Data source and study population
This study is embedded in the Chinese Longitudinal Healthy Longevity Survey (CLHLS), which is a national cohort focusing on older Chinese people and is the largest cohort of centenarians in the world. A detailed study design of CLHLS has been published elsewhere [9]. Briefly, a multistage cluster sampling approach was used to ensure its representative of the general elderly Chinese population. After randomly selecting half the total number of countries and cities from 22 provinces, all centenarians living in the sampled area were invited to the participant and the response rate was 97.7%. For each participating centenarian, one octogenarian and one nonagenarian living in the same community or village were randomly sampled. According to systematic assessments of the randomness of attrition, credibility and validity of the measurement scale, and accuracy of reported age, the quality of the data for CLHLS was high. The CLHLS study was approved by the Research Ethics Committee of Peking University (IRB00001052-13074), and all participants or their proxy respondents provided written informed consent.
The baseline survey of the current study was conducted in 2005, with follow-up waves conducted in 2008, 2011, 2014, and 2018. A further extension of the cohort was initiated in the 2008 and 2011 waves following the same study protocol. An overview of the study population is shown in e-Figure 1.
Exposures and outcome
As shown in e-Figure 2, the primary exposures were mean SBP and SBPV, measured between baseline and second wave. The primary outcome was 3-year all-cause mortality, identified between second and third waves.
Mean SBP was assessed by calculating the updated arithmetic mean of SBP in the consecutive two waves from 2005 onwards ((Second wave-Baseline)/2). Within-individual SBPV between two sequential waves was defined as the difference in SBP between two waves divided by the mean ((Second wave-Baseline)/mean). To account for slightly different visit intervals, this measurement was further scaled to the average variation per year, assuming a constant rate of variation between the two waves [10].
Both exposure measurements (mean SBP and SBPV) were assessed as time-varying exposures, first assessed at the 2008 wave using SBP values from 2005 and 2008 waves, and then updated at the 2011 wave using SBP values from 2008 and 2011 waves, and so on.
Covariates
Information on covariates was collected at baseline using a structured questionnaire. One section included sociodemographic characteristics (age, sex, body mass index (BMI), educational level, economic income (high vs medium/low)), smoking status (current, past, or never smoker), and alcohol consumption (current vs former/never). Visual status was defined as “good” or “poor” according to whether participants could identify the break in the image of a circle held before them. Cognitive function was measured by the Chinese version of Mini-Mental State Examination (MMSE) and we defined mild cognitive impairment (MCI) based on both MMSE score and education level: <18 for those without formal education, < 21 for those with 1–6 years of education, < 25 for those with more than 6 years of education [11]. Restriction in daily living activities was defined as a participant being dependent on toileting, bathing, indoor activities, dressing, eating, or continence. Comorbidity was defined according to the number of the self-reported disease, including diabetes mellitus, cardiovascular disease, stroke, respiratory disease, and cancer.
Frailty was assessed by the adjusted osteoporotic fracture index [12, 13], which including three components: (1) underweight (BMI < 18.5 kg/m2); (2) participants having trouble standing up from a chair without the assistance of arms; and (3) a positive response to the question “how many times suffering from serious illness in the past two years”. We categorized frailty status into: frail (two or three components), pre-frail (one component), and robust (no component).
Statistical methods
Primary analyses. Our analysis focused on the association between control of SBP (mean SBP and SBPV), assessed over two sequential study waves, and 3-year all-cause mortality, among oldest-old. Person-time accumulated from the second wave (first assessment of mean SBP and SBPV) until the date of death, date of loss to follow-up, or end date of follow-up (the updated third wave), whichever came first.
We first investigated the association between continuous mean SBP and SBPV and 3-year all-cause mortality using Cox proportional hazards models with penalized splines, which examine the potential non-linear or irregular shape of the hazard functions. Other covariates, such as BMI, could also exert a non-linear effect here. Following the suggested procedure [14], we obtained the corresponding multivariable degree of freedom based on the corrected Akaike information criterion and biological plausibility. Then, we further stratified mean SBP and SBPV into quintile with the reference group defined as the middle quintile to facilitate understanding.
All Cox models were adjusted for baseline covariates, collecting from the updated baseline: age, sex, BMI, educational background, economic income, smoking status, alcohol consumption, visual status, MCI, restriction in activities of daily living, comorbidity, and cohort. The proportional hazard assumption was assessed by visual inspection of the scaled Schoenfeld residuals plot.
Secondary analyses. Given the previously reported terminal decline in SBP at the end-of-life [15], we also checked the potential impact of reverse causality by repeating the above analyses using 1-year and 2-year lag periods, separately. Also, to identify the potential effect modification, we stratified the analyses by self-reported doctor-diagnosed hypertension and frailty status at baseline.
Sensitivity analyses. To test the robustness of the main findings, we performed the following analyses: (1) excluding participants who contributed to more than one cohort; (2) reporting the associations using multiple imputations to reduce potential selection bias caused by missing covariates; and (3) estimating how strong residual confounding would need to be to explain away the observed association using the E-value [16].