The public health context for PM2.5 and ozone air quality trends
- First Online:
- Cite this article as:
- Fann, N. & Risley, D. Air Qual Atmos Health (2013) 6: 1. doi:10.1007/s11869-010-0125-0
- 647 Views
Tropospheric ozone (O3) and particulate matter (PM2.5) are associated with adverse health effects, including premature mortality. Regulation of these pollutants by the US Environmental Protection Agency has resulted in significant improvements in air quality over the last decade, as demonstrated by a national network of air quality monitors. However, ambient trends provide limited information regarding either the change in population exposure to these pollutants or how fluctuations in the levels of these pollutants might affect public health. We leverage the spatially and temporally extensive monitoring network in the US to estimate the improvements to public health associated with monitored air quality changes over a 7-year period. We estimate the impacts of monitored changes in ozone and PM2.5 on premature mortality using health impact functions based on short-term relative risk estimates for O3 and long-term relative risk estimates for PM2.5. We spatially interpolate the O3 and PM2.5 data and utilize ozone air quality data that are adjusted for meteorological variability. We estimate that reductions in monitored PM2.5 and ozone from 2000 to 2007 are associated with 22,000–60,000 PM2.5 and 880–4,100 ozone net avoided premature mortalities. The change in estimated premature mortality can be highly variable from 1 year to the next, sometimes by thousands of deaths. The estimate of avoided ozone-related mortalities is sensitive to the use of meteorologically-adjusted air quality inputs. Certain locations, including Los Angeles and Houston see an opposing trend between mortality impacts attributable to ozone and PM2.5.We find that improving air quality over the past 7 years has reduced premature mortality significantly.
KeywordsAir quality trendsOzonePM2.5Health impacts
The US Environmental Protection Agency (EPA) sets National Ambient Air Quality Standards (NAAQS) for six common “criteria” pollutants (PM, ozone, Pb, CO, NOx, and SO2) at levels adequate to protect public health and the environment. A national network of air quality monitors measures the ambient levels of each pollutant to determine if a community meets each of the six NAAQS. Observations from this national network are also used to characterize trends in the daily and annual changes of these pollutants (US EPA 2008a). Information about long-term air quality trends can be useful in determining the extent to which EPA’s air quality management program is helping reduce concentrations of pollutants to the levels specified by the health-based standards.
However, ambient trends alone provide limited information regarding the public health burden of air pollution. The overall public health burden of pollution levels is governed by the combination of air pollution levels, population exposure to those pollutants, and population susceptibility to air pollution health impacts. Recognizing this, we employ the health impact assessment (HIA) approach, which combines changes in air quality, population exposure, baseline health status, and concentration–response relationships drawn from the epidemiological literature to relate changes in monitored air quality to the incidence of adverse health outcomes. In this paper, we use the HIA technique to estimate impacts attributable to changes in two key ambient pollutants: fine particulate matter (PM2.5) and ground-level ozone (O3). In particular, we assess the level of premature mortality incurred or avoided as a result of changes in the level and distribution of PM2.5 and O3 air quality nationwide.
HIA’s are a well-established approach for estimating the retrospective or prospective change in adverse health impacts resulting from population-level changes in exposure to pollutants (Levy et al. 2009; Tagaris et al. 2009; Hubbell et al. 2009). PC-based tools such as the environmental Benefits Mapping and Analysis Program (BenMAP) can systematize health impact analyses by applying a database of key input parameters, including concentration–response functions, population projections, and baseline incidence rates. Analysts have applied the HIA approach to estimate health impacts resulting from hypothetical changes in pollutant levels (Hubbell et al. 2005; Davidson et al. 2007; Tagaris et al. 2009). EPA and others have relied upon this method to predict future changes in health impacts resulting from the implementation of regulations affecting air quality (US EPA 2008b). However, to our knowledge, this technique has not been used to quantify the health impacts occurring from long-term historical changes in monitored ambient levels of pollutants. Applying the HIA approach in this way can yield important information regarding the public health impacts associated with historical ambient levels of key criteria pollutants—and in so doing, serve as one measure of the effectiveness of EPA air quality programs.
The article is structured as follows: first, we describe the process for estimating air pollution health impacts. This discussion includes a description of how we created air quality surfaces from monitored ozone and PM2.5 levels for the period between 2000 and 2007, estimated population-level exposures, and applied health impact functions to quantify the change in ozone and PM2.5-related premature mortality. Next, we summarize the results of our health impact analysis, presenting both the cumulative and year-to-year changes in premature mortality occurring between 2000 and 2007. We then discuss the sensitivity of our results to assumptions regarding the air quality inputs. Finally, we consider the policy relevance of our findings.
Materials and methods
Overview of the health impact analysis
The HIA approach used here involves three basic steps: (1) estimating the spatial distribution of ambient air quality changes resulting from past changes in air quality, (2) determining the change in population-level exposure, and (3) calculating health impacts by applying concentration–response relationships found in the epidemiological literature (Hubbell et al. 2009) to this change in population exposure. As we discuss further below, we hold constant key factors such as population levels, baseline mortality rates, and other variables to better isolate the impact of monitored air quality changes on public health.
Of the six criteria pollutants, we limited our health impact assessment in this analysis to PM2.5 and ozone for two reasons. First, among the criteria pollutants, the monitoring network for PM2.5 and ozone is most extensive, particularly within highly populated urban areas. As we discuss further below, the ozone and PM2.5 monitoring networks provide the greatest spatial coverage, allowing us to better estimate national-level population exposures. Second, in part because of the national coverage of monitoring data, ozone and PM2.5 have been the subject of large-scale multi-city and long-term cohort studies (Bell et al. 2004; Krewski et al. 2009) this increases our level of confidence in estimating national-level health impacts. We discuss this issue in greater depth below. Next, we describe how we performed the health impact analysis, beginning with the selection of air quality data.
Observed changes in ambient air quality levels
Monitored O3 and PM2.5 concentrations are from the US EPA’s Air Quality System (EPA 2008c) for the years 2000—2007. For O3, we use concentrations from May 1 through September 30 to be consistent with EPA determination of O3 NAAQS attainment. We utilized monitors that both included observations for at least one half of this 153-day period and included at least 9 hourly measurements from 0800 to 1950 h. These criteria reflect our attempt to balance between selecting a sufficient number of monitors that provide adequate spatial and temporal coverage, while excluding those that that, due to incomplete data, may introduce significant bias to the analysis.
PM2.5 monitors generally capture daily mean PM2.5 concentrations every third or sixth day throughout the year. We selected PM2.5 monitors that included at least 11 observations each quarter out of a possible 15 (assuming at least once every 6-day monitoring). We did not exclude any PM2.5 or ozone data for exceptional events and included only those monitors classified by states as Federal Reference Method or Federal Equivalence Method.
Number of PM2.5 and ozone monitors in the continental US over time
Number of monitors
We interpolated the PM2.5 and ozone monitor concentrations to create a 12 × 12 km gridded spatial surface of concentrations for each year from 2000 to 2007. Interpolation was performed using Voronoi neighbor averaging (VNA), which estimates ambient concentrations by selecting the closest neighboring monitors surrounding the center of each grid square and then calculates the inverse distance weighted average of the monitor values for the selected neighboring monitors (Chen et al. 2004; Gold 1997). Predicted values are more reliable in urban areas due to the relatively high density of monitors (Hubbell et al. 2005) and we aggregated our health impact estimates to a 100-km radius around each ozone and PM2.5 monitor (this includes approximately 96% of the total US population). We test the sensitivity of our results to this approach by counting impacts within 50 km, 200 km, and an unlimited distance of each monitor, finding that the results are generally insensitive to the size of the radius (Supplement Fig. S2). When generating an air quality surface to estimate population exposure for ozone and PM2.5, we quantified changes in the daily 8-h maximum concentration and the annual mean, respectively.
The question of whether trends in ozone concentrations should be adjusted for meteorology hinges in part on the policy objective of the analysis—applying unadjusted ozone values in the health impact assessment more directly relates observed air quality changes with health impacts. Conversely, applying ozone data adjusted for meteorology may provide more policy-relevant information regarding the overall direction of key ozone-related health impacts. Below, we report estimated health impacts using the meteorologically-adjusted ozone values and explore the influence of this adjustment in a sensitivity analysis. No peer-reviewed approach exists to account for the influence of meteorology in PM2.5 concentrations and so we do not adjust these air quality values.
Estimated change in population exposure
Next, we estimate how much the population was exposed to each pollutant. BenMAP contains US Census block-level populations (Geolytics Inc. 2002) which it then aggregates up to a user-specified modeling domain—in this case, a national 12 km modeling domain. The software stratifies the population by the relevant age/sex/race/ethnicity categories that correspond to the demographic classifications considered in the health impact functions (see below). The 2000 population of each group is then projected to the relevant analysis year in five year increments, using an economic forecasting model (Woods and Poole Economics Inc. 2001). Given that this analysis considers health impacts occurring in each year between 2000 and 2007, we selected a single projection year of 2005 to control for the influence of population growth over space and time and changes in the demographic composition of the population; this approach also helps isolate the impact of air quality changes alone.
Calculated ozone and PM2.5-related premature mortality
We calculate changes in premature mortality using a health impact function, combining each of the above elements with estimates of baseline incidence and effect coefficients to quantify the avoided (or incurred) incidence of adverse health outcomes. The US EPA PM2.5 Integrated Science Assessment (US EPA 2008d) and Ozone Criteria Document (US EPA 2006a) summarize an extensive number of epidemiological studies that quantify the changes in population risk for a variety of adverse health outcomes. These adverse health outcomes include premature mortality, chronic bronchitis, acute myocardial infarctions, respiratory and cardiovascular hospitalizations, exacerbation of asthma, work loss days and school loss days, among others. For the purposes of this analysis, we elect to assess the change in PM2.5 and ozone-related premature mortality alone for two key reasons: (1) the epidemiological studies estimating PM2.5 and ozone-related mortality generally observe large populations in a large number of urban areas. Conversely, the epidemiological studies assessing morbidity impacts from ozone and PM2.5 exposure tend to consider populations living in a single city or some small number of cities. Given our analytical goal of estimating national-level health impacts, we elect to assess only mortality as a means of reducing the overall level of uncertainty in the analysis. (2) Premature mortality is the most severe health impact associated with PM2.5 and ozone air pollution exposure, and so arguably serves as the most relevant indicator of the public health impacts associated with air pollution trends. By contrast, EPA Regulatory Impact Analyses (RIA) utilize risk coefficients from an array of single- and multi-city studies to assess a variety of morbidity impacts including non-fatal heart attacks, hospitalizations and emergency department visits. This approach is consistent with the goal of the RIA, which is to characterize as completely as possible the benefits and costs of regulatory actions.
When estimating PM2.5-related mortality we select effect coefficients drawn from epidemiological studies based on data from two prospective cohort groups, often referred to as the Harvard Six-Cities Study, or “H6C” (Dockery et al. 1993; Laden et al 2006) and the American Cancer Society “ACS” study (Pope et al. 1995, 2002, 2004; Krewski et al. 2009); these studies have found consistent relationships between fine particle indicators and premature mortality across multiple locations in the USA.
We select an all-cause mortality risk estimate drawn from the Krewski et al. (2009) extended analysis of the ACS cohort. In particular, we apply an estimate based on the random effects Cox model that controls for 44 individual and seven ecological covariates, using average exposure levels for 1999–2000 over 116 US cities (RR = 1.06, 95% confidence intervals 1.04—1.08 per 10 μg/m3 increase in PM2.5). We elect to quantify all-cause mortality because among the strongest evidence exists for PM-related all cause mortality. We also select an all-cause mortality risk estimate drawn from the Laden et al. (2006) reanalysis of the Harvard Six cities cohort (RR = 1.16 , 95% confidence intervals 1.07—1.26 per 10 μg/m3 increase in PM2.5).
There are strengths and weaknesses inherent in each of these two PM mortality studies. For this reason, we use risk estimates drawn from both analysis when quantifying PM-related mortality. While ACS-based study considers a much larger population than the H6C study and covers a broader geographic area, this population is less diverse, better educated, and more affluent than the national average. Conversely, while the H6C cohort population is more representative of the US, it assesses urban areas in a small number of eastern US cities, where PM2.5 is generally comprised of a larger percentage of sulfate than it is in the western US. To the extent that PM2.5-related mortality depends strongly on particle composition, applying a H6C-based risk coefficient nationwide may bias our estimated PM mortality in the west. On the other hand, using an ACS-based risk estimate nationwide may provide an incomplete characterization of PM mortality impacts in the eastern US.
To estimate the change in ozone-related premature mortality, we drew upon short-term studies of ozone mortality. Time-series ozone mortality studies including the Schwartz (2005) analysis in Houston have strengthened the conclusion of previous studies regarding the relationship between short term ozone exposure and premature death. An analysis of the National Morbidity, Mortality, and Air Pollution Study data set by Bell et al. (2004) and meta-analyses by Bell et al. (2005); Ito et al. (2005), and Levy et al. (2005) specifically sought to disentangle the roles of ozone, PM, weather-related variables, and seasonality (US EPA 2008b).
For this analysis, we elect to apply the Bell et al. (2004) estimate of ozone-related non-accidental mortality (RR = 1.0052, 95% confidence intervals 1.0027—1.0077 per 10 ppb ozone increase) because it is broadly cited and, as a multi-city study, is not subject to the same risk of publication bias as the metaanalyses which draw upon risk estimates within the published literature. We also apply the Levy et al. (2005) estimate of all-cause mortality (RR = 1.0021, 95% confidence intervals 1.0016—1.0026 per 10 μg/m3 increase in ozone) to quantify an upper-bound estimate of ozone-related mortality. We estimate confidence intervals using a Monte Carlo analysis, sampling the standard error reported in each epidemiological study.
Baseline incidence rates
Epidemiological studies generally calculate changes in risk relative to a baseline rate that is matched to the health endpoint assessed. For this reason, when using this risk estimate to calculate health impacts, it is necessary to specify a baseline incidence rate for the health endpoint of interest. Ideally, the incidence rate should also be matched to the geographic area of interest so that it describes the health status of the population of interest. In this analysis we apply a 3-year average of 2006–2008 all-cause county level mortality rates from the CDC-WONDER database; we use a 3-year average to mitigate the influence of year-to-year variability in the mortality rates.
The transfer of risk coefficients from epidemiological studies that examine a subset of urban areas to a national health impact analysis introduces an important source of uncertainty in the analysis (Hubbell et al. 2009). When using risk coefficients drawn from these epidemiological studies, we have taken care to ensure that we match the same general demographic characteristics of the study population, including age. We also follow a similar approach to estimating population exposure by using monitored concentrations. However, unobserved differences between the populations assessed in the epidemiological study and those considered in this analysis likely introduces some uncertainty to our results.
Results and discussion
PM2.5- and ozone-related health impact estimates nationwide
Estimated annual change in premature mortality by pollutant (95% confidence intervals)
Mortality estimate (Bell et al. 2004)
Mortality estimate (Levy et al. 2005)
Mortality estimate (Krewski et al. 2009)
Mortality estimate (Laden et al. 2006)
−87 (−400 to 220)
−48 (−77 to 18)
−220 (−152 to −281)
−12,000 (−7,500 to −16,000)
−32,000 (−47,000 to −16.000)
−10 (−8.3 to 29)
−46 (−85 to −8)
−3,900 (−5,300 to −2,500)
−11,000 (−16,000 to −5,300)
Direct comparisons between the PM and ozone health impact estimates are made difficult by the fact that we base the ozone-related estimates on meteorologically adjusted air quality estimates, while the PM analysis is based on unadjusted air quality inputs (though PM2.5 concentrations are known to be generally less sensitive to meteorology that ozone concentrations) (US EPA 2008a).
Relationship between air quality trends and health impacts by location
The same unit change in PM2.5 or ozone may yield significantly different health impacts depending on the size of the exposed population and the susceptibility of that population to air pollution health impacts. For example, between 2000 and 2007, Palo Alto county in Iowa experienced a net decrease in the meteorologically-adjusted average summer season 8 h maximum value of approximately 14.5 ppb between 2000 and 2007; this air quality improvement results in an estimated decrease of <1 premature mortality. Conversely, during this same time period, an 11-ppb reduction in ozone levels in Harris County, Texas yielded a reduction of between 70 and 100 ozone-related premature mortalities. While both counties experience a substantial improvement in ozone air quality, the greater population density and baseline mortality rate in Harris County produces a larger overall ozone-related mortality benefit.
Key analytical uncertainties
Any complex health impact analysis is subject to a variety of uncertainties, some of which we have noted above. Many of these uncertainties are endemic to all health impact assessments, and are well-described in other analyses (Roman et al. 2008; US EPA 2006b). We performed a sensitivity analysis to characterize the influence of assumptions that we believe to be most important to this particular analysis, including: (1) the application of ozone monitoring data that have been adjusted for meteorological variability; (2) the use of ozone and PM2.5 concentrations drawn from a monitoring network whose geographic scope varies in each year of the analysis; (3) the use of a 100-km range within which to aggregate estimated health impacts. Of these three sensitivities, we found that the meteorological adjustment of the ozone concentrations exerted the strongest influence on the results, and so we present the results of this analysis below; the remaining two may be found in the supplemental materials.
Meteorological adjustment of ozone concentrations
As we describe above, the formation of ozone in the atmosphere is highly sensitive to meteorological conditions that can vary significantly from year-to-year. When estimating the health impacts resulting from long-term trends in air quality, it is arguably preferable to rely on air quality inputs that have been adjusted to account for this annual variability. Because such an adjustment may influence the magnitude of the health impact assessment, we assess the sensitivity of the results to this input. Meteorologically adjusted air quality inputs based on a peer-reviewed method were not available for PM2.5 and so we do not provide a sensitivity analysis for this pollutant.
Sensitivity of PM2.5 mortality estimate to geographic extent of 2007 air quality monitor interpolation
Krewski et al. 2009
Laden et al. 2006
Population (age 25 and older)
Percentage of population (%)
In this analysis, we estimate the change in air pollution-related health impacts attributable to monitored air quality changes between 2000 and 2007. We find that the annual change in the excess number of premature mortalities avoided or incurred is highly variable. We estimate an overall reduction in the number of PM2.5-related premature mortalities of between 22,000 and 60,000 and a decrease in the number of ozone-related premature mortalities of between 880 and 4,100. These health benefits are substantial, and are comparable to the projected mortality benefits of large EPA air quality regulations including the US Acid Rain Program and the attainment of the PM2.5 National Ambient Air Quality standards (Chestnut and Mills 2005; US EPA 2006b).
This analysis also indicates a sizable year-to-year and geographic variability in these estimated health impacts—suggesting that while the overall trend in health impacts is declining, certain locations benefit more than others. In 3 of the 7 years, we estimate national ozone-related premature mortalities to increase; and in 1 year, we estimate national PM2.5-related mortality to increase. And, in the remaining years, the estimated decrease in avoided cases of premature mortality is highly variable. The county-level mortality impact estimates indicate a high degree of spatial variability as well. While most counties experience some change in health impacts and a large percentage of these counties experience a health benefit associated with improved air quality, the magnitude and direction of the health effect estimate differs by region of the country and urban area. This variability may be attributable to fluctuation in industrial activity affecting emissions of PM2.5 and ozone precursors or meteorology (which we have controlled for, in part, in the case of ozone).
One intriguing finding is that the 2000–2007 county-level ozone and PM2.5-related mortality impacts demonstrate an inverse relationship in certain areas. Certain large urban areas, including LA, San Francisco, Chicago, and Houston, experience a large reduction in premature mortality due to a reduction in the levels of one pollutant and a large increase in premature mortality from an increase in another. Differences in background concentrations, regional transport, local emissions, and local atmospheric conditions may influence the relative levels of ozone and PM2.5 in these urban areas. We have identified this topic as a potential subject of future analysis.
We also found that that the estimates of ozone-related mortality impacts were sensitive to the use of ozone concentrations adjusted for meteorological variability. We elected to use these adjusted results when calculating our central ozone health impact estimates because we believed that controlling for the effects of meteorology was important when estimating the health impacts of air quality trends. However, as our sensitivity analysis illustrates, the use of non-adjusted data produces a far different estimate of ozone-related mortality impacts.
Another possible limitation of our approach is the potential to count the same avoided PM2.5-related mortality more than once. When applying risk coefficients drawn from long-term cohort studies, some proportion of the avoided premature deaths resulting from air quality changes in 1 year will occur within the same year as the air quality change, while others will “lag” a year or more into the future (Schwartz et al. 2008). For example, an incurred premature mortality resulting from an air quality degradation in 2001 and estimated for that year might be counted again (to the same individuals) from another air quality degradation in 2002—when in fact only one death would have occurred. We believe that the potential for this type of double-counting is small, however. In each year, the air quality changes are highly variable over space and time and so the risk of affecting the same individual in the same location from one year to the next is very low.
While subject to important uncertainties and limitations, the results from this analysis would be useful to decision makers as they characterize the magnitude and spatial and temporal distribution of impacts of air quality trends. Estimates of health impacts might be combined with information regarding long-term trends in both air quality and emission levels to provide a clearer link between air quality policy, emission changes, air quality levels, and public health impacts.
This paper has not been subjected to EPA peer and administrative review; therefore, the conclusions and opinions contained herein are solely those of the authors, and should not be construed to reflect the views of the EPA. The authors thank V. Rao, B. Cox, and E. Baldridge for providing technical guidance regarding the air quality inputs to this analysis. We thank S. Anenberg for providing valuable editorial support.