The CHARLS, initiated in 2011, is a longitudinal, nationally representative survey conducted in China.¹ The study investigated individuals aged 45 years and older together with their spouses in 150 counties/districts and 450 villages/urban communities across 28 provinces (including autonomous regions and municipalities) through a multistage stratified probability-proportional-to-size sampling method (18). The collected data encompass basic information, family composition, physical and mental health status, healthcare coverage, employment and pension, financial resources and laboratory examinations. CHARLS commenced in 2011, with subsequent follow-ups conducted in 2013, 2015, 2018, and 2020. The Ethical Review Committee of Peking University granted approval for the CHARLS study (IRB00001052–11015). All participants obtained informed consent at the beginning of their study.

The current investigation retrospectively analyzed data from the CHARLS 2011 and 2018. The baseline survey conducted in 2011 involved a total of 17,708 participants. The inclusion criteria for the cross-sectional study in 2011 were as follows: (1) complete information on individuals aged 45 years or older, including gender and other covariates; (2) complete data on night/total sleep duration; (3) complete data on the frailty index (Figure 1). Individuals who did not meet all of the above criteria were excluded. The inclusion criteria for the longitudinal study between 2011 and 2018 were as follows: (1) individuals aged 45 years or older with complete information on covariates at baseline; (2) complete data on night/total sleep duration; (3) individuals without frailty in 2011; (4) individuals with complete data on frailty index in 2018 (Figure 1). Individuals who did not meet all of the above criteria were excluded. Multiple imputations by chained equations with 5 imputations were conducted to impute missing covariate information in the sensitivity analysis. Estimates were obtained by pooling the results. The cross-sectional study in 2011 ultimately included 10,258/10,250 participants (night/total sleep duration), while the longitudinal study between 2011 and 2018 included 4,770/4,768 participants (night/total sleep duration).

The night sleep duration was evaluated through the following question: “During the past month, how many hours of actual sleep did you get at night (average hours for one night)?” The duration of nap time was evaluated using the following inquiry: “During the past month, how long did you take a nap after lunch?” We determined total sleep duration by adding together the duration of night sleep and nap time. Night sleep and total sleep were divided into three categories: short sleep duration (<6 h), normal sleep duration (6-9 h), and long sleep duration (≥9 h).

The FI was developed by Rockwood et al. and is used for assessing frailty by quantifying the accumulation of age-related health deficits. The index can be customized and developed as needed by adhering to the health deficit selection principle (6). It is recommended to include a minimum of 30 health deficit items. The FI in this study is created from 33 terms, encompassing 14 chronic diseases, 5 disabilities, 12 limitations in activities of daily living (ADLs) and instrumental activities of daily living (IADLs), cognitive function and depression (Supplementary Table S1). The selection process is based on the established standard procedure and previous literature (8, 19–21). Chronic diseases identified in CHARLS include hypertension, diabetes, dyslipidemia, cancer, chronic lung disease, heart disease, stroke, psychiatric disorders, memory-related diseases, rheumatism, liver disease, kidney disease, digestive diseases, and asthma (22). Disabilities include physical disabilities, brain damage/mental retardation, vision problem, hearing problem and speech impediment (21, 22). Functional limitations are evaluated based on six questions for ADLs and six for IADLs (19). The cognitive function assessment encompasses an orientation test, a picture drawing test, a calculation test, and evaluations of both immediate and delayed memory capabilities, totaling 21 questions (21). The final cognitive score is a continuous variable obtained by dividing the total score by 21, with higher scores indicating worse cognitive function. Remaining terms were also assigned values from 0 to 1, with 0 indicating no health deficit and 1 indicating a health deficit. FI was calculated by dividing the total health deficit scores by 33 and multiplying by 100, with a range between 0 and 100. Higher scores suggest a higher level of frailty. As indicated in prior research, participants were categorized into physically robust (FI ≤ 10), pre-frailty (10 < FI < 25), and frailty (FI ≥ 25) groups (19, 23).

The analysis was adjusted for demographic characteristics, health behaviors, and anthropometric measurements. The covariates used in this study included age, sex, education level, marital status, residence, yearly expenditure, BMI, waist circumference, smoking status, and drinking status. Age was divided into two groups: 45 to 60 years old and 60 years old or older. Education level was classified as illiterate, middle school or below (home school, primary school and middle school) and high school or above (high school, vocational school, college/associate degree, bachelor’s degree, master’s degree and doctoral degree). We classified the marital status as either married or any other status, including unmarried, separated, divorced, or widowed. BMI was classified into three categories: underweight (BMI < 18.5 kg/m²), normal weight (18.5 kg/m² ≤ BMI < 24 kg/m²), and overweight or obesity (BMI ≥ 24 kg/m²). We determined the smoking status based on the individual’s smoking history. The drinking status of individuals was assessed based on their alcohol consumption within the previous 12 months.

Baseline characteristics were summarized according to sleep duration. The categorical variables were presented as n (%) and were compared using the chi-squared test. The continuous variables were presented as the mean ± standard deviation (SD) and were compared using one-way ANOVA.

In both cross-sectional and longitudinal study, linear regression and RCS analysis were used to assess the association between sleep duration and FI, while multinomial logistic regression analysis was used to examine the relationship between sleep duration and the risk of frailty. Multinomial logistic regression belongs to the generalized linear model family. This model uses a link function to mathematically connect the linear combination of input variables and their coefficients (linear predictor) to the expected probability of each categorical outcome. This transforms the typical S-shaped curve of logistic models into a linear relationship for analysis. Restricted cubic spline is a way to identify non-linear relationship, which essentially is a piecewise cubic polynomial (24). Each spline connects smoothly at each knot. To improve model fitting and smoothness while reduce overfitting, four knots were set at the 5th, 25th, 75th, and 95th percentiles of latitude, as recommended by Harrell (25). The selection of covariates was based on existing literature, and three models were used in the analysis (22). Model 1 is a crude model without any adjustments; Model 2 included adjustments for age, sex, educational level, marital status, residence and yearly expenditure; Model 3 adjusted for age, gender, educational level, marital status, residence, yearly expenditure, BMI, waist circumference, smoking status, and drinking status. Sensitivity analysis was conducted on the dataset after the imputation of incomplete covariate information. Subgroup analysis and interaction analysis were performed to investigate potential effect modifications of covariates on the association between sleep duration and frailty. Some subgroups were not statistically significant due to insufficient sample size such as “High school or above,” “Unmarried” and “Urban residence” subgroups in the longitudinal study, which comprised fewer than 1,000 individuals. All statistical analysis was performed by R software (Version 4.3.2). All p values were two-sided, and the statistical significance was defined as p < 0.05.

Table 1 presents the baseline characteristics of the cross-sectional study population grouped by different night sleep durations. A total of 10,258 participants were included in the cross-sectional study of night sleep duration and FI/frailty. This sample consisted of 5,738 middle-aged adults aged between 45 and 60 years and 4,520 older adult adults aged 60 years and above (mean [SD] age: 59.1 [9.3] years). In the sample, 51.3% of the participants were female, 62.2% had middle school or below education, 88.3% were married, 80.5% resided in rural areas, 52.5% had normal weight, 59.1% were non-smokers, and 67.0% never consumed alcohol. The prevalence of frailty in the overall population is 7.9%.

The cross-sectional study comprised 2,999 participants with short sleep duration, 6,456 with normal sleep duration, and 803 with long sleep duration. Participants in the short night sleep group were more likely to be female, older, less educated, single, with lower annual expenses, underweight, with smaller waists, less likely to smoke or drink and more prone to pre-frailty and frailty compared to the normal group. The characteristics of the participants in the longitudinal study of night sleep were approximately similar (Table 2). This is also manifested in cross-sectional and longitudinal studies of total sleep (Supplementary Tables S2, S3).

In the cross-sectional study, a negative correlation was identified between both night and total sleep duration and FI in the crude model (Model 1, night sleep: β = −1.04, p < 0.001, total sleep: β = −0.83, p < 0.001) (Table 3). After adjusting for demographic characteristics (Model 2), and further for health habits and anthropometric measurements (Model 3), the association remained significant (Model 2, night sleep: β = −0.81, p < 0.001, total sleep: β = −0.63, p < 0.001; Model 3, night sleep: β = −0.83, p < 0.001, total sleep: β = −0.66, p < 0.001). After categorizing sleep duration into three groups, short night or total sleep duration increased FI in unadjusted and adjusted models compared to normal sleep duration (Table 3). Long sleep duration initially increased FI in models of night sleep but not total sleep. The findings suggest that short night and total sleep duration are associated with a higher FI.

In the cross-sectional study, the risk of frailty was negatively associated with both night and total sleep duration (night sleep: OR [95 CI%], 0.98 [0.98, 0.98], p < 0.001; total sleep: 0.99 [0.98, 0.99], p < 0.001) (Table 3). The association retained statistical significance after adjusting for confounding variables (Model 2, night sleep: 0.99 [0.98, 0.99], p < 0.001, total sleep: 0.99 [0.99–0.99], p < 0.001; Model 3, night sleep: 0.99 [0.98, 0.99], p < 0.001, total sleep: 0.99 [0.99, 0.99], p < 0.001). After categorizing sleep duration into three groups, short sleep duration is associated with an increased risk of frailty compared to normal sleep duration in three models. Long sleep duration only increased the risk of frailty in the unadjusted model and this significance dissipated following adjustment. This suggests that short night sleep duration, as well as short total sleep duration, is associated with an increased risk of frailty.

In the longitudinal study spanning from 2011 to 2018, FI exhibited a negative correlation with both night and total sleep duration in the crude model (Model 1, night sleep: β = −0.75, p < 0.001, total sleep: β = −0.62, p < 0.001) (Table 4). After adjusting for demographic characteristics, as well as health habits and anthropometric measurements, the association continued to exhibit statistical significance (Model 2, night sleep: β = −0.57, p < 0.001, total sleep: β = −0.43, p < 0.001; Model 3, night sleep: β = −0.59, p < 0.001, total sleep: β = −0.47, p < 0.001). The FI was significantly higher in short sleep group compared to normal sleep group (Table 4). The association still existed in the sensitivity analysis (Supplementary Table S4). This result is consistent with cross-sectional study indicating that shorter sleep duration may lead to increased FI.

The longitudinal study also revealed a negative association between both night and total sleep duration and the risk of frailty across all three models (Model 1, night sleep: 0.99 [0.98, 0.99], p < 0.001, total sleep: 0.99 [0.99, 0.99], p < 0.001; Model 2, night sleep: 0.99 [0.99, 1.00], p < 0.001, total sleep: 0.99 [0.99, 1.00], p < 0.01; Model 3, night sleep: 0.99 [0.99, 1.00], p < 0.001, total sleep: 0.99 [0.99,1.00], p < 0.01) (Table 4). The incidence of frailty was increased in short sleep group compared to normal sleep duration. The relationship persisted in the sensitivity study (Supplementary Table S4). This suggests that short night sleep duration, as well as short total sleep duration, may increase the risk of frailty.

While a linear relationship between sleep duration and FI was revealed both in the cross-sectional and longitudinal analysis, a non-linear relationship was also detected (P-nonlinear <0.001). Restricted cubic spline regression showed a non-linear (U-shaped) relationship between both night and total sleep duration and the risk of frailty using the fully adjusted model (Figures 2, 3). The OR value of 1 is depicted by a dotted horizontal line, which intersects the curve at 6.5 h and 8.1 h for night sleep, and at 7.0 h and 9.0 h for total sleep. In the cross-sectional study, the risk of frailty decreases as sleep duration increases up to 7.2 h per night or 8.1 h per day, followed by a reversal trend beyond these thresholds (Figure 2). In the longitudinal study, the curve remained almost the same except the right half becoming insignificant (turning point: night, 7.9 h; total, 8.6 h) (Figure 3). This suggests that both short and long sleep duration may increase the risk of frailty.

The result of long sleep duration in restricted cubic spline regression is different from the insignificant result of the fully adjusted linear model. In comparing the linear and non-linear models, the AIC and BIC values suggest that the linear model performed better than the non-linear model. Therefore, our study tends to believe that long sleep duration is less likely to increase the risk of frailty. The discrepancy may be attributed to the relatively smaller sample size in the long sleep duration group, which complicates the interpretation of the results. Nonetheless, both results have supporting evidence. A study from China Kadoorie Biobank which conducted in 10 regions across China found that short sleep duration, but not long sleep duration, increased the risk of exacerbating frailty (16). This study included only a brief assessment of depression and did not incorporate a cognitive evaluation. A study in the Netherlands found that both short and long sleep durations were associated with an increased risk of physical frailty (9). However, the association with long sleep duration lost significance after further adjustment. This study also found that both short and long sleep duration were associated with psychological frailty but not cognitive frailty. This suggests that the significant result of long sleep duration in our RSC analysis may be attributed to the relatively comprehensive assessment of depression or cognitive function in FI.

To investigate whether sleep duration has varying effects on the risk of frailty among different subgroups, we performed a stratified analysis. The associations were robust in subgroup analysis in terms of age, gender, marital status, education level (excluding high school or above), residence, yearly expenditure, BMI (excluding underweight individuals), smoking and drinking status in the cross-sectional study of night sleep duration (Figure 4A). The insignificance observed in specific subgroups may be attributed to the relatively small sample sizes within these subgroups. Night sleep duration seems to have a greater impact on frailty in specific populations, such as females, individuals aged 60 and above, single individuals, those with limited literacy, and nondrinkers (Figure 4A). Likewise, significant interaction effects were also observed within subgroups of gender, marital status, education level, BMI, and drinking status, but not age in the cross-sectional analysis of total sleep duration (Figure 4B).

The longitudinal stratification analysis of night sleep duration showed consistent results with the main findings. However, for certain subgroups such as males, individuals aged 60 and above, illiterate, with a high school education or above, single, living in urban areas, with low or high annual expenses, underweight, overweight or obese, and smokers, the observed relationship was not statistically significant (Figure 5A). This may arise from the limited sample size in these subgroups as a result of lost follow-up. As a consequence, no significant interactions were observed among subgroups based on age, marital status, education level, residence, yearly expenditure, BMI and drinking status (p for interaction >0.05). However, a significant interaction was noted for smoking status. The analysis of total sleep duration yielded similar results (Figure 5B).