Final Report: An Evaluation of Confounders in PM10/Mortality Associations

EPA Grant Number: R825271
Title: An Evaluation of Confounders in PM10/Mortality Associations
Investigators:
Institution: New York University Medical Center
EPA Project Officer: Chung, Serena
Project Period: November 25, 1996 through November 24, 1999 (Extended to November 24, 2000)
Project Amount: $363,394
RFA: Air Quality (1996) RFA Text |  Recipients Lists
Research Category: Air Quality and Air Toxics , Air

Objective:

For almost four decades since the early London, England, data analyses, researchers have been correlating and regressing mortality/morbidity on air pollution variables. Researchers' basic interpretation of the results from these time-series observations has remained the same: the strength of association of a pollutant and a health outcome simply reflects its strength of causality. This is reasonable only if all of the exposure variables are measured equally accurately. Potential influence of differential measurement errors among pollutants on their significance of regression coefficients has been speculated, but virtually nothing has been done about it. Also, available information on the mortality records such as multiple/contributing causes has not been exploited fully to test the causal inference of apparent associations between particulate matter (PM) and mortality. It is possible that certain subsets of mortality may be more closely associated with PM. The objective of this study was to answer the critical questions regarding the role of potential confounding variables in the PM10-mortality associations reported in the past time-series epidemiological studies. Attempts were made to answer these questions by systematic evaluation of the available air pollution, weather, and mortality data. We investigated these issues?during the study period 1985-1994?in five major U.S. cities where extended (n>1000 days) PM10 data were available: (1) Cook County, IL; (2) Detroit, MI; (3) Houston, TX; (4) Minneapolis-St. Paul, MN; and (5) St. Louis, MO. For the investigation of monitor-to-monitor temporal correlation, we expanded the area coverage to seven north-central states.

Summary/Accomplishments (Outputs/Outcomes):

This study investigated the issues that are crucial to the interpretation of mortality risk estimates of PM10 and other gaseous criteria pollutants. This is probably the first study to report the difference in monitor-to-monitor temporal correlation among PM10 and all the gaseous criteria pollutants, as well as weather variables, for a large geographic coverage (seven north-central states). It was found that the monitor-to-monitor temporal correlation among the air pollution/weather variables within a 100-mile separation distance in the combined study areas could be generally ranked into three groups: (1) temperature, dew point, relative humidity (r>0.9); (2) O3, PM10, NO2 (r: 0.8-0.6), with O3 showing slightly higher correlation than the other two; and (3) CO and SO2 (r<0.5), with SO2 generally showing smaller correlation (~0.3) than that for CO (~0.4). This result suggests that if data from a single monitor were to be used for health effects analysis, then the extent of error that may attenuate the association with health outcomes may be different among the exposure variables. In fact, in the investigation of the relationship between a monitor's correlation with other monitors and corresponding estimated relative risk for that monitor in the five cities, there was suggestive evidence that the monitors with low correlation with other monitors tend to yield lower estimated relative risks. In another component of this study, the sensitivity of PM10 and other gaseous pollutants' mortality relative risks to model specifications were examined. It was found that model specification could differentially affect each pollutant's estimated relative risks. For example, the estimated relative risks for O3 often were more sensitive to alternative weather model specifications than other pollutants. In the two pollutant models, the reductions in PM10 or gaseous pollutants' relative risks from those in the single pollutant models were sometimes substantial, but could be explained by relatively high correlation between the two pollutants, especially when they were included in the model using the same lag days. Finally, the question of biological plausibility was investigated in the five cities, as well as in New York City (NYC), by examining the role of contributing respiratory causes in the association between air pollution and nonrespiratory mortality (i.e., circulatory and cancer). It was found that the pollution relative risks for nonrespiratory underlying causes of death were higher when respiratory-contributing causes of death were present, supporting the hypothesis that air pollution is causally related to circulatory mortality via an effect on the respiratory system. While this pattern appeared to be consistent for PM10, gaseous pollutants also showed a similar pattern. Overall, this project has provided important information in interpreting the PM-health outcome associations found in recent years. The framework developed in this project can be applied to larger databases to obtain more comprehensive information in minimizing the uncertainty in the quantitative interpretation of PM-mortality associations.

Component-Specific Summary of Findings

To address the issues described in the objectives, the investigations were conducted in five components:

(1) Monitor-to-monitor temporal correlation of air pollution and weather variables in the north-central United States;

(2) An examination of the relationship between monitor-to-monitor temporal correlation of air pollution and their association with daily mortality in five U.S. cities;

(3) Sensitivity of PM10-mortality associations to model specifications in five U.S. cities;

(4) Contributing respiratory disease as a cause of nonrespiratory mortality associations with ambient air pollution in five U.S. cities; and

(5) Contributing respiratory disease as a cause of nonrespiratory mortality associations with ambient air pollution in NYC.

It should be noted that in the first component (examination of monitor-to-monitor temporal correlation), we expanded the study area from the (originally proposed) five cities to seven contiguous north-central states that include many cities in addition to the five cities. By including a larger number of monitors, we expected to obtain a more comprehensive picture of the difference in monitor-to-monitor temporal correlation between air pollution and weather variables. In the second component, we examined the relationship between monitor-to-monitor temporal correlation of air pollution and their association with daily mortality in five U.S. cities. In the third component, the sensitivity of PM and other pollutants' estimated relative risks to model specifications were examined using the five cities' data. In the fourth and fifth components, we refined our original question and shifted our focus from simply looking at various causes of mortality to examining the role of contributing respiratory causes in the association between PM (and other gaseous pollutants) and nonrespiratory underlying causes of deaths (i.e., cardiovascular and cancer). We chose to shift the focus of cause-specific mortality analyses because the up-to-date review of existing literature confirmed larger relative risks for respiratory mortality than for total mortality and also suggested the effects on cardiovascular system, raising the issue of the mechanism of the apparent cardiovascular effects. In the following, the summary findings of these five components are presented.

(1) Monitor-to-monitor temporal correlation of air pollution and weather variables in the north-central United States.

Issues and Specific Aim. Numerous time-series studies have reported associations between daily ambient concentrations of air pollution and morbidity or mortality. Recent personal exposure studies also have reported relatively high longitudinal correlation between personal exposures to PM and ambient PM concentrations, lending support to the health effects reported in time-series studies. However, the question remains as to how well the temporal fluctuations in the air pollution levels observed at a given monitor represent the temporal fluctuations in the population exposures to pollution of outdoor origins in the city, and how such representativeness affects the size and significance of risk estimates. Also, such monitor-to-monitor temporal correlations would vary from pollutant to pollutant, likely influencing their relative significance of statistical associations with health outcomes. Thus, the specific objectives of this component were to estimate "average" monitor-to-monitor correlation for each of the pollution and weather variables that are typically included in the time-series studies of air pollution health effects; and to quantitatively describe, for each pollutant, the monitor-to-monitor correlation as a function of separation distance and qualitative site characterization (e.g., the location setting of a site). Knowing the relative extents of monitor-to-monitor correlation among the explanatory variables is important because they are likely to influence their corresponding relative significance of associations with health outcomes. Understanding what factors make a monitor poorly correlated with others is useful in selecting a site as being representative for epidemiological studies. We analyzed data for seven north-central contiguous states, as described below.

Approach. To examine monitor-to-monitor correlation in a large area, we chose seven central and eastern contiguous states: Illinois (IL), Indiana (IN), Michigan (MI), Ohio (OH), Pennsylvania (PA), Wisconsin (WI), and West Virginia (WV). These states cover 312,968 square miles and contain over 56 million people. The study area also includes major cities, such as Chicago, Cincinnati, Cleveland, Columbus, Detroit, Indianapolis, Milwaukee, Philadelphia, and Pittsburgh. Air pollution data for PM10, SO2, O3, NO2, and CO were retrieved from the EPA's Aerometric Information Retrieval System (AIRS) for the study period from 1988 to 1990 for these states. Because most of the PM10 data in this region were collected on an every sixth day sampling schedule (gaseous pollutants and weather variables were collected every day) at most sites, the data analyses were to be conducted for the PM10 samplings days only, in order to use comparable sample sizes. Therefore, we eliminated those sites that had less than 80 percent of the PM10 every sixth day sampling schedule (183 days x 0.8 = 146 days). This elimination process and eliminating "Agricultural," "Forest," or "Rural" sites resulted in 97, 137, 99, 26, and 71 stations for PM10, SO2, O3, NO2, and CO, respectively. We chose weather variables that are often analyzed in time-series mortality/morbidity studies: temperature, dew point, and relative humidity. The daily average values of these variables were retrieved from the Earth Info (EarthInfo, CO) database, which compiled the National Climatic Data Center's (NCDC) surface hourly observations. There were 45 weather stations in these states.

The AIRS database contains site characteristic data elements associated with each air pollution monitor. These elements include: Land Use (Residential, Commercial, Industrial, Agricultural, Forest, Desert, or Mobile); Location Setting (Urban, Suburban, or Rural); and Monitoring Objective (Maximum Concentration, Population Exposure, Background, or Source). The AIRS database also identified each site's Air Quality Control Region (AQCR). There were 17 "major" AQCRs that contained at least seven monitors for PM10 or SO2 data in this dataset. These characteristics were used in data analysis as group indices or indicator variables.

Before the examination of monitor-to-monitor temporal correlation, the air pollution or weather time-series from each monitor was detrended using smoothing splines of time?with a period corresponding to approximately 1 month and longer?to eliminate the influence of seasonal cycles. Correlation over time for each pair of monitors was then computed, resulting in n*(n-1)/2 correlation coefficients when n sites were available. We then characterized the resulting monitor-to-monitor temporal correlation as a function of: (1) separation distance; (2) associated EPA qualitative site characterization (land use, location setting, and monitoring objective);
(3) AQCR; and (4) variance (the lowest and highest 10 percentiles).

We first smoothed the monitor-to-monitor correlation over the separation distance between each pair of monitors using locally weighted regression. Because we were particularly interested in characterizing monitor-to-monitor correlation within the geographic scale of a Metropolitan Statistical Area (MSA) or city (the scale used to aggregate health outcomes in time-series studies), we repeated the procedure using correlation with a separation distance less than 100 miles.

To quantitatively summarize factors that explain monitor-to-monitor correlation, we next conducted regression analyses of the correlation coefficients using the data subset for separation distance less than 100 miles. Forward step-wise regressions were used to select variables that significantly affected monitor-to-monitor correlation coefficients using the candidate explanatory variables, including the separation distance, site characteristics (indicator variables), a high or low variance (indicator variables for the highest and lowest 10 percentiles), and indicator variables for the 17 major AQCRs.

Findings. Without taking into account the site characteristics and regional difference, the monitor-to-monitor temporal correlation among the air pollution/weather variables within a 100-mile separation distance in the combined study areas could be generally ranked into three groups: (1) temperature, dew point, relative humidity (r>0.9); (2) O3, PM10, NO2 (r: 0.8-0.6), with O3 showing slightly higher correlation than the other two; and (3) CO and SO2 (r<0.5), with SO2 generally showing smaller correlation (~0.3) than that for CO (~0.4). Within the 100-mile separation distance, the monitor-to-monitor correlation for the pollution variables declined as a function of separation distance, with much of the drop observed within the first 30 miles.

The correlation coefficients were higher for the weather variables?especially for temperature and dew point?and also tighter in their spread than for pollution variables. Among the weather variables, relative humidity's spatial correlation declined more steeply as a function of distance than that for temperature and dew point.

There were a relatively small number of very low correlation coefficients (i.e., outliers) for PM10. Most of the very low correlation coefficients were found to be those from the "Industrial" sites. However, the medians of the correlation coefficients were not appreciably different between those that included industrial sites and those that did not include industrial sites, suggesting that the "Industrial" sites generally do not explain low correlation coefficients. Monitoring objectives ("Population exposure" and "Maximum concentrations") did not contribute to the difference in PM10 correlation. However, for SO2, the smoothed correlation was somewhat smaller for "Maximum concentration" than for "Population exposure." For NO2, the sites with location setting = "Urban" showed higher correlation than those with "Suburban." For CO, the sites with land use = "Residential" showed higher correlation than the "Commercial" sites.

In the forward step-wise multivariate regression results of correlation on various site characteristics, several factors were significant predictors of monitor-to-monitor correlation. The intercepts showed essentially the same ranking/grouping as those for the correlation smoothed over the distance discussed above: PM10, O3, and NO2 showing similar intercepts (~0.7 to 0.75), followed by CO (0.47), then SO2 (0.32). The piece-wise slope variable within the 30-mile distance was selected in the step-wise regression for all of the pollutants. The computed decay in correlation as a function of distance at the 30-mile separation point was substantial for PM10 and NO2 (-0.17 and -0.20, respectively), but more modest for O3 (-0.08), CO (-0.07), and SO2 (-.06). While sometimes significant and mostly consistent with the expectation (i.e., "Industrial" sites reduce correlation), their magnitude of impacts (mostly < |0.1|) were not substantial. The large variance of temporal fluctuations (indicator for the highest 10 percentile in variance) for PM10 reduced correlation, while the small variance (the lowest 10 percentile in variance) for CO reduced monitor-to-monitor correlation.

In the forward step-wise multivariate regression results, the magnitudes of some of the regression coefficients for certain AQCRs were substantial. Of all the pollutants, the largest positive regression coefficient for the AQCR was that for the SO2 sites in metropolitan Philadelphia (0.444). The largest negative regression coefficient for the AQCR was that for the SO2 sites in metropolitan St. Louis (-0.176), making the already generally low monitor-to-monitor correlation for SO2 even poorer. Thus, the regional difference in monitor-to-monitor correlation, as modeled with the AQCR indicator variables, can be substantial.

Discussion. This is probably the first study to show the monitor-to-monitor correlation for all the criteria pollutants in a large area. As expected, PM10 and O3 showed high monitor-to-monitor correlation. There are some possible explanations to the ranking of pollution monitor-to-monitor correlation observed in this study. SO2 and CO are both primary pollutants and likely are highly influenced by local sources, whereas O3 and NO2 are largely secondary pollutants and likely are more uniformly distributed within the scale of a city or MSA. PM10, in this U.S. region in the summer months, contains a large fraction (from 50 to 80% in the summer) of sulfate, which is a secondary pollutant derived from SO2.

Site-to-site temporal correlation potentially is a useful indicator of ecologic level exposure characterization error. However, the knowledge of monitor-to-monitor temporal correlation of a pollutant alone does not allow correction or adjustment of the health outcome regression coefficients in a time-series study because: (1) the extent of the error due to person-to-monitor needs to be considered; (2) the classical additive error may not be applicable; and (3) the effects of "error" in a multiple regression model requires consideration of intercorrelation among covariates. What we can infer from these results is that for the pollutants that have poor monitor-to-monitor correlation, such as SO2 and CO, we have, on the average, a worse chance of getting accurate population exposure estimates using a single or a few monitoring sites compared with other pollution and weather variables.

Despite the limitation of the use and interpretation of monitor-to-monitor temporal correlation, the ranking of monitor-to-monitor correlation among variables appears to be, to some extent, consistent with these predictors' relative significance in recent time-series mortality studies reviewed in which temperature often was the most significant predictor of mortality; PM indices and O3 often were the most significant predictors among air pollution indices; and SO2 and CO often did not significantly explain mortality variation.

(2) An examination of the relationship between monitor-to-monitor temporal correlation of air pollution and their association with daily mortality in five U.S. cities.

Issues and Specific Aims. The issues and specific aims in this component are the same as those in the first component, except that in this component we have computed the monitor-to-monitor temporal correlation and also examined the relationship between the monitors' median correlation with others and corresponding relative risks for mortality using the five U.S. cities where daily PM10 and gaseous pollutants' data were available for multiple monitors.

Methods. The air pollution, weather, and mortality data were prepared for five cities, which are either a single county or a MSA that consists of multiple counties: (1) Cook County, IL; (2) Detroit, MI Primary MSA (PMSA); (3) Houston, TX PMSA; (4) Minneapolis-St. Paul, MN-WI MSA; and (5) St. Louis, MO-IL MSA.

Total daily nonaccidental deaths (International Classification of Diseases, 9th edition (ICD9) codes <800) were aggregated for each of the five cities for the 3,652 days during the study period from January 1, 1985 to December 31, 1994. Only residents of the MSAs whose deaths occurred in the respective MSA of interest were included in the study.

Air pollution data were obtained from the EPA's AIRS database for each of the five cities (1985-1994). Daily values were obtained for PM10, whereas hourly readings were obtained for each of the gaseous pollutants, and daily averages were computed. For analyses of PM10 and each of the gaseous pollutants, we retained the monitors that had over 300 days of observations. This process resulted in the numbers of multiple sites in each of the five cities as follows, in the order of PM10, O3, SO2, NO2, and CO: Chicago, 11, 17, 15, 19, and 14; Detroit, 8, 10, 21, 4, and 12; Houston, 5, 17, 11, 7, and 5; Minneapolis-St. Paul, 8, 8, 27, 6, and 15; and St. Louis, 14, 20, 23, 10, and 11. Many of the PM10 monitors collected data every sixth day, and therefore had substantial fractions of days missing. Case-wise deletion of such data sets would make the resulting sample size too small for multivariate regression, and therefore was prohibitive. Filling in missing values based on usual regression prediction from other variables also was problematic; while the mean may not be biased, the variance from such prediction tends to be deflated. Such bias is undesirable in our application in which effects of exposure characterization error, which is affected by the variance, are being investigated. Therefore, we decided to employ multiple imputation techniques that retain the variances and covariance structure (Schafer JL. Analysis of incomplete multivariate data. London, Chapman and Hall, 1997). The methodology utilizes EM algorithm and Markov chain Monte Carlo. We used Splus function developed by Schafer. Because all of the pollutants were log-normally distributed, we first took log-transformation of the original data (a few zeroes were first converted to 0.5); the resulting variables with missing values were imputed using Schafer's "norm" library and then transformed (exponentiated) back to the original unit. The imputation was done for each pollutant separately. We prepared 10 sets of imputed datasets and conducted the monitor-to-monitor correlation and mortality regression analyses as described below. We then computed the median of the resulting correlation coefficients of the 10 imputed data. For regression coefficients, we computed weighted averages and estimated corresponding standard error?taking into account both within and between the replicates' variability. Random seed was specified in these 10 imputation runs so that the results can be reproduced.

Monitor-to-monitor correlation of the multiple sites in each city was computed as was done in the first component. Short-term air pollution effects on total mortality were modeled using Poisson Generalized Additive Models (GAM) (Hastie and Tibshirani. Generalized Additive Models. New York, Chapman and Hall, 1990), allowing for overdispersion in the mortality data (i.e., the regression coefficients were adjusted for overdispersion). For simplicity, we applied the same regression specification for all of the pollutants. That is, the Poisson model included smoothing spline of time function with about a 1-month period span, locally estimated smooth surface (LOESS) smoothing of temperature on the same day (for heat effects) with span=0.5, LOESS smoothing of temperature from 2 days ago (for cold effects) with span = 0.5, day-of-week variables, and each of the five pollutants at lag 0, 1, 2, and 3 days, separately.

Findings. In these five cities, PM10 monitors generally had the shortest range of separation distance between monitors (<20 miles), whereas SO2 (and O3) had the longest range (>40 miles), possibly partially explaining the higher correlation for PM10.

The general order of monitor-to-monitor correlation across air pollutants was similar to that found in the seven contiguous states (Component 1): PM10 and O3 (r ~ 0.7) > NO2 (r ~ 0.3-0.7) > CO (r ~ 0.3-0.5) > SO2 (r < 0.35) in their median values.

When the estimated mortality relative risks per comparable distributional increment of individual sites were compared with corresponding median monitor-to-monitor correlation of each site with others, in Chicago, Detroit, and Minneapolis-St. Paul, those with higher median correlation tended to have larger (and more significant) relative risks. Across pollutants, O3, PM10, and NO2 often showed higher relative risks than CO or SO2. Even within pollutants, when there was a sufficient range of median correlation (e.g., range ~ 0.4), the monitors with higher median correlation tended to have larger relative risks. These patterns were apparent in Chicago, Detroit, and Minneapolis-St. Paul, but not in Houston or St. Louis.

Discussion. These results at least partially support the hypothesis that the data from a monitor that reflects the citywide common fluctuations of air pollution (i.e., higher temporal correlation with other sites) are more likely to be correlated with mortality. Such a pattern was generally seen in three major cities studied in this component. However, the results also indicate that there may be other factors that influence the relationship between the measured ambient air pollution levels at a monitor and daily fluctuations of mortality in that city. Such factors may include air conditioning and the distribution of monitors to possibly susceptible populations. A larger number of cities with a range of these factors would be required to examine these issues.

(3) Sensitivity of PM10-mortality associations to model specifications in five U.S. cities.

Issues and Specific Aims. Aside from the issue of relative exposure estimation errors among air pollutants, there remains the issue of the sensitivity of PM mortality relative risks to model specifications. In this component, sensitivity of PM and gaseous pollutants' relative risks were examined in the five U.S. cities where daily PM measurements were available. The specific model specification issues examined were: (1) influence of weather specification; (2) influence of controlling differing extent of seasonal cycles; and (3) influence of co-pollutants.

Methods. The study cities are the same as those used in Component 2. Total daily nonaccidental deaths (ICD9 codes <800) were aggregated for each of the five cities for the study period 1985-994. Only residents of the MSAs whose deaths occurred in the respective MSA of interest were included in the study.

Air pollution data were obtained from the EPA's AIRS database as described in Component 2. Unlike Component 2, area-wide multiple-site averages of daily PM10 values were calculated for each of the five cities. The number of PM10 monitors with over 300 days of observations ranged from 5 monitors for Houston, TX, to 14 monitors for St. Louis, MO. After taking multiple sites' averages, daily PM10 values were available for a minimum of 1,778 days (Houston, TX) to a maximum of 3,157 days (Minneapolis-St. Paul, MN). Generally, the daily (24-hour) averages for the gaseous pollutants were available for almost 100 percent of the study period, or 3,652 days, except for Detroit, MI, which had daily averages available for 71 percent of the study period for O3 and 90 percent of the study period for NO2. Weather variables (temperature and relative humidity) from the NCDC were extracted from the EarthInfo (Boulder, CO) compact disks. Data from the nearest weather station to each of the five cities were used.

Short-term air pollution effects on total mortality were modeled using Poisson GAMs, allowing for overdispersion in the mortality data (i.e., the standard error of regression coefficients were adjusted for overdispersion). In evaluating the sensitivity of estimated relative risks to the model specification, three separate issues were considered: (1) the extent of seasonal cycle control; (2) weather model specifications; and (3) inclusion of co-pollutants. Obviously, the combinations of these factors could be considered, resulting in unfeasible numbers of model specifications. Instead, when each parameter was changed, other parameters were set at only one condition, a condition that we considered as "epidemiologically most reasonable." Thus, to adjust for seasonal cycles, we considered controlling cycles down to approximately 1 month to accommodate influenza epidemics (i.e., smoothing spline function of days with 12 degrees of freedom per year) (the base model). In addition, we also considered smoothing spline function of days with 24 degrees of freedom per year, as well as 7 degrees of freedom per year. For weather modeling, based on our past experience with U.S. mortality data, we considered:
(1) two piece-wise linear temperature terms for heat (same day) and cold (with 2-day lag);
(2) smooth function of same-day temperature for immediate heat effects and smooth function of 2-day lagged temperature for delayed cold effects (base model); (3) smooth function of same-day temperature, average of 1- to 3-day lagged temperature, same-day dew point, average of 1- to 3-day lagged dewpoint; (4) smooth function of same-day temperature and average of 1- to 3-day lagged temperature; and (5) smooth interaction function of same-day temperature with relative humidity, as well as 2-day lagged temperature and 2-day lagged relative humidity. Two-pollutant models considered the lags at which each of the pollutant showed the largest coefficient, which could be the same or different lags.

Findings. Unlike the analysis in which an individual monitor's data were analyzed (Component 2), when the average of multiple monitors' data were used in this analysis, SO2 and CO often were significant predictors of total mortality. In Chicago, Detroit, Minneapolis-St. Paul, and St. Louis, the estimated relative risks often were significant at multiple lags. In Houston, O3 was not a significant predictor of total mortality in most model specifications.

The effects of using different extent of control for seasonal cycles (degrees of freedom of 7, 12, and 24 per year) did not substantially affect the estimated relative risks for any of the pollutants, and the direction of change was not consistent across pollutants.

The sensitivity of estimated relative risks to different weather model specifications was largest for O3, often resulting in more than 50 percent difference between the largest and smallest regression coefficients for the alternative weather model specifications examined. The alternative weather model specifications generally gave the same "optimum" lag at which the estimated relative risk was highest for a given pollutant and a given city.

In some cases, the estimated PM10 relative risks were substantially affected by an inclusion of a gaseous pollutant, but the extent of the change in the estimated relative risks could be explained by the correlation and the lags used. For example, when the same lags were chosen for the two pollutants, especially between PM10, SO2, and NO2, the impact was larger than when different lags for the two pollutants were chosen.

Discussion. The difference in the effects of model specifications across cities likely depends on the difference in correlation among pollutants. An analysis of data with a larger number of cities, taking into account the difference in interpollutant correlation, would provide a more comprehensive explanation of relative importance of factors affecting the sensitivity of PM relative risks.

(4) Contributing respiratory disease as a cause of nonrespiratory mortality associations with ambient air pollution in the five U.S. cities.

Issues and Specific Aim. Although numerous past time-series studies have found associations between ambient particle levels and circulatory mortality, it is not clear why those with circulatory conditions are at risk for the adverse effects of ambient particle levels. Because the circulatory-related deaths account for a substantial fraction of total mortality, the public health implication of air pollution effects on the circulatory system could be important. Several researchers suggested that observed circulatory effects may be due to a direct effect of air pollution on the circulatory system, while others suggested that air pollution affects the respiratory system that then could have an impact on the circulatory system.

Our objective was to determine whether individuals with contributing respiratory disease who had a nonrespiratory underlying cause of death were more sensitive to the adverse effects associated with ambient particle levels than those without contributing respiratory causes of death. This was accomplished by comparing the relative risk estimate for underlying circulatory conditions that had contributing respiratory causes with that for circulatory conditions that had no contributing respiratory causes. To determine whether the presence of co-existing respiratory conditions increases an individual's risk of death among those in poor health, but in the absence of a circulatory condition, underlying cancer deaths with versus without contributing respiratory causes also were analyzed.

Methods. Total daily nonaccidental deaths (ICD9 codes <800) were aggregated for each of the five cities for the 3,652 days during the study period from January 1, 1985 to December 31, 1994. Only residents of the MSAs whose deaths occurred in the respective MSA of interest were included in the study. Death counts were aggregated into a daily time series for underlying circulatory (ICD9 code: 390-459) and cancer deaths (ICD9 code: 140-239). In addition, daily counts were aggregated for each of the underlying mortality categories with or without contributing respiratory (ICD9 code: 460-519) causes of death.

Air pollution and weather variables used for the five cities were the same as those described in Component 3.

Short-term air pollution effects on mortality were modeled using Poisson GAMs, allowing for overdispersion in the mortality data. Because each of the specific cause mortality series may have different seasonal cycle patterns as well as different effects of temperature, different model specifications in terms of the extent of seasonal cycles and temperature/humidity specifications were initially explored, and the "best" model was chosen based on biological plausibility and Akaike's information criterion. The optimum weather model that was selected for most of the mortality series considered was used for all outcomes in that city. For most cities, the weather model involved a LOESS smooth of same day and an average (1- to 3-day) lag of temperature. Due to the possibly different underlying mechanisms involved for each of the mortality outcomes, mortality regression models were analyzed for each lag (0-4 days) for each pollutant.

To compare the level of risk between the two groups (e.g., circulatory deaths with and without contributing respiratory causes), a one-sided t-test assuming unequal variances of the two populations was used. Lags were considered optimum if underlying deaths with contributing respiratory causes had significantly greater risk compared with those without contributing respiratory causes, and this effect was observed to decrease on either side of that lag.

Findings. Generally, across all pollutants considered in each of the five cities, there was an identifiable optimum lag(s) in which underlying circulatory or cancer deaths with contributing respiratory causes had significantly greater effect estimates compared with those without contributing respiratory causes.

Larger effect estimates for cancer deaths with contributing respiratory causes of death were observed at a single or at contiguous lags for all pollutants across all cities, with the exception of O3 (Minneapolis-St. Paul and St. Louis), NO2 (Houston), and PM10 (Chicago). This effect modification was found most consistently for PM10 in four out of the five cities considered.

Among all the cities, with the exception of Detroit where no effect modification was observed, circulatory deaths with contributing respiratory causes had significantly greater PM10 effect estimates at a single lag or at multiple lags that were contiguous. When all of the other pollutants were considered, the effect modification was observed at a single or contiguous lag for all pollutants, with the exception of SO2 (Chicago) and NO2 (St. Louis and Chicago). Larger effect estimates due to the presence of contributing respiratory causes were most consistently observed in four out of five cities for PM10, CO, and NO2.

For all circulatory and cancer outcomes considered (although there were no consistent lags that were observed for a particular pollutant or city), in most cases the effect modification was observed at a single or at contiguous lags. In addition to the larger effect estimates for deaths with contributing respiratory causes, deaths without contributing respiratory causes also were associated with air pollution at optimum lag(s), particularly for circulatory deaths for PM10.

Discussion. The results of this study suggest that in individuals whose primary diagnosis was a nonrespiratory condition (such as circulatory or cancer), those with contributing respiratory causes had a higher risk of death associated with ambient air pollution levels. For cancer deaths, the effect modification was only observed consistently across most cities for PM10 and O3. While the effect modification was detected for circulatory deaths across all pollutants, only PM10, CO, and O3 had single or contiguous optimum lags consistently observed. Although no consistent lags could be determined for a particular pollutant or city, clearly defined optimum lag(s)?where the detection of the effect modification declined on either side of that lag?were observed for PM10 and O3 in most of the cities. For both cancer and circulatory deaths, when the raw pollutant was used in the regression model, PM10 and O3 effect estimates were least sensitive to how the pollutant was modeled. Due to the moderate correlation among the pollutants in these five U.S. cities, it will be difficult to ascribe observed effects to individual pollutants. However, it is interesting to note that although consistent effects were observed for PM10 and O3, these pollutants were not the most highly correlated pollutants in these cities.

Because circulatory and respiratory conditions are closely related physiologically, possible misclassification between circulatory and respiratory causes of death is quite possible, but realistic levels of misclassification between circulatory and respiratory categories found in the literature would not be expected to alter the conclusions of this study. However, if there were some consistent pattern to the misclassification of contributing causes of death (e.g., consistent underreporting of respiratory versus circulatory conditions), then the interpretation of the effect estimates from the mortality subcategories would be limited.

The results of this study suggest that individuals with nonrespiratory underlying causes of death are particularly sensitive to the adverse effects of air pollution when co-existing respiratory conditions are present, especially for PM10 and O3. These results were consistently observed across the five U.S. cities. While there was limited evidence for direct circulatory effects, the results of this study mainly suggest that respiratory diseases have an important role in explaining observed associations between air pollution and nonrespiratory underlying causes of death. These results are consistent with the hypothesis that air pollution is causally related to circulatory mortality via an effect on the respiratory system, and suggest that direct PM effects on the heart also may be significant.

(5) Contributing respiratory disease as a cause of nonrespiratory mortality associations with ambient air pollution in NYC.

Issues and Specific Aims. The basic background and issues in this component are the same as those in Component 4. We conducted analyses similar to the ones in Component 4 for NYC, which includes Richmond, Kings, Queens, Bronx, and New York counties, for the years 1985-1994. This city was selected because it has a large population (~7.3 million people), which increased the power to detect associations with the smaller mortality subcategories. Thus, in addition to the examination of contributing total respiratory causes, additional issues were examined using this large dataset. First, while the presence or absence of co-existing respiratory conditions could act as an effect modifier to the observed air pollution effect, advanced age also may modify the observed risk due to increasing severity and complexity of underlying disease. Therefore, we examined how the air pollution effect estimates for deaths with or without contributing respiratory disease were modified by age by comparing the effect in an older group (?75 years) with a younger group (<75 years). Secondly, more specific cardiovascular underlying causes of death, as well as specific respiratory-contributing causes of deaths were examined.

Methods. Total daily nonaccidental deaths (ICD9 codes <800) were aggregated for the 3,652 days during the study period from January 1, 1985 to December 31, 1994. Only residents of NYC whose deaths occurred in NYC were included in the study. Death counts were aggregated into a daily time series for underlying circulatory (ICD9 code: 390-459) and cancer deaths (ICD9 code: 140-239). In addition, daily counts were aggregated for each of the underlying mortality categories with or without respiratory (ICD9 code: 460-519) contributing causes of death. The age distribution of the mortality data was such that approximately half of the individuals were above the age of 75 and half were below the age of 75. Therefore, daily time series for these two age groups were separately compiled and analyzed to determine whether the effect of air pollution on those with contributing respiratory disease was modified by age. In addition, more specific underlying circulatory categories were analyzed: ischemic heart disease (IHD, ICD9: 410-414); acute myocardial infarction (AMI, ICD9: 410); heart failure (HF, ICD9: 428); cerebrovascular disease (CV, ICD9: 430-438); dysrhythmia (DR, ICD9: 427); and diseases of the peripheral circulation (DPC, ICD9: 440-459). Specific contributing respiratory disease also were analyzed: pneumonia (ICD9: 480-487) and chronic obstructive pulmonary disease (COPD, ICD9: 490-496). Age stratification was not used in these specific cause analyses because of small sample sizes.

Air pollution data were obtained from the EPA's AIRS database for NYC (1985-1994). Daily values were obtained for PM10, whereas hourly readings were obtained for each of the gaseous pollutants, and daily averages were computed. For each of the gaseous pollutants, the daily average was computed using data from all available sites with over 300 days of observations. An area-wide average of every sixth day PM10 values was calculated using all available data from the 17 monitors that measured PM10 during the study period. Daily PM10 values were available for 585 days after taking multiple sites' averages, and daily (24-hour) averages for the gaseous pollutants were available for more than 96 percent of the entire study period, or 3,500 days. Weather variables (temperature and relative humidity) from the NCDC were extracted from the EarthInfo (Boulder, CO) compact disks. The data from the New York LaGuardia International Airport weather station were used.

The statistical approach was the same as that used in Component 4.

Findings. For all ages combined, those with contributing respiratory causes were at greater risk compared with those without contributing respiratory causes for underlying circulatory deaths (e.g., PM10 relative risk [RR] = 1.054 per interquartile range [IQR], 95% CI: 1.023-1.087 compared with RR = 1.026 per IQR, 95% CI: 1.015-1.037) and cancer deaths (e.g., PM10 RR = 1.051 per IQR, 95% CI: 0.998-1.107 compared with RR = 1.011 per IQR, 95% CI: 0.996-1.026).

When the data were stratified by age, increased risk based on the presence of co-existing respiratory conditions was observed for those 75 years, but not for those <75 years.

At both lag 0 and 1 day, most of the specific underlying circulatory deaths as well as underlying respiratory deaths were positively and significantly associated with PM10 (except DR, DPC, and HF, which were associated with PM10 at lag 0 day only). At lag 0 day, the relative risks for underlying circulatory deaths with or without contributing respiratory and pneumonia or COPD were essentially the same, indicating no effect modification. However, at lag 1 day, the relative risks for circulatory with contributing respiratory, pneumonia, and COPD were significantly larger than those without (significant for respiratory and pneumonia). A similar pattern was seen for IHD.

For underlying cancer deaths, the relative risks for those with contributing respiratory causes were significantly higher than those without contributing respiratory causes at lag 0 day, but not on lag 1 day.

In the two pollutant models (PM10 and each one of the gaseous pollutants), the PM10 relative risks for the underlying circulatory and IHD without respiratory (and without pneumonia or COPD) were not affected by the addition of gaseous pollutants. However, the PM10 relative risks for the underlying circulatory and IHD without respiratory (and without pneumonia or COPD) were reduced by the addition of CO, SO2, and NO2 (pollutants that had relatively high correlation with PM10 , r ~ 0.6-0.7), but not by addition of O3 (correlation with PM10 was ~0.4).

Discussion. In NYC, it was found that underlying circulatory deaths (all-cause and IHD) with contributing respiratory causes had higher PM relative risk estimates than those deaths without contributing respiratory causes. Both COPD and pneumonia, as contributing conditions, increased the PM effect, with pneumonia having somewhat higher relative risks. These results are consistent with the hypothesis that air pollution is causally related to circulatory mortality via an effect on the respiratory system.


Journal Articles on this Report : 2 Displayed | Download in RIS Format

Other project views: All 7 publications 2 publications in selected types All 2 journal articles
Type Citation Project Document Sources
Journal Article De Leon SF, Thurston GD, Ito K. Contribution of respiratory disease to nonrespiratory mortality associations with air pollution. American Journal of Respiratory and Critical Care Medicine 2003;167(8):1117-1123. R825271 (Final)
R827351 (2002)
R827351 (Final)
  • Abstract from PubMed
  • Full-text: ATS-Full Text HTML
    Exit
  • Abstract: ATS-Abstract
    Exit
  • Other: ATS-Full Text PDF
    Exit
  • Journal Article Ito K, Thurston GD, Nadas A, Lippmann M. Monitor-to-monitor temporal correlation of air pollution and weather variables in the North-Central U.S. Journal of Exposure Analysis & Environmental Epidemiology 2001;11(1):21-32. R825271 (Final)
    R827351 (Final)
    R827351C001 (2000)
    R827351C001 (2002)
    R827351C001 (2003)
    R827351C001 (Final)
  • Abstract from PubMed
  • Full-text: JESEE-Full Text PDF
    Exit
  • Abstract: JESEE-Abstract
    Exit
  • Supplemental Keywords:

    ambient air, exposure, particulate matter, risk, health effects, epidemiology., RFA, Health, Scientific Discipline, Air, Geographic Area, particulate matter, air toxics, Epidemiology, State, Risk Assessments, Atmospheric Sciences, Ecological Risk Assessment, ambient air quality, copollutant exposures, particle size, particulates, PM10, Minnesota, MN, co-factors, human health effects, exposure and effects, synergistic effects, chronic health effects, human exposure, Illinois (IL), confounding variables, Acute health effects, acute toxicity, epidemiological studies, Missouri (MO), mortality, atmospheric chemistry

    Progress and Final Reports:

    Original Abstract
  • 1997
  • 1998
  • 1999