Long-term Exposure to Nitrogen Dioxide and Mortality: A Systematic Review and Meta-analysis

Abstract

Background.

Ambient air pollution is among the greatest environmental risks to human health. However, little is known about the health effects of nitrogen dioxide (NO₂), a traffic-related air pollutant. Herein, we aimed to conduct a meta-analysis to investigate the long-term effects of NO₂ on mortality.

Methods.

We conducted a systematic search for studies that were published up to February 2020 and performed a meta-analysis of all available epidemiologic studies evaluating the associations between long-term exposure to NO₂ with all-cause, cardiovascular, and respiratory mortality. Overall pooled effect estimates as well as subgroup-specific pooled estimates (e.g. location, exposure assessment method, exposure metric, study population, age at recruitment, and key confounder adjustment) and 95% confidence intervals were calculated using random-effects models. Risk of bias assessment was accessed by following WHO global air quality guidelines. Publication bias was accessed by visually inspecting funnel plot and Egger’s liner regression was used to test of asymmetry.

Results.

Our search initially retrieved 1,349 unique studies, of which 34 studies met the inclusion criteria. The pooled hazard ratio (HR) for all-cause mortality was 1.06 (95%CI: 1.04–1.08, n=28 studies, I²=98.6%) per 10 ppb increase in annual NO₂ concentrations. The pooled HRs for cardiovascular and respiratory mortality per 10 ppb increment were 1.11 (95%CI: 1.07–1.16, n=20 studies, I²=99.2%) and 1.05 (95%CI: 1.02–1.08, n=17 studies, I²=94.6%), respectively. The sensitivity analysis pooling estimates from multi-pollutant models suggest an independent effect of NO₂ on mortality. Funnel plots indicate that there is no evidence for publication bias in our study.

Conclusion.

We provide robust epidemiological evidence that long-term exposure to NO₂, a proxy for traffic-sourced air pollutants, is associated with a higher risk of all-cause, cardiovascular, and respiratory mortality that might be independent of other common air pollutants.

1. Introduction

Ambient air pollution is among the greatest environmental risks to human health, and was reported to be responsible for 4.2 million deaths in 2016 worldwide [1]. Over the past decades, mounting epidemiological evidence has documented the adverse effects of particulate matter and ozone on human health [2–8]. Recently, there has been increased interest in nitrogen dioxide (NO₂), another traffic-related air pollutant.

Although NO₂ has multiple ecological sources, combustion of fossil fuels and motor vehicle emissions represent the primary source of NO₂ in the environment [9]. Previous work has considered NO₂ as an indicator of traffic pollution, given its strong correlation with other components of mobile exhaust. Additionally, NO₂ levels have been used to characterize other ambient air pollutants, such as NOx and ozone [10–12]. More recent work has focused on NO₂ as a possible independent contributor to adverse health effects. A growing body of evidence has reported associations between NO₂ and respiratory and cardiovascular disease-related mortality [13–15]. For example, Jerrett, Finkelstein [16], reported that long-term exposure to NO₂ is positively associated with cardiovascular (RR=1.45, 95%CI: 1.11–1.91) and respiratory mortality (RR=1.06, 95%CI: 0.79–1.43). However conflicting results have emerged, as Bentayeb, Wagner [17] reported null associations between NO₂ and cardiovascular mortality (HR=0.88, 95%CI: 0.51–1.54) and respiratory mortality (HR=0.76, 95%CI: 0.48–1.18). It should be noted that such discrepancies in the data is highly dependent upon location of measurement, time of measurement, sample sizes, study population, study designs, as well as other factors [18]. Thus, a more robust estimate is needed to improve the understanding of the relationship between NO₂ and overall mortality, in addition to respiratory and cardiovascular disease-related mortality.

To date, three meta-analyses, Faustini et al. (2014)[19], Atkinson et al. (2018)[9] and Huangfu et al. (2020)[20], integrated existing studies published prior to January 2013, October 2016, and January 2018, respectively, all of which reported a link between long-term exposure to NO₂ (e.g., annual mean or multiple-year average) and overall and cause-specific mortality. However, previous studies mainly focused on cohorts in the United States and Europe, and cohorts in Asia and Oceania were limited. Recently, an emerging interest in the health effects of NO₂ has motivated the study and publication of NO₂-exposed cohorts that provide a more global representation of the affected populations. Given this increased interest, to date, the latest epidemiological studies on long-term NO₂ have not been incorporated in any systematic review, presenting a serious gap in our understanding of the current data.

In the present study, we performed a systematic literature search with no location restriction and performed a meta-analysis of all available up-to-date epidemiological studies to examine the association between long-term exposure to ambient NO₂ and mortality endpoints, including all-cause, cardiovascular, and respiratory mortality. We have incorporated 6 new studies compared with Huangfu et al. (2020)[9, 20]with a total of 15 million study population in this meta-analysis that have not been included in the previous ones.

2. Methods

2.1. Search strategy

We conducted a systematic search using both PubMed and EMBASE to identify epidemiologic studies that evaluated long-term exposure to NO₂ and mortality. We restricted our search to all-language studies that were published up to February 29, 2020.

We used the following search terms: (“nitrogen dioxide” OR “NO₂” OR “NO_X” OR “nitrogen oxide” OR “traffic-related air pollution” OR “traffic related air pollution”) AND (“mortality” OR “cardiovascular mortality” OR “respiratory mortality”) AND (“epidemiology” OR “epidemiological” OR “epidemiologic” OR “cohort” OR “case-control” OR “case control”). We limited our search to human studies. Synonyms of NO and mortality were included using Medical Subheadings (MeSH) terms.

2.2. Study selection

We excluded toxicity studies, in vitro studies, book chapters, commentaries, letters to the editor, conference abstracts, review articles, meta-analyses, and studies that were not written in English. We also excluded epidemiology studies that did not provide risk estimates for NO₂ exposure, or that reported extremely high risk estimates (HR>5), or that did not evaluate all-cause, cardiovascular, or respiratory mortality. In addition, for studies examining the same cohort, we included only the most updated and comprehensive study.

The included population-based studies of all ages exposed to long-term concentration of NO₂ (> 1 year). Outcomes included in our study were all-cause, cardiovascular, and respiratory mortality. Following the PRISMA guidelines, four authors (initial name SH, HL, MW, YQ) independently evaluated titles and abstracts found in the 2 databases (n=1,774). Reference lists of review articles and meta-analyses were also reviewed to further identify epidemiology studies of NO₂ exposure and mortality (n=1). This resulted in a total of 159 potentially relevant articles for full-text review by four independent authors (SH, HL, MW, YQ). The eligibility of each study was independently assessed by two investigators (SH and YQ, or HL and MW) and the discrepancy was resolved through discussion with a third investigator. The protocol was registered at OSF and the link is provided at the bottom of the Figure 1. Detailed description of PECOS question is provided in Supplemental Table S1.

Figure 1.

Flow chart of the study selection process.

2.3. Data extraction

Data extraction and accuracy assessment were done by four independent authors (SH, HL, MW, YQ) on July 2020. Extracted information was entered into an Excel database, which included titles, authors’ names, publication year, country, study design, study period, cohort name, sample size, age range, sex distribution, time period of exposure assessment, exposure assessment method, exposure levels, exposure increment, effect measure, effect estimate and its standard error, and co-pollutant adjustment as well as adjustment for other confounders. For each study, we extracted the effect estimates from the main model or with the most stringent adjustment of potential confounders. Several studies employed both single- and multiple-pollutant models. In this situation, we extracted estimates from both models, and used estimates from the former in the main analyses and the other in sensitivity analysis.

2.4. Statistical analysis

After data extraction, all effect estimates were converted to HR (95%CI) per 10 ppb increase in NO₂ concentrations. Unit conversion is followed by Air Pollution Information System[21] and assumed ambient pressure of 1 atmosphere and a temperature of 25 degrees Celsius (1 ppb = 1.88 μg/m³).Forest plots were used to display the brief study information and HRs in each study graphically as well as the pooled results.

We tested for heterogeneity in the reported effect estimates, and we provided the p-values of the I²-based Cochran Q test and the I² metric of inconsistency [22]. We considered I² >50% to represent substantial heterogeneity [23]. An inverse variance random effects model was used to provide the pooled estimates. We performed stratified analyses to explore potential sources of heterogeneity by either cohorts or methodological characteristics. These included (1) study location: we divided study locations into four regions including North America, Europe, Asia and Oceania; (2) exposure assessment method: we separated the exposure assessment method to fixed monitor sites, land use regression (LUR) and other exposure assessment methods; (3) exposure metrices: annual single year average verses multiple year average; (4) study population: general population cohorts versus cohorts using subjects with preexisting disease; (5) age at recruitment; (6) key confounders adjustment for individual measures of BMI and smoking versus no adjustment. We screened for publication bias using funnel plot analysis with standard error as the measure of study size and Egger’s liner regression test of asymmetry [24, 25].

We conducted a series of sensitivity analyses to assess the robustness of results. We added back the studies with duplicated cohorts (n=8) and with extremely high HRs (n=2) and reran the meta-analysis. We also extracted estimates from the multi-pollutant models in the sensitivity analysis, if both single- and multi-pollutant models were fit. Moreover, we also excluded the articles that include high risk for each domain if available and rerun the model. Additionally, if I² ≥ 50%, we fit Hartung-Knapp-Sidik-Jonkman random effects models and compare the results with those from the DerSimonian-Laird random effects models. Statistical significance was assessed at the α = 0.05 level, unless otherwise reported. All statistical analyses were conducted in R version 4.0.1 using packages “meta”, “metagen” and “robvis”.

2.5. Risk of bias assessment

A Risk of Bias (RoB) tool was developed by a working group convened by WHO for the assessment of cohort studies in air pollution epidemiology [26]. The tool consists of six domains: confounding, selection bias, exposure assessment, outcome assessment, missing data and selective reporting, each divided to one to four subdomains. In total, there are 13 sub-domains each potentially rated as low, moderate, or high risk of bias [27]. If any single sub-domain is rated medium or high RoB then the domain is rated similarly. RoB was applied to each NO₂-outcome pair for studies included in a meta-analysis. For all-cause mortality, assessment of RoB for the confounding sub-domain “Were all confounders considered adjusted for in the analysis?” important confounders were: age, sex, body mass index (BMI) and individual- or area-level socio-economic status (SES). For respiratory mortality we also added smoking.

3. Results

3.1. Characteristics of the eligible studies

Our study selection process is presented in Figure 1, which represents the PRISMA Flowchart. A total of 159 peer-review articles were identified for our search in PubMed and EMBASE. 125 studies did not meet with the inclusion criteria and were excluded. The reason for the exclusion were: 9 studies not related to NO₂ concentration; 75 studies not in correct endpoint of interest; 15 studies not in qualified results; 8 studies published as editorial pieces or conference; 1 review study; and 7 studies not published in English. Of these, after title and abstract screening, we identified a total of 44 articles that fulfilled our initial inclusion criteria, of which eight were excluded because the same cohort was analyzed in other more recent publications (i.e., duplicated cohorts). Particularly, two studies reporting extreme estimates (HR>5) based on the Shenyang cohort were excluded for the further analyses, given that there were expressed concerns about the validity of the results (Supplemental Table S2) [28, 29]. As a result, 34 studies based on 32 separate cohorts were included in the final meta-analysis (3 studies were based on the same cohort but reported the risk estimates on three endpoints separately), comprising 10 studies from North America (7 from USA and 3 from Canada), 17 studies from Europe, 5 studies from Asia, and 2 studies from Oceania (Table 1). The measure of association reported in most studies was a hazard ratio along with 95% confidence intervals, while two studies reported relative risks (and 95% confidence intervals). Of the 34 studies, 28 studies examined all-cause mortality, 20 studies examined cardiovascular mortality, and 17 studies examined respiratory mortality. Particularly, Sanyal et al. (2018) [30] reported results for two partial-overlapping cohorts, including the ESPS survey data (Health, Health Care and Insurance Survey) and the CépiDc database (French Epidemiology Centre on Medical Causes of Death). We extracted results based on the CépiDc database due to the larger sample size and is therefore relatively more representative of the study. The study period, exposure assessment method, and exposure levels varied across the included studies. Most studies included both sexes, but two studies recruited only females [31, 32] and three studies recruited only males [33–35]. One article [36] studied males and females separately. Table 1 summarizes detailed characteristics for the studies included in the final meta-analysis.

Table 1.

Descriptive characteristics of cohorts included

Country	Study	Study period	Total population	Mean age (SD) or range (yrs)	Exposure assessment	Mean annual exposure (SD) or range	Study population
North America
USA	Lipfert et al. (2006)	1997–2001	26,843	51 (12)	Air monitoring sites	21.5 (6.1)ppb	Washington University-EPRI Veterans cohort
	Hart et al. (2011)	1985–2000	53,814	42.1 (9.9)	Spatial smoothing exposure model	14.2 (7.1) ppb	US Trucking Industry cohort
	Lipsett et al. (2011)	1997–2005	12,366	≥ 20	Air monitoring stations	33.59 (9.63) ppb	California Teachers Study (CTS)
	Hart et al. (2013)	1990–2008	84,562	30–55	Generalized additive models	13.9 ppb*	Nurses’ Health Study (NHS)
	Eckel et al. (2016)	1988–2009	352,053	69.3 (11.0)	Air monitoring stations	21.9 (10.2) ppb	Lung cancer patients
	Turner et al. (2016)	1982–2004	669,046	≥ 30	Land use regression	11.6 (5.1) ppb	American Cancer Society’s Cancer Prevention Study II (ACS CPS-II)
	Eum et al. (2019)	2000–2008	14.1million	65–120	Air monitoring stations	14.2 ppb*	Medicare cohort
Canada	Jerrett et al. (2009)	1992–2002	2,360	60*	Land use regression	22.9 ppb	Toronto respiratory cohort
	Chen et al. (2013)	1982–2004	205,440	35–85	Land use regression	12.1–21.7 ppb#	The Ontario Tax Cohort
	Crouse et al. (2015)	1991–2006	2,521,525	25–89	Land use regression	11.6 (6.7) ppb	Canadian Census Health and Environment Cohort (Can CHEC)
Europe
Norway	Næss et al. (2007)	1992–1998	77,891	51–70	Air dispersion model	NA	Oslo cohort
Netherlands	Beelen et al. (2008)	1987–1996	120,852	58–67	Interpolation, regressions, and GIS&	36.9 (8.2) μ/m3	The Netherlands Cohort Study on Diet and Cancer (NLCS)
	Fischer et al. (2015)	2004–2011	7,218,363	≥ 30	Land use regression	31 μ/m3*	The Dutch Environmental Longitudinal Study (DUELS)
UK	Maheswaran et al. (2010)	1995–2006	3,320	70.3 (14.6)	Air monitoring Sites	41 (3.3) μ/m3	South London Stroke cohort
	Carey et al. (2013)	2003–2007	830,429	40–89	Air dispersion model	22.5 (7.4) μ/m3	Clinical Practice Research Datalink
	Tonne et al. (2013)	2004–2010	154,204	68(13)	Gaussian dispersion model	18.8 μ/m3	Myocardial Ischaemia National Audit Project (MINAP)
	Halonen et al. (2015)	2003–2010	≥8,000,000	≥ 25	KCL urban dispersion model	38.9 (6.21) μ/m3	London cohort
	Dehbi et al. (2017)	1989–2015	7,529	48.45 (7.0)	Land use regressio	28.80 μ/m3*	National Study of Health and Development (NSHD) + Southall and Brent Revisited (SABRE)
Italy	Cesaroni et al. (2013)	2001–2010	1,265,058	≥ 30	Land use regression	43.6 (8.4) μ/m3	Rome Longitudinal Study (RoLS)
Denmark	Hvidtfeldt et al. (2019)	1993–2015	49,564	50–64	THOR/AirGIS dispersion model	25.0μ/m3*	The Diet, Cancer and Health cohort
France	Bentayeb et al. (2015)	1989–2013	20,327	43.7 (3.5)	CHIMERE chemistry-transport model	23 (12.1)μ/m3	Gazel cohort
	Sanyal et al. (2018)	1999–2012	13,239	≥ 15	CHIMERE chemistry-transport model	4.55 46.96 μ/m3	French cohort
Spain	de Keijzer et al. (2016)	2009–2013	44,561,414	NA	CALIOPE air quality forecasting system	9.48 μ/m3	Spain cohort
	Nieuwenhuijse n et al. (2018)	2010–2014	792,649	50.9 (18.3)	Land use regression	53.42 μ/m3	SIDIAPcohort
Multi-countries	Beelen et al. (2014)	1985–2007a	367,251	All ages	Land use regression	5.2–59.8 μ/m3	European Study of Cohorts for Air Pollution Effects (ESCAPE)
	Beelen et al. (2014)	1985–2007a	367,383	All ages	Land use regression	5.2–59.8 μ/m3	European Study of Cohorts for Air Pollution Effects (ESCAPE)
	Dimakopoulou et al. (2014)	1985– 2007a	307,553	All ages	Land use regression	5.2–59.8μ/m3	European Study of Cohorts for Air Pollution Effects (ESCAPE)
Asia
Japan	Katanoda et al. (2011)	1983–1995	63,520	≥ 40	Air monitoring stations	1.2–33.7 ppb	Three-prefecture Cohort Study
	Yorifuji et al. (2013)	1999–2009	13,412	74 (5.4)	Land use regression	22 (15)μ/m3	The Shizuoka elderly cohort
China	Chen et al. (2016)	1998–2009	39,054	44.29 (13.95)	Air monitoring stations	40.66 μ/m3	Four Northern Chinese city
	Yang et al. (2018)	1998–2011	66,820	70.2 (5.5)	Land use regression	104 (25.6) μ/m3	Hong Kong Elderly Health Service Cohort
South Korea	Kim et al. (2017)	2007–2013	136,094	42.05 (14.83)	Air monitoring stations	34.45 (12.92) ppb	National Health Insurance Service-National Sample NHIS-NSC) cohort
Oceania
Australia	Dirgawati et al. (2019)	1996–2012	11,627	72.1 (4.4)	Land use regression	13.4 (4.1) μ/m3	Health in Men Study (HIMS)
	Hanigan et al. (2019)	2007–2015	75,145	45–79	Satellite-based spatial regression model	17.75 (4.80) μ/m3	“45 and up study” Cohort

Notes:

^abaseline study period;

^*median;

^#mean annual exposure concentrations are 12.1 ppb in Windsor, 15.5 ppb in Hamilton, and 21.7 ppb in Toronto;

^&Sum of regional (interpolation), urban (regressions), and local traffic (GIS).

SIDIAP: Sistema d’Informació pel Desenvolupament de la Investigació en Atenció Primària NA indicates Not Applicable, SD standard deviation.

3.2. Risk of bias assessment

The risk of bias assessment for each study is shown in traffic light plot (Figure 2) in six different domains. The traffic light plot indicates that the quality for all studies was moderate to high. None of our studies had a ‘high’ or ‘probably high’ risk rating in all the key elements (exposure assessment, outcome assessment, and confounding) and therefore no studies were excluded from the analyses. Detailed rationale for each domain for subdomain of each study are provided in supplement material Table S3 a–c.

Figure 2.

Traffic light plot of risk of bias assessment for each study.

3.3. Results of the meta-analysis

Table 2 presents the pooled effect estimates and heterogeneity for each of the three endpoints of interest. Despite substantial heterogeneity across studies, and the fact that estimates vary by region and exposure assessment method, the results generally suggest an association of NO₂ with all three endpoints. Figures 3–5 respectively summarize the studies examining all-cause, cardiovascular, and respiratory mortality associated with traffic-related air pollution as measured by NO₂.

Table 2.

Pooled effects of nitrogen dioxide on all-cause, cardiovascular, and respiratory mortality

	All-cause mortality			Cardiovascular mortality			Respiratory mortality
	Studies (n)	HR (95% CI)	I2 (%)	Studies (n)	HR (95% CI)	I2 (%)	Studies (n)	HR (95% CI)	I2 (%)
Full meta-estimate	28	1.06 (1.04, 1.08)	98.6	20	1.11 (1.07, 1.16)	99.2	17	1.05 (1.02, 1.08)	94.6
Continent
North America	9	1.06 (1.03, 1.09)	97.9	7	1.09 (1.05, 1.12)	88.8	5	1.03 (1.02, 1.04)	0
Europe	13	1.03 (1.02, 1.05)	88.4	10	1.05 (l.00, 1.09)	86.9	9	1.04 (0.98, 1.09)	92.8
Asia	4	1.13 (0.83, 1.54)	99.1	3	1.39 (1.02, 1.88)	92.8	3	1.16 (l.00, 1.34)	63.0
Oceania	2	1.12 (1.01, 1.23)	0	-	-	-	-	-	-
Exposure assessment method
Fixed-site monitor	7	1.10 (1.04, 1.16)	99.3	3	1.24 (0.96, 1.60)	97.0	3	1.06 (0.96, 1.18)	95. 4
Land use regression	10	1.05 (1.04, 1.06)	61.3	9	1.09 (1.05, 1.13)	82.1	7	1.04 (1.01, 1.07)	34.9
Other	11	1.02 (1.01, 1.03)	80.3	8	1.07 (l.00, 1.15)	81.5	7	1.03 (0.89, 1.18)	92.4
Exposure metric
Single year	10	1.06 (1.04, 1.08)	76.4	8	1.09 (1.06, 1.12)	81.9	7	1.10 (1.03, 1.17)	86.5
Multiple year	16	1.06 (1.03, 1.07)	98.8	11	1.14 (1.07, 1.22)	99.5	9	1.03 (0.99, 1,07)	96.3
Study population
General population	24	1.05 (1.04, 1.07)	98.3	19	1.11 (1.06, 1.15)	99.2	16	1.05 (1.02, 1.08)	94.9
Preexisting disease	4	1.14 (1.02, 1.28)	85.5	1	NA	NA	1	NA	NA
Age at recruitment
≥60	8	1.08 (1.02, 1.14)	98.1	4	1.26 (1.02, 1.55)	87.6	4	1.10 (0.94, 1.28)	60.7
All age	20	1.05 (1.04, 1.07)	97.4	16	1.09 (1.05, 1.13)	94.3	13	1.05 (1.00, 1.10)	93.7
Key confounders adjustment for individual measures
BMI
Yes	16	1.08 (1.04, 1.11)	97.3	11	1.14 (1.09, 1.20)	90.7	10	1.09 (1.02, 1.16)	83.2
No	12	1.05 (1.02, 1.08)	98.9	9	1.06 (1.02, 1.11)	92.3	7	1.02 (0.97, 1.08)	95.3
Smoking
Yes	21	1.03 (1.01, 1.05)	97.8	14	1.07 (1.02, 1.24)	99.5	14	1.06 (1.02, 1.09)	95.3
No	7	1.11 (1.06, 1.17)	99.4	6	1.20 (1.12, 1.30)	94.0	3	1.04 (1.02, 1.06)	0

Figure 3.

Forest plot of study-specific hazard ratio (HR) of all-cause mortality associated with a 10-ppb increase in exposure to NO₂. The meta-estimate and weights in the forest plot are estimated from random effects meta-analyses.

Figure 5.

Forest plot of study-specific hazard ratio (HR) of respiratory mortality associated with a 10-ppb increase in exposure to NO₂. The meta-estimate and weights in the forest plot are estimated from random effects meta-analyses.

3.4. All-cause mortality

The overall pooled meta-estimate for all-cause mortality was 1.06 (95%CI: 1.04–1.08, n=28 studies) per 10 ppb increase in long-term NO₂ exposure (Figure 3). The pooled HRs for studies in Asia (HR=1.13, 95%CI: 0.83–1.54, n=4 studies) and Oceania (HR=1.12, 95%CI: 1.01–1.23, n=2 studies) were larger than that in North America (HR=1.06, 95%CI: 1.03–1.09, n=9 studies) and Europe (HR=1.03, 95%CI: 1.02–1.05, n=13 studies). The estimated heterogeneity across all 28 studies was substantially high, with I² of 98.6% (P<0.05). We also observed considerable heterogeneity across the studies in North America (I²=97.9%), Europe (I²=88.4%), and Asia (I²=99.1%). Notably, there is no heterogeneity for studies in Oceania partially because of insufficient study numbers (I²=0%, n=2 studies).

Four studies investigated the associations with mortality in cohorts selected on the basis of preexisting disease: STEMI [37], respiratory disease [16], stroke [38] and lung cancer [39]. Meta-analysis gave a summary HR of 1.14 (95%CI: 1.02, 1.28) compared with 1.05 (95%CI: 1.04, 1.07) from the general population. Eight studies recruited old population (age>60 years) and twenty studies recruited all age population. Meta-analysis reported substantial difference in HR for different age at recruitment, 1.08 (95%CI: 1.02, 1.14) versus 1.05 (95%CI: 1.04, 1.07), respectively. Meta-analytic summary estimates stratified by BMI and smoking status adjustment are also reported in Table 2. Moderate heterogeneity was also observed in the studied that used exposure estimates derived from LUR (I²=61.3%).

3.5. Cardiovascular mortality

The overall meta-estimate for cardiovascular mortality was 1.11 (95%CI: 1.07–1.16, n=20 studies) per 10 ppb increase in long-term NO₂ exposure (Figure 4). The pooled estimate was higher in studies from Asia (HR=1.39, 95%CI: 1.02–1.88, n=3 studies), compared to the studies in North America (HR=1.09, 95%CI: 1.05–1.12, n=7 studies) and Europe (HR=1.05, 95%CI: 1.00–1.09, n=10 studies), which was marginally significant. The overall heterogeneity between 21 studies was significantly high, with I² of 99.2% (P<0.05). Like all-cause mortality, there was also considerable heterogeneity across the studies in North America (I²=88.8%), Europe(I²=86.9%), and Asia(I²=92.8%). Larger summary of HRs was observed in cohorts with an older age (age>=60 years, HR=1.26, 95%CI: 1.02–1.55); and in studies by using fixed-site monitor (HR=1.24, 95%CI: 0.96–1.60). Meta-analytic summary estimates stratified by BMI and smoking status adjustment are also reported in Table 2.

Figure 4.

Forest plot of study-specific hazard ratio (HR) of cardiovascular mortality associated with a 10-ppb increase in exposure to NO₂. The meta-estimate and weights in the forest plot are estimated from random effects meta-analyses.

3.6. Respiratory mortality

The overall meta-estimate for respiratory mortality was 1.05 (95%CI: 1.02–1.08, n=17 studies) per 10 ppb increase in long-term NO₂ exposure (Figure 5). The stratified analysis by continent disclosed that the pooled estimate in Asia (HR=1.16, 95%CI: 1.00–1.34, n=3 studies) was larger than the estimates in North America (HR=1.03, 95%CI: 1.02–1.04, n=5 studies) and Europe (HR=1.04, 95%CI: 0.98–1.09, n=9 studies). The overall heterogeneity between 17 studies was significantly high, with I² of 94.6% (P<0.05). In addition, we observed a substantial heterogeneity across the studies in Europe (I²=92.8%) and a moderate heterogeneity in Asia (I²=63.0%), while the heterogeneity for North America was null (I²=0%). Larger summary of HRs was also observed in cohorts with an older age (age≥60 years, HR=1.10, 95%CI: 0.94–1.28); in studies using fixed-site monitor (HR=1.06, 95%CI: 0.96–1.18); and in cohorts with BMI (HR=1.09, 95%CI: 1.02–1.16) as well as smoking (HR=1.06, 95%CI: 1.02–1.09) adjustment (Table 2). We observed low heterogeneity in the studied that used LUR exposure assessment method (I²=34.9%) as well as the studies that did not adjust for smoking (I²=0%). We also found moderate heterogeneity for the studies that recruited at an elder age (age≥60, I²=60.7%) for the respiratory mortality (Table 2).

3.7. Publication bias

The funnel plots were visually symmetrical for cardiovascular mortality as well as respiratory mortality endpoint and asymmetrical for all-cause mortality endpoint (Figure 6). To further quantify the funnel asymmetry, we performed the Egger’s linear regression test. The P-value was 0.26 for all-cause mortality, 0.21 for cardiovascular mortality, and 0.17 for respiratory mortality, indicating no evidence of publication bias in all three endpoints.

Figure 6.

Funnel plots for (a) all-cause mortality, (b) cardiovascular mortality, and (c) respiratory mortality.

3.8. Sensitivity analysis

Table 3 summarizes the sensitivity analyses for long-term NO₂ exposure and mortality. We calculated pooled effect estimates only for studies using multi-pollutant models, and the results were essentially the same. Adding back the studies that reported extremely high HRs, the meta-estimates for cardiovascular and respiratory mortality were moderately elevated as expected and were nearly identical for all-cause mortality. We also removed the articles that identified as “high risks” for each domain and performed a subgroup analysis and the results were also identical for every domain (supplement Table S4).

Table 3.

Sensitive analysis of NO₂ and all-cause, cardiovascular, and respiratory mortality

	All-cause mortality			Cardiovascular mortality			Respiratory mortality
	Studies (n)	HR (95% CI)	I2 (%)	Studies (n)	HR (95% CI)	I2(%)	Studies (n)	HR(95% CI)	I2(%)
All countries
Including duplicate cohorts (a)	35	1.07 (1.05, 1.08)	98.3	25	1.13 (1.08, 1.17)	99.0	20	1.06 (1.03, 1.09)	93.7
Including extremely high effect estimates (b)	28	1.06 (1.04, 1.07)	98.6	21	1.20 (1.15, 1.26)	99.4	18	1.12 (1.08, 1.17)	97.8
Including both (a) and (b)	35	1.07 (1.05, 1.08)	98.3	26	1.21 (1.16, 1.26)	99. 2	21	1.12 (1.08, 1.17)	97.4
Studies using multipollutant model	9	1.05(1.0 2, 1.08)	98.4	7	1.09 (1.02, 1.16)	99. 7	5	1.00 (0.97, 1.03)	97.0
Two-pollutants model	2	1.00 (0.93, 1.09)	99. 7	1	NA	NA	2	1.01 (0.96, 1.06)	95.4
Three-pollutants model	5	1.03 (1.01, 1.07)	91. 1	5	1.02 (0.99, 1.06)	86. 9	3	0.99 (0.99, 0.99)	0
Hartung-Knapp-Sidik-Jonkman model
Full meta-estimate	28	1.07 (1.03, 1.12)	98.6	20	1.15 (104, 1.27)	99.2	17	1.04 (0.96, 1.13)	94.6
Continent
North America	9	1.06 (1.00, 1.12)	97.9	7	1.13 (0.93, 1.37)	88.8	5	1.03 (0.99, 1.06)	0
Europe	13	1.05 (0.99, 1.12)	88.4	10	1.09 (0.96, 1.25)	86.9	9	1.01 (0.88, 1.17)	92.8
Asia	4	1.13 (0.79, 1.61)	99.1	3	1.39 (0.68, 2.82)	92.8	3	1.16 (0.77, 1.75)	63.0
Oceania	2	1.12 (1.11,1.13)	0	-	-	-	-	-	-

4. Discussion

In this systematic review, we identified 34 studies from 32 separate globally representative cohorts that evaluated the effect of long-term exposure to NO₂ on all-cause, cardiovascular, and respiratory mortality. Our study provides evidence that long-term NO₂ exposure is positively associated with all three endpoints, with the largest effect estimates in Asia. No evidence of publication bias was observed, and none of our studies had a ‘high’ or ‘probably high’ risk rating within the risk of bias assessment, therefore, no studies were excluded from the meta-analysis.

The sensitivity analysis in which we replaced the results from the single-pollutant models with those from the multi-pollutant models, when available, presented nearly identical results. This suggests that NO₂ has independent effects on each of the health outcomes defined in this study.

Three recent meta-analyses have respectively evaluated studies published prior to January 2020[20], October 2016 [9], and January 2013 [19, 20], all of which reported substantial heterogeneity and significant associations between long-term exposure to NO₂ and all-cause, cardiovascular, and respiratory mortality, consistent with our present findings. Our study updates existing evidence by incorporating 6 new studies published from January 2018 through February 2020, indicating a growing evidence base. These 6 new studies include 15 million additional participants, which represent a 170% increase in sample size analyzed in previous meta-analyses, creating the largest evidence base to date. Moreover, 3.5 million, 1.6 million, and 0.4 million deaths were newly included for all-cause mortality, cardiovascular, and respiratory mortality, respectively. There are two studies conducted in Australia [34, 40], where for the first time long-term NO₂-mortality associations in Oceania were investigated, and a meta-estimate of 1.12 (95CI%: 1.01–1.23) with no heterogeneity (I²=0) was reported. These results were not included in previous studies; therefore, our analysis covers a broader geographical area. Compared to previous reviews, our overall meta-estimates are slightly larger with the addition of new studies with updated cohorts, longer follow-up periods, or better exposure estimates. We also examined the publication bias which had rarely been done in previous NO₂-mortality meta-analysis [9, 19], and we found that all our eligible studies lie symmetrically around our pooled effect sizes.

We registered our protocol to OSF and followed our a-priori decisions as reflected in the protocol. We did not register our protocol to PROSPERO which is the major registration platform of systematic reviews and this could be a limitation of our study.

Consistent with previous meta-analyses, our study observed a large degree of heterogeneity for all NO₂-outcomes pairs across enrolled studies, indicating a significant variation among results that could not be expected by chance alone. Although such heterogeneity does not impact our determination of consistency in causal inference, it is still essential to explore why the results are so disparate with each other [41]. We explored possible sources of heterogeneity by performing subgroup analyses using variables that can biologically or based on prior knowledge drive these associations. For instance, we stratified analyses by exposure assessment method (i.e. fixed-site monitor versus LUR versus other) and we found that the monitor-based studies had the highest estimates for all three endpoints, which is consistent with Atkinson et al. (2018) [9]. One possible reason is that amongst the limited studies that based on fixed monitors, the study populations were mainly comprised of elderly population that are typically vulnerable to air pollution [38, 39]. Apart from these sources, high statistical heterogeneity could attribute to methodological diversity or differences in outcome assessments. The methodological diversity is due to, first, substantial variation of sample size across different studies, ranging from 2,000 to 44.5 million [42]. Other possible factors further relate to the variation of study demographics, such as location of the study population, study population, and NO₂ concentration levels and different levels of confounding adjustment, which made the studies suffer from different degrees of bias and lead to diverse estimates. For instance, Atkinson et al. (2013) [43] adjusted for 4 covariates including age, sex, Body mass index, and smoking status, and reported an effect on all-cause mortality of 1.13 (95%CI: 1.07, 1.20) per 10 ppb increase in NO₂ levels. In contrast, Katanoda et al. (2011) [44] adjusted for 17 potential confounders such as lifestyle, dietary, socioeconomic status, marital status, and medical history and reported an estimated effect on all-cause mortality of 0.97 (95%CI: 0.91, 1.04) per 10 ppb increase in NO₂ levels. Moreover, the ICD code for cause-specific mortality can be slightly different, which may also result in high heterogeneity. Even though the high heterogeneity suggests that the studies are not all estimating the same quantity, it does not necessarily suggest that the true exposure effect varies.

Accurate exposure estimates are crucial for environmental epidemiology studies. Recently, satellite data have been widely used in high-resolution air pollution level predictions, such as daily 1-km PM_2.5 and ozone prediction [45, 46]. However, high-resolution NO₂ prediction models utilizing satellite information are very sparse [47], and only two latest cohort studies from Denmark and Australia included in our meta-analysis were able to integrate satellite retrieved NO₂ estimates [40, 48]. The NO₂ levels in studies included in the meta-analysis tended to be derived from fixed-site monitors, LUR and CHIMERE chemistry transport models that yielded larger exposure measurement errors, as compared to more advanced exposure techniques such as machine learning. Further, fixed-site monitors are usually insufficient to adequately capture the spatial and temporal variability within a large area. Though geospatial statistical methods (such as LUR and Kriging) allow characterizing the spatial variation of exposure, they do not generally capture temporal variability in exposure, because they are commonly averaged over on a year, bi-annually or more. Chemical transport models usually increase the spatial variability to a few hundred kilometers, still not comparable to the satellite-based approach. Epidemiology studies with finer-resolution NO₂ exposure estimates (and consequent less exposure measurement error) are in urgent need.

Toxicological studies suggest possible mechanisms via which NO₂ might contribute to mortality. For example, NO₂ is related to increased levels of oxidative free radicals and inflammation [49, 50]. Numerous experimental studies demonstrated that air pollution promotes a systemic vascular oxidative stress reaction [51]. Radical oxygen species can cause endothelial dysfunction, monocyte activation, and certain pro-atherosclerotic changes in lipoproteins, thereby initiating plaque formation, exacerbating disease, and increasing mortality [51].

Previous epidemiologic reviews concluded that even though some evidence between NO₂ and mortality is suggestive, it is still not sufficient to infer a causal relationship between long-term exposure to ambient NO₂ and mortality [9, 52]. In most of the epidemiological studies, ambient NO₂ is positively related to mortality, but we cannot rule out the possibility that such an association may be due to confounding variables, such as socioeconomic status (SES), behavioral factors, and co-pollutants (e.g., O_3, PM_2.5, PM₁₀, SO₂). Because different studies may adjust for different confounding variables as well as co-pollutants and such different adjustment could affect the result of effect estimates. Moreover, substantial high heterogeneity between the study results can also weaken the causality argument. Therefore, we believe that based on current evidence, the causal association for estimating the burden of NO₂ on mortality and life expectancy is still moderate. Although this study cannot provide a confirmed causal relationship between NO₂ and mortality, it can still help to evaluate the scientific debate as the meta-analysis improves the precision and validity of estimates as increased amounts of data are utilized. The pooled effect estimates we provided can also be useful for future health impact assessment.

In conclusion, we provide robust epidemiological evidence that long-term exposure to NO₂, a proxy for traffic-sourced air pollutants, is associated with a higher risk of all-cause, cardiovascular, and respiratory mortality that might be independent of other criteria air pollutants. This finding can inform public health policy regarding the health effects of traffic pollution on taking appropriate measures to reduce exposure to traffic pollution, especially in vulnerable populations.

Highlights.

We performed a systematic literature search and a meta-analysis of all available up-to-date epidemiological studies to examine the association between long-term exposure to NO₂ and all-cause, cardiovascular and respiratory mortality.

Our study incorporates 6 new studies compared with Huangfu et al. (2020) with 15 million newly included subjects, creating the largest evidence base to date.

NO₂ has an independent effect on mortality and reinforced the conclusion that long-term NO₂ exposure, a proxy for traffic-sourced air pollutants, could increase the risk of mortality.