Our analyses found that in New York City, citywide health impact estimates were relatively insensitive to the geographic level (county compared to neighborhood) used to average air quality and health outcome incidence data. However, estimated pollutant-attributable mortality and morbidity varied widely across neighborhoods with differing socio-economic status. Baseline mortality and morbidity rates varied by season, with largest temporal variability found in the rates of emergency department visits for asthma and respiratory hospitalizations. Comparing the use of annual incidence rates versus seasonal rates for a citywide health impact analysis revealed the largest differences in ozone-attributable asthma hospitalizations and emergency department visits because high ozone concentrations occur during times of lower than average baseline incidence rates.
Across all pollutants and health endpoints, we found similar estimates in the burden when calculating at the UHF or county level then aggregating to a citywide estimate. This can be explained by the fact that the change in air quality associated with the rollbacks was relatively poorly correlated geographically with the baseline incidence counts and UHF-level pollutant gradients were not large within individual counties or across the city as a whole. For pollutant/health endpoint combinations where there were limited associations between baseline incidence counts and pollutant levels at the neighborhood (UHF) level, we saw slight differences in the estimates. In other locations where there may be large exposure gradients and strong associations between neighborhoods with high pollution levels and density of susceptible populations, averaging data to a coarser spatial scale for an impact analysis could potentially bias estimates of citywide impacts. Additional factors such as monitor locations and population density will affect exposure assignment and should be considered in developing an appropriate spatial scale for an analysis.
Neighborhood baseline incidence rates serve as one surrogate for relative vulnerability to air pollution health impacts. In New York, a city with very affluent and poor neighborhoods, variation in baseline rates, rather than air quality, account for most of the disparities in pollutant-attributable health where, for example, high-poverty neighborhoods experience 4.5 times higher burden PM2.5-attributable asthma department visits as compared to low-poverty neighborhoods. Although not generally available, relative risks may also vary between neighborhoods based on spatial differences in exposure, co-pollutants, and susceptibility including modification of effect by neighborhood traffic density (Ito et al. 2009). Future improvements in air pollution benefits analyses could come from identification of neighborhood modifiers of air pollution C–R functions and improvements in spatial resolution of air quality monitoring data.
Air quality managers relying on health incidence data available within the BenMAP tool or from publicly accessible sources would typically only able to conduct a county-level analysis. Our analysis of the gradients in pollutant-attributable health impacts found that county-level assessments may not reflect the wider range and greater disparity in impact within-city revealed using neighborhood-scale data.
This level of analysis can be particularly important when evaluating the benefit of control strategies where population susceptibility or air quality improvements may be unevenly distributed within an urban area. For example, in estimating the benefits associated with controls on power plants in the Washington DC area, Levy et al. (2002) reported that an impact assessment that stratified baseline mortality rates and PM2.5 relative risk by population susceptibility did not result in significantly different citywide mortality benefits as compared to unstratified analysis. However, the model that included stratification by education status highlighted the disproportionate impact on less-educated populations. Multipollutant risk-based strategies being developed by EPA have underscored the importance of fine scale, local data in assessing both the magnitude and distribution of benefits associated with air quality improvement strategies, demonstrating that control strategies focused on maximizing benefits in susceptible populations increased overall benefits by almost two-fold while reducing disparities across the population (Fann et al. 2011b). Fine-scale analyses that best reflect neighborhood health conditions are also appropriate in evaluating local initiatives that are aimed at reducing emissions in neighborhoods with high morbidity, such as efforts in New York City that prioritize cleaner fuel boiler conversions in schools in neighborhoods with high asthma rates (NYC 2011).
While environmental justice concerns have traditionally focused on the gradients in air quality exposure, our findings highlight the importance of including gradients in susceptibility, as reflected by baseline morbidity and mortality rates, among groups of differing socioeconomic status (SES). In this analysis, we found no significant gradient in PM2.5 exposures between neighborhoods of differing poverty status, while ozone levels were slightly higher in higher SES communities due to elevated NOx concentrations in areas of Northern Manhattan and the Bronx that increase ozone scavenging (US EPA 2006a). Despite the lack of PM2.5 exposure gradients and negative associations between ozone levels and poverty status, we found wide disparities across SES groups in pollutant-attributable health events due to differences in population susceptibility. Prior US national-level analyses have noted that across the country, the percent poverty of a county was positively associated with PM2.5 levels while an opposite relationship was observed for ozone, with indication that the relationship between county poverty and air quality can vary by geographic region (Miranda et al. 2011). Similarly, European studies have found that while there are not consistent patterns in exposure gradients between SES groups, individuals of low SES were subject to greater health effects of ambient air pollution (Deguen and Zmirou-Navier 2010).
When varying the averaging time of the baseline incidence rates, we found that differences in PM2.5-attributable health impacts were relatively small. The insensitivity of the estimates to temporal resolution of the baseline incidence data is partly due to limited variability in PM2.5 levels across seasons and the fact that seasonal patterns in PM2.5 and health incidents rates were not strongly associated (either positively or negatively).
Conversely, we found that applying an annual incidence rate likely overestimates O3-attributable asthma hospitalizations and emergency department visits and would similarly overestimate the benefits of reducing O3 concentrations. This bias can be explained by the opposing patterns of the baseline rates and ozone concentrations, where peak ozone levels in the third quarter correspond to low baseline rates of asthma-related hospitalizations and emergency department visits. Seasonally stratified analyses also indicated that the majority of ozone-attributable emergency department visits in New York City occur in the earlier portion of the ozone season (April–September).
The seasonality of baseline incidence rates should be considered particularly when the impact of control strategies varies by season. For example, regulatory impact analyses for ozone NAAQS that have applied the readily available annual average baseline incidence rates (US EPA 2008) may overestimate asthma-related impacts due to lower summertime asthma incidence as compared to the annual average. Prior evaluations of ozone-season specific emissions trading strategies such as NOx SIP Call (Burtraw et al. 1998; Environmental Protection Agency 1998) included similar limitations. Air quality advisories for ozone, during which sensitive populations are encouraged to limit exposures and all are encouraged to limit driving, tend to occur later in the ozone season, corresponding to peak ozone concentrations. Our findings suggest that more consideration of springtime ozone impacts is needed in developing air quality management and public health protection strategies. Additionally, health impact assessments of measures to reduce emissions from heating fuels such as those developed in New York City through PlaNYC 2030 (NYC 2011) should account for seasonal variation in emissions and health incidence rates.
A significant limitation in this work and any health impact analysis is uncertainty in underlying data and assumptions, many of which are difficult to quantify. By using local data on neighborhood health events, we have improved upon prior work that assume local rates can be approximated using national, regional, or county data. However, the magnitude of our pollutant-attributable estimates is limited by the uncertainty in the risk estimates derived from the epidemiological literature. For the short-term risk estimates, we have attempted to reduce this uncertainty by applying C–R functions from studies conducted on New York City populations where these estimates have been available, presumably better reflecting underlying susceptibility, local air pollutant mixtures, and PM2.5 composition. However, for endpoints without published epidemiological studies on local populations, we have relied on effect estimates from studies either conducted in other cities or from larger multicity studies, which may result in additional uncertainty in the pollutant-attributable impact estimates (Hubbell et al. 2009). For example, in our analyses, we calculated PM2.5-attributable long-term mortality effects using the Krewski et al. (2009) analysis of the American Cancer Society (ACS). Although this is the largest and most recent study on the effects of PM2.5 on mortality, the ACS population has a smaller proportion of low income and minority participants than the New York City population. Our estimate of the PM2.5 burden would have been more than twice as large had we applied a concentration–response function based on the Laden et al. (2006) analysis of the more diverse Harvard Six Cities cohort (NYCDOHMH 2011a).
An additional limitation in our analysis is that we have assumed that the same risk estimates apply across all neighborhoods. For the long-term mortality effects examined, the Krewski et al. (2009) study used the citywide average PM2.5 concentrations across cities as the exposure contrast (i.e., the subjects in the same city are assigned the same PM2.5 level), and thus no within-city exposure variations were considered. Although Krewski et al. (2009) did examine modification by level of education, finding that mortality risk estimates increased with decreasing level of education, the published data do not include sufficient information to derive New York City neighborhood-specific concentration response functions.
Similarly, for the short-term effects studies, sub-urban C–R functions are generally not available for use in health impact assessments. National, multicity studies can examine effect modification by city- or county-level characteristics and provide evidence that socioeconomic status may modify short-term risks of ozone-attributable mortality (Bell and Dominici 2008) but do not quantify how neighborhood level concentration response functions vary across New York City neighborhoods. Within many cities, population sizes at the neighborhood level limit power for time-series or case-crossover analysis, resulting in larger uncertainty in risk estimates. Ongoing work is currently exploring the use of spatially stratified time-series models and other approaches for developing neighborhood level CR function estimates (Ito et al. 2009) that better reflect neighborhood differences in susceptibility to air pollution effects.
Without neighborhood level data on air pollution exposures that corresponded to neighborhood health incidence data, we elected to characterize subcounty exposures to ozone and PM2.5 as the inverse distance weighted average of nearby regulatory monitors (EPA VNA methodology). We recognize that regulatory monitoring networks are subject to spatial limitations that may not adequately characterize fine-scale concentration gradients found in urban areas. Ongoing and future work may help reduce spatial uncertainties in exposure, including applying data from high density monitoring networks with land-use regression (LUR) modeling (NYCDOHMH 2011b) or atmospheric modeling results at fine spatial scales (Wesson et al. 2011) that can supplement monitoring data. These methods are subject to their own limitations stemming from emissions inventories, source surrogate data, and fine-scale meteorology data. While beyond the scope of this paper, we have found that LUR-based neighborhood level exposure estimates for PM2.5 mass are more variable than the estimates based on inverse distance weighting (IDW) of regulatory monitors used in this paper (coefficient of variation = 0.13 and 0.07 for LUR-based and IDW estimates, respectively; NYC 2012). Future investigations will apply these estimates as multiple years of data become available and models are developed for other pollutants.
Conducting an air quality health burden analysis such as presented here includes the assumption that the same relationship between pollutant concentrations and health risk exists at levels well below the lowest measured levels in the epidemiological literature. While this introduces additional uncertainty in the shape of the dose–response curve at lower levels, available data does not suggest a health effect threshold in the range of concentrations relevant to our analysis. While this paper examined how overall health burden of PM2.5 and ozone varied with choices of spatial and temporal resolution, we recognize that our overall estimates of pollution-attributable impacts are sensitive to other method choices, including uncertainties explored elsewhere (Hubbell et al. 2009; Fuentes 2009).