Short-Term Associations of Nitrogen Dioxide (NO 2 ) on Mortality in 18 French Cities, 2010–2014

: We present an analysis of short-term associations between ambient NO 2 and mortality according to cause, age-group, and period (cold and warm) in 18 areas in metropolitan France for the 2010–2014 period. Associations were estimated in each area using a generalized additive Poisson regression model, and effects were summarized in a meta-analysis. The percentage increase in mortality rate was estimated for a 10 µ g m − 3 increase in the NO 2 level in each area for each complete calendar year and for cold (November to April) and warm periods (May to October) in each year. We found that the NO 2 increase (lag of 0–1 days) was associated with a 0.75% increase of non-accidental mortality for all age-groups (95% confidence interval (CI): (0.4; 1.10)). During the warm period, this NO 2 increase was associated with a 3.07% increase in non-accidental mortality in the ≥ 75 years old group (95% CI: 1.97; 4.18). This study supports the short-term effects of NO 2 as a proxy of urban traffic pollution on mortality, even for concentrations below the maximum guideline of 40 µ g m − 3 set down by the European Air Quality Standards and the World Health Organization (WHO).


Introduction
Nitrogen dioxide (NO 2 ) is an air pollutant resulting from combustion processes and is mainly emitted by road traffic. In France, annual urban background and near-road traffic concentrations of NO 2 in ambient air have decreased since 2000. However, the maximum concentration set down by European air quality standards for the protection of human health (40 µg m −3 ) is exceeded every year in several urban areas (near-traffic), and near-traffic levels are twice as high as urban background levels. In 2016, the annual average urban background NO 2 concentration was 19 µg m −3 , compared with 39 µg m −3 for near-road traffic. Furthermore, numerous episodes of NO 2 pollution are observed every year during the cold period (November to April) because anticyclonic conditions with cold temperatures limit the dispersion of pollutants [1].
Several epidemiological studies have shown short-term associations between exposure to NO 2 and various health outcomes, in particular respiratory and cardiovascular diseases and mortality [2][3][4]. More specifically, in 2008, associations were found between ambient NO 2 concentrations and mortality in nine areas in France [5]. Moreover, in 2013, epidemiological and toxicological evidence in the Review of evidence on health aspects of air pollution (REVIHAAP) report [3] suggested a short-term causal relationship, particularly for respiratory effects, which prompted the World Health Organization (WHO) to conclude that it was reasonable to consider that NO 2 had direct effects on health. The WHO also considered that there was sufficient evidence to warrant an estimation of the short-term health impact of NO 2 on all-cause mortality and hospitalizations for respiratory diseases. In 2016, the United States Environmental Protection Agency (US-EPA) considered that there was robust evidence for an In 2016, the United States Environmental Protection Agency (US-EPA) considered that there was robust evidence for an association between short-term NO2 exposure and exacerbation of asthma, and that the findings from recent studies also suggested an effect on cardiovascular health and mortality [6].
In order to estimate the short-term health impact of NO2, concentration-response functions (CRFs) are necessary. However, available CRFs for NO2 in Europe have primarily been derived from data from studies conducted outside the continent. The handful of European-based studies providing CRFs include the APHEA [7] and EpiAir [8] projects in Italy and studies in Switzerland [9] and Austria [10]. Moreover, due to having a higher proportion of diesel-fueled vehicles than elsewhere in the world, the specificity of vehicle emissions in France (and indeed Europe) most probably influences the relationship between NO2 and health effects. This was supported by the meta-analysis by Mills et al. in 2015 [2], which identified strong geographical heterogeneity between the effects estimated for different regions of the world.
All the above elements highlight the relevance of studying the short-term links between exposure to ambient NO2 and mortality in France in order to update existing CRFs with a view to using them in future health impact assessments (HIA).
Accordingly, in the present study, we conducted a retrospective time series analysis to evaluate the association between NO2 as one proxy of urban traffic pollution and the risk of daily nonaccidental and cardiovascular mortality in 18 cities in metropolitan France over the period 2010-2014.

Study Period and Study Area
We conducted a daily time-series study of the relationship between NO2 exposure and causespecific mortality in 18 French areas (Figure 1) (accounting for 499 municipalities) from 2010 to 2014 based on the availability of mortality data.   For each area, daily mortality data were obtained from the French National Institute of Health and Medical Research (CépiDc) for the study period. Causes of death were coded according to the International Classification of Diseases, 10th Revision (ICD-10). More specifically, data collected included daily counts of total non-accidental (ICD-10: A00-R99), cardiovascular (ICD-10: I00-I99), and respiratory (ICD-10: J00-J99) mortality for all ages. We dichotomized data by age group (≥75 years) to compare with similar studies.

Pollution Data
In each study area, pollution data were obtained from local air quality monitoring networks. NO 2 was measured using the chemiluminescence method, a tapered element oscillating monitor filter dynamic measurement system (TEOM-FDMS) was used for PM 10 , while ozone (O 3 ) was measured using the standard reference ultraviolet absorption method. Urban background monitors were used to create daily indicators of NO 2 and PM 10 exposure, while urban and suburban background monitors were used to create daily indicators of O 3 . Depending on the area, between 1 and 15 monitors were available for NO 2 , one to seven for PM 10 , and 1 to 13 for O 3 (Table 1). All study areas were constructed in such a way that the average exposure level of the population could be accurately estimated from the monitoring stations [11]. We built homogeneous study areas because, in time-series studies, the principle of analysis assumes that, on average, all individuals in the population are exposed to the same levels of air pollution each day.

Meteorological Data
We obtained daily minimum, mean, and maximum temperature data from the French national meteorological service (Météo-France), with one reference monitoring station per area studied. Airport stations were used for 15 areas and city center stations were used for the other 3 areas.

Statistical Analyses
The association between NO 2 and daily mortality was investigated using a time-series analysis with generalized additive Poisson regression models (GAM) allowing for overdispersion [12].
In each area, the daily death count was regressed on NO 2 levels, while the following possible confounders were controlled: temperature, long-term trends, seasonal trends, day of the week, and bank holiday effects. Seasonality was taken into account using a penalized spline function. The smoothing parameter of the spline was selected in order to minimize the absolute value of the partial autocorrelations of the residuals over 30 days by imposing a minimum of 3-degrees of freedom per year [13]. The mean levels of NO 2 observed during the current and previous days (i.e., lag 0-1 days) were introduced as a linear term.
Minimum and maximum temperatures were introduced into the model as a natural spline with 3-degrees of freedom. Daily minimum temperatures at lag 0 and daily maximum temperature at lags 1-7 were also introduced as natural splines with 3-degrees of freedom. We assumed that minimum temperatures at lag 0 would represent a possible short-term heat effect, while maximum temperatures at lags 1-7 would represent a possible delayed cold effect [14,15].
The regression equation is formulated below: where Yt denotes the daily mortality counts on day t, dow is the day of the week, hol is a bank holiday, pol is the mean level of NO 2 observed during the current and previous days, and δ1, δ2, and δ3 are the corresponding regression coefficients. s(time) is a penalized spline function of calendar time designed to control for trend and seasonality, ns(tmin) and ns(tmax) are natural splines of temperature designed to control weather, tmin is the daily minimum temperatures at lag 0, and tmax is the daily maximum temperature at lags 1-7.
For each area, we tested the assumptions of the model using the following set of graphical and statistical tools: (1) The observation of the residual plot helped verify that, after modeling, no particular structure (trend, seasonality) persisted; (2) The residuals were assimilated to white Gaussian noise. Residual autocorrelation was particularly influenced by the choice of smoothing parameter of the spline function modeling seasonality; (3) The normality of the residuals was checked graphically (histogram, QQPLOT); (4) The statistical significance of the partial autocorrelations was verified graphically on the partial correlogram; (5) The comparison of the graphs of the observed and predicted values allowed the quality of the model to be assessed; (6) The partial effect of each factor on the health variable was shown graphically. The graphs helped check the coherence of the effects (Appendix A-Figures A1 and A2: example for Paris and Montpellier).
Air pollution effects were estimated for each year and each season (warm period: May to October; cold period: November to April) for each cause of mortality. An interaction between the pollution indicator and the season was added to the models to study seasonal effects on pollution.
To investigate the confounding of NO 2 measurement by PM 10 and O 3 , we fitted two-pollutant models. We also fitted polynomial distributed lag models over a period of 6 days (0-5 days) to investigate the shape of each lagged association [16].
In the second stage, we pooled area-specific estimates to perform a meta-analysis using random-effect models [17]. We tested for heterogeneity and reported it using the I 2 statistic [18]. Association estimates were expressed as excess relative risk (ERR) of mortality for a 10 µg m −3 increase in air pollutant concentrations.
The time-series analysis was conducted using the R package mgcv (which provided generalized additive modeling functions), dlnm (which contained functions to specify and interpret distributed lag linear and non-linear models), and mvmeta (this function performed fixed and random-effects multivariate and univariate meta-analysis and meat regressions) (version 3.2.3, http://cran.r-project.org/).  Table 2 summarizes population and mortality data for each area included in the analysis. The 18 areas covered a combined total of over 15 million inhabitants. The largest area was Paris, with more than 6.7 million inhabitants. With 8861 inhabitants per km 2 , the population density was two to eight times higher than in the other areas (Appendix B- Table A1). Lille had the second largest population, but a density below 1900 inhabitants per km 2 , while Nice had one of the densest populations with more than 3600 inhabitants per km 2 , despite its mid-size population (see Appendix B- Table A1).

Mortality Data
In the 18 French areas studied, the daily mean number of non-accidental deaths ranged from 4% in Rennes (one of the smallest areas) to 104.5% in Paris, the biggest area ( Table 2). The percentage of non-accidental deaths in people aged 75 and over ranged from 62% in Lens-Douai, Le Havre, and Lille to 74% in Nice. The daily mean number of deaths for cardiovascular diseases ranged from 1.2% in Dijon to 23.8% in Paris. For respiratory mortality, the very small numbers of deaths prevented us from being able to perform statistical analyses of the links between air pollution and death with sufficient statistical power. Respiratory mortality was, therefore, not considered in this study.

Pollution Data
Mean levels of air pollution indicators (NO 2 , PM 10 ) over the study period in each area are presented in Table 1.
There were large variations in NO 2 levels across the 18 areas over the study period. The highest mean levels (36.0 µg m −3 in Paris and 32.7 µg m −3 in Marseille) were approximately twice as high as the lowest (17.9 µg m −3 in Nantes and 18.7 µg m −3 in Rennes). NO 2 levels were higher during the cold period (November-April) than the warm period (May-October). During the cold period, mean levels ranged from 23 µg m −3 in Nantes to 42 µg m −3 in Paris. During the warm period, they ranged from 13 µg m −3 in Nantes to 30 µg m −3 in Paris.

Temperature Data
Between 2010 and 2014, the minimum mean annual temperatures ranged from 6.5 (Nancy) to 13.1 • C (Nice), and maximum mean annual temperatures ranged from 14.1 (Le Havre) to 20.7 • C (Marseille). Mean annual temperatures ranged from 5.6 (Nancy) to 11.5 • C (Nice) during the cold period and from 15.4 (Rouen) to 21.4 • C (Marseille) during the warm period (Table 3). Temperature was taken into account in the analysis as a confounding factor.

Concentration-Response Functions
3.2.1. NO 2 Analysis of the Same Day and the Day before (lag 0-1) Table 4 presents, for each full calendar year and for each period (i.e., warm and cold), the ERR estimates for non-accidental and cardiovascular deaths for a 10 µg m −3 increase in NO 2 levels on the same day and the day before (lag 0-1). This table also presents sensitivity analyses with two-pollutant models for PM 10 and O 3 .
We found a significant link between NO 2 levels and non-accidental and cardiovascular deaths for each full calendar year and for the warm period each year, irrespective of the age group considered. For those aged 75 years and over, ERR estimates were slightly higher than for the total population. This difference was more marked for non-accidental mortality: 1.14% (95% confidence interval (CI): (0.63; 1.66)) and 0.75% (95% CI: (0.40; 1.10)), respectively. No age difference was observed in the ERR for cardiovascular deaths: 1.15% (95% CI: (0.40; 1.91)) and 1.13% (95% CI: (0.37; 1.90)), respectively. Table 4. Excess relative risk (ERR) and 95% confidence intervals (95% CI) of non-accidental and cardiovascular deaths associated with a 10 µg m −3 increase of NO 2 concentrations (lag 0-1 day) according to age (total population versus 75 years and over) using the single-and two-pollutant models ( All ERR values were also higher for the warm period than for the whole calendar year, irrespective of the cause of death or age group. The highest ERR was 3.17% (95% CI: (1.53; 4.84)) for cardiovascular deaths among those aged 75 years and over in the warm period. In the cold period, there was no association.
After adjusting the models for PM 10 , the ERR estimates remained unchanged in the annual period and in the cold period and were lower in the warm period. The ERR estimates also remained unchanged after adjusting the model for O 3 . Figure 2 shows the distribution of the ERR for non-accidental and cardiovascular deaths over 0 to 5 days measured separately (lag 0, lag 1, lag 5) and 0 to 5 cumulative days (lag 0-1, lag 0-2, lag 0-5) for a 10 µg m −3 increase in NO 2 . In cumulative lags, 0-5 effects were cumulated over 0 to 5 days (i.e., all contributions from day 0-5 were summed). An increase in ERR values was observed for the cumulative lags, irrespective of the cause of death or age group. The point estimate for the ERR for the cumulative lags for non-accidental deaths was slightly higher for those aged 75 years and over than for the total population. More specifically, the ERR for the cumulative 0-5 lags was 2.35% (95% CI: (1.35; 3.37)) for those aged 75 years and over and 1.74% (95% CI: (1; 1.48)) for the total population.  Figure 2 shows the distribution of the ERR for non-accidental and cardiovascular deaths over 0 to 5 days measured separately (lag 0, lag 1, lag 5) and 0 to 5 cumulative days (lag 0-1, lag 0-2, lag 0-5) for a 10 µg m −3 increase in NO2. In cumulative lags, 0-5 effects were cumulated over 0 to 5 days (i.e., all contributions from day 0-5 were summed). An increase in ERR values was observed for the cumulative lags, irrespective of the cause of death or age group. The point estimate for the ERR for the cumulative lags for non-accidental deaths was slightly higher for those aged 75 years and over than for the total population. More specifically, the ERR for the cumulative 0-5 lags was 2.35% (95% CI: (1.35; 3.37)) for those aged 75 years and over and 1.74% (95% CI: (1; 1.48)) for the total population.

Discussion
In this study, we found an increase in the numbers of non-accidental and cardiovascular deaths associated with a 10 µg m −3 increase in NO2 for both the total population and people aged 75 years

Discussion
In this study, we found an increase in the numbers of non-accidental and cardiovascular deaths associated with a 10 µg m −3 increase in NO 2 for both the total population and people aged 75 years and over, for the entire study period (2010-2014), for warm and cold periods, and for the warm period alone. These study results complement the ERR values published by Blanchard et al. in 2008 [5] as they include new urban agglomerations that were integrated into the French national public health agency's air and health program in 2011. Indeed, the addition of these areas with different profiles (population, pollution, and climate), helped us increase the statistical power of this meta-analysis.
The deleterious effects on health of NO 2 have been examined by the US EPA Integrated Science Assessment report on the oxides of nitrogen. The report states that the cumulative body of evidence on NO 2 indicates that short-term exposure can affect health, in particular, effects related to asthma exacerbation, and may be associated with cardiovascular effects and premature mortality. However, the report states that to assign a causal effect of NO 2 on mortality requires more research to separate NO 2 exposure effects from those of other traffic-related pollutants [6]. The REVIHAAP report by WHO [3] also states that the observed associations in epidemiological studies may not be completely attributable to NO 2 per se, as NO 2 may also include other constituents (which have adverse health effects) not represented by currently regulated pollutants. This is why NO 2 is used as one possible tracer, indicator, or proxy of traffic air pollution among others in this study.
Using a single-pollutant (such as NO 2 ) as a proxy for complex mixtures of pollutants, especially in urban areas, often misrepresents the real pollution mixture, as well as the associated health outcomes (which may depend upon a different component of the mixture). In this study, we were not able to distinguish which components of the mixture represented by NO 2 were responsible for the observed effects. NO 2 is typically formed through the following photochemical process. Due to ultraviolet radiation, NO 2 dissociates into nitrogen oxide (NO) and atomic oxygen (O). Atomic oxygen (O) combines with dioxygen (O 2 ) very quickly and produces the main photochemical pollutant, ozone (O 3 ). This reaction occurs in the low layers of the atmosphere and favors sunshine and warm temperatures. Ozone can also react with NO to regenerate NO 2 (https://www.aeroqual.com/meet-the-nitrogenoxide-family). Relative toxicity of air pollution mixtures has been widely studied [19], and the effects of photochemistry on the toxicity of air pollution are well documented [20]. It is known that photochemical reactions can release secondary organic aerosols that increase inflammatory responses, and this phenomenon enhances the toxicity of air pollution [21,22].
The potential causal role of NO 2 and the pathophysiological mechanisms by which NO 2 may affect health have been studied. As per the REVIHAAP report [3], there is evidence of small effects on inflammation and increased airway hyperresponsiveness with NO 2 per se in the range from 0.2 to 1 ppm (380 to 1880 µg m −3 ) from chamber studies (under a broad range of exposure conditions, with exposure durations of 15 min to 6 h, with some inconsistency in results), with more marked, consistent, responses observed from 1 ppm (1880 µg m −3 ). Newer review reports suggest weak to moderate lung cell changes in animals at one-hour concentrations of 0.2 to 0.8 ppm (380-1500 µg m −3 ). These concentration ranges are not far from concentrations that occur at the roadside or in traffic for multiple hours. The WHO report concludes that there is some mechanistic support for causality, particularly for respiratory outcomes. This is confirmed by Petit et al. [23], who state that NO 2 is a toxic gas that can damage the lungs. When inhaled, it oxidizes protective antioxidants within the epithelial lining fluid and triggers extracellular damage in the airways. The presence of NO 2 within the epithelial lining fluid triggers oxidative stress, possibly leading to edema, bronchoconstriction, and a reduced forced expiratory volume in 1 s.
Despite results in international studies showing the associations of NO 2 with respiratory mortality, we were not able to carry out an analysis of this indicator because of the low number of deaths attributed to respiratory mortality in the study areas.
The present study showed that NO 2 exposure was associated with an increased number of non-accidental and cardiovascular deaths. Excess relative risks estimated in the present study for a 10 µg m − 3 increase in NO 2 were lower than those published in the previous meta-analysis in 2008, although the comparison was limited by the different statistical methods used in both studies, and the higher number of areas included in this work (18 vs 9) [5].
The study results for the general population were consistent with those in the literature and especially with those from multicenter and meta-analysis studies. Two similar studies, the first conducted in 2015 by Perez et al. in 21 cantons in Switzerland [9] and the second in 2016 by Carugno et al. in 18 areas of the Lombardy region of Italy [24], found a significant increase in all-cause deaths for a 10 µg m −3 increase in NO 2 concentrations. The results of a meta-analysis of several studies performed throughout the world, conducted by Mills et al. in 2015, showed a similar risk, and when focusing only on European studies, the ERR was slightly higher [2]. Another study carried out in 2015 by Renzi et al. in the Rome area showed a higher risk than in the present study [25]. A study of the Vienna area by Neuberger et al. found a similar ERR to this study [10], and finally, a Canadian study by Crouse et al. showed a significantly lower risk than the present study [26] (Table 5). Table 5. Excess relative risk (ERR) and 95% confidence intervals (95% CI) of non-accidental and cardiovascular deaths associated with a 10 µg m −3 increase of NO 2 concentrations (lag 0-1 day). Most international studies to date also examined the relationship between cardiovascular mortality and NO 2 concentrations but found a slightly lower excess risk than in this study (1.13% [0.37-1.90]). In Europe, the meta-analysis results of Mills et al. in 2015 [2] and the Carugno et al. study [24] both showed a significant increase in cardiovascular mortality for a 10 µg m −3 increase in NO 2 , with an ERR close to this study. In contrast, the Swiss study by Perez et al. in 2015 [9], the worldwide study by Mills et al. in 2015 [2], and the Canadian study by Crouse et al. in 2015 [26] showed a significantly lower risk than the present study (Table 5).

Non-Accidental Deaths
We observed slightly higher ERR in people aged 75 and over than in the total population for non-accidental (1.14% vs 0.75%) and cardiovascular (1.15% vs 1.13%) deaths. Chinese studies have shown higher risks among people aged 75 and over, although age-related differences were negligible [27,28]. The Italian study by Carugno et al. in 2016 showed a higher risk for cardiovascular deaths but not for non-accidental deaths [24] in the same group. However, other studies have shown a lower risk for this sub-population [9] (Table 5).
NO 2 associations were more evident in the warm period (May-October). Similar results have been observed in studies in several countries, for example, in Italy [24,29], South Africa [30], Canada [31], and China [28]. One likely explanation for this is that during the warm period, the concentrations measured by the monitoring stations better represent real exposure as people spend more time outdoors in the warm period and keep windows open, leading to ambient air pollutants entering their homes. Another hypothesis is that during the warm season, the increase in photochemical activity leads to a change in the composition of the mixture in automobile emissions, with more volatile organic compounds (VOC) and finer particles than in the cold period. As these pollutants have been identified as having health effects [3,30], the mixture of pollutants from automobile emissions, with NO 2 as a proxy, would potentially have a larger health effect in the warm season.
In order to dissociate the effect of NO 2 from that of other pollutants, in particular from traffic, we adjusted models for PM 10 and O 3 . These were the only pollutants we could adjust for, despite PM 2.5 , SO 2 , and CO also being analyzed in other articles. In France, SO 2 is not measured anymore in urban areas because concentrations are extremely low (except for hotspots), and for CO, we do not have continuous measurements. PM 2.5 concentrations were not available in all the study areas, and, in most cases, data was not yet consolidated and came from only one monitoring station for all areas. It takes time for emitted particles to agglomerate and/or grow via uptake of secondary pollutants before they reach the size of PM 10 . However, because PM 10 and PM 2.5 indicators are highly correlated, and because it is preferable to work with robust data from several monitoring stations in order to limit the impact of possible measurement errors or missing data on an indicator, PM-adjusted models were only built for PM 10 . The estimates of ERR for NO 2 remained unchanged after adjustment for PM 10 , and this result was consistent with the literature. Several studies reviewed in REVIHAAP by the WHO [3] also found significant associations between NO 2 and all-cause mortality after adjustment for PM 10 [29,[32][33][34].
The present results remained unchanged after adjusting for O 3 . A Brazilian study [35] showed that the ERR for circulatory mortality differed little between single-pollutant NO 2 analysis and analysis adjusted for the other pollutants (PM 10 , CO, and O 3 ). A meta-analysis conducted in 2016 [36] identified that most studies showed a PM 10 -independent effect of NO 2 . More specifically, for all-cause mortality, a 10 µg m −3 increase in NO 2 was associated with a 0.78% increase in the risk of death (95% CI: (0.47; 1.09)). This increase was 0.60% (95% CI: (0.33; 0.87)) after adjustment for particulate matter.
Finally, we observed an increase in ERR for cumulative lags, irrespective of the cause of death or the age group. This result was consistent with those found in international studies. The associations were larger on multi-day lags (0-5 cumulative lags) than on single-day lags (0-1 lags) [27,28,32,37].

Conclusions
This study found an association between short-term exposure to NO 2 concentrations as a proxy of urban traffic pollution and the risk of death, with a higher risk for people aged 75 years and over. The results support the strengthening of measures to reduce traffic air pollution sources in Europe to protect the most vulnerable. Partial effect graphs Figure A2. Assessment of the modeling assumptions for Montpellier. Figure A2. Assessment of the modeling assumptions for Montpellier.