Literature searches

MEDLINE, Embase, CENTRAL, Global Health and Toxline databases were searched using terms developed by a librarian (S1 Table). The search strategy underwent Peer Review of Electronic Search Strategies (PRESS) [22]. Searches were last updated February 25, 2020. Inclusion criteria were as follows: Participants/population: Humans; Intervention(s), exposure(s): Exposure to outdoor NO₂ (and other oxides of nitrogen); Comparator(s)/control: Lower levels of exposure; Main outcomes: Mortality from all/ natural causes or specific causes; Study design: cohort. Publications in abstract form only were excluded. Publications in English or French were included and there were no restrictions on publication date. Effect measures considered were: mortality effects reported as regression coefficients, hazard ratios (HR) or relative risks associated with exposures over a period of years, expressed per specified increment in exposure. The present review is one part of a series of reviews of effects of NO₂, all of which were included in the original search. Other reviews pertain to non-asthma respiratory morbidity and ischemic heart disease morbidity related to short term exposure [21]. Studies were selected for the present review if reported outcomes matched the inclusion criteria specified above.

Screening, data extraction and risk of bias assessment

Screening and data extraction were completed independently by two reviewers in DistillerSR. Conflicts between reviewers were resolved through consensus and/or involvement of a third reviewer. All studies retrieved from literature searches were screened for relevance based on title and abstract according to the above inclusion criteria. Where relevance could not be determined based on abstract and title, the full text was reviewed. Manual searches were also completed of reference lists of all relevant studies. Bibliographic data, study location and timing, population age group(s), sample size, cause of death (including the International Classification of Diseases (ICD) code(s) if available), method of exposure assessment, pollutant (including name, units, descriptive statistics), type of regression model, effect measure and standard error or confidence interval, model covariates (potential confounders) and their specification were extracted from all studies meeting inclusion criteria. When single pollutant results were presented for multiple exposure periods, we extracted the most highly statistically significant result (regardless of the direction of the association), or that reported by the authors as their primary finding. Results from multi-pollutant models that resulted in the greatest reduction in magnitude of effect compared to single pollutant results were selected in order to bracket the magnitude of effect from each study. Our objective in this regard was not to assess the magnitude of the association with NO₂ in the presence of a single co-pollutant common to all studies, but rather to determine the maximum potential for confounding of the association of mortality with NO₂ by any co-pollutant(s). Results expressed per pollutant increment expressed in μg/m³ were converted to parts per billion [23], and those based on nitrogen oxides (NO_x) were converted by multiplying the log(HR) by 2.31 (the average ratio of log(HR) based on NO₂ to log(HR) based on NO_x in four studies based on 25 cohorts [24–27]). Where required data were not provided, authors were contacted by e-mail. In some instances Engauge Digitizer [28] was employed to extract numeric results presented only in graph form. The Navigation Guide systematic review methodology [29] was employed to evaluate risk of bias at the study level according to the following domains: selection bias, exposure assessment, confounding, outcome assessment, completeness of outcome data, selective outcome reporting, conflict of interest (including sources of funding) and other sources of bias. Criteria are detailed in S1 Fig. In the confounding domain, we considered age, sex, smoking and socioeconomic status as critical potential confounders. As a sensitivity analysis, we treated body mass index (BMI) rather than smoking as a critical confounder, as specified in the World Health Organization (WHO) risk of bias criteria [30]. Assessment of risk of bias was completed independently by two reviewers and conflicts between reviewers were resolved through consensus and/or involvement of a third reviewer.

Data analysis

Pooling of HRs across studies was conducted using random effects models computed using Restricted Maximum Likelihood (REML) estimation, with sensitivity analyses employing Dersimonian and Laird and Empirical Bayes estimators [31]. Heterogeneity among included studies was assessed using Cochran’s Q and I² measures, and sources of heterogeneity were evaluated using meta-regression [31]. Sensitivity of pooled estimates to individual studies was examined using Leave One Out analysis and publication bias was evaluated using Funnel plots, Begg’s and Egger’s tests, and trim and fill [31]. Subgroup analyses were conducted by cause of death, region, risk of bias characterization (pre-specified in protocol) and single vs. multi-pollutant models. Analysis was conducted in R version 3.6.0 [32] using the metafor package [31].

Overall rating of quality and strength of evidence

We applied the Navigation Guide methodology [33] and the causality determination framework used by the US EPA and Health Canada [12] (S2 and S3 Tables) to judge the overall quality and strength of the evidence, and likelihood of a causal relationship. Following the Navigation Guide methodology, given the observational nature of the evidence, the starting point for rating overall quality was “moderate,” which was downgraded or upgraded to “low” or “high” based on the criteria summarized in S2 Table [33]. Identification of multiple upgrading or downgrading factors does not necessarily mean that overall quality is upgraded or downgraded multiple levels. Rather, the overall degree of upgrading or downgrading is based on reviewer judgement of the overall importance and impact of all factors considered [33]. The Navigation Guide characterizes strength of evidence of toxicity as “sufficient”, “limited”, “inadequate” or “indicative of a lack of toxicity” (S2 Table), based on the overall quality of the evidence, the direction of effect, confidence in the effect and any other factors identified as germane by the reviewers [33]. Given the parallels with the USEPA/Health Canada causality determination criteria (S3 Table), while we did not conduct a systematic review of other lines of evidence, we drew upon summaries of other lines of evidence from a recent assessment document [11], supplemented by findings from more recent mechanistic studies, in order to characterize the likelihood of a causal relationship.

Effect estimates and pooled effect estimates

All effect estimates from individual studies, including from single and multi-pollutant models, and by population and outcome subgroup are provided in forest plots by region in S3 Fig. Of these, we excluded estimates from pooling if they were superseded by other studies encompassing the same geographic area or time period, e.g. in subsequent multi-city studies or those spanning a longer study duration, results were provided only from multi-pollutant models, or there were too few studies of the outcome, leaving 53 studies included in the meta-analysis (see S4 Table). Pooled estimates based on single pollutant model results by cause of death are summarized in Table 1. Pooled estimates were largest for respiratory and cerebrovascular causes of death, similar for cardiovascular and ischemic heart disease, and smallest for all/natural cause and lung cancer. All pooled estimates were characterized by a high degree of heterogeneity.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Pooled estimates by cause of death and sensitivity analyses.

https://doi.org/10.1371/journal.pone.0246451.t001

Meta-regression and sensitivity analysis

Meta-regression revealed that the magnitude of the log HR for cardiovascular, respiratory and ischemic heart disease mortality was significantly larger for studies conducted outside North America and Europe (p = 0.002, 0.025, 0.021 respectively). For lung cancer mortality, standard deviation (p = 0.0009) and range (p = 0.006) of NO₂ exposure were significantly positively associated with log HR. Probably high or high risk of bias in the confounding domain was significantly positively associated with log HR for all/ natural cause (p = 0.040), respiratory (p = 0.006) and cerebrovascular (p = 0.036) mortality. In a sensitivity analysis, treating BMI rather than smoking as a critical potential confounder, as specified in the WHO risk of bias criteria, probably high or high risk of bias in the confounding domain was not associated with log HR. Risk of bias in the exposure assessment domain, study interquartile range NO₂ and timing of study primarily before or after 2000 were not significant predictors of the log HR. Residual heterogeneity remained high (I² >80%) even after accounting for significant predictor variables.

Based on meta-regression findings, studies with probably high or high risk of bias in the confounding domain (n = 12) were excluded, resulting in substantially smaller pooled HRs for cardiovascular, respiratory and cerebrovascular mortality (Table 1). Pooled estimates were significantly smaller for cerebrovascular mortality (p = 0.031) and significantly larger for ischemic heart disease mortality (p = 0.002), than for all/ natural cause mortality. Fig 2 presents a forest plot of HRs from individual studies and pooled estimates for all/ natural cause mortality, after exclusion of studies with probably high or high risk of bias in the confounding domain. The pooled estimate for North America was smaller than those for Europe and other countries, and heterogeneity among primary studies was lower for North American studies. Forest plots for other causes of death are provided in S4 Fig. Applying the WHO risk of bias criteria (treating BMI rather than smoking as a critical potential confounder) resulted in a somewhat larger pooled estimate for all/natural cause mortality, considerably larger pooled estimates for cardiovascular and cerebrovascular mortality, together with much greater heterogeneity for the latter, and a somewhat smaller pooled estimate with narrower confidence intervals and considerably less heterogeneity for respiratory mortality (Table 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Hazard ratios from single pollutant models from individual cohort studies and pooled estimates by region, for all/ natural cause mortality (AC/NC).

https://doi.org/10.1371/journal.pone.0246451.g002

Results were generally insensitive to the exclusion of studies employing NO_x as the exposure metric [45–47], with the exception of all/natural cause mortality, for which the pooled estimate was somewhat smaller (Table 1). Findings were also insensitive to excluding cohorts of individuals with pre-existing disease [36, 39, 41, 42] (pooled estimate of HR for all/natural cause mortality 1.050, 95% CI 1.022–1.078). Pooled estimates were not sensitive to employing alternative estimators except that confidence intervals tended to be slightly narrower based on Dersimonian and Laird, and slightly wider based on Empirical Bayes (S1 Text). In leave one out analyses, heterogeneity among respiratory and ischemic heart disease studies was sensitive to exclusion of results from Katanoda et al. [48] and Turner et al. [49] respectively (S6 Table). Egger’s test revealed significant funnel plot asymmetry for all/ natural cause (p = 0.0022) and lung cancer (p = 0.031) mortality, while Begg’s test indicated significant funnel plot asymmetry only for lung cancer mortality (p = 0.022). Trim and fill analysis estimated that there were 4 missing studies with log HR below the pooled estimate for lung cancer mortality (S5 Fig), reducing the pooled estimate slightly to 1.066 (95%CI 1.024–1.110).

Effects of co-pollutants, noise and green space

A forest plot of nine paired estimates of associations with all/ natural cause mortality from single and multi-pollutant models from the same study, after exclusion of studies with probably high or high risk of bias in the confounding domain, is shown in Fig 3. The pooled estimate from single pollutant models was higher than that from multi-pollutant models with 1–4 co-pollutants, and the confidence interval for the multi-pollutant pooled estimate included 1, indicating no association. However, the difference between pooled estimates for single and multi-pollutant models was not significant (p = 0.13). Fewer paired estimates from single and multi-pollutant models were available for other causes of death. Pooled estimates from multi-pollutant models were consistently smaller in magnitude than those from single pollutant models, and the pooled estimates based on multi-pollutant models included 1, indicating no association, with the exception of ischemic heart disease mortality which was based on only three studies (Table 2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Hazard ratios from individual cohort studies and pooled estimates from single and multi-pollutant models for all/ natural cause mortality (AC/NC).

https://doi.org/10.1371/journal.pone.0246451.g003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Pooled estimates from single and multi-pollutant models by cause of death.

https://doi.org/10.1371/journal.pone.0246451.t002

Few studies have jointly examined effects of NO₂ and noise. Sørensen et al. found that NO₂ exhibited a significant positive association with fatal stroke when modelled jointly with traffic noise [50], while in the same cohort, Hvidtfeldt et al. found that NO₂ remained significantly associated with natural cause mortality after adjustment for traffic noise, although the magnitude of the association was reduced, and NO₂ was no longer significantly associated with cardiovascular mortality after adjustment for traffic noise [51]. Héritier et al. [52] also reported that the association of NO₂ with myocardial infarction mortality was no longer significant after adjustment for traffic, rail and aircraft noise. Klompmaker et al. found no association between NO₂, PM_2.5, greenness or traffic noise and mortality [53], and Tonne et al. observed a negligible change in the association between NO₂ and mortality with additional adjustment for traffic noise [40]. Nieuwenhuijsen et al. [54] found that the association of NO₂ with mortality was reduced in magnitude in models with noise and/or green space.

Shape of exposure-response relationship

Twenty-nine studies evaluated the shape of the exposure-response relationship between NO₂ or NO_x and mortality by examining the association by pollutant quantile [24, 41, 44, 47, 55–57], plotting the association using a non-linear function of pollutant concentration and/or testing the significance of the difference between linear and non-linear models [5, 25, 27, 45, 46, 50, 51, 53, 57–69], or plotting associations by city [48]. Of these, 19 studies found a linear association [27, 41, 44, 50, 51, 53, 55–57, 60, 63–65, 67–69], in some instances only in subsets of the data by cause of death [5, 45, 46], and five studies concluded that the association was supralinear [25, 45, 47, 59, 66]. Three found evidence of a threshold [48, 61, 62], but only one study identified specific numeric values, ranging from 20–40 μg/m³, depending on age and cause of death [62]. Two studies reported no association, based on analysis by quantiles [24], and using a non-linear function of NO₂ [58]. While there was some variation in findings by cause of death in individual studies, overall there was evidence of an exposure-response relationship across all six causes of death.

Other lines of evidence

The 2016 US Environmental Protection Agency (EPA) Integrated Science Assessment on oxides of nitrogen concluded that toxicological evidence suggested there were several possible mechanisms through which long term exposure to NO₂ could contribute to adverse respiratory and cardiovascular effects, including pulmonary inflammation, oxidative stress, pulmonary injury, changes in lung morphology, impaired respiratory immune defences, atherosclerosis, autonomic dysfunction, and changes in lipid metabolism [11]. Examples of findings from relevant studies include the modulation of alveolar macrophage activity in rats in association with NO₂ exposure [74], and increased mortality from respiratory infection in mice in two older studies of NO₂ exposure [75, 76]. With respect to cardiovascular effects, in one study, long term exposure to NO₂ was associated with reduced heart rate variability, but only in women [77]. In a Spanish cross sectional study, long term NO₂ exposure was associated with increased blood pressure [78] and subclinical atherosclerosis [79]. A more recent study in Germany found that those with higher long term NO₂ exposure had significantly higher odds of elevated ankle-brachial index, reflective of arterial stiffening, although the magnitude of the association was reduced in a joint model with PM_2.5 absorbance (a proxy for elemental carbon) [80]. Overall, however, the evidence base was considered limited, associations were inconsistent, and it was difficult to separate effects of NO₂ from those of co-pollutants [11]. With respect to short term exposure to NO₂ and mortality, a recent systematic review concluded that there was a high degree of certainty of the evidence linking 24 hour average exposure and mortality [81]. While there may be a degree of overlap in the effects captured by studies of short term and long term exposure, the overall health impact captured by each type of design is not identical [82], and certainty of evidence regarding effects of short term exposure does not necessarily imply certainty regarding effects of long term exposure.

Overall rating of quality and strength of evidence

In their 2016 Science Assessments, both the US EPA and Health Canada concluded that the evidence was suggestive of, but not sufficient to infer, a causal relationship between long term NO₂ exposure and mortality, based on a smaller number of studies, and fewer studies examining the impact of adjustment for co-pollutants than considered here, as well as limited and inconsistent supporting mechanistic evidence from human and animal studies [11, 12]. Applying the Navigation Guide methodology [33] and the causality determination framework used by the US EPA and Health Canada [12] to our current findings, several factors are considered downgrading factors in interpreting the overall quality of evidence. These include the significant heterogeneity among studies even after accounting for sources of heterogeneity, and the relatively large proportion of studies rated as probably high or high risk of exposure assessment bias (57.1%) (even though presence of this bias was not associated with magnitude of association in meta-regression). There was also evidence of publication bias, particularly for lung cancer mortality, but it did not have a substantial impact on pooled estimates of association. Risk of bias from residual confounding was evaluated both in relation to inclusion of critical potential confounders in statistical models, as well as impacts of co-pollutants and other co-exposures on the magnitude of associations. Pooled estimates indicated that NO₂ remained significantly associated with all/natural cause, cardiovascular, lung cancer, respiratory and ischemic heart disease mortality after exclusion of 12 studies with probably high or high risk of bias in the confounding domain. However, after excluding these studies, only 9 studies of all/natural cause mortality provided estimates based on both single and multi-pollutant models, and the pooled estimate indicated that NO₂ was no longer significantly associated with mortality after adjusting for co-pollutants. Fewer paired estimates from single and multi-pollutant models were available for other causes of death. Pooled estimates from multi-pollutant models were consistently smaller in magnitude than those from single pollutant models, and the pooled estimates based on multi-pollutant models included 1, indicating no association, with the exception of ischemic heart disease mortality, which was based on only three studies. Multi-pollutant models should be interpreted with caution in that the sensitivity of the effect of one pollutant to inclusion of other pollutants in a joint model is affected by factors such as the correlation among pollutants and their relative degree of exposure measurement error [83]. There is nonetheless evidence of confounding by co-pollutants of the association of long term NO₂ exposure with mortality. Few studies jointly modelled NO₂ with traffic noise. In a recent review, Tétrault et al. concluded that cardiovascular effects of long term air pollution exposure were probably independent of noise, but this was based on only nine studies, including only one study of air pollution and mortality [15]. One study found that that the association of NO₂ with mortality was reduced in magnitude in models with both noise and/or green space [54]. Specifically with respect to cerebrovascular mortality, imprecision is also considered a downgrading factor, as the 95% CI on the pooled estimate overlapped the null. In contrast to these downgrading factors, characterization of the exposure-response relationship as linear, supralinear, or linear with a threshold in 27 of the 29 studies in which this was evaluated, is considered an upgrading factor applicable to all six causes of death. Huangfu and Atkinson also downgraded the evidence in relation to heterogeneity for all cause and respiratory mortality, and upgraded it in relation to evidence of an exposure-response relationship [20]. Based on these considerations, we conclude that for all causes of death other than cerebrovascular disease, the overall quality of the evidence from cohort studies is moderate, the strength of evidence of toxicity is limited, and the overall evidence continues to be suggestive of, but not sufficient to infer, a causal relationship between long term NO₂ exposure and mortality. In the case of cerebrovascular disease mortality, owing to the smaller number of primary studies and the non-significant smaller magnitude association based on the pooled estimate, we conclude that the overall quality of the evidence from cohort studies is low, the strength of evidence is inadequate, and the overall evidence is inadequate to infer a causal relationship. Upgrading to a conclusion that there is sufficient evidence for a causal relationship would require more conclusive evidence ruling out potential confounders as well as consistent supporting animal toxicological and human clinical evidence. Future studies could address uncertainties related to confounding by co-pollutants by more consistently examining their correlations and effects in multi-pollutant models, in particular adjusting for other traffic-related pollutants and concomitant exposures like noise, green space and stress. Only about one third of the studies we reviewed addressed the shape of the concentration-response relationship, therefore examination of this issue in future studies would also be informative. While the evidence reviewed does not support the unequivocal conclusion that long term exposure to outdoor NO₂ causes an increased risk of death, identifying the true causal agent is of major importance to public health. Vehicle emissions are one of the main sources of NO₂, but vehicle emissions, and secondary pollutants arising from vehicle emissions also include numerous other potentially toxic pollutants such as carbon monoxide, particulate matter, benzene, formaldehyde, acetaldehyde, 1,3-butadiene, ozone, nitrates and organic and inorganic acids [10]. If the true causal agent is not NO₂, control measures which specifically reduce NO₂ will not reduce mortality risks. Conversely, identification and control of the true causal agent will have considerable public health benefits.