension (e.g. BP), CHD, stroke and cancer; and calendar year. We also collected and categorised the duration of diabetes into four groups of 0–5, 6 to ≤10, 11 to ≤20 and >20 years. Age at diagnosis of diabetes was classified into two groups (≤45 or >45 years of age [young and old, respectively]) [23].

Medication status was stratified into four categories: insulin only, oral glucose-lowering medication only, both and no medication.

Descriptive statistics were used to report the distribution of participants’ baseline characteristics by sleep duration. Continuous variables were displayed as mean ± SE, and categorical variables were displayed as percentage (%). Continuous data were compared with analysis of t test and variance, while Pearson’s χ² tests were used to compare differences in baseline characteristics (including the exposure and covariates). Mortality rates were age-standardised to the overall NHIS using the age groups of 18–44 years, 45–64 years and 65 years or older. The graphical assessment of log–log plots was used to assess the proportional hazards assumption and the assumption was met in each of the models [24]. We used a multivariate Cox proportional hazards regression model to estimate HRs with corresponding 95% CIs to test the association between sleep duration categories and all-cause and cause-specific mortality. The reference group was 7 h/day. In multivariate models, we adjusted for age (as a timescale), sex, race, education level, household income, duration of diabetes (years), BMI, physical activity, smoking status, alcohol drinking status, hypertension, CHD, stroke, cancer and calendar year. Moreover, we also conducted subgroup analyses for all-cause and CVD-related mortality stratified by sex, duration of diabetes and age at diagnosis. Adjusted Wald test accounting for complex multistage sampling design was used to test for interaction term. To examine the robustness of the results, we performed sensitivity analyses excluding individuals with a history of CHD and stroke or cancer. Sampling weights were used to account for the unequal probabilities of selection. SEs were calculated by Taylor series linearisation. A p value <0.05 was considered statistically significant (two-tailed). All data analyses were performed using STATA version 12.0 (Stata Corp, College Station, TX, USA).

Among the 248,817 adults without diabetes, there were 17,060 deaths during the mean 6.70 years of follow-up (1.62 million person-years). A total of 4593 deaths (2331 for women, 2262 for men) were recorded during a mean follow-up of 5.96 years (0.14 million person-years) among the 24,212 participants with diabetes. Baseline characteristics of participants with or without diabetes by sleep duration are presented in Table 1 and ESM Table 2, respectively.

Figure 1 shows age-standardised mortality rate per 10,000 person-years stratified by the presence of diabetes and sleep duration. Mortality rate was highest with extreme sleep duration (≤5 h/day and ≥10 h/day) and lowest with sleep duration of 7 h/day, regardless of the presence of type 2 diabetes (Table 2).

Extremes of sleep duration were associated with increased risk in mortality among adults with and without diabetes (Fig. 2). Individuals with type 2 diabetes who reported the shortest and longest sleep duration had higher all-cause mortality risk than the non-diabetic group who slept for 7 h/day (≤5 h/day, HR 1.63 [95% CI 1.24, 2.13]); ≥10 h/day, HR 2.17 [1.72, 2.73]) (Table 2). There was a non-significant interaction between sleep duration and the presence of diabetes (p for interaction = 0.08).

Among individuals with type 2 diabetes, compared with the reference group (7 h/day sleep), both shorter and longer sleep duration were associated with increased mortality risk (≤5 h/day, HR 1.24 [1.09, 1.40]; ≥10 h/day, HR 1.83 [1.61, 2.08]) (Table 3 and Fig. 3). A subgroup analysis showed a similar J-shaped curve among women and men, respectively (Table 3 and Fig. 4a).

People with type 2 diabetes who slept for ≤5 h/day, 8 h/day and ≥10 h/day had 41%, 26% and 59% greater risk of cancer mortality, respectively, compared with those who slept for 7 h/day (Table 3 and Fig. 4b). The association between sleep duration and CVD mortality risk in individuals with type 2 diabetes was only statistically significant for people who slept for ≥10 h/day (HR 1.74 [1.33, 2.28]) (Table 3 and Fig. 4c). For heart disease mortality specifically, the excess risk was only statistically significant for ≥10 h/day sleep duration (HR 1.51 [1.11, 2.07]). Of all the CVD types, the most prominent sleep–mortality association was observed for stroke mortality (sleep duration ≥10 h/day, HR 2.95 [1.63, 5.34]). For other causes of death, we observed a statistically significant association between sleep duration and mortality risk from kidney disease, Alzheimer’s disease and CLRD (Table 3). Although the subgroup analysis by sex showed higher HRs among men, the test for interaction term was not statistically significant due to wider CIs (Table 3).

Sensitivity analysis was performed to test the robustness of the main effects models. After eliminating participants with a history of CVD and/or cancer, results remained largely unchanged for all associations (ESM Table 3).