|Home | About | Journals | Submit | Contact Us | Français|
Few cohort studies of the health effects of urban air pollution have been published. There is evidence, most consistently in studies with individual measurement of social factors, that more deprived populations are particularly sensitive to air pollution effects.
Records from the 1996 New Zealand census were anonymously and probabilistically linked to mortality data, creating a cohort study of the New Zealand population followed up for 3 years. There were 1.06 million adults living in urban areas for which data were available on all covariates. Estimates of exposure to air pollution (measured as particulate matter with an aerodynamic diameter less than 10 μm, PM10) were available for census area units from a previous land use regression study. Logistic regression analyses were conducted to investigate associations between cause-specific mortality rates and average exposure to PM10 in urban areas, with control for confounding by age, sex, ethnicity, social deprivation, income, education, smoking history and ambient temperature.
The odds of all-cause mortality in adults (aged 30–74 years at census) increased by 7% per 10 μg/m3 increase in average PM10 exposure (95% CI 3% to 10%) and 20% per 10 μg/m3 among Maori, but with wide CI (7% to 33%). Associations were stronger for respiratory and lung cancer deaths.
An association of PM10 with mortality is reported in a country with relatively low levels of air pollution. The major limitation of the study is the probable misclassification of PM10 exposure. On balance, this means the strength of association was probably underestimated. The apparently greater association among Maori might be due to different levels of co-morbidity.
Anthropogenic air pollution is thought to cause a substantial global burden of disease, even if conservative assumptions are used for the analysis.1 Most epidemiological studies of air pollution effects use a time series design, in which day-to-day changes in morbidity or mortality (usually in a single city) are related to day-to-day changes in air pollution exposure. These studies have the advantage of controlling for time invariant (or slowly varying) confounding factors by design. Time series studies provide extensive and detailed evidence of short-term air pollution effects, but are not the best basis for public health policy on air pollution.2 This is because cohort studies suggest that long-term exposures have much greater public health impact than short-term exposures. Because it is difficult to estimate air pollution exposure for large populations over long periods, relatively few cohort studies have been reported and few such studies have been conducted outside north America and Europe.3–18
We report here the findings from a longitudinal study in New Zealand, based on the national census. In New Zealand, major sources of urban air pollution are home heating and motor vehicle emissions and levels of pollution are low by international standards. Typical annual average levels of particulate air pollution (particulate matter with an aerodynamic diameter less than 10 μm, PM10) at monitoring stations in the main urban centres are in the range 15–25 μg/m3. In this study, we analyse the spatial association between average PM10 exposure and mortality over the years 1996–9. We also investigate potential modification by social factors (age, sex, ethnicity, income, education and area level deprivation).
We hypothesise a priori that effects of PM10 on mortality will be stronger for cardiorespiratory causes of death than for all diseases, because these causes are most sensitive to air pollution effects.
The New Zealand Census-Mortality Study is a population-wide cohort study, in which the cohort consists of the entire 1996 resident population (N=3 732 000) and the outcome of interest is mortality. For this analysis, records from the 1996 census were anonymously linked to 3 years of subsequent mortality data, creating a cohort study of the New Zealand population followed up for 3 years.19 20 In New Zealand, urban areas are defined as cities having a population of at least 30 000 people; about half of the approximately 2000 census areas are classified as urban. Because the definition of ‘urban areas’ in New Zealand includes small towns and suburbs, many urban census areas have low levels of air pollution.
The method of record linkage has been described in detail elsewhere.19 Briefly, probabilistic record linkage is a process used to link two files of records, in which records in one file have a corresponding record in the other file. Records from the first file are compared with records from the second file in order to find ‘matching’ record pairs (ie, two records belonging to the same individual). Census and mortality records were linked using date of birth (day, month and year as separate matching variables), country of birth, sex, ethnicity and address of usual residence. (Geocoded addresses were the most discriminatory matching variable.) Linkage was restricted to individuals aged 74 years or below at the time of the census. The proportion of mortality records linked overall was 78%,20 and varied by sex, age, ethnicity and the deprivation index.19 Weights were therefore applied to adjust for linkage bias, by strata of sex, age, ethnicity, deprivation, rurality and cause of death.21 For example, if 20 of 30 Maori men who died aged 45–64 years and living in moderately deprived (see below) rural areas of New Zealand were linked to a census record, each of the 20 linked records received a weight of 1.5 (30/20). The estimated proportion of links being true links was 97%.
Air pollution monitoring data are typically only available for a few sites and may not be representative of exposure in other areas. Detailed atmospheric dispersion modelling can be used to provide estimates of air pollution exposure for small areas, but these models are difficult to apply to large areas due to data and computer processing limitations. The method of air pollution exposure assessment has been described in detail elsewhere.22 Briefly, the approach to modelling long-term average PM10 was as follows. Atmospheric dispersion modelling results were available for one city (Christchurch, population 300 000). These data were assumed to be representative of the spatial pattern of annual average PM10 exposures for small areas (census area units) in Christchurch.23 Data on meteorological variables and indicators of air pollutant emissions for these areas were used as predictors of PM10 exposure in Christchurch, using regression models.
The predictors for the regression models were: census data on domestic heating; estimates of industrial emissions; and vehicle kilometres travelled within small areas. Because data for these predictor variables were available for all of New Zealand, we were able to extrapolate the empirical results for the Christchurch regression model to urban census area units throughout the country.
The resulting exposure estimates agreed well with multiyear averages of PM10 based on routine monitoring data for urban centres (1995–2001) (N=43; r2=0.87). While we acknowledge that there may be considerable exposure misclassification, we assume that the PM10 estimates are representative of long-term average spatial patterns of exposure during the 1990s.
Ideally, for analyses with the New Zealand Census Mortality Study data we would have used the continuous PM10 estimates for individual census areas. However, because the New Zealand Census Mortality Study data contain census information at unit record level, access to the data is restricted and additional data can only be added if confidentiality is not compromised as a result. Assigning continuous PM10 estimates at census area level would have permitted many census areas to be uniquely identified. In order to maintain confidentiality, it was necessary to aggregate the PM10 estimates into quintiles before they could be merged with the census data. Quintiles of exposure were calculated based on the estimates for individual census areas. For this purpose, we assigned a PM10 estimate of 0 μg/m3 to rural census areas where PM10 estimates were unavailable. The average PM10 level for all New Zealand census areas was 8.3 μg/m3 (SD 8.4 μg/m3) and the cut-off values between PM10 quintiles were 0.0, 0.5, 12.5 and 15.4 μg/m3.
Individuals were assigned to quintiles of PM10 exposure depending on their census area of residence on census night. In order to avoid bias affecting analyses in smaller rural areas due to migration to cities following the development of disease, we restricted the analyses to the urban population. This resulted in 1 065 645 observationsi with complete data. Following restriction to the urban population, the lowest two quintiles of exposure had relatively few observations (table 1). For this reason, and given the small variation in PM levels between quintiles 1 and 2, we combined the two lowest quintiles of PM10 to produce four categories in total. The estimated mean PM10 levels for these categories were 0.1, 7, 14 and 19 μg/m3 (long-term averages).
Ethnicity and socioeconomic position are strong predictors of mortality in New Zealand, and potential confounders of the association of air pollution with mortality. We assigned each respondent to a mutually exclusive ethnic group using a prioritisation system commonly used in New Zealand: Maori, if any one of the responses was Maori; Pacific, if any one response was Pacific but not Maori; and the remainder non-Maori non-Pacific (mostly New Zealand European). Smoking status was reported in the census using the categories: never smoker, ex-smoker, or current smoker.24 Socioeconomic position was characterised as: total household income, with adjustment for the number of children and adults in the household to allow for economies of scale, log-transformed having first set all values of less than NZ$1000 to equal NZ$1000;20 highest educational qualification (higher than school, school, or none); and neighbourhood deprivation measured by the NZDep index (in quintiles).25 This index of deprivation within small geographical areas was calculated using census data on socioeconomic characteristics (eg, car access, tenure and receipt of benefits) at aggregations of approximately 100 people, and assigned to mortality data by use of address.
As climate is correlated with PM10, and associated with mortality (independently of PM10), temperature has the potential to confound air pollution effects. Therefore, we also included estimates of long-term average minimum temperature (in addition to the above sociodemographic factors) in the models, also in quintiles based on place of residence. These data were derived from an interpolated temperature surface.26 Mean values for the minimum temperature quintiles were 0.2, 2, 4, 5 and 7°C.
Logistic regression analyses were conducted to investigate associations between all-cause and cause-specific mortality rates and average exposure to PM10, with control for confounding by age, sex, ethnicity, social deprivation, income, education, smoking history and average minimum temperature. As death is a rare outcome over 3 years for the age groups included in the analysis, logistic regression results differ very little from those from either Poisson or Cox proportional hazards modelling. Initial models incorporated PM10 categories as dummy variables, and given a reasonably linear dose–response, final models used the estimated mean PM10 for these four categories (see above). In order to facilitate comparison with overseas findings, we report results for adults aged 30–74 years. In final models, age was included as a linear plus a squared term, and the income variable was natural-log transformed. All other covariates were included as dummy variables (table 2).
Any association of air pollution with mortality might be modified by social and environmental factors, due to different vulnerability or intensity of exposure to air pollution. We tested for interaction both by stratification and by inclusion of interaction terms.
Finally, we conducted a sensitivity analysis by restricting the dataset for analysis to those people who had lived in the same geographical unit at the 1991 census as they did at the time of exposure assignment (1996) in this cohort study. This restriction should reduce exposure misclassification by excluding those people who have not been exposed to the same level of PM10 for at least 5 years.
When PM10 was modelled using dummy variables, there was an approximately linear increase in mortality with increasing average PM10 exposure (figure 1). Three models are shown: (1) adjusting for sex, age and ethnicity on all observations (N=1 364 454); (2) the same model, but restricted to observations with non-missing data on other covariates (N=1 065 645; a test of any selection bias compared with model 1); (3) fully adjusted model (N=1 065 645; a test of possible confounding compared with model 2). There is no meaningful difference between models 1 and 2, suggesting no selection bias when restricting analyses to those with full data on covariates. Adjusting for potential confounders in model 3 attenuated the OR for all non-referent groups, although the relative differences between quintiles 3, 4 and 5 were not much reduced. That is, adjusting for confounders mostly closed the gap between the referent group and quintile 3.
Considering the preferred model 3, the OR increased monotonically and linearly with increasing PM10 level, and the 95% CI for the two highest quintiles excluded 1.0. Given the approximately linear association of PM10 with mortality described above, subsequent models incorporated PM10 as a linear term, using the average PM10 for each category) (tables 2 and and3).3). Table 2 shows the same model as in figure 1, except for the continuous treatment of PM10. The OR of 1.007 corresponds to a 1 unit increase in PM10, which equates to an increase of 7% (95% CI 3% to 10%) in the odds of all-cause mortality in adults (aged between age 30 and 74 years at census) per 10 μg/m3 increase in long-term average PM10 exposure.
We found stronger effects of PM10 among people who lived in the same census area in 1991 (at the time of the previous census), 8% (4% to 12%) per 10 μg/m3 increase in PM10. By cause of death, the association was similar for all natural causes, 7% (3% to 10%), but substantially stronger for respiratory deaths (including lung cancers), 14% (5% to 23%) and for lung cancers, 16% (4% to 29%). For cardiovascular disease, 6% (1% to 12%) and for ‘other and unspecified’ causes of death, 5% (1% to 10%), the association was marginally significant; while for accidental deaths and injuries, the association was non-significant, 4% (−9% to 20%) (table 3).
Considering interaction with social variables, there was an apparently stronger association of PM10 with all-cause mortality among Maori, 20% (7% to 33%). However, the 95% CI for Europeans overlapped that for Maori, 7% (3% to 10%). A Wald test for interaction between ethnicity and PM10 was not significant (p=0.12). There were no statistically significant interactions with age, sex, income, deprivation, educational status or average temperature. Mortality was lower among people living in warmer census areas, and the difference between the lowest and the highest quintile of annual average temperature was statistically significant.
Using linked mortality and census data, we report a significant positive association between estimated long-term exposure to air pollution (PM10) and mortality in New Zealand urban areas. This setting includes approximately 75% of the New Zealand population, who are exposed to relatively low levels of PM10 compared with other countries.
Selection bias and confounding seem unlikely to explain our results. There is no evidence of selection bias (figure 1). The results persist after controlling for plausible confounders, including multiple measures of socioeconomic position and smoking. It seems unlikely that mismeasured or unknown confounders might explain the remaining association.
There is likely to be substantial misclassification of the air pollution exposure. Most of this misclassification was probably non-differential by mortality risk (meaning we have probably significantly underestimated the true strength of association). There is potential differential misclassification of PM10 by mortality risk in our study, because our assessment was based on modelling in one city using proxies, including domestic heating, estimates of industrial emissions and vehicle kilometres travelled within small areas. If these proxies are not such reliable predictors of PM10 in other cities, and are (say) correlated with socioeconomic position, then it may be that our PM10 estimates are also capturing aspects of socioeconomic exposure. However, the fact that an association remained after extensive control of socioeconomic factors, including individual level income and education, makes this an unlikely explanation of the results. The use of modelled estimates of PM10 exposure will tend to smooth the data and reduce the resulting CI. However, this should not affect the central effect estimates.
It is possible that less healthy people might migrate towards health services (or other service amenities) that happen to be in more polluted areas—a form of reverse causation or endogeneity. However, we think this is unlikely to be an important factor as New Zealand cities are relatively small (maximum 1.4 million), and most suburbs in New Zealand's main cities have relatively good access to hospitals.
The odds of all-cause mortality in adults (aged between 30 and 74 years at census) increased by 7% (95% CI 3% to 10%) per 10 μg/m3 increase in average PM10 exposure. Our observations are consistent with an increasing number of studies of long-term exposure to particulate matter and mortality.3–10 13–18
The original US Six Cities Study reported an adjusted mortality rate ratio of 1.27 (95% CI 1.08 to 1.48) for the most polluted compared with the least polluted city, corresponding to 18.2 and 46.5 μg/m3 PM10, respectively3—equivalent to an increase in mortality of approximately 10% per 10 μg/m3 PM10. In the US Nurses Health Study, there was a 16% (5% to 28%) increase in all-cause mortality per 10 μg/m3 PM10.14 It is not yet clear to what extent the heterogeneity in reported dose response in those studies is related to differences in the accuracy of exposure measurement, to differences in the toxicity of complex mixtures of pollutants at differing levels of exposure, and/or differences in the sensitivity of exposed populations.
The exposure measures in other studies are not directly comparable. Recent studies use the more specific measure PM2.5 (particulate matter with an aerodynamic diameter less than 2.5 μm) rather than PM10, whereas several European studies use black smoke or total suspended particulates as the exposure measure. The association between particulate air pollution exposure and mortality is usually found to be strongest for finer fractions, such that the dose response for PM2.5 is greater than for PM10, which in turn is greater than for total suspended particulates. An extended follow-up of the US Six Cities Study reported a 14% (6% to 22%) increase in mortality per 10 μg/m3 PM2.5,9 whereas the American Cancer Society Study reported 6% (2% to 11%) and the Nurses Health Study 26% (2% to 54%), while coarse particulate matter exposure (PM10–2.5) was not associated with an increase in mortality in that study.27 An analysis based on electoral wards in the UK found a 1.3% (1% to 1.6%) increase in all-cause mortality per 10 μg/m3 increase in black smoke.10 In 18 regions in France, the PAARC study reported a 7% (3% to 10%) increase per 10 μg/m3 increase in black smoke and a 5% (2% to 8%) increase for total suspended particulates.7 In The Netherlands, there was a 5% (0% to 10%) increase in mortality per 10 μg/m3 increase in black smoke.
Our study assessed the association of PM10 and mortality over 3 years. In the US Nurses Health Study, mortality was most strongly associated with average PM10 exposures in the 24 months before death. In the UK, the association, although weak, was stronger for exposure in the 4 years before death. However, a re-analysis of the American Cancer Society Study found no clear effect of exposure period.17
A priori, we hypothesised that the association of PM10 with mortality in our study would be stronger for a cohort restricted to those census respondents who lived in the same census area at the time of the 1991 census. We also hypothesised that the association of PM10 with mortality would be stronger for cardiorespiratory deaths. Our results were consistent with these a priori hypotheses, strengthening the ability to make causal inference.
There is some evidence, most consistently in studies with individual measurements of social factors, that more deprived populations are particularly sensitive to air pollution effects.5 11 12 28 29 Our ability to detect a true difference by ethnicity in sensitivity to air pollution was limited by the relatively small Maori population. Although not significant, the difference in our central effect estimates for European and Maori was substantial: 7% versus 20%. The reasons for this difference, if real, are not clear. This might reflect a higher prevalence of pre-existing cardiorespiratory disease among Maori, or a difference in the toxicity of air pollution to which different ethnic groups are typically exposed. Alternatively, this finding may reflect biases in exposure estimates. We found no other suggestion of interaction of social factors with PM10 in the association with mortality.
In this longitudinal study we report an association of PM10 with mortality, consistent with that reported elsewhere, in a country with low levels of air pollution. The major limitation of our study is the probable misclassification of the PM10 exposure. On balance, this means we have probably underestimated the strength of association. The study design has several strengths, including national population coverage and good control of confounding. We found that the association was, as hypothesised a priori, stronger for cardiorespiratory deaths and people with less residential mobility.
Relatively few cohort studies of the health effects of urban air pollution have been published, particularly outside north America and Europe. We report here findings from a longitudinal study in New Zealand, based on the national census. There is evidence, most consistently in studies with individual measurement of social factors, that more deprived populations are particularly sensitive to air pollution effects.
We found an association of PM10 with mortality in a country with relatively low levels of air pollution. There was some evidence that Maori may have greater susceptibility to life-shortening effects of air pollution. This might reflect a higher prevalence of pre-existing cardiorespiratory disease among Maori.
The authors would like to thank June Atkinson for preparing the data for analysis.
Funding: The study was funded by the National Institute of Water and Atmospheric Research (NIWA). The New Zealand Census Mortality Study is conducted in collaboration with Statistics New Zealand and within the confines of the Statistics Act 1975. The New Zealand Census Mortality Study study was funded by the Health Research Council of New Zealand, and receives continuing funding from the Ministry of Health. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: None.
Contributors: TB conceived the analysis and contributed to the design, interpretation and revision of drafts. SH performed the analyses, drafted the text of the paper and contributed to the design and interpretation of results. SH had full access to the data, and takes responsibility for the integrity of the analysis and the accuracy of the results. AW contributed to study design, the interpretation of results and revision of drafts.
Provenance and peer review: Not commissioned; externally peer reviewed.
i: All counts were randomly rounded to a multiple of 3 in order to preserve confidentiality, as per Statistics New Zealand protocol. For this reason, subtotals may not match across categories. Note that regression analyses, however, were undertaken on actual unit record data.