2.3. Health Impact Analysis
The evaluation of health impact follows the impact pathway chain [
24] presented in
Figure 3. Using source-specific exposure estimates, the number of premature deaths was attributed to different emission sectors (see
Table 1).
Calculations were performed for the baseline in 2011 as well as for the 3 abatement strategies. Monte Carlo simulations were used to estimate the uncertainty range of number of premature deaths (see
supplementary S1 for source code). When aggregating uncertainties related to the different sources, the different ERFs were assumed to be independent. The margin of error for exposure estimates was assumed to be 20%. The baseline mortality for natural deaths representing the baseline year was acquired from the Swedish Cause of Death Register at The National Board of Health and Welfare.
In addition to the ERF for PM
2.5 and mortality recommended by WHO, there is a growing number of alternative assumptions. Data from the very large American Cancer Society’s (ACS) Cancer Prevention Study II (CPS-II) have been used in many influential studies of PM
2.5 concentrations and mortality, where most studies have used monitor data and described associations of metropolitan-level air pollution (“between-city contrasts”) and mortality [
25,
26]. In those studies, for many years dominating assessments of PM
2.5 effects, exposure data were derived at the metropolitan scale, relying on central monitor data. The increase HR in natural mortality using between-city contrasts in PM
2.5 has been estimated around 1.06 % per 10 µgm
−3. However, Jerrett et al. [
27] also used the ACS CPS-II data but constructed small-area exposure measures (using zip-code) in Los Angeles, California, by interpolation from 23 PM
2.5 monitors and observed effects nearly 3 times greater than in the models relying on comparisons between communities.
Turner et al. [
28] studied 669,046 participants from the ACS Cancer Prevention Study CPS -II with PM
2.5 concentrations estimated using a national-level hybrid land use regression and Bayesian maximum entropy interpolation model. Estimates of PM
2.5 were decomposed into near-source and regional components. Ozone and nitrogen dioxide concentrations were also modeled and included in the analyses. In the multi-pollutant model, the hazard ratio (HR) per 10 µgm
−3 for regional PM
2.5 became 1.04 (1.02–1.06), whereas for near-source PM
2.5 it was 1.26 (1.19–1.34).
In a cohort of 635,539 individuals from the US National Health Interview Survey (NHIS), Lefler et al. [
29] studied whether the PM
2.5–mortality relationship differs according to scale of spatial variability. Modeled air pollution exposure estimates for PM
2.5, other criteria air pollutants, and spatial decompositions (<1 km, 1–10 km, 10–100 km, >100 km) of PM
2.5 were assigned at the census tract-level. PM
2.5 mass was largely composed of regional and mid-range components, likely most secondary particles, while the neighborhood and local components contributed a relatively small fraction of PM
2.5 (23%). The PM
2.5–mortality association was observed across all 4 spatial scales of PM
2.5, with higher but less precisely estimated HRs observed for local (<1 km) and neighborhood (1–10 km) variations, scaled by 10 µgm
−3 1.299 (95% CI 1.014–1.664) and 1.279 (95% CI 1.173–1.395), respectively, from a joint model with all 4 scales. In a 2-pollutant model with total PM
2.5 and PM
2.5–10, the all-cause mortality HR associated with a 10 μgm
−3 increase in PM
2.5 was 1.12 (95% CI 1.09–1.15), whereas the HR associated with a 10 μgm
−3 increase in PM
2.5–10 was 1.02 (95% CI 1.00–1.04). In the most complex model with total PM levels (with no decompositions), the HR for an IQR increase in PM
2.5 (3.12) was 1.045 (95% CI 1.030–1.061) and in PM
2.5–10 it was 1.025 (95% CI 1.011–1.038) per IQR (5.43). This corresponds for the coarse fraction to 1.05 (95% CI 1.02–1.07) per 10 µgm
−3, and for PM
2.5 to 1.15 per 10 µgm
−3.
Spatial variation in particle levels within urban areas is commonly caused by local traffic emissions. When a meta-regression technique was used to investigate the heterogeneity between the studies and whether the study population or analytic characteristics modified the association between PM
2.5 and mortality, Vodonos et al. [
30] found that geographical locations with higher percent of PM
2.5 coming from traffic were significantly associated with higher estimates with and an HR 1.0205 (95% CI 1.0189–1.0181) per μgm
−3.
Published reviews often present quantitative summaries of effect size as estimated across studies regardless of the many differences in exposure levels and exposure assessment methods. However, the meta-regression technique used by Vodonos et al. [
30] described based on 53 studies with 135 estimated how the PM
2.5 coefficient decreased in a manner inversely proportional to the mean concentration, and when restricted to studies with mean concentrations below 10 μg/m
3, the meta-regression estimated an HR of 1.024 (95% CI 1.008–1.04) per 1 μgm
−3. Less error-prone exposure assessments and greater control for socioeconomic status were also factors associated with larger effect size estimates. Non-linear ERFs which level off at high concentrations have also been suggested when examining the shape of the association between PM
2.5 and non-accidental mortality applied in the Global Burden of Disease Study [
31]. The resulting Global Exposure Mortality Model (GEMM) builds on data from 41 cohorts from 16 countries but does not consider differences between exposure measures or particle sources.
It is becoming more and more apparent that the risk increase per µgm
−3 PM
2.5 is greater in areas with low total PM
2.5 concentrations, for local source contribution, and for traffic emissions than for regional PM
2.5. Meta-regressions showed that studies with more accurate exposure assessment methods reported larger effect size estimates for PM
2.5 [
32]. Furthermore, within large cohorts the scale of spatial variability in concentrations is important for estimated mortality HRs. Turner et al. [
28] observed a more than six times higher HR per absolute increase in concentration for near-source PM
2.5 in comparison with regional PM
2.5. Lefler et al. [
29] found similar patterns and concluded that regressions using spatially decomposed PM
2.5 suggest that more spatially variable components of PM
2.5 may be more toxic. Traffic [
30] and low PM
2.5 exposure [
30,
32] are factors associated with high HRs. The HR estimated for PM
2.5 and natural mortality in the Swedish SCAC study [
33] falls in the same range as reported by others with exposure variability driven by local sources and is based on the same exposure modeling as this health impact analysis (HIA). Many studies have concluded that there is no evidence of a threshold and that no safe level of PM can be determined [
4]. Some assessments have included a threshold to reflect insufficient data at low total concentrations. For this reason, a cut-off of 2 μgm
−3 for LRT PM
2.5 was applied, corresponding to the lowest exposure level with significant associations [
5], whereas no cut-off was applied for the anthropogenic local contributions.
Based on the above-mentioned literature, 3 different approaches for the HIA are compared:
In approach B, different HRs are applied for near-source and long-range (regional) contributions to PM
2.5. In
Figure 4, HRs from previously mentioned studies have been labeled according to the spatial resolution of the exposure data on which they are based, showing a clear tendency of higher risk estimates for within-city contributions to PM
2.5. An overview of the different HRs can be found in
Table A1. The study by Sommar et al. [
33] is based on the same exposure data as in this study. However, Turner et al. [
28] reported an HR for near-source PM
2.5 with a very similar value but a smaller uncertainty range and presented an HR for regional PM
2.5. Furthermore, the HR from Turner et al. was adjusted for NO2 and ozone. Therefore, the mortality HRs 1.26 (CI 95% 1.19–1.34) and 1.04 (CI 95% 1.02–1.06) reported by Turner et al. per 10 μgm
−3 were used for the local and regional contribution to PM
2.5, respectively.
Since only PM
2.5 was used as an indicator in approach A and B, changes in coarse PM (larger than 2.5 μm) were not reflected at all in the calculated health effects. This is problematic when evaluating abatement scenarios including significantly larger changes in exposure to coarse PM than to PM
2.5. In approach C, coarse PM was also included when estimating the impact by using PM
10 as an indicator for road wear PM with a HR by Sommar et al. [
33]. Furthermore, in this approach BC was used instead of PM
2.5 to estimate health effects related to vehicle exhaust. Since local emissions from vehicle exhaust are an important, often dominating, source of BC in urban areas [
34], this choice of indicator is more specific for vehicle exhaust than using HR based on bulk PM
2.5. The study by Sommar et al. [
33] was used since it includes the same geographical areas as this assessment and is based on exposure data with “within-city” contrasts resolved. For other sources of PM than road traffic, the same HR as in approach B was used.