Study design and population
The participants were derived from a public open database – CLHLS , which was a national investigation recruited elderly aged over 65 years and older from 23 of 31 province/ autonomous region / municipality in mainland China. Since the first investigation (1998), the CLHLS has conducted eighth investigations until 2018 wave. Detailed introduction of CLHLS was published elsewhere .
To protect the participants’ privacy, the home address information was removed in this open-public database. Thus, we selected fifth (2008) wave as baseline due to the information of residence units (at counties/districts level) of the participants resided were available from community investigation questionnaire . At baseline, participants were excluded if they were aged under 80 years, missing on relevant information (demographic characteristics, blood pressure and exposure assessment), and had hypertension (systolic blood pressure (SBP) ≥140 mm Hg 140 or diastolic blood pressure (DBP) ≥90 mm Hg or normal blood pressure but diagnosed with hypertension by Ⅰ&Ⅱ hospital). The participants were followed up in sixth (2011/2012), seventh (2014), and eighth (2018) investigation. The detailed flow chart of current study is shown in Figure 1. Finally, we focused on 5253 participants who older than 80 years old with available information and were followed up in the study. Only 1.3% of participants changed their residence (county/district) in our study. Thus, the sample population was highly stationary because they were less likely to move to another address with advanced age and hukou household registration system . We also performed sensitivity analysis by excluding participants changed residence (Table S4).
The Research Ethics Committees of Peking University and Duke University granted approval for the Protection of Human Subjects for the Chinese Longitudinal Healthy Longevity Survey, including collection of the data used for present study. The survey respondents gave informed consent before participation.
The duration of follow-up was from the date participants enrolled in the project to the date of hypertension event occurrence or death or end of follow-up, whichever occurs first. Person years were calculated by precise method (days/365). The follow-up ended when hypertension event occurrence or discontinued their participation (i.e. censoring including death, lost follow-up and not identified as hypertension).
As a part of CLHLS, in baseline and each interview, the participants’ SBP and DBP were measured between five minutes rest interval by mercurial sphygmomanometer and performed by trained physician. Through two times repeated measurement (between five or longer rest interval) of blood pressure, the average value of SBP and DBP were calculated. Specifically, for bedbound participants, blood pressure measurements were obtained in a recumbent position. Simultaneously, participants were also asked to respond to the following question: “Are you suffering from hypertension and diagnosed by hospital or not ?”. According to guideline, in our study, hypertension was defined as SBP≥140 mm Hg 140 or DBP ≥90 mm Hg or normal blood pressure value but self-reported be diagnosed with hypertension by Ⅰ&Ⅱ hospital. Similar study used same definition of hypertension basing on the CLHLS database was published previously .
The Normalized Difference Vegetation Index (NDVI) and enhanced vegetation index (EVI) sourced from the Moderate Resolution Imaging Spectroradiometer (MODIS) Terra were used as estimation of greenness. They were recorded per 16-day at a spatial resolution of 250, 500, and 1000 m, respectively. The two satellite-derived vegetation indices reflected vegetation of ground with range from -1 to 1. In general, NDVI of 0.1 or below reflected barren areas of rock, sand or snow; NDVI between 0.2 to 0.4 reflected shrub and grassland. The higher value of NDVI/EVI, the denser greenness. Referring to previous study , based on NDVI/EVI free of cloud of 16-day recorded for 500 m buffer, the 1-year average of NDVI/EVI value before the year of last interview (only for censoring) or hypertension event occurrence was treated as greenness exposure in our study. Because this window period had greatest impacts on hypertension incidence (Table S1). In particular, due to the individual privacy protection, exact address of residence was unavailable, thus, the exposure was assessed at a community-based unit level (districts/counties) the elderly resided. Those cloud-free of NDVI/EVI value were available downloaded from the Google Earth Engine (https://developers.google.cn/earth-engine) and they were extracted through ArcGIS 10.6 (ESRI, Redlands, CA, USA) in this study. Specific, the one year average of NDVI/ EVI before event at a spatial resolution of 16-day with 250 m and 1000 m buffer were used to sensitivity analyses.
Potential confounding variables and mediators
Some of the baseline characteristics were used to control confounding in the study. Including: age, gender (female, male), residence (rural/town, urban), geographical regions (eastern China, central/western China), pension (yes, no), living arrangements (with family member, nursing home/alone), education attainment (uneducated, educated), marital status (married/living together, widowed, and /single/divorced/separated), smoking status at the present (yes, no), drinking status at the present (yes, no), exercising habit at the present (yes, no), self-reported heart disease (yes, no), self-reported diabetes (yes, no), and ADL (Activities of Daily Living). In particular, ADL was evaluated by self-report regarding physical ability in basic self-care tasks, which was consist with six items including bathing, dressing, eating, toileting, continence, and transferring [28, 29]. If an participant is able to perform an activity, he/she gets score 1, and if he/she is limited to do and unable to do so, will get score 2 and 3, respectively. The more scores of the individual, indicates poor ADL ability.
According to prior studies, we selected contemporaneous PM2.5 concentration, BMI index, and leisure activity at the baseline as mediators in the analysis. Among them, baseline PM2.5 concentration in 652 units of the subjects resided was collected from a public open database built by Atmospheric Composition Analysis Group from University of Washington . Researchers used satellite combining Aerosol Optical Depth (AOD) retrievals from multiple satellite products (MISR, MODIS Dark Target, MODIS and SeaWiFS Deep Blue, and MODIS MAIAC) to report ground-level of PM2.5 concentration with 1.1x1.1 km resolution.
Leisure activity was collected from individual questionnaire included following activity: housework ; personal outdoor activities; garden work; read newspapers/books; raising domestic animals; playing cards and/or mahjongg; watching television and/or listening to the radio; organized activities. The checklist was ratings using a 5-point Likert scale. Each opinion was scored on a scale of 1 (never), 2 (not every month, but sometimes), 3 (not every week, but at least once a month), 4 (not every day, but at least once a week), or 5 (almost every day). As recommend, we calculated total points and divided by the full score to measure leisure activity (range = 0-1) .
Descriptive statistic analysis was performed firstly. The continuous variables with normal distribution were reported as mean ± standard deviation (SD). Median (interquartile range) was presented for variables with non-normal distribution. The categorical variables were reported as number and percentage.
The Cox proportional hazards regression models of penalized spline with different degree of freedom (2/3/4) was conducted to examine the potential non-liner association between greenness and hypertension incidence basing on minimum AIC (Akechi information criterion) value.
Given the set clustering of the study designs, to evaluate the associations of greenness exposure with the occurrence of hypertension, the random-effects Cox proportional hazards regression models were performed. In this analysis, the 652 units was set as a random effects term in the models. We first explored the links between 0.1-unit increment in greenness and hypertension. Furthermore, due to the non-liner association between greenness exposure and hypertension (Figure S2 and S3). We categorized NDVI/EVI into three categories basing on tertile: first tertile (lowest and as reference, ≤P50), second tertile (P50-P75), and third tertile (highest, >P75). The crude model only included NDVI/ EVI (model 1). The model 2 was developed by further adjusting for age, gender, residence, geographical regions, pension, living arrangements, education attainment, marital status, smoking status at the present, drinking status at the present, exercising habit at the present, self-reported heart disease, self-reported diabetes, PM2.5, leisure activity, and BMI (model 2). We also built trend analysis between greenness exposure and hypertension risk by setting NDVI/EVI tertile as continuous variables in the models. For a robustness check, the EVI/ NDVI in 250 and 1000 m buffer was used to test the sensitivity of our results. In addition, we excluded the participants who were self-reported had diabetes/heart disease/dead during follow-up to verify the effects of residential greenness exposure and hypertension incidence (Table S4). Subgroup and interaction effects analysis were performed to identify vulnerable population affected by greenness exposure. The interaction effects of NDVI/ EVI and four variables on hypertension incidence by perform a cross-product term.
We conducted a one-mediator model to explore the mediation effects of each 0.1-unit increment in greenness exposure and hypertension incidence mediated by mediators (PM2.5, leisure activity, and BMI) based on the survival outcomes. The detailed theoretical foundation, R codes, and calculation principle were published in the study of Huang and Yang . In the model, the product of the two effects in the hazard ratio scale: TEHR = DEHR × IEHR, or equivalently, TElogHR = DElogHR + IElogHR. The mediation proportion was calculated as the natural log of indirect effect divided the total effect and expressed as a percentage .
Effect estimates are showed as hazard ratio (HR) and corresponding 95% confidence intervals (CIs). The map of China was derived from National Geomatics Center of China (http://www.ngcc.cn/ngcc/), the statistics map was generated by ArcGIS Geospatial Analyst module v10.6 (ESRI, Redlands, CA, USA). All of analysis were 2-sided level of significance was set at 0.05. All of analysis were performed using R studio (R 4.0.5, R Foundation for Statistical Computing, Vienna, Austria).