Exposure to ambient nitrogen dioxide (NO2) has been linked to increased mortality in several epidemiological studies but the question remains of whether NO2 is directly responsible for the health effects or is only an indicator of other pollutants, including particulate matter. The aim of the present review was to provide pooled estimates of the long-term effects of NO2 on mortality, which are potentially useful for health impact assessment.
We selected 23 papers, published from 2004 to 2013, evaluating the relationship between NO2 and mortality, also including an assessment of the effect of particulate matter exposure. A random-effects meta-analysis was carried out on 19 studies.
The pooled effect on mortality was 1.04 (95% CI 1.02–1.06) with an increase of 10 μg·m−3 in the annual NO2 concentration and 1.05 (95% CI 1.01–1.09) for particulate matter <2.5 μm in diameter (PM2.5) (10 μg·m−3). The effect on cardiovascular mortality was 1.13 (95% CI 1.09–1.18) for NO2 and 1.20 (95% CI 1.09–1.31) for PM2.5. The NO2 effect on respiratory mortality was 1.03 (95% CI 1.02–1.03) and 1.05 (95% CI 1.01–1.09) for PM2.5. Four bipollutant analyses with particulate matter and NO2 in the same models showed minimal changes in the effect estimates of NO2.
There is evidence of a long-term effect of NO2 on mortality as great as that of PM2.5. An independent effect of NO2 emerged from multipollutant models.
Any health impact assessment overlooking NO2 would neglect some adverse effects of today’s air pollution mixture http://ow.ly/sPKsR
The long-term effects on mortality of exposure to nitrogen dioxide (NO2) were assessed in the World Health Organization (WHO) Air Quality Guidelines in 2005  and the overall evidence was considered limited, given the small number of studies available. Neither the development of chronic diseases nor lung cancer was clearly associated with NO2 in any of the studies taken into consideration in the 2005 WHO report. However, short-term mortality studies, studies of the impairment of lung function growth in children and investigations of recurrent respiratory problems all suggested a NO2 effect. Recent documents on traffic exposure from the American Thoracic Society  and the Health Effects Institute  considered NO2, among other traffic-related pollutants, and concluded that there is evidence suggesting that these pollutants have a causal role in mortality and the development of chronic respiratory diseases.
The recent review of the health effects of air pollution by a WHO working group (Review of Evidence on Health Aspects of Air Pollution (REVIHAAP))  evaluated the most recent studies on long-term exposure to NO2 and both natural (non-accidental)/total and cause-specific mortality. The question of whether NO2 is directly responsible for the health effects or is only an indicator of other pollutants was specifically considered, including particulate matter and specific constituents such as metals, polycyclic aromatic hydrocarbons and other organic matter. The conclusion was that it is difficult “to judge the independent effects of NO2 in the long-term studies because, in those investigations, the correlations between concentrations of NO2 and other pollutants are often high, so that NO2 might represent a mixture of traffic-related pollutants. In this case, chamber studies do not apply and toxicological evidence is limited. However, some epidemiological studies do suggest associations of long-term NO2 exposures with respiratory and cardiovascular mortality and with children’s respiratory symptoms and lung function that were independent of [particulate matter] mass metrics. As with the short-term effects, NO2 in these studies may represent other constituents. Despite this, the mechanistic evidence, particularly on respiratory effects, and the weight of evidence on short-term associations suggest a causal relationship.” In other words, the WHO working group suggested that the response to the causality question cannot be derived only from the epidemiological studies but requires the integration of exposure science, toxicology and human studies.
At the time of the REVIHAAP project, a quantitative systematic review of the long-term effects on mortality of NO2 was not available. A subsequent review by Hoek et al. , considering the association between several pollutants and long-term mortality, showed a significant association of all-cause mortality with NO2 in 15 longitudinal studies. Their work did not permit comparison of the effects of NO2 with those of other pollutants, namely particulate matter, when the same studies were considered or when the results of the multipollutant models were available.
To contribute to the discussion of the role of NO2 from an epidemiological perspective and to provide pooled estimates of the long-term effects of NO2 on mortality, which are potentially useful for assessing the health impact of air pollution, we systematically examined the studies that investigated the long-term effects on mortality of both outdoor NO2 and particles in the same population, and carried out a meta-analysis of both NO2 and particulate matter effects.
The search strategy focused on studies reporting NO2 (or nitrogen oxides (NOx)) effects on total and specific-cause mortality. The databases PubMed, Web of Science and LUDOK were searched. The search was restricted to the period from January 2004 to January 2013 and to studies in humans. Search terms were “nitrogen dioxide” (as “nitrogen dioxide/adverse effects” Medical Subject Headings in PubMed and as “nitrogen oxides” in LUDOK), together with “cohort study” or “case–control study”, and “mortality”, “cardiovascular mortality” and “respiratory mortality”.
Articles were included if long-term effects on natural/total or specific (cardiovascular or respiratory) mortality were assessed and the estimates of association with the pollutants were reported; the exposure needed to have been measured as at least an annual mean concentration in the study area. Papers were excluded when short-term effects were estimated, no health effect was measured, incidence or prevalence was assessed, no original data were analysed or, finally, when no control for confounding due to individual factors was carried out. Reviews and methodological papers were excluded as well. Studies carried out on groups restricted to at-risk people were, however, not excluded.
The first screening of the papers was carried out by checking the titles and abstracts. The texts of the remaining papers were read in full, so as to select those that also included effect estimates for particulate matter <10 μm in diameter (PM10) or PM2.5, total suspended particles (TSPs) and black smoke.
The following characteristics were reported for each paper: authors, location and year of publication; study population; study design and study period (including both the enrolment and the follow-up period, if it was required by the study design); exposure assessment, including the method employed to assess levels, the mean concentration of pollutants and their exposure metrics; traffic exposure indicators (such as distance from the nearest high-traffic road and traffic density); and effect estimates of both total and specific-cause mortality.
Before pooling data, we converted the effect estimates related to NOx into NO2 effects, using a conversion factor of 0.75 . The NO2 concentrations expressed in parts per billion were converted into micrograms per cubic metre; we used a conversion factor of 1.88 (at 25°C and 1013 mb) for both NO2 and NOx. The effects related to TSP were converted into PM10 effects, using a factor of 0.75 , and the effects related to PM10 were converted into PM2.5 effects, using a factor of 0.7 . Diesel particulate matter and black smoke were analysed as indicators of PM2.5 . In addition, the effects expressed as interquartile (or quintile or percentile) differences were converted into effects of a 10-μg·m−3 increase of each pollutant.
We used hazard ratios, odds ratios and relative risks (RRs) of mortality in a random-effects meta-analysis  (Stata, version 10; StataCorp, College Station, TX, USA), thus including the risk estimates in the pooled analysis irrespective of the study design. The risk estimates and 95% confidence intervals from each study were reported, having been adjusted for all the factors the authors assessed as confounders. The standard error of the effect was calculated from the risk estimates and the confidence intervals (se = (ln RR - ln lowest limit CI)/1.96 or se=(ln highest limit CI - ln RR)/1.96). The random pooled effects were reported as hazard ratios, as most studies used Cox’s survival analysis. We assessed heterogeneity across studies by using the Chi-squared test (Cochran’s Q) and inconsistency (I2) , which represents the proportion of total variation in effect estimates due to heterogeneity across the studies. Finally, we stratified the results by geographical area (Asia, North America and Europe).
To compare the relative effects of NO2 and PM2.5, we carried out a meta-analysis using the interquartile range (IQR) as the exposure metric instead of the fixed increment (10 μg·m−3). We included only studies where the IQRs were available or could be derived by using the mean concentrations and their standard deviations. The pooled estimate was obtained for a median IQR of all the single IQRs from those reported or specifically estimated.
Our initial search led to the selection of a total of 180 articles that were seen to deal with NO2 (or NOx) exposure. We excluded 146 articles on the basis of their titles and abstracts: nine reviews, 12 articles that did not analyse original data, two methodological studies, eight with study designs that made it impossible to adjust for individual confounders, 33 articles that did not assess health effects or human outcomes, 46 articles dealing with short-term effects, 26 articles with long-term effects other than mortality, one article studying only lung cancer mortality and nine articles with a mean exposure to air pollution <1 year. Thereafter, from the 34 studies that had been identified as warranting reading in full, we selected 23 studies [12–34] that assessed particulate matter effects in addition to NO2 or NOx effects. All the selected studies and their characteristics are reported in detail in online supplementary table S1. The 11 excluded studies are listed in table S2.
14 studies [12, 13, 17, 19–24, 26, 30–32, 34] evaluated natural or total mortality; 17 studies [12, 13, 15, 17–20, 22, 23, 25–29, 32–34] considered cardiovascular mortality (including four studies [13, 23, 32, 34] assessing cardiopulmonary mortality); 13 studies considered respiratory mortality [12–14, 16, 17, 19, 20, 22, 23, 26, 28, 32, 34] (including the four with cardiopulmonary mortality as a single outcome [13, 23, 32, 34], and a further one  that assessed lung cancer and respiratory mortality as a single outcome). Only 10 studies [12, 13, 17, 19, 20, 22, 23, 26, 32, 34] evaluated both total and cause-specific mortality, although data were not always exhaustively reported. There were 21 cohort studies [12–24, 26–28, 30–34] and two case–control studies [25, 29].
Cohort studies involved adults of both sexes but one study selected only adult males  and four selected only adult females [13, 19, 27, 32], and two further articles [28, 33] studied both males and females but separately. Among the other studies, one studied only females . Five studies focused on specific at-risk groups: patients with stroke , respiratory diseases , hypertension  and first myocardial infarction [25, 29].
Several confounders were assessed in each article. Sex, age and social class (identified by different indicators such as education, deprivation index, occupation, area-level socioeconomic position (SEP) or income) were included in most studies. Only Dong et al. , Katanoda et al.  and Lipfert et al.  included no SEP indicator in their final analysis. Habitual smoking was frequently adjusted for [15–17, 21, 26, 27, 32, 33], whereas comorbidities were rarely assessed [12, 18, 21].
Asia was represented by four studies (one from Japan  and three from China [14, 15, 17]), North America contributed nine studies (two from Canada [18, 22] and seven from the USA [19, 20, 23, 24, 30, 31, 33]) and Europe contributed 10 studies [12, 13, 21, 25–29, 32, 34].
NO2 was a main exposure in almost all studies, but two [17, 24] used NOx. The exposure metric was mostly expressed in micrograms of NO2 per cubic metre, but one study from Asia , one from Canada  and all seven studies from the USA [19, 20, 23, 24, 30, 31, 33] reported the mean levels of NO2 in parts per billion. The unit increase of NO2 was usually reported (16 studies from all the geographic areas) as interquartile or centile range rather than as per 10 μg·m−3 [12, 13, 18–20, 22–25, 26–29, 31–33].
The mean yearly level of NO2 was 41.6 μg·m−3 in Asia, 42.2 μg·m−3 in North America and 36.2 μg·m−3 in Europe.
Particulate matter was studied for the most part as PM10 (11 studies) [13–15, 19–21, 25, 27–29, 32] with the effect estimated per 10 μg·m−3 (four studies [14, 15, 19, 21]) or IQR, expressed in micrograms per cubic metre. PM2.5 was studied in 11 cases [12, 16, 18, 22–24, 26, 30, 31, 33, 34] with exposure expressed most frequently as an IQR increase of concentrations (six studies [12, 18, 22, 23, 31, 33]). TSPs (expressed per 10 μg·m−3) was the particle metric in one study . The mean yearly level of PM2.5 was 87.4 μg·m−3 in Asia, 14.8 μg·m−3 in North America and 25.4 μg·m−3 in Europe.
The correlation between NO2 (or NOx) and PM10 was reported in eight studies out of 11, with values ranging from 0.5  to 0.9 [14, 15, 29]. The correlation between NO2 and PM2.5 was reported in nine studies out of 11, with values ranging from 0.3 in the Japanese study  and in the Californian (USA) study  to 0.88 in the Norwegian study . The study using only TSPs  did not indicate the correlation with NO2.
Traffic exposure indicators were reported in nine publications. Two evaluated proximity to a road with heavy traffic and traffic density [12, 26], the studies in Toronto, Canada  and in Germany (the Study on the Influence of Air Pollution on Lung, Inflammation and Aging) [13, 27, 32] investigated only proximity to traffic, and the US Veterans Affairs studies [24, 30, 31] evaluated traffic density only.
We included 19 studies in the meta-analysis. Four of the 23 included in the systematic review have been further excluded from the meta-analysis, as three of them reported more complete analyses 3 years later [29–31] and one other  reported mortality only for respiratory diseases and lung cancer as a single outcome. In contrast, two of the included papers [28, 33] carried out separate analyses for males and females, thus producing 21 effect estimates actually analysed.
The pooled estimate of natural/total mortality was performed for 12 studies [12, 13, 17, 19–24, 26, 32, 34], the pooled estimate of cardiovascular mortality made use of 18 effect estimates [12, 13, 15, 17–20, 22, 23, 25–28, 32, 33, 34] (as two studies analysed males and females separately), the meta-analytic estimate of respiratory mortality included nine effect estimates [12, 16, 17, 19, 20, 22, 26, 28] (as one study analysed males and females separately ).
The pooled estimate of NO2 effects on natural/total mortality (table 1 and fig. 1) was RR 1.04 (95% CI 1.02–1.06) per 10 μg·m−3 NO2 and the corresponding effect of PM2.5 was RR 1.05 (95% CI 1.01–1.09) per 10 μg·m−3 PM2.5. However, there was high heterogeneity across the studies in North America: the NO2 effects were not statistically significant while the effects of PM2.5 were larger and statistically significant (in spite of the fact that the levels of PM2.5 observed there were the lowest we found). The greatest effect on natural/total mortality was observed in Europe for both NO2 and PM2.5 but heterogeneity observed across the studies was still high. It should be noticed that a part of this heterogeneity could be explained by just one study  on stroke patients.
The pooled estimate of the NO2 effect (table 1 and fig. 2) on cardiovascular mortality was RR 1.13 (95% CI 1.09–1.18), while the effect of PM2.5 was RR 1.20 (95% CI 1.09–1.31), per 10 μg·m−3 of each pollutant. In Asia, the effects of both NO2 and PM2.5 were greater than those in North America and Europe. Notably, the results were affected by high heterogeneity in all areas. In Asia, where the overall NO2 estimate did not attain statistical significance (and heterogeneity showed values near 99%), one study  provided very high effect estimates, much greater than for the other studies. In North America, the study in Toronto , carried out on patients from a respiratory clinic, provided effects on cardiovascular mortality as high as 64% for NO2, in contrast with a mean increase of 1% in the other studies.
To better evaluate the heterogeneity of the pooled effect estimates of NO2 on total and cardiovascular mortality, we conducted subgroup analyses considering studies not involving at-risk groups, i.e. excluding the studies by Maheswaran et al. , Jerret et al. , Lipfert et al.  and Rosenlund et al. , and studies with better exposure assessment (land use regression (LUR), dispersion models or other methods at the address level) [12, 18–22, 25, 26, 28]. The heterogeneity of the effects generally remained in these subgroups when they were considered separately (table S3). However, when we restricted our analysis to studies conducted in the general population (not on at-risk groups) and with a better exposure assessment, the heterogeneity for cardiovascular mortality decreased (p=0.690 and I2=0%, based on seven studies) corresponding to an effect estimate of 1.03 for NO2.
The pooled estimate of NO2 effects (table 1 and fig. 3) on respiratory mortality was RR 1.02 (95% CI 1.02–1.03) while the PM2.5 effect was RR 1.05 (95% CI 1.01–1.09). The results for NO2 were homogeneous across the studies. The large effect of PM2.5 on respiratory mortality was notable in Europe, where an effect of 8% was detected with a heterogeneity as high as 40%.
The results of the pooled estimates using IQRs as the exposure metric are presented in table 2. The number of studies with available estimates is slightly lower than for the analysis using a fixed increment, as four of them [16, 17, 25, 34] did not provide IQR data or figures to estimate them. The median IQRs were different for the outcomes considered and the variability of NO2 levels was greater than for PM2.5 (e.g. 14.1 μg·m−3 for NO2 and 5.4 μg·m−3 for PM2.5 in natural/total mortality studies). The estimates based on IQRs showed greater effects of NO2 than of PM2.5 on total mortality (6% versus 3% for PM2.5) and on cardiovascular mortality (29% versus 16% for PM2.5). The effects were similar for respiratory mortality (4% for both NO2 and PM2.5).
The results of the multipollutant analyses (as assessed in the original papers) are reported in table S4. Only seven studies [12, 17, 18, 20, 22, 24, 32] among those included in the meta-analysis performed bipollutant analyses, including those by Gehring et al. , Jerrett et al.  and Lipfert et al. , who analysed the NO2 effects only with a traffic indicator. Cao et al.  in Asia and Cesaroni et al.  in Europe found no change in the effects of NO2 on total mortality when it was analysed together with particulate matter, whereas the reduced effect observed by Hart et al.  in the USA relates to a model including three pollutants (NO2, PM10 and sulfur dioxide). Among the three studies that analysed NO2 and particulate matter effects on cardiovascular mortality [17, 18, 20], two [18, 20] found independent effects of NO2. Three studies analysed NO2 effects with a traffic indicator: those by Lipfert et al.  in the USA and Jerret et al.  in Toronto found reduced effects of NO2 on total and cardiovascular mortality, but no change was reported in the European study .
We found similar risk estimates for total mortality in studies investigating the long-term effects of both NO2 and PM2.5 (4% versus 5% increase, respectively) using an exposure metric of 10 μg·m−3. In Europe, there was a 7% increase in total mortality for both NO2 and fine particles. Hoek et al. , in their recent review, found a worldwide pooled estimate of 5.5% (95% CI 3.1–8.0%) for NO2; their study, unlike the present one, included estimates from studies that did not analyse particles. We found greater effects of PM2.5 than of NO2 for cardiovascular (20% versus 13%) and respiratory (5% versus 2%) mortality, per 10 μg·m−3 of pollutants. Per IQR, estimates of cardiovascular mortality were larger for NO2 than for PM2.5 (29% versus 16%). In the recent review by Hoek et al. , NO2 effects ranging from 3% to 36% per 10 μg·m−3 have been reported for cardiovascular mortality in European cohorts as well as an effect of 12% for respiratory mortality. All our estimates were rather heterogeneous, with the sole exception of respiratory mortality. In the few studies with bipollutant models, the effects of NO2 and particles (of different sizes) seemed independent.
The comparison of studies carried out in worldwide geographical areas is beset by many challenges. The methods used to assess exposure play an important role. The NO2 levels observed in the USA, Asia and Europe are very similar. In contrast, PM2.5 levels are lower in the USA than in Europe, and both are much lower in Asia. The methods used in assessing exposure (such as dispersion modelling, LUR models and inverse distance weighting) as well as the methods used in pollution monitoring (such as the number and location of fixed monitors as well as the daily sampling intervals) could all influence the quality of the exposure assessment causing a possible differential misclassification. However, as lower levels of PM2.5 in the USA produced effects on total mortality comparable to those observed in Europe, a different toxicity of the fine particle components between areas could also play a role.
The length of the exposure (or modelling) period was not always taken sufficiently into consideration. This is an important drawback when risk estimates for long-term mortality are to be compared. The length of exposure assessment ranged from 1 [21, 23] to 12 years  when reported as measured levels of pollutants, and from 2 to 40 years when estimated by means of models (mainly dispersion or LUR models). In addition, length of exposure is mostly assessed on the basis of length of residence, assuming that the pollution levels monitored in a shorter time than residence period represents the real exposure of the whole residence period. Only a few investigations have addressed the long-term stability of the spatial contrasts in the exposures used in the epidemiological analyses [35, 36]
Latency differs among studies, often not consistently with the study design. Three studies [15, 17, 19] had an overlap between the exposure and the follow-up periods; others started the follow-up in the same year the pollutant exposure was assessed [24, 28], although they prolonged the follow-up for up to 30 years. In a few studies [12, 20, 21, 34] exposure was assessed in a few years included within the follow-up or even after the follow-up . Finally, eight studies [16, 18, 23, 25, 26, 27, 32, 33] carried out a follow-up from at least 1 year after the assessed exposure for up to 6–20 years thereafter. Therefore, the differences in assessing the temporal relationship between exposure and outcome across the studies represent a possible challenge for the interpretation of the results. Doubts could also arise when the exposure has been assessed after starting follow-up, since strong assumptions were necessary about the length of residence and the corresponding levels of pollutants.
Some methodological choices in carrying out the meta-analysis need to be discussed. The first relevant point for the interpretation of the results is the reliability of the meta-analytic estimates by geographic area. The European and North American estimates for total and cardiovascular mortality, as well as the European estimates for respiratory mortality, are more reliable than Asian estimates because at least four articles were included. However, the heterogeneity across the studies affects the reliability [10, 37] of the estimates of total and cardiovascular mortality, in all areas and for both pollutants. The sole exception is the effect of NO2 on respiratory mortality in both Europe and the USA (with three studies). It should be noted that the heterogeneity for cardiovascular mortality was strongly reduced when we restricted to studies with a better quality of exposure assessment and evaluating the entire population and not at at-risk groups .
A second methodological point for the interpretation of results is related to the pollutant exposure metric for analysing the effects of different pollutants. Using a fixed increase of 10 μg·m−3 in pollutant levels makes it possible to compare the effects of a single pollutant across different studies and different countries. In contrast, if one wishes to compare the effects of two or more pollutants, one needs another metric that takes into account the different distributions of pollutant levels. A pollutant with a broader exposure distribution will show smaller effect estimates per unit increase than a pollutant with a less wide exposure distribution in the same population. The additional meta-analyses we performed using the IQRs of NO2 and PM2.5 as the exposure metric show that the NO2 effects on mortality are larger than those of PM2.5, at least for total and cardiovascular mortality.
There is obviously a need for caution in drawing conclusions about the independence of NO2 effects from PM2.5 effects from the multipollutant models. First, not all countries were included (e.g. Asia was less represented in this analysis). Second, the high correlation between NO2 and PM2.5 (around 0.7–0.8) in these studies still suggests the possibility that the NO2 effects could be due in part to confounding from particulate matter. Third, uncertainty is also due to the limited number of bipollutant estimates available. It should be noted, however, that only minimal difference between the single-pollutant and the multipollutant results for NO2 were found.
Finally, although this study cannot respond exhaustively to the causality question about NO2, given that it was a meta-analysis, it contributes to the scientific debate on this topic because it assesses the respective role of NO2 and PM2.5 as they emerge from prospective studies  and provides a pooled effect estimate to be used for future impact health assessment.
In conclusion, the magnitude of the long-term effects of NO2 on mortality is at least as important as that of PM2.5. These results hold when using either 10 μg·m−3 or IQR as the metric of choice. The results of the multipollutant models suggest that the role of NO2 is independent of that of particles. All these elements may be useful when discussing the causality issue for NO2. Moreover, NO2 is the only widely regulated pollutant (apart from carbon monoxide) that indicates exposure to combustion pollution. Health impact assessments relying only on PM2.5, and not considering NO2, would be neglecting an important source of the adverse effects of today’s pollution mixture.
We thank K. Christenfeld (freelance medical editor, Rome, Italy) for editorial revision and S. Ricci (Dept of Epidemiology, Regional Health System, Lazio Region, Rome, Italy) for her help with the figures.
This article has supplementary material available from erj.ersjournals.com
Conflict of interest: None declared.
- Received July 5, 2013.
- Accepted December 19, 2013.
- ©ERS 2014