Risk of Long-term Exposure to Fine Particulate Matter
Risk of Long-term Exposure to Fine Particulate Matter
Of the 2.1 million subjects, 49% were men (Table 1), and the mean age at baseline was 45.3 years. There were ~ 21.7 million person-years of follow-up in the cohort and nearly 200,000 nonaccidental deaths. Generally, concentrations of PM2.5 were highest in urban areas, along the corridor between Windsor and Quebec City, and in the prairies of southern Saskatchewan and Alberta (Figure 1). Among all subjects, the minimum concentration was 1.9 μg/m, the median was 7.4 μg/m, the mean was 8.7 μg/m, the maximum was 19.2 μg/m, and the interquartile range was 6.2 μg/m. The three ecological covariates (i.e., percent adults without a high school diploma, percent individuals in the lowest income quintile, percent unemployed adults) were negatively correlated with PM2.5 (Table 2).
(Enlarge Image)
Figure 1.
Mean satellite-derived estimates of PM2.5 across Canada, 2001–2006, and the mean concentrations in the 11 cities included in our subcohort analysis. P.E.I., Prince Edward Island.
Among the 11 Canadian cities for which ground-based observations were available, we found very similar patterns (i.e., r = 0.84) of concentrations of PM2.5 when comparing the 2001–2006 remote sensing-based observations (mean, 9.4 μg/m) with the 2001–2006 ground-based observations (mean, 8.9 μg/m) and with the 1987–2001 ground-based observations (mean, 11.2 μg/m; r = 0.89). We found almost identical associations for nonaccidental mortality in this subgroup of the cohort using the standard Cox model (adjusted only for individual-level covariates) with ground-based observations (HR for an increase of 10 μg/m = 1.11; 95% CI: 1.07, 1.15) and with satellite-derived estimates (HR 1.11; 95% CI = 1.09, 1.13). These latter results included subjects in only 11 cities and were very similar to the estimate for the entire cohort (standard Cox, individual-level covariates only, HR = 1.13; 95% CI: 1.12, 1.14).
We present in Table 3 the HRs and 95% CIs (from standard Cox and from spatial random-effects models adjusted for personal and contextual covariates) for the associations between PM2.5 and selected cardiovascular causes of death among subjects in the full cohort. The HR estimate for all nonaccidental causes from the fully adjusted standard Cox model was 1.15 (95% CI: 1.13, 1.16), and the corresponding HR from the random-effects model was 1.10 (95% CI: 1.05, 1.15). We estimated the strongest association with ischemic heart disease: the HR from the fully adjusted standard Cox model was 1.31 (95% CI: 1.27, 1.35), and the HR based on the random-effects model was 1.30 (95% CI: 1.18, 1.43). We estimated positive and almost identical associations with mortality from cardiovascular and circulatory disease using both model structures. We found no evidence of a linear association with cerebrovascular disease (standard Cox model: HR = 1.04; 95% CI: 0.99, 1.10; random-effects model: HR = 1.04; 95% CI: 0.93, 1.16). This lack of association for cerebrovascular mortality was also supported by the natural spline representation, which did not display a clear increasing mortality risk with PM2.5 concentrations (Figure 2D).
(Enlarge Image)
Figure 2.
Concentration–response curves (solid lines) and 95% CIs (dashed lines) based on natural spline (ns) models with 4 df, standard Cox models stratified by age and sex, adjusted for all individual-level covariates, urban/rural indicator, and ecological covariates. (A) Nonaccidental causes. (B) Cardiovascular disease. (C) Ischemic heart disease. (D) Cerebrovascular disease. The tick marks on the x-axis identify the location of the PM2.5 concentrations.
Modeling PM2.5 using natural splines did not improve model fit (based on BIC) relative to models that assumed linearity for nonaccidental, cardiovascular, or cerebrovascular deaths (Figure 2A,B,D). However, using a natural spline model with 4 df yielded lower BIC than other alternatives for ischemic heart disease mortality (Figure 2C).
Following Krewski et al. (2009), who reported that the logarithmic function of PM2.5 was a better predictor for mortality from ischemic heart disease than was the linear model in analyses of the ACS cohort, and based on the shape of the concentration–response curve (see Figure 2C), we fitted a model with ln(PM2.5 + 1). This model specification yielded a lower BIC than each of the spline representations. The mean ± SE coefficient for our logarithm model for ischemic heart disease was 0.3031 ± 0.0152 for the standard Cox survival model and 0.2894 ± 0.0440 for the random-effects survival model. This logarithmic model predicted a relative risk in ischemic heart disease mortality of 1.20 associated with a change in PM2.5 exposure from 5 μg/m to 10 μg/m and a relative risk of 1.12 based on the change in exposure from 10 μg/m to 15 μg/m.
Results
Of the 2.1 million subjects, 49% were men (Table 1), and the mean age at baseline was 45.3 years. There were ~ 21.7 million person-years of follow-up in the cohort and nearly 200,000 nonaccidental deaths. Generally, concentrations of PM2.5 were highest in urban areas, along the corridor between Windsor and Quebec City, and in the prairies of southern Saskatchewan and Alberta (Figure 1). Among all subjects, the minimum concentration was 1.9 μg/m, the median was 7.4 μg/m, the mean was 8.7 μg/m, the maximum was 19.2 μg/m, and the interquartile range was 6.2 μg/m. The three ecological covariates (i.e., percent adults without a high school diploma, percent individuals in the lowest income quintile, percent unemployed adults) were negatively correlated with PM2.5 (Table 2).
(Enlarge Image)
Figure 1.
Mean satellite-derived estimates of PM2.5 across Canada, 2001–2006, and the mean concentrations in the 11 cities included in our subcohort analysis. P.E.I., Prince Edward Island.
Subcohort Analysis
Among the 11 Canadian cities for which ground-based observations were available, we found very similar patterns (i.e., r = 0.84) of concentrations of PM2.5 when comparing the 2001–2006 remote sensing-based observations (mean, 9.4 μg/m) with the 2001–2006 ground-based observations (mean, 8.9 μg/m) and with the 1987–2001 ground-based observations (mean, 11.2 μg/m; r = 0.89). We found almost identical associations for nonaccidental mortality in this subgroup of the cohort using the standard Cox model (adjusted only for individual-level covariates) with ground-based observations (HR for an increase of 10 μg/m = 1.11; 95% CI: 1.07, 1.15) and with satellite-derived estimates (HR 1.11; 95% CI = 1.09, 1.13). These latter results included subjects in only 11 cities and were very similar to the estimate for the entire cohort (standard Cox, individual-level covariates only, HR = 1.13; 95% CI: 1.12, 1.14).
Main Findings
We present in Table 3 the HRs and 95% CIs (from standard Cox and from spatial random-effects models adjusted for personal and contextual covariates) for the associations between PM2.5 and selected cardiovascular causes of death among subjects in the full cohort. The HR estimate for all nonaccidental causes from the fully adjusted standard Cox model was 1.15 (95% CI: 1.13, 1.16), and the corresponding HR from the random-effects model was 1.10 (95% CI: 1.05, 1.15). We estimated the strongest association with ischemic heart disease: the HR from the fully adjusted standard Cox model was 1.31 (95% CI: 1.27, 1.35), and the HR based on the random-effects model was 1.30 (95% CI: 1.18, 1.43). We estimated positive and almost identical associations with mortality from cardiovascular and circulatory disease using both model structures. We found no evidence of a linear association with cerebrovascular disease (standard Cox model: HR = 1.04; 95% CI: 0.99, 1.10; random-effects model: HR = 1.04; 95% CI: 0.93, 1.16). This lack of association for cerebrovascular mortality was also supported by the natural spline representation, which did not display a clear increasing mortality risk with PM2.5 concentrations (Figure 2D).
(Enlarge Image)
Figure 2.
Concentration–response curves (solid lines) and 95% CIs (dashed lines) based on natural spline (ns) models with 4 df, standard Cox models stratified by age and sex, adjusted for all individual-level covariates, urban/rural indicator, and ecological covariates. (A) Nonaccidental causes. (B) Cardiovascular disease. (C) Ischemic heart disease. (D) Cerebrovascular disease. The tick marks on the x-axis identify the location of the PM2.5 concentrations.
Modeling PM2.5 using natural splines did not improve model fit (based on BIC) relative to models that assumed linearity for nonaccidental, cardiovascular, or cerebrovascular deaths (Figure 2A,B,D). However, using a natural spline model with 4 df yielded lower BIC than other alternatives for ischemic heart disease mortality (Figure 2C).
Following Krewski et al. (2009), who reported that the logarithmic function of PM2.5 was a better predictor for mortality from ischemic heart disease than was the linear model in analyses of the ACS cohort, and based on the shape of the concentration–response curve (see Figure 2C), we fitted a model with ln(PM2.5 + 1). This model specification yielded a lower BIC than each of the spline representations. The mean ± SE coefficient for our logarithm model for ischemic heart disease was 0.3031 ± 0.0152 for the standard Cox survival model and 0.2894 ± 0.0440 for the random-effects survival model. This logarithmic model predicted a relative risk in ischemic heart disease mortality of 1.20 associated with a change in PM2.5 exposure from 5 μg/m to 10 μg/m and a relative risk of 1.12 based on the change in exposure from 10 μg/m to 15 μg/m.
