Numerous studies have identified a relationship between fine particulate air pollution and its impact on human health [1]. Particles with an aerodynamic diameter of 2.5 μm or less (PM_2.5) are small enough to invade air pathways in the body, and have been known to cause adverse health effects [2]. Several epidemiologic and human clinical studies have examined the cardiovascular and respiratory health effects of both acute and long term exposures to PM_2.5[3-5]. The Medicare Air Pollution Study (MCAPS), a multi-city study in the United States (U.S.), reported a short-term increase in hospital admission rates associated with elevated ambient PM_2.5 concentrations, for health outcomes such as ischemic heart disease, heart failure, chronic obstructive pulmonary disease and respiratory infection [6]. The MCAPS study also concluded that the cardiovascular risks, estimated at the county level, tended to be higher in the eastern U.S. Similarly, the extended follow-up of the Harvard Six Cities study showed that cardiovascular and lung cancer mortality were positively associated with long-term ambient concentrations of PM_2.5[7].

To quantify the health impacts of PM_2.5, and to track population exposure to PM_2.5, accurate and timely data collected on an ongoing basis are needed at the sub-state level. The Pew Environmental Health Commission report released in 2000 found that the state of environmental public health systems were fragmented and not robust enough to respond to environmental threats [8]. Based on the recommendations of the Pew Commission, the U.S. Congress funded the Centers for Disease Control and Prevention (CDC) to establish a National Environmental Public Health Tracking Program. The cornerstone of this program is the National Environmental Public Health Tracking Network (Tracking Network) which provides nationally consistent data and metrics (indicators and measures) to monitor relationships among hazards, exposures, and health effects [9]. The CDC, U.S. Environmental Protection Agency (EPA), and state local health departments funded by the CDC have been collaborating in the development of air quality metrics for PM_2.5 for integration into the Tracking Network (http://ephtracking.cdc.gov/showAirData.action). In July 2009 during the initial launch of the Tracking Network, only Federal Reference Method (FRM) Air Quality System (AQS) monitor data were incorporated into the Network to provide county level air quality metrics.

While AQS monitor data are viewed as the “gold standard” for characterizing ambient air quality and determining compliance with National Ambient Air Quality Standards (NAAQS), such data are limited in space and time (Figure 1A) [10,11]. During the year 2005, fewer than 20% of counties in the conterminous U.S. (leaving out 32% of the resident population) were monitored for PM_2.5 and most monitors operated every third day [12]. As a result, the AQS-based metrics on the Tracking Network are adjusted to account for missing days in monitor data (http://ephtracking.cdc.gov/showIndicatorPages.action). Also, when AQS data are available from multiple monitors for a given county and day, the highest 24-h average (daily) concentration among all the monitors is selected for purposes of calculating the measures. These adjustments are consistent with EPA practices and were adopted by the Tracking Network to ensure that the AQS-based metrics for PM_2.5 do not understate the air quality problem in any given area [13].

In order to better understand air quality for areas and days without monitor data, the CDC collaborated with the EPA on the development of a hierarchical Bayesian (HB) model to predict daily PM_2.5 concentrations for use in the Tracking Network. The HB model integrates AQS monitor data with results from the EPA’s Community Multiscale Air Quality (CMAQ) model to generate predicted PM_2.5 concentrations for a 36-km grid (individual predictions for 36-km × 36-km grid cells) across the conterminous U.S. and for a 12-km grid across the eastern half of the U.S. (http://www.cmaq-model.org). The statistical approach incorporates prior knowledge of the model parameters in the hierarchical Bayesian model, which results in improved estimation of the PM_2.5 concentrations in areas (also covered by the CMAQ grid) and at times (days) that are not monitored [14]. The model also quantifies the prediction error associated with the predicted daily concentrations for each grid cell. In general, the HB model utilizes monitor data, and bias-adjusted CMAQ model output for non-monitored areas and days. Background documents for the HB model can be found on the Tracking Network and at the EPA webpage (http://www.epa.gov/heasd/sources/projects/CDC/index.html).

The remainder of this paper describes the utility of HB predictions for public health surveillance purposes and their use in the development of county level PM_2.5 metrics for the Tracking Network. Using monitor and model data for the year 2005, we considered the following:

1. creating a spatial relationship between grid cells and counties so that daily county level estimates can be generated from HB predictions;

2. evaluating whether HB predictions at the 12- or 36-km resolution PM_2.5 should be used for the calculation of daily county level estimates;

3. comparing the resultant daily county level HB estimates with AQS county level monitor data;

4. comparing county level annual averages of PM_2.5 based on HB estimates with those based on AQS monitor data.

For 2005, 587 (19%) of counties in the conterminous U.S. had PM_2.5 monitors that operated year-round. Most of these PM_2.5 monitors only operated every third day while some operated every sixth day. A few monitors operated on an every-day schedule. HB predictions available at the 36-km grid resolution were available for 11266 grid cells covering the entire conterminous U.S. It should be noted that CMAQ estimates dominated the 36-km HB predictions in the western areas where few monitors are located. The 12-km HB predictions were available for the eastern U.S. with 66960 grid cells, out of which 66123 grid cells overlapped and were aligned with the 36-km grid.

Spatio-temporal gaps in monitoring can result in uncertainty in the ascertainment of population-level exposures, especially when fluctuations in PM_2.5 concentrations tend to occur at frequencies not detectable given existing monitor sampling schedules. Monitors that operate for regulatory purposes are not usually sited very close to sources, where high concentrations of PM_2.5 can be observed. Rather they are sited in places to measure levels of PM_2.5 that represent average concentration levels over large areas. Lack of PM_2.5 concentration measurements over continuous spatial and temporal scales limits our ability to link air quality levels with health effects data. Thus, modeled predictions may provide a suitable alternative for use in public health surveillance.

CMAQ model output as well as output from any model that relies on CMAQ output has a spatial unit, the grid cell that differs from the spatial unit of health and demographic data, which are often available at geo-political resolutions, such as county, census tract, etc. Geo-imputation procedures are necessary to convert grid-level data to county level estimates, which are needed to generate metrics for environmental public health surveillance through the CDC Tracking Network and for linkage to spatially comparable health data. The three geo-imputation methods mentioned in this paper are routinely used in public health and other allied fields. In our study, a population-weighted county centroid containment approach performed best among the methods considered for translating grid-level HB predictions to county level estimates. Also, a population-weighted county centroid denotes a spatial mean of the underlying population distribution within each county and estimates of PM_2.5 generated using this method are intended to represent the ambient concentration levels to which most of the population are potentially exposed.

In the context of linking with health data, the spatial scale of modeled predictions was a very important consideration in developing county level PM_2.5 estimates. For 2005, HB predictions of PM_2.5 were available at a 12-km resolution for the eastern U.S., whereas HB predictions of PM_2.5 were available at a 36-km resolution for the entire conterminous U.S. County level estimates of PM_2.5 derived from 36-km predictions correlate more strongly with AQS-based PM_2.5 concentrations than do estimates derived from 12-km predictions. The predominant reason for the difference in performance of 12- and 36-km HB predictions was the underlying CMAQ estimates. The input needed to generate the 12-km and 36-km CMAQ estimates were processed with different assumptions and, for certain inputs, resolving to a finer spatial scale could add uncertainty to the final model output [25]. We developed county level estimates of PM_2.5 from 36-km HB predictions since our primary goals were to have the HB-based metrics approach the values of the AQS-based metrics, and generate metrics for the entire conterminous U.S.

The strength and consistency of the relationship between daily HB- and AQS-based PM_2.5 county level estimates are acceptable at concentrations below the daily NAAQS and, at these concentrations, differences between HB- and AQS-based PM_2.5 estimates are reasonable for most census regions. For 2005, less than 2% of the measurements available from AQS monitoring reflected concentrations greater than 35 μg/m³. At these higher concentrations, HB-based county level estimates are more likely to under predict AQS concentrations, with the largest differences observed for the western region of the U.S. Some of these differences can be explained based on model features. The HB model fuses monitor data when available with CMAQ estimates and, for most locations, days with only CMAQ estimates outnumber days with both CMAQ and AQS data. CMAQ estimates are primarily used for predicting background concentrations and do not adequately capture spikes in air quality levels as a result of exceptional events [26]. While the bias in the CMAQ estimates is adjusted using a global (national-level) regression approach, and the AQS data measurement error is accounted for in the HB model , the daily HB predictions can be different from the coincident AQS measurements when CMAQ estimates greatly differ from the AQS data. Additionally, there are relatively fewer monitor-based observations available for the western U.S. and CMAQ estimates under predict AQS concentrations in the western U.S., especially when the terrain is mountainous [27]. Hence, HB estimates rely heavily on CMAQ in the western U.S. and we see larger absolute and relative differences between county level HB and AQS estimates with increasing PM_2.5 concentrations (Table 2). Users of HB-based PM_2.5 estimates should be aware of the limitations of these data as well as the benefits of having data over continuous spatio-temporal scales.

Annual county level metrics of PM_2.5, such as annual averages, provide a useful summary of prevailing concentration levels. However, averages created from AQS-based PM_2.5 concentration measurements are limited to counties with monitors and therefore do not provide a complete picture of prevailing air quality throughout the conterminous U.S. Moreover, PM_2.5 annual metrics derived from AQS data based on a sampling frequency that is predominantly every third day can be taken to accurately characterize general conditions only under the assumption that days included in the sample fairly represent the full calendar. Given that the HB-based estimates are available for every day of the year, annual averages incorporating these estimates can be interpreted without any assumptions concerning days without data.

The benefits of employing HB predictions should be considered in light of the added uncertainty which they introduce. As noted, the annual county level HB-based annual averages can understate or overstate the air quality problem in specific areas compared to averages based on AQS concentrations. At higher concentrations, especially near the annual NAAQS—15 μg/m³, and in the western U.S., HB-based annual averages are more likely to fall below monitor-based measurements. Notably, combining HB-based estimates with AQS-based concentrations results in annual averages that comport well with annual averages created using AQS data exclusively; however, a few counties have lower annual averages when compared with the annual averages obtained exclusively from AQS-based concentrations.

In summary, we characterized the difference between HB-based estimates and AQS-based concentrations with the intent that the results can guide public health professionals and researchers on the appropriate use of the county level estimates of PM_2.5. Our analysis of daily differences between AQS-based concentrations and HB-based estimates of PM_2.5 indicate that the differences can vary across census regions and divisions, and that generally HB-based county level estimates tend to under predict at higher concentrations. This needs to be considered when using daily county level HB-based estimates to identify exceedances of the daily NAAQS in different parts of the country or to conduct studies that assess health outcomes related to short-term PM_2.5 exposures. The annual averages developed by combining HB- and AQS-based PM_2.5 data show less variation with AQS-based annual averages. Given the need to correctly characterize air quality levels and minimize the discrepancy with county level annual averages created from AQS data that are commonly used, we suggest that the county level annual averages of PM_2.5 for the Tracking Network be calculated using AQS data when and where they are available and using HB-based estimates for days and locations without such data.

Poor air quality is a worldwide problem. The global burden of disease report (2010) identifies air pollution as a major contributor to premature mortality [28]. Measurements from monitors are limited in space and time, and modeled data can be an alternative to ascertain population level exposures. Our evaluation of HB predictions and AQS measurements explores the utility of modeled data from a public health perspective. Using the HB-based predictions to augment gaps in AQS data provides a more spatially and temporally consistent basis for calculating the metrics deployed on the Tracking Network. Further, “data fusion” techniques, combining monitor and modeled data, are being used in many countries to produce grid-level air quality predictions [29]. The evaluation framework presented in this paper is robust and can be extended to areas outside the U.S.

Converting grid-level predictions to estimates by geo-political units, such as counties or county equivalents, is needed to link health and population information with air pollutant data. This manuscript suggests that assigning modeled predictions to counties using a population-weighted centroid containment method is an effective approach for making the translation between a fixed grid and county-specific geography. Counties or similar administrative units exist in several European counties, for example, and the geo-imputation methods suggested in the paper can be adopted seamlessly to obtain pollutant concentrations to which most of the population is potentially exposed.

AD: Absolute difference; AQ: Air quality; AQS: Air quality system; CDC: Centers for disease control and prevention; CMAQ: Community multi-scale air quality; EPA: US environmental protection agency; ESRI: Environmental systems research institute; FRM: Federal reference method; HB: Hierarchical Bayesian; IQR: Interquartile range; KM: Kilometer; MCAPS: Medicare air pollution study; NAAQS: National ambient air quality standards; PM: Particulate matter; RD: Relative difference; SAS: Statistical analysis system; U.S.: United states

The authors declare that they have no competing interests.

AV acquired the data, analyzed and interpreted the data, contributed to design of the study, and drafted and revised the manuscript. WFD contributed to acquiring the data, interpretation of data, and design of the study and drafted parts of the manuscript and made revisions to the manuscript. SRK made contributions to the interpretation of data, and drafted parts of the manuscript and made critical revisions to the manuscript. JRQ made contributions to design and revision of the manuscript. All authors read and approved the final manuscript.