Interaction between residential greenness and air pollution mortality: analysis of the Chinese Longitudinal Healthy Longevity Survey

Summary

Background

Both air pollution and green space have been shown to affect health. We aimed to assess whether greenness protects against air pollution-related mortality.

Methods

We used data from the 2008 wave of the Chinese Longitudinal Healthy Longevity Survey. We calculated contemporaneous normalised difference vegetation index (NDVI) in the 500 m radius around each participant's residence. Fine particulate matter (PM_2·5) concentration was calculated using 3-year average concentrations in 1 km × 1 km grid resolution. We used Cox proportional hazards models to estimate the effects of NDVI, PM_2·5, and their interaction on all-cause mortality, adjusted for a range of covariates.

Findings

The cohort contained 12 873 participants, totalling 47 884 person-years. There were 7426 deaths between 2008 and 2014. The mean contemporaneous NDVI was 0·42 (SD 0·21), and the mean 3-year average PM_2·5 was 49·63 μg/m³ (13·72). In the fully adjusted model, the mortality hazard ratio for each 0·1-unit decrease in contemporaneous NDVI was 1·08 (95% CI 1·03–1·13), each 10 μg/m³ increase in PM_2·5 was 1·13 (1·09–1·18), and the interaction term was 1·01 (1·00–1·02) with a p value of 0·027. We observed non-linear associations in our stratified analyses: people living in urban areas were more likely to benefit from greenness, and people living in rural areas were more likely to be harmed by air pollution.

Interpretation

Our study showed some indication of a synergistic effect of greenness and air pollution, suggesting that green space planning and air pollution control can jointly improve public health.

Funding

Bill & Melinda Gates Foundation, National Institutes of Health, National Key R&D Program of China, National Natural Science Foundation of China.

Introduction

Air pollution in China is the fourth largest cause of death and disability combined, behind dietary risks, tobacco, and high blood pressure.¹ Although some efforts have been aimed at reducing air pollution, fine particulate matter (PM_2·5) will likely remain a contributor to high burdens of disease in the most populous country in the world for the foreseeable future.² Several large-scale population studies in China have shown the relationship between air pollution and premature mortality.^{3, 4, 5, 6} A study using the Chinese Longitudinal Healthy Longevity Survey (CLHLS) reported that each 10 μg/m³ increase in the past 3-year average PM_2·5 was associated with 8% higher mortality in adults aged 65 years or older, and extrapolated that 1 765 820 premature deaths among Chinese older adults were caused by ambient air pollution.⁷

Since 2008, China has made efforts towards controlling air pollution. However, these efforts might be offset by urbanisation, which has intensified the health effects of air pollution as people move from rural areas with lower ambient air pollution into urban areas with higher ambient air pollution.² By 2013, the number of rural-to-urban migrants in China was 245 million, or a fifth of the total population.⁸ In 2017, more than half of the population lived in cities.⁹ Urban areas typically have worse air quality, exposing residents to more air pollution.¹⁰ Meanwhile, the urban sprawl has created more built-up areas to serve the increasing urban population, and green spaces are sometimes part of city planning and infrastructure development.^{11, 12} Average green space coverage increased from 17% in 1989 to 37% in 2009.¹¹ In some areas (Shaanxi, Shanxi, Ningxia, Henan, Shandong, Qinghai, and Gansu), the amount of green space increased, whereas in others (northeast Inner Mongolia, south Tibet, Jiangsu, and Shanghai), it decreased due to agricultural reclamation, climate change, and urbanisation.¹²

Greenness has generally been thought to improve health, per the biophilia hypothesis.¹³ This hypothesis proposed that seeking out connections with nature and other forms of life is an innate human trait with an underlying biological basis. Vegetation in nature has been reported to confer health benefits, including documentation of improved physical and mental health and improved birth outcomes,¹⁴ increased convenience of engaging in physical activities,¹⁵ and promotion of social contact and reduced stress through social determinant of health pathways.¹⁶ Concerning mortality, protective effects for residential greenness have been shown in cohorts in North America, Europe, and Asia.^{17, 18, 19, 20, 21} In China, greater greenness was associated with lower mortality in an older cohort.²²

Research in context.

Evidence before this study

We searched PubMed and Google Scholar for studies on residential greenness, air pollution, and mortality published in English from database inception until Aug 1, 2019. We used a combination of search terms, including “residential greenness”, “greenness”, “green space”, “normalized difference vegetation index”, “air pollution”, “fine particulate matter”, “PM_2·5”, and “mortality”. We found four studies exploring the effects of residential greenness and air pollution on mortality. They were done in high-income countries (Canada, the USA, Spain, and South Korea); the study designs ranged from ecological to cohort. These studies found greenness modulates air pollution in Canada, Spain, and South Korea, but not in the USA. Additionally, the direction of interaction varied between studies.

Added value of this study

Our study found that there was an interaction between residential greenness and air pollution and mortality in a large prospective cohort. Our study used a national cohort of adults aged 65 years and older, whose vulnerability to air pollution and residential greenness probably differs from other age groups. To our knowledge, this was the first study of this kind done in a non-high-income country.

Implications of all the available evidence

Our findings suggest that greenness and air pollution have a synergistic effect, which implies that green space planning and air pollution control can jointly improve public health.

Greenness could interact with exposure to air pollutants.²³ Trees and vegetation are hypothesised to act to clean the air of pollution, mitigating the health effects. Airborne chemistry evidence is suggestive of vegetation's ability to intercept airborne particles or to promote the uptake of gaseous pollutants through leaf stomata on the plant surface.²⁴ Increasing green space could yield a positive return on investment relative to the cost when taking health into account in cost–benefit analyses. Thus, the effect of residential greenness and air pollution on mortality is of particular interest, but the published literature is not clear on their connection. A study in Massachusetts, USA, of 179 986 adults with a mean age of 80 years reported no consistent interaction between PM_2·5 and NDVI on cardiovascular mortality; greenness attenuated some of the effects of PM_2·5 on cardiovascular mortality, but only in the communities with low socioeconomic status.²⁵ A study in Canada of 2·4 million adults aged 25–89 years showed that the effect estimates of PM_2·5 on non-accidental mortality decreased as greenness increased.²⁶ A study of seven metropolitan cities in South Korea observed that the effects of PM₁₀ on non-accidental mortality were smaller in less green areas, whereas the effects of PM₁₀ on cardiovascular-related mortality were smaller in greener areas.²⁷ A study of 4·4 million residents from 2148 small areas in Spain found that the interaction between PM_2·5 and greenness and all-cause mortality differed by urban or rural residence. In rural areas, the effects of PM_2·5 on mortality were stronger in less green areas, whereas in urban areas, the effects of PM_2·5 on mortality were stronger in greener areas.²⁸ No study has been done in a developing country. We, therefore, aimed to assess the interactive effect of residential greenness and air pollution using a longitudinal cohort of Chinese older adults who reside throughout the country.

Methods

Study design and participants

We used data from the 2008 wave of CLHLS, a prospective cohort study on the determinants of healthy longevity. The CLHLS applied a multistage, stratified cluster sampling design to recruit participants from 22 of the 31 provinces in China. 631 cities and counties were randomly selected, representing about 85% of the Chinese population. This study started with a focus on centenarians, matched with octogenarian, nonagenarian, and 65–79 year olds.²⁹ The survey recruited individuals aged 80 years or older in 1998 and 2000; participants aged 65–79 years were included from 2002. We excluded participants lost to follow-up in the 2011 survey, with missing or invalid death dates, living in regions where greenness could not be calculated, with missing PM_2·5 values, and younger than 65 years. For the 2008 wave, the survival status was assessed in follow-up surveys done in 2011 and 2014. The investigators collected information on various determinants of health, including demographic and socioeconomic characteristics, lifestyle, physical capacity, cognitive function, and psychological status. More details on sampling design and data quality can be found elsewhere.²⁹

CLHLS was approved by the Institutional Review Board, Duke University (Pro00062871), and the Biomedical Ethics Committee, Peking University (IRB00001052–13074). All participants provided written, informed consent.

Procedures

We used remote sensing to calculate normalised difference vegetation index (NDVI) in the 500 m radius around participants’ residential addresses. We extracted longitude and latitude for each residential address. Then, the NDVI values at the corresponding coordinates for residential addresses were generated. NDVI is the ratio of the difference between the near-infrared region and red visible reflectance to the sum of these two measures, ranging from −1·0 to 1·0. Negative NDVI values are often thought of as blue space or water, whereas larger values indicate denser green vegetation.^{18, 30} Values of 0·1 and below reflect barren areas of rock, sand, or snow; values of 0·2 to 0·4 represent shrub and grassland; higher values indicate temperate and tropical rainforests.³¹

We measured NDVI values from the Moderate-Resolution Imaging Spectro-Radiometer (MODIS) in the National Aeronautics and Space Administration's Terra Satellite.³² MODIS has a temporal resolution of 16 days and varying spatial resolution up to 500 m. Evidence has shown that 0·25 miles (about 400 m) is an optimum radius size to measure accessible greenness in residential neighbourhoods.³³ The 500 m radius is closest to 0·25 miles that can be obtained using MODIS. Additionally, the 500 m radius is also widely used in other residential greenness studies.^{17, 34}

We calculated two NDVI values every month for each season, from 2008 to 2014. Furthermore, we computed contemporaneous NDVI to reflect acute exposure to greenness, which is the NDVI value at the residential address of the participants at the time closest to an event. NDVI values were estimated at death dates for deceased participants, and the last interview dates for living participants and participants lost to follow-up. If values were equal, NDVI values from the earlier date were included. For example, if a participant died on April 15, 2010, the length between the death date and April 7 and April 23, 2010 (dates for our extracted seasonal NDVI), would be the same, so we would use their contemporaneous NDVI values assessed on April 7, 2010.²² We used 0·1-unit increments of NDVI for the statistical analysis.

On the basis of the participants’ residential address, the ground-level concentrations of PM_2·5 were calculated from the Atmospheric Composition Analysis Group. It combined the remote sensing from National Aeronautics and Space Administration's Moderate Resolution Imaging Spectroradiometer, Multiangle Imaging SpectroRadiometer, and Sea-viewing Wide Field-of-view Sensor satellite instruments, vertical profiles derived from the GEOS-Chem chemical transport model, and calibration to ground-based observations of PM_2·5 using geographically weighted regression. Annual PM_2·5 estimates were calculated from 1998 to 2014, at 1 km² spatial resolution, which was the longest and the highest resolution exposure dataset available.⁷ Additionally, our estimations were highly consistent with out-of-sample cross-validated concentrations from monitors (R²=0·81) and another exposure dataset (R²=0·81) in China.⁷ A previous study⁷ found that 3-year average PM_2·5 before death or the end of the study had the strongest association with mortality. We used 3-year average PM_2·5 to reflect air pollution.

The primary outcome was all-cause mortality. Mortality information was obtained from the follow-up survey done in 2011 and 2014.

Covariates were chosen as potential confounders between exposures and outcomes or predictors of outcomes. We assessed a range of demographic, behavioural, and socioeconomic covariates, including age, sex, ethnicity, marital status, geographical region, urban or rural residence, education, main occupation before 60 years of age, financial support, social and leisure activity, smoking status, alcohol consumption, and physical activity. The percentages of missing values of variables coding socioeconomic status and health behaviour were relatively small, less than 0·3%. We assigned the missing values to the mode for the categorical variables for model convergence.

We calculated age on the basis of interview dates and self-reported birth dates, which were verified by family members, genealogical records, identification cards, and household registration booklets. We divided ethnicity into two categories: Han Chinese (the majority in China), and ethnic minorities (Hui, Korean, Manchurian, Mongolian, Yao, Zhuang, and others). We dichotomised marital status to married and not married at the time of interview (separated, divorced, or widowed or never married). We categorised the participants into seven geographical regions: central China (Henan, Hubei and Hunan provinces), east China (Anhui, Fujian, Jiangxi, Jiangsu, Shandong, Shanghai, and Zhejiang provinces), northeast China (Heilongjiang, Jilin, and Liaoning provinces), north China (Hebei, Shanxi, and Tianjin provinces), northwest China (Shaanxi province), south China (Guangdong, Guangxi, and Hainan provinces), and southwest China (Chongqing and Sichuan provinces).

We divided residence into urban and rural areas on the basis of governmental administrative categories.³⁵ Given that any educational attainment was scarce in 20th century China, we dichotomised education to formal education (≥1 year of schooling), and no formal education (<1 year of schooling). We defined the main occupation before 60 years of age as professional work (professional and technical personnel, government, and management), and non-professional work (agriculture, fishing, service, industry, and housework). Financial support was assessed depending on whether the participants were financially independent, with their own work and retirement wage, or financially dependent on other family members.

We measured social and leisure activity index by taking into consideration seven activities: gardening, personal outdoor activities excluding exercise, raising poultry or pets, reading, playing cards or mah-jong, listening to the radio or watching TV, and participating in organised social activities.³⁶ Each included activity scored zero (never) or one (monthly or more frequent), resulting in an index that ranged from zero to seven. We evaluated smoking status, alcohol consumption, and physical activity through the questions of “do you smoke/drink/exercise or not at present”.

Statistical analysis

We used Cox proportional hazard models to estimate mortality, adjusted for age, sex, ethnicity, marital status, geographical region, urban or rural residence, education, main occupation before 60 years of age, financial status, social and leisure activity, smoking status, alcohol consumption, and physical activity. The survival time was measured in months from the first interview date to the recorded death date or last interview date up to 2014. We created separate models to examine the independent effect of contemporaneous NDVI, PM_2·5, and their interaction. The interaction term was calculated for each 0·1-unit decrease in contemporaneous NDVI and each 10 μg/m³ increase in 3-year average PM_2·5 concentrations. Hazard ratios (HRs) and 95% CIs were reported to reflect the effect magnitude of NDVI and PM_2·5 on mortality. Stratified analyses were done by sex, urban or rural residence, financial status, smoking status, and physical activity. We did a sensitivity analysis by using time-varying annual average PM_2·5 and NDVI. We also did an additional analysis using the survey-weighted regression to obtain population-representative effect estimates. We also plotted mortality trends by different tertiles of contemporaneous NDVI and 3-year average PM_2·5 to explore non-linearity, by using three knots cubic splines.³⁷

We did additional analyses to verify the results. We compared baseline characteristics between the included participants and all participants in the 2008 CLHLS to assess the sample's representativeness. We used the survey-weighted regression to obtain population-representative effect estimates. We did a sensitivity analysis by using time-varying annual average PM_2·5 and NDVI. We adjusted the regression models by using more informative covariates and by excluding negative contemporaneous NDVI and covariates with missing values. We compared the effects of cumulative NDVI on the associations with the contemporaneous NDVI. We also plotted the geographical distribution of the included participants.

We used STATA version 14.0 for all statistical analyses.

Role of the funding source

The funders of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.

Results

16 954 participants were interviewed in the 2008 CLHLS. Our final sample consisted of 12 873 participants (appendix p 8), after exlusion of 2894 participants who were lost to follow-up in the 2011 CLHLS, 88 with missing or invalid death dates, 201 who lived in regions where greenness could not be calculated, 764 with missing PM_2·5 values, and 134 who were younger than 65 years. Participants lost to follow-up did not significantly differ from our study population, except that they were more likely to reside in urban areas (appendix pp 2–3). The mean age was 87·01 years (SD 11·34), and 9359 (72·7%) were aged 80 years and older (table 1). We had slightly more female participants (7385; 57·4%), and more people residing in rural areas (10 893; 84·6%; table 1). Participants living in areas with more residential greenness were more likely to be ethnic minorities or from rural areas, have non-professional work, not have formal education, and be financially dependent. The mean contemporaneous NDVI was 0·42 (SD 0·21) and the mean 3-year PM_2·5 was 49·63 μg/m³ (13·72).

Table 1.

Baseline characteristics

		Total	Contemporaneous NDVI	3-year average PM2.5(μg/m3)
Total		12 873	0·42 (0·21)	49·63 (13·72)
	Range	..	−0·14 to 0·96	9·21 to 109·37
	Median (IQR)	..	0·41 (0·25 to 0·59)	49·04 (39·86 to 59·34)
Age, years		87·01 (11·34)	..	..
Age group
	65–79 years	3514 (27·3%)	0·45 (0·20)	48·84 (13·76)
	80–89 years	3388 (26·3%)	0·43 (0·21)	48·79 (13·46)
	90–99 years	3503 (27·2%)	0·40 (0·21)	49·53 (13·89)
	≥100 years	2468 (19·2%)	0·40 (0·22)	52·04 (13·47)
Sex
	Male	5488 (42·6%)	0·42 (0·21)	49·30 (13·83)
	Female	7385 (57·4%)	0·42 (0·21)	49·87 (13·63)
Ethnicity
	Han Chinese	12 015 (93·3%)	0·42 (0·21)	50·34 (13·72)
	Ethnic minorities	858 (6·7%)	0·50 (0·22)	39·63 (9·02)
Marital status
	Married	3987 (31·0%)	0·43 (0·21)	49·23 (13·69)
	Not married	8886 (69·0%)	0·42 (0·21)	49·80 (13·73)
Residence
	Urban area	1980 (15·4%)	0·25 (0·16)	54·14 (15·55)
	Rural area	10 893 (84·6%)	0·45 (0·20)	48·81 (13·19)
Main occupation before 60 years of age
	Professional work	820 (6·4%)	0·36 (0·21)	51·23 (15·88)
	Non-professional work	12 053 (93·6%)	0·43 (0·21	49·52 (13·55)
Education
	Formal education	4664 (36·2%)	0·41 (0·21)	48·90 (13·82)
	No formal education	8209 (63·8%)	0·43 (0·21)	50·04 (13·64)
Financial support
	Financial independence	3022 (23·5%)	0·39 (0·21)	50·76 (14·21)
	Financial dependence	9851 (76·5%)	0·43 (0·21)	49·28 (13·54)
Social and leisure activity index		2·03 (1·53)	..	..
Smoking status
	Yes	2309 (17·9%)	0·43 (0·21)	49·66 (12·74)
	No	10 564 (82·1%)	0·42 (0·21)	49·62 (13·92)
Alcohol consumption
	Yes	2307 (17·9%)	0·43 (0·21)	49·64 (12·93)
	No	10 566 (82·1%)	0·42 (0·21)	49·62 (13·88)
Physical activity
	Yes	3425 (26·6%)	0·39 (0·21)	49·30 (14·39)
	No	9448 (73·4%)	0·43 (0·21)	49·74 (13·46)
Region
	Central China	2192 (17·0%)	0·47 (0·21)	55·83 (11·04)
	East China	4896 (38·0%)	0·42 (0·21)	54·00 (12·52)
	Northeast China	858 (6·7%)	0·25 (0·17)	41·72 (10·45)
	North China	515 (4·0%)	0·29 (0·16)	70·53 (19·68)
	Northwest China	111 (0·9%)	0·40 (0·17)	47·61 (12·52)
	South China	2569 (20·0%)	0·46 (0·21)	37·90 (6·95)
	Southwest China	1732 (13·5%)	0·43 (0·21)	44·63 (7·62)

Data are n (%) or mean (SD) unless otherwise specified.

Participants who were Han Chinese (non-ethnic minority), were living in urban areas, had professional work, did not have formal education, and who were financially independent tended to live in areas with more PM_2·5. More urban areas had greater air pollution owing to greater numbers of emissions sources, and less greenness owing to the impervious surfaces of city build-up.

In the fully adjusted model (Model 4), the mortality HRs for each 0·1-unit decrease in contemporaneous NDVI was 1·08 (95% CI 1·03–1·13), each 10 μg/m³ increase in PM_2·5 was 1·13 (1·09–1·18), and the interaction term was 1·01 (1·00–1·02, p=0·027; table 2). We did not find appreciable differences in our effect estimates when incorporating population-representative sampling weights (appendix p 5). Additional sensitivity analysis also reported consistent findings, except for cumulative NDVI (appendix pp 1, 4, 6, 7). In our stratified analysis, we saw some indications that study participants residing in urban areas were more likely to benefit from greenness, and people in rural areas were more likely to be harmed by air pollution (table 3). There was no effect modification by sex (table 2).

Table 2.

Hazard ratios for each 0·1-unit decrease in contemporaneous NDVI, each 10 μg/m³ increase in 3-year average PM_2.5, and mortality

		Model 1 (contemporaneous NDVI)	Model 2 (3-year average PM2.5)	Model 3		Model 4
				Contemporaneous NDVI	3-year average PM2.5	Contemporaneous NDVI	3-year average PM2.5	Interaction	p value
All participants		1·13 (1·12–1·15)	1·09 (1·07–1·12)	1·13 (1·12–1·14)	1·09 (1·07–1·12)	1·08 (1·03–1·13)	1·13 (1·09–1·18)	1·01 (1·00–1·02)	0·027
Stratified by sex
	Male (n=5488)	1·13 (1·11–1·15)	1·09 (1·05–1·13)	1·13 (1·11–1·15)	1·08 (1·05–1·12)	1·07 (1·00–1·14)	1·13 (1·06–1·20)	1·01 (1·00–1·02)	0·10
	Female (n=7385)	1·13 (1·12–1·15)	1·10 (1·07–1·14)	1·13 (1·12–1·15)	1·10 (1·07–1·14)	1·09 (1·03–1·15)	1·14 (1·08–1·20)	1·01 (1·00–1·02)	0·16
Stratified by residence
	Urban area (n=1980)	1·22 (1·17–1·28)	1·06 (1·00–1·12)	1·22 (1·17–1·28)	1·04 (0·98–1·10)	1·00 (0·84–1·19)	1·13 (1·03–1·24)	1·04 (1·01–1·07)	0·022
	Rural area (n=10 893)	1·12 (1·11–1·14)	1·10 (1·08–1·13)	1·12 (1·11–1·14)	1·10 (1·08–1·13)	1·07 (1·02–1·12)	1·15 (1·10–1·21)	1·01 (1·00–1·02)	0·024
Stratified by financial status
	Financial independence (n=3022)	1·20 (1·15–1·24)	1·05 (0·99–1·12)	1·19 (1·15–1·24)	1·04 (0·98–1·11)	1·09 (0·96–1·23)	1·11 (1·00–1·23)	1·02 (1·00–1·05)	0·12
	Financial dependence (n=9851)	1·12 (1·11–1·14)	1·10 (1·08–1·13)	1·12 (1·11–1·14)	1·10 (1·08–1·13)	1·07 (1·02–1·12)	1·14 (1·09–1·19)	1·01 (1·00–1·02)	0·042
Stratified by smoking status
	Yes (n=2309)	1·18 (1·15–1·22)	1·13 (1·06–1·21)	1·19 (1·15–1·22)	1·14 (1·06–1·22)	1·03 (0·91–1·17)	1·28 (1·13–1·45)	1·03 (1·00–1·05)	0·029
	No (n=10 564)	1·12 (1·11–1·14)	1·09 (1·06–1·11)	1·12 (1·11–1·14)	1·09 (1·06–1·11)	1·09 (1·04–1·14)	1·12 (1·07–1·16)	1·01 (1·00–1·02)	0·14
Stratified by physical activity
	Exercise (n=3425)	1·16 (1·13–1·19)	1·14 (1·09–1·20)	1·16 (1·13–1·19)	1·13 (1·08–1·19)	1·09 (0·99–1·20)	1·18 (1·09–1·28)	1·01 (0·99–1·03)	0·21
	Do not exercise (n=9448)	1·12 (1·11–1·14)	1·08 (1·06–1·11)	1·13 (1·11–1·14)	1·08 (1·06–1·11)	1·08 (1·03–1·13)	1·12 (1·07–1·17)	1·01 (1·00–1·02)	0·079

All models were adjusted for a number of covariates, including age, sex, ethnicity, marital status, urban or rural residence, education, main occupation before 60 years of age, financial support, social and leisure activity, geographical region, smoking status, alcohol consumption, and physical activity at baseline. p values are for the test for interaction between each 0·1-unit decrease in contemporaneous NDVI and each 10 μg/m³ increase in 3-year average PM_2·5in Model 4. Model 1 tested the main effect of each 0·1-unit decrease in contemporaneous NDVI on mortality. Model 2 tested the main effect of each 10 μg/m³ increase in 3-year average PM_2·5 on mortality. Model 3 tested the main effects of each 0·1-unit decrease in contemporaneous NDVI and each 10 μg/m³ increase in 3-year average PM_2·5 on mortality. Model 4 tested the interaction between each 0·1-unit decrease in contemporaneous NDVI and each 10 μg/m³ increase in 3-year average PM_2·5 on mortality. NDVI= normalised difference vegetation index. PM_2·5=fine particulate matter.

Table 3.

HRs for 3-year average PM_2.5 and mortality under different tertiles of NDVI

		Tertile 1 (range −0·14 to 0·30)		Tertile 2 (range 0·30 to 0·53)		Tertile 3 (range 0·53 to 0·96)
		HR (95% CI)	p value	HR (95% CI)	p value	HR (95% CI)	p value
Per 10 μg/m3 increase in 3-year average PM2·5		1·10 (1·06–1·14)	<0·0001	1·11 (1·07–1·16)	<0·0001	1·07 (1·02–1·11)	0·0030
By residence
	Urban area	1·04 (0·98–1·11)	0·20	1·12 (0·96–1·30)	0·17	0·80 (0·61–1·06)	0·12
	Rural area	1·12 (1·08–1·17)	<0·0001	1·12 (1·08–1·17)	<0·0001	1·07 (1·03–1·12)	0·0014

HR=hazard ratio. PM_2·5=fine particulate matter. NDVI=normalised difference vegetation index.

We compared air pollution mortality within different tertiles of greenness; we did not find a monotonic trend of higher or lower air pollution mortality (table 3). We also did not find a monotonic trend in the effects of residential greenness on mortality within different tertiles of PM_2·5 (table 4). These findings can also be seen in the spline analysis (Figure 1, Figure 2).

Table 4.

HRs for NDVI and mortality under different tertilesof 3-year average PM_2·5

		Tertile 1 (range 9–43)		Tertile 2 (range 43–55)		Tertile 3 (range 55–109)
		HR (95% CI)	p value	HR (95% CI)	p value	HR (95% CI)	p value
Each 0·1-unit decrease in contemporaneous NDVI		1·12 (1·10–1·15)	<0·0001	1·08 (1·05–1·10)	<0·0001	1·20 (1·18–1·23)	<0·0001
By residence
	Urban area	1·18 (1·08–1·29)	0·0003	1·17 (1·09–1·26)	<0·0001	1·33 (1·23–1·44)	<0·0001
	Rural area	1·12 (1·10–1·14)	<0·0001	1·06 (1·03–1·08)	<0·0001	1·19 (1·17–1·22)	<0·0001

HR=hazard ratio. NDVI=normalised difference vegetation index. PM_2·5=fine particulate matter.

Figure 1.

Association curve for 3-year average PM_2·5 and mortality under different tertiles of contemporaneous NDVI

(A) Tertile 1 of contemporaneous NDVI. (B) Tertile 2 of contemporaneous NDVI. (C) Tertile 3 of contemporaneous NDVI. HR=hazard ratio. NDVI=normalised difference vegetation index. PM_2·5=fine particulate matter.

Figure 2.

Association curves for contemporaneous NDVI and mortality under different tertiles of 3-year average PM_2·5

(A) Tertile 1 of 3-year average PM_2·5. (B) Tertile 2 of 3-year average PM_2·5. (C) Tertile 3 of 3-year average PM_2·5. HR=hazard ratio. NDVI=normalised difference vegetation index. PM_2·5=fine particulate matter.

Discussion

In this study, we found that greenness and air pollution had significant effects on mortality, consistent with previous studies. Our study showed a protective effect for greenness in China, a developing country, in line with other cohort studies in developed countries: Australia, Canada, Lithuania, Italy, Spain, Switzerland, and the USA.^{17, 18, 19, 22, 34, 38, 39, 40, 41} In high-income countries, individuals with better health outcomes tended to live in more desirable areas with more greenness, increasing the chance of residual confounding by socioeconomic status. In contrast, the possible residual confounding in our study is in the opposite direction, because the study participants who reside in areas with more greenness had lower socioeconomic status owing to China's urbanisation and economic development.⁹ Therefore, our effect estimates might gravitate towards the null, and finding an association between greenness and premature mortality reinforces the strength of association between residential greenness and health.

Air pollution contributes to premature mortality; the effect estimate for mortality was 1·09 (HR) for each 10 μg/m³ increase in 3-year annual average PM_2·5 (Model 2). Because our study population covers a wide range of geographies in China, we could study the effect in areas with varying degrees of economic development and other regional characteristics. Generally, concentrations of PM_2·5 tended to be higher in northern provinces than in southern provinces (range 9·21 μg/m³–109·37 μg/m³), due to energy use in industries and heating. The effect on mortality of PM_2·5 is different in areas of lower air pollution versus areas with greater air pollution. We saw that individuals residing in rural regions had more detrimental effects from air pollution (HR 1·10, 95% CI 1·08–1·13) compared with urban areas (1·06, 1·00–1·12). This finding indicates a higher marginal effect dose-response relationship at lower levels of air pollution, similar to the findings of larger scales from the global burden of disease studies.³

Our study's main contribution to the published literature is the statistically significant interaction between residential greenness and air pollution and mortality, albeit with a small effect estimate. In our fully adjusted model, individuals living in areas with more greenness appear to be affected more by air pollution, indicating a non-linear dose-response relationship for air pollution on mortality. This small effect indicates that increasing greenness and lowering air pollution might have a synergistic effect on the health of elderly populations, although the exact mechanism needs to be elucidated. Green space and vegetation could filter out air pollutants aerodynamically, or there could be in-vivo biological interactions. In our model, when multiplying the two exposure variables to create the interaction term, 0·1-unit decrease in NDVI and 10 μg/m³ increase in PM_2·5, the HR was 1·01 (p=0·027). This finding suggest that decreasing air pollution and increasing contemporaneous greenness have a synergistic effect on mortality. However, the effect size of the interaction depends on exposure windows considered for greenness and air pollution, because we did not find an association when using cumulative NDVI as the exposure of residential greenness (appendix p 4). We did not find a clear association between cumulative NDVI and mortality, possibly owing to some residual bias, because cumulative NDVI might reflect degree of urbanisation, whereas contemporaneous NDVI indicates acute exposure to greenness.²² Although older adults could be more vulnerable to short-term exposure than long-term exposure, further study is needed to explore the type of greenness and its interaction with PM_2·5 and mortality.

We did not observe monotonic trends in the effects of PM_2·5 on mortality within different amounts of exposure to residential greenness, and this finding reinforces the complex and unproven relationship between greenness and air pollution on health outcomes. Other studies have had mixed findings with no consensus on whether greenness does indeed act to mediate air pollution-related exposure pathways.^{25, 26, 27, 28} Our findings were slightly different from previous research. We found that air pollution mortality was lowest in the greenest areas, but only in rural areas, which also have greater greenness. A certain amount of vegetation might be needed to cleanse air pollution.⁴² The popularly proposed mechanisms include trees absorbing nitrogen oxides, ammonia, sulphur dioxide, and ozone and filtering of particulates by trapping them in their leaves and bark.^{42, 43} Thus, the effects might be stronger in rural areas with a greater density of vegetation. Adding to the mechanistic complexity, there might be differences in the sources, composition, or oxidative potential of PM_2·5, the structure of green space, and the types of vegetation between urban and rural areas, and among different areas. For example, large-scale population-based studies have shown that the health effects of green space differed by different structures of green space.^{44, 45} Structure difference in PM_2·5 and green space might be partly contributing to the non-linear trends. Restricted by data availability, we could not verify this hypothesis. Improved exposure assessment capabilities will be needed to further continue this research.

A strength of our study is that it is relatively large, and one of the most comprehensive in the region, including 12 873 participants from 22 provinces in China. We had detailed demographic and behavioural information, allowing us to do stratified analyses by sex, smoking status, and urban versus rural residence. This study had several limitations. First, we were only able to get an overall area of greenness coverage through NDVI, and we do not have information on the specific type of vegetation. Second, we lost 17·1% of participants to follow-up, and those living in more urban areas were more likely to be lost. Third, our mortality outcome did not include detailed information on the specific cause of death, so we could not analyse cardiovascular or pulmonary mortality, which are more directly related to air pollution. Fourth, there might be selection effects such as migration and neighbourhood self-selection. However, we believe our study population is highly stationary because of the advanced age and their social benefits, which are linked to the hukou household registration system.⁴⁶ Neighbourhood self-selection could be affected by individual and community socioeconomic status. In China, we found that larger amounts of greenness were associated with lower socioeconomic status. Thus, neighbourhood self-selection is unlikely to bias our analysis. Additionally, the resolution of MODIS images could introduce some, albeit minor, non-differential classification of greenness exposure, but there is a consensus among previous research of using 500 m radius for NDVI, as shown in meta-analyses.⁴⁷ At the same time, our PM_2·5 data are of the highest resolution exposure dataset available in China. Fifth, if we had better measurement of physical activity beyond a self-reported response to a single question, we would be able to use a mediation analysis to study behavioural connections of greenspace and air pollution. Finally, because our participants were from 22 provinces in China, we included the variable of geographical region to account for different dietary and other behavioural differences. Regional characteristics might have different effect sizes on the associations. Cluster analysis by geographical regions, with detailed vegetation type, is needed in future research.

In conclusion, we found that greenness and air pollution interacted non-linearly with mortality. Our study suggests that the effect of air pollution might be one mechanism in which greenness can protect health. However, relative to the main effect of air pollution and greenness, the interactive effect size is relatively small. Our finding provides evidence for urban planners, as greenness can be a protective measure in places with high population density and air pollution. Both national and local policies have emphasised a balanced economic development agenda that accounts for ecological civilisation. Our findings of interactions in the effects of greenness and air pollution can be used for both public health stakeholders and development agencies. Our stratified analysis showed that people living in urban areas were more likely to benefit from greenness, and people in rural areas are more likely to be harmed by air pollution. Sex did not modify the effect, indicating that greenness and air pollution affects women and men equally. Further study of the type of vegetation and the patch size of the green space is needed for a more precise understanding of this relationship.