Long-term Exposure to Fine Particulate Matter Air Pollution and Mortality Among Canadian Women

Abstract

Background: 

Long-term exposure to fine particulate matter (PM_2.5) has been associated with increased mortality, especially from cardiovascular disease. There are, however, uncertainties about the nature of the exposure–response relation at lower concentrations. In Canada, where ambient air pollution levels are substantially lower than in most other countries, there have been few attempts to study associations between long-term exposure to PM_2.5 and mortality.

Methods: 

We present a prospective cohort analysis of 89,248 women who enrolled in the Canadian National Breast Screening Study between 1980 and 1985, and for whom residential measures of PM_2.5 could be assigned. We derived individual-level estimates of long-term exposure to PM_2.5 from satellite observations. We linked cohort records to national mortality data to ascertain mortality between 1980 and 2005. We used Cox proportional hazards models to characterize associations between PM_2.5 and several causes of death. The hazard ratios (HRs) and 95% confidence intervals (CIs) computed from these models were adjusted for several individual and neighborhood-level characteristics.

Results: 

The cohort was composed predominantly of Canadian-born (82%) and married (80%) women. The median residential concentration of PM_2.5 was 9.1 μg/m³ (standard deviation = 3.4). In fully adjusted models, a 10 μg/m³ increase in PM_2.5 exposure was associated with elevated risks of nonaccidental (HR: 1.12; 95% CI = 1.04, 1.19), and ischemic heart disease mortality (HR: 1.34; 95% CI = 1.09, 1.66).

Conclusions: 

The findings from this study provide additional support for the hypothesis that exposure to very low levels of ambient PM_2.5 increases the risk of cardiovascular mortality.

Exposure to air pollution particulate matter with an aerodynamic diameter of less than 2.5 μm (PM_2.5) has been identified as one of the top 10 health risks worldwide.¹ Both short-term episodic increases in pollution levels as well as long-term exposure to elevated air pollution concentrations adversely affect human health, although the overall mortality impact from longer term exposures is substantially larger.² Long-term exposure to fine particulate matter (PM_2.5) has consistently been associated with an increased risk of cardiovascular disease and nonaccidental mortality. These associations first gained widespread attention from findings reported in two US cohorts: the American Cancer Prevention Cancer Prevention II (ACS-CPS II)³ and the Harvard Six Cities⁴ studies. Since then, other US cohorts,^5–7 as well as some from Europe^8,9 and Canada,¹⁰ have reported positive associations between residential measures of PM_2.5 and mortality. A systematic review by Chen et al.¹¹ concluded that a 10 μg/m³ increase in PM_2.5 increased the risk of nonaccidental and cardiovascular mortality by approximately 6%, and cardiovascular disease by between 12% and 14%.

In urban areas, the concentrations of ambient PM_2.5 in Canada are among the lowest in the world.¹² Crouse et al.¹⁰ recently investigated the association between PM_2.5 and mortality using mandatory census data collected from 20% of Canadian households (in the Canadian Census Health and Environment Cohort [CanCHEC]) and found that their residential median exposure to PM_2.5 was 7.4 μg/m³. In the CanCHEC cohort of 2.1 million adults, who were individually linked to mortality data, the hazard ratios (HR) in relation to a 10 μg/m³ increase in PM_2.5 were 1.15 (95% confidence interval [CI] = 1.13, 1.16) for nonaccidental mortality and 1.31 (95% CI = 1.27, 1.35) for ischemic heart disease mortality. As highlighted in an accompanying editorial in the same journal issue,¹³ the findings observed in this Canadian cohort indicated that “exposure even to very low amounts of PM_2.5 over long periods may pose a greater risk to human health than originally assumed.” The exposure–response plots produced by Crouse et al.,¹⁰ which demonstrated associations with mortality at less than 7 μg/m³, have important implications for the comparative analyses undertaken in the Global Burden of Disease initiative where the calculations assumed no adverse mortality impacts from PM_2.5 at these low concentrations.¹

We sought to evaluate in a second Canadian cohort whether there were similar associations between long-term exposure to PM_2.5 and cause-specific mortality. This was done using a prospective cohort of Canadian women who enrolled in a randomized trial of breast cancer screening between 1980 and 1985,¹⁴ and whose mortality was ascertained through 2005. Individual-level risk factor data were collected from all women at baseline. Apart from their role as possible confounders, individual-level characteristics are also important to identify individuals who may be more susceptible to the long-term effects of PM_2.5. Hoek et al.¹⁵ contend that findings from previous cohort studies of air pollution indicate that education, obesity, and sex modify associations between exposure to fine particulate matter and air pollution. Moreover, Puett et al.¹⁶ recommend that the next generation of longitudinal studies should try to identify the characteristics that make individuals more susceptible to the harmful effects of air pollution. The size and scope of this Canadian cohort of women is similar to that of the US Nurses’ Health Study,¹⁶ which has provided important insights about the impact of long-term exposure to air pollution on health.

METHODS

Study Population

Our study population included participants of the Canadian National Breast Screening Study (CNBSS), which was a randomized controlled trial of screening for breast cancer. Its study design has been described in detail elsewhere.^17,18 In brief, the study included 89,835 women who were recruited from the general population between 1980 and 1985, and were between 40 and 59 years of age. Women were healthy and free of cancer at the time of enrollment. We used a self-administered questionnaire to collect information on their demographic, reproductive, and lifestyle characteristics. Participants’ height, weight, and skinfold thickness were measured by a trained health professional. Body mass index (BMI) was calculated using these weight and height measurements. Residential address information was obtained at baseline, and six character postal codes were collected for all participants. In urban areas, six-character postal codes represent one block face between two intersecting streets.

The participants of the study provided informed consent. Approval for this current study was provided by Health Canada’s Research Ethics Board.

Exposure to Fine Particulate Matter Air Pollution

To assign long-term exposure to ambient air pollution, we used satellite-based estimates of surface concentrations of PM_2.5.^19,20 The satellite-based concentrations were derived from aerosol optical depth data from the Moderate Resolution Imaging Spectroradiometer, Multiangle Imaging Spectroradiometer, and Sea-viewing Wide Field-of-view Sensor instruments.^19,20 Using satellite data collected from January 1, 1998 to December 31, 2006, long-term average concentrations of PM_2.5 were derived at a resolution of approximately of 10 km × 10 km. These measures cover nearly all of North America except for some coastal and mountain areas where retrieval of images is not possible. The sampling effects of snow- and cloud-cover on satellite-derived PM_2.5 estimates are adjusted for using the temporal variation simulated by the GEOS-Chem chemical transport model, as represented by the ratio of mean PM_2.5 simulated during successful satellite retrievals, and PM_2.5 simulated during the entire time period.¹⁹ These satellite-based estimates of long-term average concentrations of PM_2.5 have been shown to correlate well with ground-based measurements at fixed site stations across North America (Pearson correlation coefficient r = 0.79, slope = 1.0, n = 1,002).²⁰ Similar satellite-based exposure surfaces have been applied previously to examine the associations between long-term exposure to air pollution and mortality.^10,21 For participants in the CNBSS, residential measures of exposure were determined by mapping the geospatial coordinates of the centroid of each postal code to the PM_2.5 surface.

Ascertainment of Mortality

We determined the vital status of each participant using a probabilistic record linkage to the Canadian Mortality Database through the end of 2005. This database provides data on all deaths occurring in Canada, as well as those among Canadian residents that occur in approximately 20 US states. The accuracy of identifying deaths has been estimated to be approximately 98%.²² Until 1999, the underlying cause of death was coded using the ninth revision of the International Classification of Diseases (ICD-9); from 2000 onward the 10th revision was used. We were interested in characterizing associations between long-term exposure to PM_2.5 and several underlying causes of death including: coronary heart disease (ICD-9: 410–414; ICD-10: I20–I25), cerebrovascular disease (ICD-9: 430–438; ICD-10: I60–I69), and all cardiovascular diseases combined (ICD-9: 400–440; ICD-10: I00–I99). We also examined mortality from all nonaccidental causes (ICD-9: <800; ICD-10: A–<V), nonmalignant respiratory (ICD-9: 460–519; ICD-10: J00–J99), cancer (ICD-9: 140–239; ICD-10: C00–C99), and lung cancer (ICD-9: 162, ICD-10: C33–C34).

Statistical Analyses

Initially, we mapped the place of residence data supplied by the participants, and tabulated the number of deaths for each underlying cause under investigation. Descriptive statistics were compiled across socio-demographic, occupational, and risk factor characteristics. Associations between residential concentrations of PM_2.5 and cause-specific mortality were investigated using the Cox proportional hazards model using calendar time as the time axis. A continuous measure of PM_2.5 was included in this model and the resulting HRs and their 95% confidence limits were scaled in relation to a 10 μg/m³ increase in PM_2.5. We chose to present HRs across a series of models so we could better understand the influence that these covariates had on the PM_2.5-mortality associations.

First, we fit a simplified model that contained PM_2.5 and age only. We then expanded this simplified model to include the same covariates used by Crouse et al.¹⁰ in their analysis of the CanCHEC. This was done in two steps where the first model included the covariates of marital status, occupation, and education attained. The second model included these covariates as well as neighborhood characteristics obtained from 1991 Canadian census data including: mean household income, proportion of individuals with high school education, percentage of low income households, and unemployment rate. Finally, we extended this model to include self-reported measures of cigarette smoking, and BMI determined from measures taken from trained staff. These covariates were not available in the CanCHEC,¹⁰ and adding them at this final stage was done to provide some insights on their possible confounding roles over and above that accounted for with other socio-demographic characteristics, and to gauge the importance (for future studies) of capturing direct measures of smoking status and other individual measures. HRs and their 95% CIs were generated for the each of the different causes of death.

We then conducted additional analyses to determine whether associations of air pollution with these causes of death were modified by place of birth (in Canada or elsewhere), whether participants had moved during the first phase of follow-up (between 3 and 5 years after baseline interview), obesity (BMI > 30 kg/m²), or whether they smoked tobacco. The p values associated with the first-order interaction terms were used to evaluate whether associations were statistically significant across strata.

We also modeled concentration–response functions to better understand the nature of the relation between PM_2.5 and mortality. This was done by modeling a natural cubic spline functions with three degrees of freedom within the fully adjusted Cox regression model. Plots of these dose–response functions were made using the R software program. These plots were made for nonaccidental mortality, cardiovascular mortality, cancer mortality, and ischemic heart disease mortality. A formal threshold analyses was undertaken to determine whether a threshold existed for these four causes of death. This was done using the approach previously outlined by Jerrett et al.²³ Specifically, for various thresholds (2,3, …, and 14 μg/m³ of PM_2.5), we assumed that there was no association between exposure to PM_2.5 and mortality below the specified threshold, and a linear association (on the logarithmic scale of the proportional hazards model) above the threshold. The fit of the models was compared using -2 × log likelihood values, and for each cause of death, we compared the best fitting model (ie, that with lowest -2 × log likelihood statistic) to the no-threshold model to determine whether the improvement in fit was statistically significant.

Finally, to investigate whether the HRs might be influenced by any possible spatial dependence, we included in our model a frailty term (random effect) for census division, identified using the SAS procedure PHREG.²⁴ This approach has been used by others who have investigated associations between long-term exposure to air pollution and health.^25–27 However, due to the relatively small number of deaths these models would not converge.

RESULTS

Of the 89,835 women in the Canadian National Breast Screening Study, residential measures of PM_2.5 could be assigned at an individual level to 99% of them (n = 89,248). Among our study participants, the median exposure to ambient PM_2.5 was 9.1 μg/m³ (standard deviation = 3.4), and the 25th and 75th percentiles were 6.4, and 12.4 μg/m³, respectively. The PM_2.5 concentrations ranged from a low of 1.3 to a high of 17.6 μg/m³. Participants’ places of residence at the time of enrollment along with their corresponding estimates of ambient PM_2.5 are depicted in Figures 1 and 2, respectively. In total, 9,419 women died during the 25-year follow-up (Table 1). Of these, 1,845 were from cardiovascular disease, whereas 5,233 were from cancer.

TABLE 1: Number of Deaths in the Canadian National Breast Screening Study, by Underlying Cause

FIGURE 1: Map showing the place of residence of the 89,292 women, with available PM_2.5 concentrations, who participated in the Canadian National Breast Screening Study.

FIGURE 2: Satellite-derived estimates of PM_2.5 concentrations for Canada, 1998–2006.

The mean age of the participants was 48.5 years (standard deviation = 5.6). Most women were Canadian born (82.2%) and married (79.8%; Table 2). Almost half (49.4%) of the women in the cohort had never smoked cigarettes. Many of these characteristics were associated with concentrations of PM_2.5. For example, on average, higher PM_2.5 values were observed among women who were: older, foreign-born, had higher educational attainment, had never married, and had a lower BMI (Table 2).

TABLE 2: Descriptive Characteristics of the 89,248 Participants of the Canadian National Breast Screening Study Whose Residential Exposure to PM2.5 Could Be Estimated

The HRs for long-term exposure to PM_2.5 exceeded unity for all causes of death that we examined except for nonmalignant respiratory mortality (Table 3). The highest fully adjusted HR was for ischemic heart disease (HR = 1.34; 95% CI = 1.09, 1.66). The HRs changed little following adjustment for marital status, occupation, and education. However, the inclusion of contextual measures obtained from Canadian census data amplified the associations between PM_2.5 and cardiovascular mortality. For example, the HR for ischemic heart disease increased from 1.26 (95% CI = 1.03, 1.53) to 1.31 (95% CI = 1.06, 1.62). The HRs increased only slightly upon further adjustment for individual-level factors of obesity and smoking status. After adjusting for age and other individual and neighborhood-level covariates, each increase of 10 μg/m³ was estimated to increase the risk of nonaccidental mortality by 12% (HR = 1.12; 95% CI = 1.04, 1.19). After adjusting for cigarette smoking, we observed no association between PM_2.5 and lung cancer mortality (HR = 0.97; 95% CI = 0.80, 1.18).

TABLE 3: Adjusted Mortality HRs and 95% CIs in Relation to an Increase of 10 μg/m3 Increase in PM2.5, by Underlying Cause of Death, Canadian National Breast Screening Study, 1980–2005

The HRs for PM_2.5 were generally higher among those who were obese (BMI ≥ 30) when compared with those who were not for most causes of death (Table 4). For cardiovascular mortality outcomes, there was little difference in the risk estimates between those who were born in Canada (HR = 1.34; 95% CI = 1.15, 1.57) and those who were not (HR = 1.33; 95% CI = 0.95–1.85). Similarly, HRs were similar among individuals who had stayed in the same residence during the first few years of the study compared with those who had moved.

TABLE 4: Adjusted Mortality HRsa and 95% CIs in Relation to an Increase of 10 μg/m3 Increase in PM2.5, by Underlying Cause of Death, by Body Mass Index, Canadian National Breast Screening Study 1980–2005

Concentration–response plots for cancer, ischemic heart disease, cardiovascular disease, and nonaccidental mortality are provided in Figure 3. We observed a nonlinear V-shaped pattern with nonaccidental mortality and to a lesser extent with cancer mortality. In contrast, the exposure–response plots for cardiovascular and ischemic heart disease were consistent with a linear trend. Our threshold analyses revealed that there was no statistically significant improvement in fit over a no-threshold linear model for any of the thresholds considered for ischemic heart disease, cardiovascular, or cancer mortality (Table 5). In contrast, for nonaccidental mortality, the best fitting model was the one with a PM_2.5 threshold of 11 μg/m³. The difference in the -2 × log likelihood between this model and the no-threshold model was 8.17 and based on a χ²₍₁₎ yielded a p value of 0.004.

TABLE 5: -2 × Log Likelihood Values Based on a Threshold Concentration Response Model^a for PM2_{.5 A}gainst Selected Causes of Death

FIGURE 3: Exposure–response relation between ambient concentrations of PM_2.5 and selected causes of death.

DISCUSSION

In this longitudinal study of 82,248 Canadian women, we found positive associations between ambient PM_2.5 and nonaccidental and cardiovascular mortality. The strength of these associations was remarkably similar to those published by Crouse et al.¹⁰ who reported in a separate large-scale Canadian cohort study an approximately 30% increased risk of cardiovascular mortality and 10%–15% for nonaccidental mortality in relation to a 10 μg/m³ increase in PM_2.5. In addition, our analyses suggest that obesity and smoking do not explain previous associations between PM_2.5 and mortality in Canada, and support the validity of these findings published by Crouse et al.¹⁰ despite the fact that they could not control for these effects directly.

There are, however, several important differences between these two studies. First, our cohort consisted exclusively of women, who were, by and large, married, white, and of upper socioeconomic status. In contrast, because one in every five households was sampled in the long-form mandatory census, the CanCHEC is truly nationally representative, although it is important to note that Crouse et al.¹⁰ excluded those who were not born in Canada. Second, with respect to exposure assignment, we were able to link ambient PM_2.5 measures to participants’ places of residence using six-character postal codes. In contrast, the approach used in the CanCHEC analysis relied on somewhat coarser spatial measures (census enumeration areas). The consistency of the risk estimates between studies suggests that the specific method used to assign exposure does not play a strong role in determining the magnitude of the PM_2.5 mortality effect.

The associations between PM_2.5 and the four selected causes of death examined (nonaccidental, ischemic heart, cancer, and cardiovascular) were all stronger among obese women than nonobese women. A recent review of 14 panel studies with short-term exposure measures of PM_2.5 and 3 cohort studies of long-term measures concluded that obesity may increase susceptibility to cardiovascular health effects.²⁸ A similar finding to that observed in our cohort was noted in the prospective study of 66,250 women enrolled in the US Nurses’ Health Study. This study found that the HR for cardiovascular disease associated with a 10 μg/m³ increase in PM₁₀ was 1.99 (95% CI = 1.23, 3.22) for women with a BMI greater than 30, but was only 1.08 (95% CI = 0.76, 1.52) for women with a BMI of less than 30.⁷ Evidence of effect modification by obesity was also observed among women in the Women’s Health Initiative,²⁹ and more recently among men in the US Agricultural Health Study cohort.³⁰

A limitation of our study is the inevitability of some exposure misclassification introduced by assigning PM_2.5 concentrations to participants’ residences at the time of enrollment using a surface that was generated using exposure data collected between 1998 and 2006. Given the length of the follow-up interval, PM_2.5 concentrations would have changed both temporally and spatially. However, analyses of Canadian data suggest that the spatial gradients in ambient PM_2.5 have remained fairly stable over time. For example, Chen et al.²⁶ examined PM_2.5 concentrations in Ontario over the period 1996–2010 and found that the variability in the concentrations in PM_2.5 was primarily spatial and not temporal. In addition, a number of studies undertaken in diverse locations across the United States have observed that spatial patterns in PM_2.5 have remained stable over the long-term.^{31, 29, 32} For these reasons, we think it reasonable to expect that the spatial contrasts in PM_2.5 during 1998–2006 are fairly reflective of spatial exposure to PM_2.5 in Canada during our follow-up interval.

There have been a number of advancements in the use of models to estimate residential-based exposure to fine particulate matter. These have included attempts to combine data from fixed site monitors, satellite data, and land-use regression surfaces to improve estimates of exposure to PM_2.5.³³ There is ongoing work to try and further improve the exposure surfaces in Canada by developing these fused surfaces. However, there are unique challenges to developing a Canadian fused surface given that there are few gravimetric PM_2.5 monitors in Canada, and ground-based measures rely largely on tapered element oscillating microbalancemonitors that have major artifacts during cold periods (approximately half the year). As a result, we have little reliable ground data to use in fusing. In addition, Canada has a relatively sparse population and there are large distances between monitors, so spatial interpolation models are severely limited and thus it is difficult to develop a reliable fusing method. The fused surface developed by Brauer et al.³³ for use in the Global Burden of Disease project performs poorly for concentrations less than 10 μg/m³. Specifically, for their fused model when concentrations less than 10 μg/m³ were calibrated to ground data, a single equation for the entire world was used, and a random value for concentrations less than 10 μg/m³ was assigned. At this time, most areas in Canada fall below 10 μg/m³. At this time, we feel that the remote sensing-derived estimates of PM_2.5 that we used are a valid and best available measure of ground-level Canadian PM_2.5 concentrations.

The residential mobility of the women in the cohort is a potential source of exposure misclassification. In addition to knowing the home address of the women at the time of enrollment, due to the nature of the study (evaluating screening for mammography), unlike many other air pollution cohort studies, we were able to determine whether they moved during the first 4 years of the study. This allowed us to evaluate whether there were differences in risk among the 19% of women who moved during this time window and those who did not. The associations for nonaccidental and ischemic heart disease mortality were virtually identical in the movers and nonmovers, thereby suggesting that associations are not unduly affected by residential mobility. Moreover, in our view, there is an equal likelihood among women in the cohort being assigned an overestimate or underestimate of true exposure, which means that our findings may be subject to nondifferential misclassification bias, suggesting the risks we report may be underestimated.

The use of a residential-based approach to assigning exposure to PM_2.5 may also introduce some exposure misclassification as individuals will move throughout the city during most days. However, data from a recent time-activity survey of Canadian adult women indicate that, on average, they spend approximately 70% of their time at home.^34,35 In addition, concentrations of ambient PM_2.5 are far less variable than other pollutants in Canadian cities.^36,37 For example, in Hamilton the 25th and 75th percentiles of PM_2.5 obtained from a dense summer sampling campaign involving 43 monitors throughout the city were 6.10 and 6.95 μg/m³.³⁶ Levy et al.³⁷ also found that there was little variation in the median PM_2.5 concentrations across 10 different areas in Montreal.

Although we noted a positive association between PM_2.5 and cardiovascular disease, we noted no such association with nonmalignant respiratory disease. This is consistent with recent analyses of 16 European cohort studies that included 307,553 participants.³⁸ Similarly, the US cohort studies^5,39,40 have not generally found associations between PM₁₀ and PM_2.5 with nonmalignant respiratory disease. It has been suggested that the lack of association observed in some studies may reflect the fact that the study populations did not include vulnerable individuals, such as the elderly. However, we observed no increased risks of respiratory mortality in any of the sub-analyses we performed including those who were older, obese, or smokers. Given that those with nonmalignant respiratory disease tend to have comorbid conditions including cardiovascular disease, the absence of a positive association between PM_2.5 with nonmalignant respiratory disease may reflect an influence from these competing risks. Additional studies that are able to examine contributing causes of death, in addition to underlying cause, or those that can evaluate incident disease are encouraged, to provide additional insights.

Our findings provide further support that low-level concentrations of ambient PM_2.5 are associated with increased mortality for cardiovascular disease, and to a lesser extent nonaccidental mortality. The authors of the recent Global Burden of Disease report assumed that there is no evidence for, or against, associations between mortality and PM_2.5 below 5.8 μg/m³ and stated that they were unsure whether associations could be detected empirically below concentrations of 8.8 μg/m³.⁴¹ They therefore created in their counterfactual scenario an uncertainty distribution using a uniform random variables U(5.8, 8.8). Our findings are consistent with those of Crouse et al.¹⁰ and suggest that long-term exposure to ambient PM_2.5 increases the risk of cardiovascular mortality below concentrations generally assumed to pose little threat to human health. They also suggest that existing methodologies⁴¹ that quantify the impact of ambient air pollution on cardiovascular endpoints may understate the health burden. In contrast, our analysis of nonaccidental mortality suggests that increases in risk are evident at a threshold value of approximately 11 μg/m³.