Data for this present prospective study was derived from the China Health and Retirement Longitudinal Study (CHARLS), which was an ongoing cohort study of middle-aged and older adults in China conducted by the National School for Development at Peking University. The multistage stratified probability sampling technique was used in CHARLS to get a representative population sample obtained from 150 counties within 28 provinces, municipal cities, and autonomous regions. The CHARLS collected information on demographics and household information, biomedical measurement, health status, and functioning, etc. with face-to-face computer-assisted personal interviews (CAPIs) for the first time in 2011–2012 and every 2 years thereafter. The CHARLS was approved by the Ethics Committee of Peking University Health Science Center. Details of the survey design have been described elsewhere (21).

The baseline questionnaires and physical examinations were conducted with 17,705 participants and 11,847 respondents (67%) completed the blood test. Of these, 10,384 participants were successfully re-interviewed in 2015–2016 and 7,648 individuals (74%) of them completed the blood test. Additionally, we excluded 1,711 individuals with missing information related to sleep duration, blood lipids, or other covariates. The final sample was composed of 5,016 adults after excluding individuals with cancer, cardiovascular disease, or stroke (n = 788), and those who were under 45 years old (n = 133). See Figure 1 for further details on the inclusion and exclusion of study participants.

Nightly duration of sleep was assessed by asking subjects the following questions, “During the past month, how many hours of actual sleep did you get at night (average hours for one night)? (This may be shorter than the number of hours you spend in bed.).” Afternoon naps were assessed by asking, “During the past month, how long did you take a nap after lunch in general?” Participants were classified as either having regular lunchtime naps or not. Triglycerides (TG), high-density lipoprotein cholesterol (HDL-C), low-density lipoprotein cholesterol (LDL-C), and total cholesterol (TC) concentrations were determined using an enzymatic colorimetric test. All respondents were asked to have fasted overnight. Sleep duration and lipids data were analyzed as continuous variables in the cross-lagged analysis.

Sociodemographic characteristics, habits and customs, and health status were evaluated as potential confounders of relationships between sleep duration and blood lipids. Age, gender, region (rural-urban), educational level, medical insurance, and marital status were obtained through a self-reported questionnaire at the baseline survey. The health-related covariates included smoking and alcohol consumption. Participants who had either completely quit smoking or had never smoked were defined as non-current smokers. Current drinkers were defined as if they drank liquor, beer, wine, or other alcoholic beverage more than once a month in the last year. The depressive symptoms of the participants were evaluated using the 10-item Center for Epidemiologic Studies Depression Scale (CESD-10) short form. Each item was scored on a 4-point scale, with a total possible score ranging from 0 to 30, and the score of 12 was used as a cutoff point to determine binary depression symptoms (22). Diastolic and systolic pressures were the averages of the three repeated measurements of blood pressures, which were measured by an Omron (Dalian, China) HEM-7200 monitor after the participants had rested for 30 min. Bodyweight, height and circumference were measured to the nearest 0.1 kg, 0.1 cm, and 0.1 cm, respectively, with subjects in light indoor clothing. Abdominal obesity was defined as a waist circumference ≥90 in man and ≥80 in women (23). The venous blood samples were collected during the investigation by trained nurses and were transported based on standard protocol to the Chinese Center for Disease Control and Prevention within 2 weeks, where the samples were stored at −80 and then were assessed at the Clinical Laboratory of Capital Medical University. The glucose was measured using the enzymatic colorimetric test, and HbA1c was measured using boronate affinity chromatography (24).

Continuous and categorical variables are presented as the mean ± SEM or as numbers and percentages, respectively. The significance of differences between BMI and age groups was assessed by non-parametric tests for skewed continuous variables, and χ² tests for categorical variables. As a form of path analysis, cross-lagged panel model was used to estimate the directional influence inter-related variables have on each other over time in the field of sociology and medicine (25–27). Figure 2 provides graphical representations of the cross-lagged panel model. Cross-lagged panel model is a kind of structural equation modeling, which can examine the causal relationship between variables. Instead of modeling separately, this approach can simultaneously detect the relationship between two variables over 2 years by testing stability paths (e.g., sleep duration at baseline to sleep duration at follow-up), concurrent paths (e.g., sleep duration at baseline and HDL-C at baseline), and cross-lagged paths (e.g., sleep duration at baseline to HDL-C at follow-up; HDL-C at baseline to sleep duration at follow-up). According to the principle that the paths in a model cannot be saturated, the residual terms of the two dependent variables (e.g., sleep duration at follow-up and HDL-C at follow-up) were assumed to be uncorrelated. The robust maximum likelihood method, which was recommended to produce efficient estimates of the parameters in the structural equation model when the observations did not satisfy the assumption of normal distribution, was used to estimate the parameters in the cross-lagged panel model (28). The model fit was evaluated using the comparative fit index (CFI) (≥ 0.90 was considered acceptable) (29), the Standardized Root Mean Square Residual (SRMR) (<0.08 was considered acceptable) (30), and Root Mean Square Error of Approximation (RMSEA) (<0.06 was considered acceptable) (30). The difference between β₁ and β₂ derived from the standardized variables (Z-scores) was examined using Fisher's Z test. Stratified analyses were performed by baseline characteristics (age [45- <60, ≥ 60 years], BMI [ ≤ 25, > 25 kg/m²] to test the difference in cross-lagged path parameters between groups. Population was divided into under-weight (BMI <18.5 kg/m²), normal weight (18.5 ≤ BMI <25 kg/m²), overweight (25 ≤ BMI <30 kg/m²) and obese (BMI ≥ 30 kg/m²) according to World Health Organization divided population. We set the BMI cut-off point to 25, as participants with a BMI below 18.5 or >30 accounted for only about 8% of the total. The cut-off point for age was set at 60 years, giving rise to a middle-aged and older adults group. The initial model was used to test the potential bidirectional associations between sleep duration and blood lipids, including triglyceride, HDL-C, LDL-C and TC. Potential confounders were adjusted for gender, smoking status, drinking status, region, marital status, education, insurance status, depression, systolic, diastolic, glucose, HbA1c, BMI (not included in the stratified analysis according to BMI) and age (not included in the stratified analysis according to age). The path analyses were conducted in R by using package lavaan, which had been developed to provide applied researchers package for structural equation modeling (31). All the statistical analyses were performed using R version 3.5.3 (R Foundation for Statistical Computing, Vienna, Austria).

Table 1 provides a descriptive summary of the variables for the entire study population categorized by BMI and age group. Nearly 85% of all participants lived in rural areas, and more than 73% of the respondents were non-illiterate. Sleep duration at baseline and follow-up differed at statistical significance among the BMI and age categories, and the proportion of participants who sleep longer (>9 h) or shorter (<6 h) was higher in the older people. Among subjects with BMI > 25 kg/m², LDL-C, TC and TG were significantly higher, and HDL-C was significantly lower compared with those with lower BMI. By contrast, only baseline blood lipid levels were significantly or marginally different between age groups. Participants over 60 years old tended to have higher glucose, HbA1c and pulse pressure. Pulse pressure was defined as the difference between systolic and diastolic measurements of blood pressure. And those with higher BMI tended to have higher systolic, diastolic blood pressure, glucose and HbA1c.

To examine whether there are linear relationships between sleep duration and blood lipid levels 4 years later as well as blood lipid levels and sleep duration 4 years later, Table 2 presents the means and confidence intervals for TG, HDL-C, LDL-C, and TC in the different sleep duration groups. TG and sleep duration at baseline were positively correlated with sleep duration and TG at follow-up, respectively. The tracking correlations of HDL-C and sleep duration from baseline to follow-up and the synchronous correlations between HDL-C and sleep duration at baseline and follow-up were both negative. The results for LDL-C and TC were similar to those of HDL-C, showing negative tracking correlations between blood lipid levels and sleep duration. The directions of correlations among the study variables showed consistency with the proposed cross-lagged models.

Figure 2 shows cross-lagged path analysis of sleep duration and blood lipid levels in the total population. After adjusting for gender, age, smoking status, drinking status, region, marital status, education, medical insurance, depression, systolic, diastolic glucose, and glycosylated hemoglobin, there was a significant bidirectional relationship between sleep duration and HDL-C level. Sleep duration was negatively associated with HDL-C 4 year later (β₁ = −0.171, P = 0.005), and HDL-C was negatively associated with sleep duration 4 year later (β₂ = −0.006, P = 0.002). In contrast, there was a positive correlation between TG and sleep duration, and the path coefficient from TG to sleep duration 4 year later (β₂ = 0.001, P = 0.018) was greater than the path coefficient from sleep duration to TG 4 year later (β₁ = 0.109, P = 0.847), with P = 0.030 for the difference between β₁ and β₂. Moreover, longer sleep duration were significantly associated lower levels of LDL-C (β₁ = −0.339, P = 0.036) and TC (β₁ = −0.506, P = 0.010) 4 year later, and the sleep duration were marginally significantly associated with LDL-C (β₁ = −0.275, P = 0.097) and TC (β₁ = −0.329, P = 0.096) 4 year later after adjusting for covariates. Fit indexes of all the cross-lagged models indicated a good fit to the observed data according to the criteria of CFI ≥ 0.90, SRMR <0.08, and RMSEA <0.06.

Tables 3, 4 provide a summary of cross-lagged path analysis of TG, HDL-C, LDL-C, TC, and sleep duration by age and BMI. Significant longitudinal association between TG at baseline and sleep duration at follow-up was observed among subjects with BMI > 25 kg/m² (β₂ = 0.001, P = 0.019) or age ≥ 60 years old (β₂ = 0.001, P = 0.026). The direction of effect between HDL-C and sleep duration which was bidirectional in all subjects differed across various subgroups. HDL-C at baseline was significantly associated with sleep duration at follow-up subgroups of BMI > 25 kg/m² (β₂ = −0.012, P = 0.001) or age ≥ 60 years old (β₂ = −0.011, P = 0.001) with P = 0.003 and P = 0.046 for the difference in the path coefficients (β₁ and β₂), respectively. In contrast, the path coefficients from sleep duration at baseline to HDL-C at follow-up in subgroups of BMI ≤ 25 kg/m² (β₁ = −0.267, P = 0.001) and aged <60 years (β₁ = −0.236, P = 0.007) were significant. In addition, HDL-C and TG were strongly correlated with BMI in the total study population. At baseline, the correlation coefficient between HDL-C and BMI was −0.284 (P < 0.001), and that between TG and BMI was 0.217 (P < 0.001).

There are several limitations that need to be considered in our study. First, the subjective self-reported sleep duration obtained through the questionnaire might be biased, which could be caused by systematic over-reporting. Second, although several potential confounders were included in our analysis, there was a lack of physical performance, diet measures and menopause status, which could be related to sleep and blood lipid levels (44). Third, measures of insomnia, obstructive sleep apnea syndrome and sleep symptoms, such as snoring, which have been suggested to be associated with metabolic syndrome (45), were not investigated in the present study. Fourth, although the number of eligible subjects in the study was large enough, the proportions of participants who were successfully followed up and were able to be collected blood samples at each time point were about 70%, which might lead to selection bias common in longitudinal studies. Lastly, although path analysis permits the exploration directional effect variables have on each other over time, it simply assumed that the research variables are linearly correlated. In the present study, there was no obvious other relationship such as “V” or “U”-shape between variables and the proposed models showed a good fit to the data; however, there may be more subtle relationship between variables, and the shape of associations may vary among the population of different ages or races. Hence, further prospective and mechanistic studies are required to validate our results on timing relationships and causality.

In conclusion, this study aimed to clarify the temporal relationship between blood lipids and sleep duration in Chinese middle-aged and older adults by employing the two-wave cross-lagged design. Our research revealed that negative associations existed between HDL-C and sleep duration and the relationships were bidirectional across the general population. TG has a positive longitudinal association with sleep duration, and sleep duration is negatively associated with LDL-C and TC. The strength and direction of the relationship are inconsistent among different ages and BMI groups. The longitudinal association between HDL-C at baseline and sleep duration at follow-up was more pronounced in individuals with BMI >25 or age≥60 years, while the association in the opposite direction was only observed in the other groups. Similarly, the significant effect of TG on sleep duration was not observed in individuals with BMI ≤ 25 or age <60 years. Further mechanism studies are necessary for providing evidence for causality.