Urban air pollution and mortality in a cohort of Norwegian men.

We investigated the association between total and cause-specific mortality and individual measures of long-term air pollution exposure in a cohort of Norwegian men followed from 1972-1973 through 1998. Data from a follow-up study on cardiovascular risk factors among 16,209 men 40-49 years of age living in Oslo, Norway, in 1972-1973 were linked with data from the Norwegian Death Register and with estimates of average yearly air pollution levels at the participants' home addresses from 1974 to 1998. Cox proportional-hazards regression was used to estimate associations between exposure and total and cause-specific mortality. During the follow-up time 4,227 men died from a disease corresponding to an ICD-9 (International Classification of Diseases, Revision 9) code < 800. Controlling for a number of potential confounders, the adjusted risk ratio for dying was 1.08 [95% confidence interval (CI), 1.06-1.11] for a 10- microg/m3 increase in average exposure to nitrogen oxides (NOx) at the home address from 1974 through 1978. Corresponding adjusted risk ratios for dying from a respiratory disease other than lung cancer were 1.16 (95% CI, 1.06-1.26); from lung cancer, 1.11 (95% CI, 1.03-1.19); from ischemic heart diseases, 1.08 (95% CI, 1.03-1.12); and from cerebrovascular diseases, 1.04 (95% CI, 0.94-1.15). The findings indicate that urban air pollution may increase the risk of dying. The effect seemed to be strongest for deaths from respiratory diseases other than lung cancer.

The current understanding of the association between long-term exposure to air pollution and mortality is based on results from a few cohort studies, of which two studies have received a lot of attention (Dockery et al. 1993;Pope et al. 1995Pope et al. , 2002. Both of these studies have assessed air pollution exposure on an aggregated (not individual) level. All participants living in the same city were given a cityspecific exposure level based on the available measurements. At least two other cohort studies have assessed exposure at an individual level (Beeson et al. 1998;Hoek et al. 2002). Results from these studies strengthen the evidence that urban air pollution is associated with increased risk of dying. Associations have varied and exposure measures have been rather crude. No single epidemiologic study can be the basis for determining a causal relation between air pollution and mortality (Health Effects Institute 2000), and there is still a strong need for exploring the associations between long-term air pollution and mortality further. For a long time Norway has had monthly updated registers for all citizens' home addresses and for deaths. This makes it possible to assess air pollution exposure at home addresses historically by geographical information systems and to link the information with information on deaths and causes of deaths for persons who have lived at these addresses. In this study we have estimated yearly air pollution levels at the home addresses for a cohort of 16,209 male citizens from 1974 to 1998 by geographical information systems Håheim et al. 1996;Leren et al. 1975;Nafstad et al. 2003). The cohort was established in [1972][1973]. The aim of the present study was to estimate associations between air pollution levels estimated at the participants' home addresses and the participants' risk of dying.
points and the definition according to , and -10 classifications are listed in Table 1.
Exposure assessment. Indicators of air pollution exposure considered in these analyses were sulfur dioxide and nitrogen oxides. The Norwegian Institute for Air Research has estimated average concentrations per year of these air pollutants at the home addresses of all participants from 1974 to 1998. Good emission data existed for the whole period, but concentration measurements were sparse and stability data existed only for short periods. Particles were not considered in this study because measurement methods changed during the period from measuring black soot from coal and heavy oil combustion to measuring size fractions of particles from traffic.
The model for SO 2 is based on detailed model calculations per square kilometer for the years 1979 and 1995. For the other years, the concentrations were calculated using the modeling results together with observed concentrations in central Oslo and emission data for industry, heating, and traffic. Observed NO x concentrations were not available for all the years, so a new procedure had to be developed; the shape of the SO 2 -concentration fields was determined from the detailed model calculations for SO 2 , and the emissions levels were determined from the SO 2 -measurements. We used this method to define separate dispersion fields (concentration divided by emission) for each year for heating and traffic. These measures were multiplied by the corresponding emissions of NO x and added with a background contribution from other sources. Annual nitrogen dioxide fields were not calculated because ground level ozone concentrations were not measured. Addresses linked to 50 of the busiest streets were given an additional exposure based on estimates of annual average daily traffic. For many of the streets, the traffic changed considerably during the 25-year period because of changes in the main road network. More detailed descriptions of the exposure estimates are given elsewhere .
The National Population Register (Skattedirektoratet, Oslo, Norway) includes home addresses for all Norwegians and is updated monthly. For a person who moved within Oslo, we used the average air pollutant concentration at the address where he lived the largest part of that year. The air pollution levels were linked to map references and then to the participants home addresses during the study period. All of the men lived in Oslo during the first years of the study, and about 90% were still living in Oslo in 1985. Air pollution exposures for persons moving from Oslo were assessed in a special way. Because Oslo is the largest city in Norway, with around 500,000 citizens, moving out of Oslo meant-with few exceptions-moving into a less-populated area with low levels of air pollution. Norway has had a network of background air pollution monitoring stations since the beginning of the 1970s, and values from these stations were used for the less-populated areas. For larger cities and industrial areas, we used available measurements to adjust background values; the annual variations were based on emission estimates from Statistics Norway. The concentration for one specific year was calculated as the region concentration multiplied by an emission index for that year and linked to the cohort information. We used the calculated values to estimate average exposure for different time windows and cumulative exposures. All available information was used in the modeling, and we were only able to evaluate the model for short periods during these 27 years (Gram F, unpublished data).
Covariates. Statistics Norway provided information on the highest level of education for each participant in the cohort: < 10 years, 10-12 years, and > 12 years of education, which reflects low, medium, and high education in Norway.
Information on all other covariates is based on the baseline screening of the cohort, which included a self-administered questionnaire, sampling of blood, and measurements of height and weight (Håheim et al. 1996;Leren et al. 1975). We used the following information in the present analyses: participants' age, smoking habits, leisure-time physical activity, occupation, height, weight, cholesterol, blood pressure, and risk groups for cardiovascular diseases. Smoking habits within the cohort were grouped as follows: nonsmoker, 1-9 cigarettes/day, 10-19 cigarettes/day, ≥ 20 cigarettes/day, smoker (no amount reported), and former smoker.
Smoking reported as grams of tobacco per week (smoked as hand-rolled cigarettes or in pipes) was converted to cigarettes per day (0.8 g tobacco = 1 cigarette). Occupation was Environmental Medicine | Mortality and air pollution Environmental Health Perspectives • VOLUME 112 | NUMBER 5 | April 2004 611 Table 1. Causes of deaths, numbers of deaths, and incidences of deaths with 95% CIs according to the ICD-8, -9, and -10.

Causes of death
No.
Statistical methods. We calculated incidences of the studied death causes per 1,000 observation years starting at the beginning of 1974, the first year with information on air pollution exposure. We used Cox proportional hazard regression models to evaluate the association between incidence of death and the indicators of air pollution. For the analyses of time to death from a specific cause, we examined each participant's observation time at the end of the follow-up (1998), at the year when the person died from another cause, or the year the participant emigrated or moved to an address without information on air pollution exposure. We used average 5-year levels of exposure in most of the analyses. Air pollution levels were included as both continuous and categorical variables. The average exposure levels were skewed with little variation in the lower quartiles (range within first to second quartile: NO x , 0-11 µg/m 3 ; SO 2 , 0-9 µg/m 3 ). The exposure levels were therefore grouped as follows: 0-9.99, 10. 00-19.99, 20.00-29.99, and ≥ 30.00 µg/m 3 . Furthermore, NO x and SO 2 were never included in the same regression model because of potential colinearity problems. To study the functional form of the association between air pollution exposure and different causes of death, we also modeled the continuous exposures as smoothed cubic splines using S-PLUS, Release 6.0 for Windows (Insightful, Seattle, WA, USA) (Therneau and Grambsch 2000). Two extreme values (outliers) were excluded from the analyses because they did not fit any reasonable statistical distribution, they seemed unreasonably high, and they could have large effects when air pollution exposure was used as a continuous variable.
All of the registered covariates were evaluated in the Cox proportional hazard model for inclusion in the final statistical models. We used different modeling strategies. The first strategy was to only include covariates that changed the log-likelihood test of the model significantly (p < 0.05). We specified different models for each of the health outcomes. The second strategy was to perform analyses with a core covariate set (education, occupation, smoking habits, leisure-time physical activity, risk group of cardiovascular diseases, and age at inclusion), which was the same for all the health outcomes. With all covariates in the model, we detected deviations from the proportional hazard rates assumption for two of the variables (education and risk groups for cardiovascular diseases), and we had to perform the analyses stratified for these variables (Harrell 2001). Risk ratios and confidence intervals (CIs) showed only minor variation between the first and second approach. This was also the case if the analyses were performed without education and risk groups for cardiovascular diseases in the model. Only results from the second approach are presented.
Our main focus was on air pollution exposure during the first 5 years of follow-up. We also performed analyses with yearly averages and with later 5-year periods; NO x and SO 2 were also treated as time-dependent variables in the Cox regression models, representing average exposure for the last 5 years before death. Each time a risk set was created for a new death case with 5 years of exposure, the average air pollutant variable for each individual in the risk set was recomputed as the mean of the yearly exposures the last 5 years before the death occurred. Table 1 presents the number of deaths and the incidences of different causes of death within the cohort during the observational period. Death from ischemic heart diseases was the dominating cause of death (n = 1,508; incidence rate 4.13/1,000 observation years), but   as many as 200 men were registered as having died from a respiratory disease other than lung cancer (incidence rate, 0.55/1,000 observation years). Distribution of air pollution exposure within the cohort has been described in more detail elsewhere Nafstad et al. 2003). The yearly averages of SO 2 levels were reduced with a factor of 7 during the study period (from 16.0 µg/m 3 in 1974 to 2.4 µg/m 3 in 1995), whereas NO x varied between 11.5 and 21.7 µg/m 3 without any clear downward or upward trends. The 5-year median average level of exposure for NO x at the participants' home address during 1974-1978 was 10.7 µg/m 3 , varying from 0.7 to 168.3 µg/m 3 ; quartiles for NO x were divided at 6.8, 10.7, and 20.4 µg/m 3 ; and median levels within the quartiles of NO x exposure were 3.8, 9.3, 15.4, and 28.8 µg/m 3 . The corresponding 5-year median average exposure for SO 2 was 9.4 µg/m 3 , varying from 0.2 to 55.8 µg/m 3 ; quartiles were divided at 4.5, 9.4, and 23.0 µg/m 3 ; and median levels within the quartiles were 2.5, 6.2, 14.7, and 31.3 µg/m 3 . The Pearson's correlation coefficient between yearly NO x and SO 2 levels was 0.63 (95% CI, 0.60-0.66). Table 2 presents the distributions of categorical covariates that in some way were significantly associated with the risk of dying. Most of the participants had sedentary work and only 21% had intermediate or vigorous physical activity during work time. A sedentary or moderate physical leisure-time lifestyle was common (80.7%). About 70% of the participants had ≥ 10 years of education, and a large proportion of the participants reported that they were smokers or former smokers (56.3% current smokers, and 24.7% former smokers). Table 2 also demonstrates how the covariates were distributed according to the NO x levels at the participants' home addresses. Most covariates were evenly distributed at all NO x levels, whereas there was a tendency for the participants with the highest level of education to have the highest NO x exposure. The mean level of systolic blood pressure, cholesterol, height, and weight were about the same in different strata of NO x ( Table 3).

Results
Several of the core covariates, including smoking habits, were significantly associated with the risk of dying (Table 4). Crude and adjusted risk ratios (aRRs) were similar except for occupation (results not shown). The association between mortality and categories of occupation was weakened when education was included in the same model. Table 5 shows the incidence and risk ratios for the different causes of death that were considered in the analyses according to average exposure to NO x from 1974 to 1978 and adjusted for the core covariates and corresponding figures for SO 2 . Exposures to NO x and SO 2 were divided into four exposure levels in model 1 and included as continuous variables (per 10 µg/m 3 ) in model 2. The risk of dying from a disease increased with increasing levels of NO x exposure. The effect was clearest for deaths from a respiratory disease other than lung cancer, both when the analyses were performed with NO x as a categorical variable and as a continuous variable. The aRR was 1.71 (95% CI, 1.09-2.68), contrasting the highest and lowest categories of NO x exposure (model 1), and aRR 1.16 (95% CI, 1.06-1.26) per 10 µg/m 3 increase in NO x exposure (model 2). We used a smoothed cubic spline that described the relationship between NO x and the log risk of death to examine if air pollution levels could be included in the analyses as a continuous variable. For respiratory mortality, there was a monotone increase in risk with increasing exposure levels up to 50 µg/m 3 . For the 82 persons with higher exposure levels, the risk ratios became unstable with wide confidence limits but still in accordance with a linear trend. A similar but somewhat weaker trend was seen for the risk of dying from lung cancer. No clear dose-response pattern was seen between deaths from ischemic heart diseases and NO x exposure when NO x was included in the model as a categorical variable. Even so, the aRR increased [1.08 (95% CI, 1.03-1.12) per 10 µg/m 3 NO x ] when the analyses were performed with NO x as a continuous variable. For cerebrovascular diseases, there was no clear association or trends between the health outcome and increasing NO x exposure. For SO 2 , there were, in general, no clear and meaningful associations between mortality and increased exposure; it was not the higher exposure levels but exposure between 10.00 and 19.99 µg/m 3 that mainly increased risk ratios.

Environmental Medicine | Mortality and air pollution
Environmental Health Perspectives • VOLUME 112 | NUMBER 5 | April 2004 613 Other analytical approaches, including cumulative air pollution exposures, exposures the 5 last years before death, and inclusion of height, weight, blood pressure, and cholesterol in the regression models, gave quite similar associations between air pollution exposures and health outcomes (results not shown). We also analyzed data according to more recent 5-year time windows for NO x exposure, excluding cases of death occurring before or during the time windows used in the analyses. This reduced the number of deaths and observation years included in the analyses (Table 6). Even so, for a 10-µg/m 3 increase in NO x exposure during the different time windows the risk estimates were quite stable also if shorter time windows for exposure were used in the analyses. The risk estimates for respiratory deaths increased somewhat, whereas they were reduced for lung cancer. None of the changes were statistically significant.

Discussion
We found that home address NO x exposure, which was estimated using a geographical information system, was associated with the risk of dying in a cohort of Oslo men. The associations were robust for different analytical approaches and adjustments for different covariates. The effect was strongest for respiratory deaths other than lung cancer [aRR 1.16 (95% CI, 1.06-1.26) for a 10-µg/m 3 increase in NO x exposure at the participants' home address during [1974][1975][1976][1977][1978]. Increased NO x exposure was also associated with deaths from lung cancer and ischemic heart diseases. The effects were smaller and the dose-response relation less clear for ischemic deaths. Analyses with more recent time windows for exposure that excluded men who died early in the follow-up gave similar results. High correlation between exposure levels for different years made it difficult to use these findings to explore latency in time between exposure and diseases. However, the findings indicated stable risk estimates that did not depend on a few cases of death. A small group of participants had NO x levels > 50 µg/m 3 , which was clearly higher than the rest. Analyses with and without these persons showed that they did not change the overall effects, and they were kept in the analyses.
Different approaches have been used to estimate exposure in studies of long-term air pollution exposure and mortality. This has varied from attributing all citizens in one city the same level of exposure (Dockery et al. 1993;Pope et al. 1995Pope et al. , 2002 to using geographical information systems, including air pollution monitoring and information on emissions to estimate more or less individual exposure (Beeson et al. 1998;Bellander et al. 2001;Hoek et al. 2001Hoek et al. , 2002Langholz et al. 2002;Raaschou-Nielsen et al. 2001;Reynolds et al. 2002). In the present study we used the last approach and developed a model that estimates individual home address air pollution exposure by historical data on air pollution measurements and emission and by meteorologic and topographic observations. To obtain the best results, we included all available information in the model . This reduced our ability to evaluate accuracy of the model. Even if we found reasonable agreement for shorter time intervals, we considered that our information was not sufficient to give overall estimates of accuracy and precision. To assess the accuracy and precision of estimates of air pollution levels at more than 16,000 addresses during 27 years would have been unrealistic. Thus, we cannot exclude the possibility of some misclassification of exposure, as will be the case in most similar studies. However, we see no reason to expect systematic misclassification because exposure was assessed independently from all the other data. We believe this model is best for estimating long-term average exposure. We also expect the accuracy to be quite good for estimating exposure for men moving out of Oslo into areas with lower pollution. Another general problem is that home address exposure may not correctly measure true individual exposure because people do not stay home all the time and often stay indoors when they are at home. However, adults, in general, spend much of their time in their home environment; this was probably even more common during the 1970s when people were less mobile and often worked near their homes. Furthermore, there were few indoor sources of NO x , and studies have found a strong correlation between indoor and outdoor exposure levels under such conditions (Emenius 2003;Magnus et al. 1998). We therefore believe that home address NO x levels are a reasonable way of estimating long-term NO x exposure of these subjects.
Our 27-year follow-up study of 16,209 middle-aged men living in one city has several strengths. It ensures a large number of observational years and many deaths and it reduces problems with age variation and between-city heterogeneity. This study also ensures a plausible relationship in time between these conditions and disease development and a low chance of reporting bias. Air pollution exposure was estimated independently. Systematic misclassification in home address exposure related to the outcomes is therefore unlikely, and random misclassification would probably dilute the associations between exposure and outcome. We have no reason to believe that the Oslo men not willing to participate would have a different health effect from air pollution exposure. The study included information on most conditions that we considered potential confounders. NO x exposure was quite evenly distributed for different categories of these conditions, including smoking, which was very common among Norwegian men in the 1970s. This reduces the chance of confounding. Furthermore, we have adjusted for these conditions in several ways in the analyses, and these adjustments gave only minor changes in associations between air pollution exposure and outcomes. It is impossible, however, to fully exclude the possibility of confounding and misclassification of exposure in this type of study. Consistent findings from different cohort studies will therefore be important to increase the understanding of the associations between mortality and long-term air pollution exposure (Health Effects Institute 2000). The associations between SO 2 exposure and mortality give further support to this view. The lack of dose-response relations and the fact that primarily moderate levels of exposure increased the risk of death should be confirmed in other studies before speculating too much about possible explanations.
Our study supports findings from other cohort studies that urban air pollution may increase mortality (Beeson et al. 1998;Dockery et al. 1993;Hoek et al. 2002;Pope et al. 1995Pope et al. , 2002. It is difficult to compare associations between specific outcomes and exposures because of differences in design, local conditions, exposure assessment methods, and   1974-1978 1979-1983 1984-1988 1989-1993 increase  which air pollutants have been measured. In this study we did not include indicators on human exposure to particulate matter, which most often has been shown to be associated with adverse health effects. Urban air pollution is always a mixture of pollutants. We believe that air pollution components assessed in this study and similar studies are exposure indicators only; it is difficult to conclude from such studies which air pollutant causes which health effects. Even so, we believe that the contrasting effect between SO 2 and NO x found in this study is interesting and might indicate that traffic-related air pollution is important. Respiratory mortality other than lung cancer was the cause of death most strongly associated with NO x exposure. Other studies have also shown that exposure to different air pollutants is associated with respiratory health effects (Archer 1990;Clancy et al. 2002;De Leon et al. 2003;Hedley et al. 2002;Lebowitz 1996;Pope 2000;Sunyer et al. 2002). A large proportion of the respiratory deaths is expected to be chronic obstructive pulmonary disease (COPD). Both COPD and lung cancer have been shown to be related to environmental conditions including tobaccosmoke exposure (Mannino 2002). It therefore seems biologically reasonable that urban air pollution may increase the risk of developing these diseases. As for the risk of death from lung cancer, the association was well in accordance with the risk of developing lung cancer, as shown by Nafstad et al. (2003).
Several time-series studies have found associations between ischemic heart diseases and deaths, and biological mechanisms for explaining these findings have been developed (Dockery 2001;Magari et al. 2002;Peters et al. 1997Peters et al. , 2001Pope 1991). This includes inflammation of the lung by cytokine release, which also affects the cardiovascular system and alterations in autonomic cardiac function (Pope 2000). Even if there was a statistically significant association between NO x exposure and deaths from ischemic heart disease, the dose-response relation was not clear and warrants caution in our conclusions. We have no good explanation for this finding except that some other studies have also shown that the association between indicators of air pollution exposure and cardiovascular outcome is weaker than between such exposures and respiratory outcomes (Archer 1990;Clancy et al. 2002;De Leon et al. 2003). These exposures may also primarily affect the risk of developing acute cardiac events and, to a lesser extent, contribute to the development of chronic heart diseases. These findings illustrate that no single epidemiologic study can be the basis for determining a causal relation between air pollution and mortality (Health Effects Institute 2000).