Short-Term Effects of Nitrogen Dioxide on Mortality and Susceptibility Factors in 10 Italian Cities: The EpiAir Study

1Nitrogen dioxide (NO2) is a strong respiratory irritant gas originating from high-temperature combustion. Main outdoor sources of NO2 include motor vehicles (particularly those equipped with diesel engines) and fossil-fuel power plants, whereas the most important indoor sources are gas heaters, stoves, and environmental tobacco smoke [Kraft et al. 2005; U.S. Environmental Protection Agency (EPA) 2008].

2Large meta-analyses of studies on the short-term health effects of NO2 have been carried out in Europe, the United States, and Canada. The results indicate a positive association between daily increases of NO2 and natural, cardiovascular, and respiratory mortality. The findings are consistent with an independent effect of NO2, although the possibility remains that NO2 acts as a surrogate for other unmeasured pollutants. Several epidemiological studies have indicated that NO2 may be a more relevant health-based exposure indicator than particulate matter (PM). Based on these observations, the U.S. EPA has recently proposed to strengthen the NO2 air quality standard that protects public health (U.S. EPA 2008).

3Despite the large body of evidence linking NO2 with daily mortality, few studies have addressed the issue of susceptibility to NO2 by performing analyses by age, sex, and other factors, including socioeconomic status (SES) and chronic morbidity. On the other hand, the evaluation of the role of susceptibility factors in modifying the effect of air pollutants is of increasing interest in order to better understand the mechanisms of NO2 health effects and to provide public health warnings to specific population subgroups. Along these lines, we have already explored the role of individual characteristics (age, sex, socioeconomic factors, and clinical characteristics) as effect modifiers of the association of PM ≤ 10 μm in aerodynamic diameter (PM10) and ozone (O3) with natural mortality. The present study is part of EpiAir, the Italian surveillance project on the health effects of air pollution whose main objective is the continuous update of the effect estimates for PM10, NO2, and O3.

4The specific objectives of the present article were to investigate the NO2–mortality relationship for specific causes of death while exploring the latency of the effects and the potential confounding role by other pollutants and to evaluate sociodemographic features and chronic or acute medical conditions as potential effect modifiers. Preliminary results of this study were presented at the 2009 Conference of the International Society for Environmental Epidemiology.

5Health data. We collected mortality data for 10 Italian cities (Bologna, Cagliari, Florence, Mestre–Venice, Milan, Palermo, Pisa, Rome, Taranto, and Turin); this data accounted for about 12% of the total Italian population (Table 1). We selected 276,205 subjects ≥ 35 years old, resident within the city at the time of death, who died between 2001 and 2005 of natural causes [International Classification of Diseases, version 9 (ICD-9), codes 1–799]. The underlying cause of death was classified as cardiac (ICD-9 codes 390–429), cerebrovascular (ICD-9 codes 430–438), and respiratory (ICD-9 codes 460–519). The resident population data (year 2001) were recovered from the census office registry.

6For all centers except Cagliari (where hospital discharge data were not available at the time of the study), we collected data at the individual level on the following susceptibility factors: age, sex, median income of the census block of residence (these data were available only for Milan, Turin, Bologna, and Rome and accounted for 75% of the study population), and median socioeconomic position of the census block of residence (these data were available only for Mestre–Venice, Pisa, Rome, Taranto, and Turin and accounted for 44% of the study population). A record linkage with the regional archives of hospital admission databases allowed us to gather data on the place of death [classified as out-of-hospital, recently discharged (within 4 weeks) from a hospital, in-hospital, nursing home] and on discharge diagnoses in the previous 2 years.

7Health conditions in the 24 months before death were classified as “chronic” or “acute” according to several criteria. Chronic diseases were those with a course consistent with clinical criteria of chronicity that were diagnosed at least 1 month before death and did not present a recent exacerbation; both primary and secondary discharge diagnoses were considered for hospitalizations that occurred between 29 days and 2 years before death. Acute conditions included not only clinical manifestations with sudden onset, short course, and high likelihood to be cured but also exacerbations of chronic diseases, provided that both these clinical forms caused a hospitalization within 1 month before death; only the primary discharge diagnoses were considered for hospitalizations that occurred in the 4 weeks before death. We based our list of diagnoses on Elixhauser’s list of comorbidities; this approach is consistent with previous work. We selected as “chronic conditions” diabetes, coagulation disorders, hypertension, myocardial infarction, cardiac ischemic diseases, diseases of pulmonary circulation, heart conduction disorders, dysrhythmias, heart failure, cerebrovascular diseases, and chronic pulmonary diseases; “acute conditions” included diseases of pulmonary circulation, dysrhythmias, heart failure, and renal failure.

8Environmental data. Air pollution data were provided through city-specific air monitoring networks managed by regional environmental agencies or local authorities. We obtained data on nitrogen dioxide (NO2; daily average, micrograms per cubic meter), ozone (O3; daily maximum 8-hr running mean, micrograms per cubic meter), and PM10 (daily average, micrograms per cubic meter) (Table 2). Air pollution data were collected according to methods already employed in several European studies. We estimated daily levels of air pollutants for each city by averaging monitor-specific daily measurements available from different monitoring stations. A previously defined algorithm was implemented to impute missing values for pollutant concentrations in each center.

9We collected data on meteorological variables (air temperature, dew point temperature, and barometric pressure) from the Italian Air Force Meteorological Service. Apparent temperature was estimated taking into account air temperature and humidity.

10Statistical analysis. We investigated the association between NO2 and mortality using a case-crossover design. To control for time trends, we selected control days using a time-stratified approach; the study period was divided into monthly strata, and the control days for each case were chosen on the same day of the week in the stratum. We applied a conditional logistic regression to data from each city. We considered as confounding factors: population decrease, holidays, influenza epidemics, barometric pressure, and apparent temperature. Day of the week and long-term and seasonal trends were adjusted for by design. For population decreases during summer vacation periods (typical of Italian cities, when urban populations are substantially reduced), we defined a three-level variable: 2 for the 2-week period around mid-August, 1 for the time period from 16 July to 31 August (with the exception of the mentioned 2-week period), and 0 (the reference category) otherwise. For holidays we defined a two-level variable: 1 for national or city-specific holidays, and 0 otherwise. For influenza epidemics we defined a two-level variable: 1 for the 3-week winter period of influenza epidemic peak (as defined for each year by the National Institute of Health), and 0 for the remaining days of the year. For barometric pressure we used a penalized spline of the original variable at lag 0. The effect of temperature on mortality was controlled for by modeling high temperatures and low temperatures separately. More specifically, high temperatures were adjusted for by calculating the average of current- and previous-day apparent temperature (lag 0–1) and by fitting a penalized spline of the lagged variable only for days with lag 0–1 apparent temperature above the city-specific median value, calculated on the year-round time series of apparent temperature. Similarly, low-winter temperatures were adjusted for by calculating the mean air temperature of the previous 6 days (lag 1–6) and by fitting a penalized spline of the lagged variable only for days with lag 1–6 air temperature below the city-specific median value, as determined from the year-round series.

11We began by modeling the effect of NO2 on cause-specific mortality for each city using unconstrained distributed lag models, cubic polynomial distributed lag models, and single-lag models to identify the best lag structure. Unconstrained distributed lag models were implemented for different lag structures a priori, defined as immediate (up to lag 1), delayed (lag 2–5), and prolonged (lag 0–5) effects, with the aim of obtaining an unbiased estimate of the cumulative effect of NO2 on mortality at the different lags. Cubic polynomial distributed lag models were fitted within each city in order to visually display the latency of the NO2 effect, being aware that single days of the polynomial curve were not directly interpretable in terms of effect size. Finally, we fitted single-lag models in order to estimate the effect of NO2 on mortality for each specific exposure day separately, up to 5 days before death. All the city-specific results (from distributed-lag and single-lag models) were finally pooled to estimate the lagged effect of NO2 across all of the cities. Pooled estimates were obtained from city-specific results by applying a random-effects meta-analysis (maximum likelihood method).

12We implemented bipollutant models in order to estimate the association of NO2 with cause-specific mortality while adjusting for PM10 or, in turn, O3 (the latter during April–September only). The lag exposures were chosen within the unconstrained distributed-lag modeling framework, selecting the lags showing the strongest association. A pooled estimate was obtained from city-specific results via random-effects meta-analysis.

13We evaluated the role of the potential effect modifiers (sex, age, SES at the census block level, chronic and acute conditions as previously defined) via conditional logistic regression models stratified by the levels of each presumed effect modifier: we compared the effect estimate for NO2 in each category of the potential effect modifier with the effect estimate in the reference category for the potential modifier, from the stratified models. Because all effect modifiers were likely to be associated with age (e.g., chronic conditions), all stratum-specific estimates were standardized by age, using the relative frequencies of the overall age distribution as weights. We formally evaluated statistical significance (at α = 0.05 level) of the effect modification and computed p-values for relative effect modification (p-REM). In particular, the relative effect modification (REM) was evaluated by analyzing the difference between the coefficient of the NO2–mortality association within a specific stratum of the effect modifier and the coefficient within the reference stratum of the same variable. The corresponding p-value (p-REM) is derived by assuming that the difference between the two coefficients follows a normal distribution with zero mean and variance equal to the sum of the two stratum-specific variances. We assumed that effect modification was “likely” when p-REM ≤ 0.05, regardless of the magnitude of the stratum-specific association estimate. The effect modification was “suggested” when 0.05 < p-REM < 0.20 and either a) the risk estimated in a specific stratum was twice the risk estimated in the reference stratum or b) the excess risk estimated for the stratum was statistically significant. In addition, in case of a possible effect modifier with more than two ordinary modalities (e.g., number of chronic condition), we considered evidence of effect modification to be “suggestive” when a dose–response trend was observed.

14Analysis of effect modification by individual characteristics was performed for each city, and pooled effects were estimated via random-effect meta-analysis. For each pooled effect estimate, we computed the Q-statistic and the p-value of heterogeneity (HET) to test for heterogeneity among city-specific estimates (against the null hypothesis that the city-specific estimates were homogeneous).

15We express all effect estimates as the percent increase in mortality, with corresponding 95% confidence intervals (CIs), associated with a 10-μg/m3 increase in NO2.

16We performed the statistical analyses using SAS (version 8.0; SAS Institute Inc., Cary, NC, USA) and R software (version 2.6.1; R Project for Statistical Computing, Vienna, Austria).

17Table 1 summarizes population and mortality data for each city included in the analysis. We considered a total of 276,205 natural (nonaccidental) deaths among those > 35 years of age. Cardiac, cerebrovascular, and respiratory mortality accounted for about 28%, 10%, and 7% of natural deaths, respectively.

18Table 2 summarizes the descriptive statistics for air pollution indicators, expressed as daily means over the time period considered, for each city. For NO2, values consistently exceeded 40 μg/m3 in six cities.

19Figure 1 shows pooled effect estimates (10 cities) for the association between NO2 and mortality by cause of death and lag (single-lag models and constrained and unconstrained distributed-lag models). The lag structure suggests a prolonged effect of the NO2 on all outcomes considered up to lag 5, whereas a delayed association was more evident for respiratory mortality, from lag 1 to 5. Based on these results, we selected lag 0–5 as the lag with the maximum estimated effect for natural, cardiac, and cerebrovascular mortality, and lag 1–5 as the lag with the maximum estimated effect for respiratory mortality.

20Pooled results for all 10 cities indicated that a 10-μg/m3 increase in NO2 was significantly associated (at α = 0.05 level) with all natural mortality, cardiac mortality, and respiratory mortality, with the strongest estimated effects for respiratory mortality (Table 3). The estimated effects of NO2 were not confounded by PM10 in bipollutant models. Associations were enhanced during the warm season (April–September) when the NO2–cerebrovascular disease association also became statistically significant, but these associations were not confounded by O3. Effect estimates for natural and respiratory mortality were heterogeneous across cities, with Rome being an outlier for natural mortality (4.41%; 95% CI, 3.38–5.45) and Turin for respiratory mortality (–1.83%; 95% CI, –3.35 to –0.28). No significant heterogeneity was present for cardiac and cerebrovascular outcomes.

21Pooled associations between NO2 and natural mortality (lag 0–5) are reported for nine cities (excluding Cagliari), overall and according to strata of selected susceptibility factors (Table 4). We observed an overall increase of 2.03% (95% CI, 0.87–3.21) in natural mortality associated with a 10-μg/m3 increase of NO2. The association was stronger for subjects > 84 years of age (3.41%; 95% CI, 2.10–4.74) than for younger subjects, but age was not a significant effect modifier (p-REM = 0.270), and associations did not follow a monotonic trend with age. Neither income nor socioeconomic position (both measured as the median of the census block of residence) significantly modified the association between NO2 and mortality, but we observed significant heterogeneity in the stratum-specific effect estimates among the cities.

22The season of death significantly modified the association between NO2 and all natural mortality, with a stronger association (4.64%; 95% CI, 3.33–5.97) during the warm season than during the rest of the year (1.18%; 95% CI, 0.20–2.16; p-REM = 0.000). The association between NO2 and all natural mortality also was significantly stronger among subjects with at least one hospital admission between 2 years and 29 days before death (2.86%; 95% CI, 1.39–4.35) than among other subjects (0.73%; 95% CI, –0.69 to 2.18; p-REM = 0.043). Similarly, the association was significantly stronger for subjects with three or more chronic conditions (3.62%; 95% CI, 2.04–5.22) than for those without chronic conditions (1.54%; 95% CI, 0.27–2.82; p-REM = 0.045), with a monotonic increase in the excess risk in relation to the number of chronic conditions. Associations between NO2 and all natural mortality also were stronger among subjects hospitalized between 2 years and 29 days before death for the specific chronic conditions examined. Estimated risks were particularly high for subjects with disorders of pulmonary circulation (8.03%; 95% CI, 3.17–13.2; p-REM = 0.014), and we found some evidence of effect modification by heart conduction disorders (5.91%; 95% CI, 1.78–10.2; p-REM = 0.064), diabetes (3.61%; 95% CI, 1.73–5.53; p-REM = 0.108), heart failure (3.40%; 95% CI, 1.43–5.40; p-REM = 0.166), and cardiac ischemic diseases (3.22%; 95% CI, 1.52–4.96; p-REM = 0.185).

23We did not find evidence of effect modification by sex, place of death, hospital admissions between 0 and 28 days before death (Table 4), or hospitalization for diseases of the pulmonary circulation, dysrhythmias, heart failure, or renal failure during the 28 days before death (data not shown).

24In this study we found statistically significant increases in mortality due to natural, cardiac, and respiratory causes associated with a 10-μg/m3 increase in NO2 regardless of season, and a significant association between NO2 and cerebrovascular mortality during the warm season. Overall, associations were strongest for exposures lagged 0–5 days and were stronger in the warm than in the cold season. Associations with NO2 appeared to be independent of PM10 and independent of O3 exposure during the warm season. Associations with total mortality were stronger for subjects with a hospital admission in the 2 preceding years. Interestingly, previous cardiovascular morbidity (changes in pulmonary circulation, heart conduction disorders, heart failure, and ischemic heart diseases) and diabetes appeared to confer a strong susceptibility.

25Excess risks estimated in the present study for a 10-μg/m3 increase in NO2 (natural causes, lag 0–5: 2.09%; 95% CI, 0.96–3.24; cardiac causes, lag 0–5: 2.63%; 95% CI, 1.53–3.65; respiratory causes, lag 1–5: 3.48%; 95% CI, 0.75–6.29) are higher than those published in previous meta-analyses, although comparisons are limited because of the use of different statistical methods, lags, populations, and metrics for NO2 exposure.

26used the most extensive European database available [Air Pollution on Health: A European Approach (APHEA-2)] to investigate the effects of NO2 on mortality. They estimated 0.30%, 0.40%, and 0.38% excess risks for natural, cardiovascular, and respiratory causes, respectively (for a 10-μg/m3 increase in NO2). The study was related to the calendar period 1990–1997, and the study population was not restricted to a specific age group. The meta-analysis of the Italian studies on short-term effects of air pollution (MISA2) reported at lag 0–1, 0.59%, 0.40%, and 0.38% excess risks for natural, cardiovascular and respiratory causes, respectively (for a 10-μg/m3 increase in NO2) for deaths during 1996–2002. Our estimates for deaths in many of the same cities during 2001–2005 are almost double (0.99%, 1.13%, and 1.19% for lag 0–1, respectively) the MISA2 estimates, despite the fact that in recent years the concentrations of NO2 are slightly reduced in the cities involved in these Italian studies. An interesting temporal and multicity analysis conducted in Canada has suggested that, despite decreasing ambient concentrations over time, mortality risks associated with NO2 appear to be increasing. One possible explanation for our findings and the Canadian results is that NO2 itself may not be causally linked to mortality and that the truly toxic components may not be changing over time, at least not at the same rate as NO2.

27In our study the estimated effects of NO2 appeared to extend over the 5 days before death; the shape of the association between lagged NO2 and cardiac mortality was comparable to that for natural mortality, with strongest associations with NO2 2 days before death. NO2 on the day of death was not associated with respiratory mortality, whereas NO2 was 1–5 days before death. This is in agreement with the APHEA-2 results.

28Analyses restricted to April–September showed stronger associations, with excess risk estimates of 4.49%, 4.77%, and 10.07% for natural, cardiac, and respiratory mortality, respectively; moreover, the risk for cerebrovascular diseases became significant (7.87%) in the warm season. The season of death was, in fact, the strongest of the effect modifiers examined for natural mortality (4.64% during the warm period vs. 1.18% for October–March; p-value of REM < 0.001), and these associations persisted after adjustment for O3. These results are in agreement with previously published European meta-analyses. There can be several possible explanations for this finding, not mutually exclusive. A likely explanation is that during the warm season the measured concentrations of the air pollutants are more representative of the true exposure of the subjects: during the summer people spend more time outside than during the cold season and are more likely to keep windows open, thus allowing ambient air pollutants to enter buildings. Furthermore, an increased individual susceptibility to the effects of air pollutants in summer may be present.

29However, other interpretations cannot be ruled out. First of all, a true synergic effect between NO2 and high summer temperature could be present, because that biochemical reaction is faster at higher temperature. In addition, the summer effect of NO2 could be enhanced by the lower average mortality during summer months, given that in Italian cities the dose–response curve is steeper at lower mortality and concentration values (Biggeri et al. 2009).

30Our results suggest an effect of NO2 on mortality independent from PM10, but the role of NO2 as a surrogate of unmeasured pollutants cannot be ruled out. In the presence of high levels of traffic, PM consists of a mixture of carbon particles (including PM in the ultrafine range), which are mainly produced by diesel engines (diesel soot). In fact, diesel engines are the main sources of both NO2 and ultrafine PM. These ultrafine particles are sulfated and nitrogenized, which may explain the high correlation observed among PM, NO2, and sulfur dioxide. From this perspective, NO2 should be considered mainly as a surrogate of ultrafine PM. In our EpiAir study, NO2 is a strong confounder of the relationship between PM10 and mortality: in the single-pollutant model, we observed a significant increase of 0.80% (lag 0–2) for the association of PM10 with natural mortality that decreased to 0.18% (not significant) in the bipollutant model with NO2. This is consistent with the results of a previous Italian meta-analysis.

31Specific groups within the general population are at increased risk of adverse effects from NO2 exposure. Factors that may influence susceptibility to the effects of air pollution include age (e.g., elderly), sex, race/ethnicity, genetic factors, and preexisting diseases or conditions (e.g., obesity, diabetes, respiratory disease, asthma, chronic obstructive pulmonary disease, cardiovascular disease, dysrhythmias, airway hyperresponsiveness, respiratory infection). In addition, exposure to air pollution may vary among population subgroups according to SES, educational level, air conditioning use, proximity to roadways, geographic location, level of physical activity, and work environment. We analyzed several factors that may confer susceptibility and/or vulnerability to air pollution, and most of our results are in agreement with previous studies, even though some of the associations could be partially spurious because of the multiple tests performed on the same database.

32Many chronic health conditions appeared to increase susceptibility to effects of NO2; in particular, people with previous hospital discharges for cardiovascular conditions appeared to be at higher risk. The estimated effect of NO2 was highest for subjects with three or more chronic health conditions. The estimated NO2 effect was stronger among the elderly—a finding that is consistent with previous studies for exposure to particles, although the association for those ≥ 85 years of age was not significantly different from that for the 35- to 64-year age group.

33The results obtained in the present study do not clarify the role of SES as an effect modifier of the association between mortality and variation in NO2. However, data on SES were not available at an individual level (only at area level), and this represents a limitation of our study. Another limitation is related to the ascertainment of the chronic health condition, which we based on hospital discharge records and suffers from the limits of accuracy of the source used. Therefore, we tried to increase the sensitivity of the definition of chronic conditions by using all hospital admissions in the 2-year period before death and by considering both primary and contributory causes.

34This study confirms a clear association between short-term exposure to NO2 and natural mortality and supports increased susceptibility among people suffering from chronic cardiovascular conditions and diabetes. These conditions should be considered when developing prevention-oriented health policies.