We estimated traffic density (TD) for each subject based on their reported street of residence and nearest cross-street. Specifically, reported residential cross-streets were geocoded using geographic information system (GIS) software and data (ArcView StreetMap 2000 version 1.1; ESRI, Redlands, CA). We then identified each subject’s probable home street segment, which had the reported nearest intersection at the center bound on both ends by the adjacent cross-streets (Meng et al. 2007). We drew a 500-foot buffer around the probable home street segment of each subject and identified all roadways within this buffer that had an AADT value based on the year 2000 Caltrans data. The 500-foot criterion is based on environmental measurement data showing that the impact of direct traffic emissions on ambient concentrations becomes insignificant at approximately this distance (Zhou and Levy 2007). However, the buffers used here were > 0.028 mi² in area (as would be the case with a 500-foot radius circle around a specific geocoded home location), because the buffers were defined not by a point but by a probable street segment and its length. On average, the buffers were 0.068 mi² in area (range, 0.031–0.199 mi²). There was not much difference in buffer size between urban areas (average, 0.067 mi²; range, 0.033–0.168 mi²) and suburban areas (average, 0.069 mi²; range, 0.031–0.199 mi²). Most subjects resided in urban areas (71%); 29% were suburban, and none lived in rural areas.

Similar to methods of Gunier et al. (2003) and Reynolds et al. (2004), we estimated the TD value for each subject by first calculating the vehicle miles traveled (VMT) for each attributed road segment within the buffered area; VMT was estimated by multiplying the AADT by the road segment length. We then calculated TD as the sum of the VMT for all road segments in the buffer divided by the area of the buffer, that is,

where TD is traffic density (vehicles × miles/day/mi²), AADT is the annual average daily traffic count (vehicles/day), L is the length of roadway segment (miles), and A_B is the area of the 500-foot buffer around the probable street segment (square miles). Subjects with no Caltrans-counted streets within their buffers (n = 47) were included in the low-traffic referent category, because we assumed that these individuals had only residential streets with low AADT near their homes.

In addition to TD values, we assigned to each subject the annual average concentrations of O₃, NO₂, PM_2.5, PM₁₀, and carbon monoxide measured at the nearest monitoring station within 5 miles of the reported residential cross-street intersection (82% of subjects). We used ArcView GIS (ESRI) software to estimate the distance between each respondent’s residential cross-street intersection or ZIP code population-weighted centroid (for the 18% of subjects whose reported cross-street could not be geocoded) and the nearest station that measured each pollutant. Subjects residing outside a 5-mile range were excluded from the pollutant analyses. Annual average concentrations were calculated for the 1-year period before the interview date. These averages were based on hourly measurements for the gaseous pollutants (CO, NO₂, and O₃) and 24-hr average measurements for PM₁₀ and PM_2.5 (with most stations recording measurements every 6 and 3 days for these pollutants, respectively). The 5-mile range was chosen to ensure a large enough radius for sufficient sample size balanced against a possible increase in exposure misclassification due to greater residential distances from monitoring stations.

We evaluated associations between air pollution and TD and asthma morbidity using logistic regression. Specifically, we examined differences in our exposure metrics for a) children with asthma reporting daily or weekly symptoms in the previous year versus those reporting less than weekly symptoms, and b) children with asthma reporting at least one asthma-related ED visit or hospitalization in the previous year versus children with asthma not reporting such visits. The analyses incorporated sampling weights that adjusted for unequal probabilities of selection into the CHIS sample. We initially grouped TD values into quintile categories according to their distributions in the total population. Because the effect estimates in the three middle quintiles were similar, we collapsed them and thus created three categories roughly based on these quintile distributions: a) low TD, defined as ≤ 20,000 daily VMT/mi² (approximately ≤ 20th percentile); b) medium TD, defined as 20,001–200,000 daily VMT/mi² (approximately 21–80th percentile); c) high TD defined as > 200,000 daily VMT/mi² (approximately the 80th percentile).

Measured air pollutants (O₃, PM₁₀, PM_2.5, CO, NO₂) were evaluated continuously, using both single and multipollutant models. We evaluated changes in point and 95% confidence interval (CI) estimates when including covariates such as age, sex, race/ethnicity, poverty level, insurance status, delays in receiving care for asthma, asthma medication use, and county in our models (Table 1). Based on a 10% change-in-estimate criterion (Rothman and Greenland 1998), race/ethnicity and poverty level were included in our final models. Some subjects were excluded because of missing data for these covariates. Final sample sizes for each model are reported in the tables.

We were able to successfully geocode reported residential cross-street intersections and calculate TD values for 500 (81.7% of 612) respondents. Only three subjects who were geocoded could not be assigned a TD value because of missing AADT data in the Caltrans file. The reasons reported locations were not successfully mapped included missing information for one or both streets (63%) and inability to find a reported cross-street in the reference street map (i.e., one or both streets were not found based on spelling provided or reported streets did not intersect, 37%). The mean TD value was 152,311 (median of 86,513), with a range of 0–1,557,027 daily VMT/mi² (Table 2). The percentage of subjects excluded from analyses based on annual average air pollution averages varied depending on pollutant (because not every station measured every pollutant, and we used a 5-mile exclusion criterion, as explained above). Approximately 26%, 47%, 36% of subjects were excluded from the O₃, PM₁₀, and PM_2.5 analyses, respectively. Mean annual average pollutant concentrations did not differ substantially for subjects included versus excluded from analyses based on this criterion; however, the percentages of children with asthma reporting daily or weekly symptoms or ED visits/hospitalizations were lower in the excluded population (by approximately 2–5% depending on pollutant and outcome).

Based on logistic regression models, we did not observe associations between our asthma symptom outcome measures and NO₂ or CO; thus, we limit the following discussion to residential TD, O₃, and particles (PM₁₀ and PM_2.5) (Tables 3 and 4). We observed an approximately 2-fold increase in the odds of daily/weekly symptoms in children with asthma for each 1 part per hundred million (pphm) increase in annual average O₃ [odds ratio (OR) = 1.96; 95% CI, 1.23–3.13], and this estimate did not change appreciably when we controlled for race/ethnicity and poverty level or when we added particle measures to the model, although the 95% CIs widened because of the inclusion of these covariates and the reduction in sample size (Table 3). There were suggestive associations between particles (PM₁₀ and PM_2.5) and daily/weekly symptoms after adjustment for O₃, but estimates were imprecise. We also observed an approximately 2-fold increase in the odds of ED visits/hospitalizations for asthma per 1-pphm increase in O₃ after adjusting for particles (particle and O₃ exposure estimates were strongly negatively correlated). Furthermore, we also estimated 2- to 3-fold increases in the odds of ED visits/hospitalizations per 10-μg/m³ increases in PM₁₀ and PM_2.5 after adjusting for O₃ (Table 3).

We observed associations between residential proximity to traffic and asthma-related ED/hospitalizations but not for daily/weekly symptoms (Table 4). We estimated an approximately 3-fold increase in the odds of ED visits/hospitalizations for children with asthma in the highest TD exposure category (TD > 200,000 daily VMT/mi²) compared with those with TD values ≤ 20,000 daily VMT/mi² after adjusting for race/ethnicity and poverty level (OR = 3.27; 95% CI, 1.08–9.89). A model that included TD, O₃, and particle measures was consistent with these findings (results not shown). Approximately 48% of the subjects with TD > 200,000 daily VMT/mi² had a freeway within their buffer. (Approximately 41% of children in Los Angeles County versus ~ 84% in San Diego County had a freeway in their buffer, yet of all children with TD values > 200,000, ~82% resided in Los Angeles.)

An advantage of the CHIS for linkage of health and environmental data is knowledge of residential locations at the cross-street (CHIS 2001) and address (CHIS 2003 and 2005) level. Knowledge of residential locations is particularly important for assessing source-specific exposures, such as exposures to certain key traffic exhaust pollutants (e.g., ultrafine particles) whose concentrations change rapidly within approximately 300 m of the source (roadway) (Zhou and Levy 2007). Mapping home locations to the census tract or even census block group level likely results in exposure misclassification for such pollutants. Here, only residential cross streets were available, which likely introduced error into the traffic exposure metric; future CHIS surveys collect information on residential addresses. Knowledge of residential locations at the cross-street and address level also allows determination of distance to existing monitoring stations; thus, subjects living far from stations can be excluded from analyses (based on the assumption that accuracy of exposure assessment decreases as one moves farther away from a station). The disadvantages of using such exclusion criteria are that they reduce sample size and may limit generalizability.

Geostatistical methods such as kriging and land use–based regression modeling are alternatives for assessing air pollution exposure [see Jerrett et al. (2005) for an in-depth discussion of these methods and others such as inverse distance weighting and air dispersion modeling]. Kriging spatially interpolates pollution levels measured at existing monitoring stations across urban areas. A major advantage is that it allows the estimation of both predicted values and their standard errors (kriging variance) at unmeasured locations (Jerrett et al. 2005). However, the predictive capability of this method is dependent on the density of monitoring stations in a given area. Reliance on existing monitors (which are generally limited in number) normally results in surfaces that oversmooth the true pattern of pollution and may introduce large errors in estimates over portions of a study area for which few observations are available (Jerrett et al. 2005). This problem may be more severe for pollutants known to vary significantly over small scales, such as NO_x and other direct traffic pollutants. For these pollutants, land use–based regression (LUR) detects small-area variations in air pollution more effectively than kriging (Briggs 1997; Lebret et al. 2000). In the LUR modeling approach, concentrations of vehicle exhaust markers such as NO_x are measured simultaneously at many locations throughout an urban area, using relatively inexpensive passive monitors (e.g., Ogawa monitors). Various GIS parameters (such as traffic and roadway density, population density, and land use) are used to predict the measured concentrations (Jerrett et al. 2007). The developed model can then be used to estimate concentrations at home and school locations of study subjects based on GIS parameter values at these locations. Although these methods allow generation of air pollution exposure estimates at a wide number of locations, a number of statistical assumptions are needed (Jerrett et al. 2005). They also require additional expertise to implement, and in the case of LUR, additional data collection, which can be labor intensive. Furthermore, LUR models developed in one area may not be readily transferable to other urban areas.

Another major issue in air pollution exposure assessment is accounting for time–activity patterns of the subjects (time spent at home, at school, in vehicles) and extent of infiltration of outdoor air pollution into indoor spaces (homes and schools). Exposure misclassification resulting from the lack of such information may be one explanation for the differences we observed in particle and TD results for the two outcomes (daily/weekly symptoms vs. ED visits/hospitalizations) in this study. Specifically, the particle associations for daily/weekly symptoms were weaker than for ED visits/hospitalizations. This may reflect more nondifferential exposure misclassification of our residence-based particle measures for older children more likely to spend time away from home during the day; although daily/weekly symptoms were more frequently reported for this age group, ED visits/hospitalizations were more common in younger children (Table 1). The same explanation could also hold for the null association we observed between residential TD and daily/weekly symptoms. Although levels of O₃ tend to be more spatially homogeneous within communities (Avol et al. 1998; Geyh et al. 2000; Monn 2001), several O₃ exposure assessment studies in Southern California children have noted the important contribution of time–activity patterns (e.g., time spent outdoors) and housing characteristics (e.g., use of air conditioning, age and size of house, use of open windows and fans for ventilation) to personal O₃ levels (Avol et al. 1998; Geyh et al. 2000; Liu et al. 1997; Xue et al. 2005). In this study, we observed associations between ambient O₃ measures and both outcomes, but point estimates were generally lower for ED visits/hospitalizations (younger children). This may suggest that station-based measures of O₃ are more accurate estimates of exposure for older children, who may spend more time outdoors and who may engage in outdoor activities that increase their exposure, compared with younger children (infants and young toddlers), who may spend more time indoors and thus generally have lower O₃ exposures. Subsequent CHIS surveys collected information on school locations so that potential differences in exposure based on location of the children can be taken into account and the impact on estimates of association can be assessed. Increased sample sizes when adding CHIS 2003 and 2005 data will also allow us to determine the robustness of our findings, which were relatively imprecise when we relied on 1 year of data for Los Angeles and San Diego Counties only. Future EPHT tracking efforts focusing on O₃-related health effects may also be improved by collecting housing information, such as air conditioning use and house size, and age and time–activity information, such as time spent at school and outdoors. Collection of such data, in concert with more spatially resolved outdoor pollution gradients (e.g., through kriging or LUR), would allow the development of more complex exposure models for air pollution (see, for example, Marshall et al. 2006 and Wu et al. 2005). This type of modeling may also be possible without the collection of time–activity and housing information directly from subjects as part of surveys, that is, by using data from the U.S. Census and existing travel demand surveys coupled with stochastic modeling. [See Marshall et al. (2006) for an example using existing Southern California Association of Governments travel demand and mobility data] However, the usefulness of such stochastic modeling for EPHT needs further investigation.

The CHIS includes a large, statewide sample, thereby facilitating statistical analyses of associations between air pollution and asthma while taking into consideration potential confounding factors. For this particular demonstration project, residential cross-street information was available only for respondents in Los Angeles and San Diego Counties; thus, our resulting air pollution effect estimates were relatively imprecise. CHIS 2003 and 2005 each included approximately 3,000 children with a reported asthma diagnosis or asthma-like symptoms and residential address information, providing much larger sample sizes for future analyses. In addition to a relatively large sample size, CHIS data include information on a number of potential confounding factors including race/ethnicity, income, health insurance status, usual source of care, type of usual source of care, experience of delays in receiving asthma care, and asthma medication use. For this study, we lacked information on some potentially important confounders of the association between air pollution and asthma morbidity, including exposure to secondhand tobacco smoke and indoor allergens such as pets, cockroaches, and molds, parental history of asthma, day care attendance, and breast-feeding history. Such factors may also be important to investigate as effect measure modifiers of the relationship between air pollution exposure and asthma exacerbation. Subsequent CHIS surveys (CHIS 2003 and 2005) will allow us to examine potential residual confounding of our outdoor air pollution effect estimates by some of these factors. Systematic collection of all of these potential confounding variables may not be feasible within the framework of EPHT. Nevertheless, within an EPHT framework it may be necessary to first establish a minimally sufficient set of variables necessary to allow estimation with little residual confounding.

The goals of EPHT are not only to quantify and examine links between exposure to environmental agents—such as air pollution—and disease, but also to determine whether increases or decreases in exposure result in subsequent changes in health indicators. From a public health perspective, such information supplies data to support policy decisions—for example, mandating cleaner-burning motor vehicles and fuels. Previous studies have capitalized on natural experiments (Friedman et al. 2001; Pope 1996) or prospectively followed schoolchildren (Avol et al. 2001; McConnell et al. 2003) to demonstrate a lowering of asthma morbidity resulting from changes (reductions) in air pollution. However, serial cross-sectional surveys such as the CHIS may be more amenable to EPHT for assessing the impact of air pollution regulations, as they provide asthma prevalence data on a more routine basis and cover a diverse and representative sample of children (in this case, for California). The specific mechanics for relating changes in regulatory reductions in air pollution to consequent changes in asthma health end points are beyond the scope of this paper. We currently have exposure metrics only for the 2001 CHIS participants. However, major factors that would need to be taken into account for such analyses include a) disentangling the effects of meteorology versus emissions reductions on air pollution levels, b) evaluating and accounting for potential confounding factors that may also change over time (e.g., changes in access to health care, use of asthma management plans) in statistical models, and c) attributing emissions reductions and subsequent health improvements to specific regulations. It may also be of interest to examine “what if” scenarios whereby the impacts of projected air pollution levels with and without the implementation of a specific emission reduction program on subsequent health end points are evaluated. This may be particularly important in areas where major changes in population or VMT are occurring.