Introduction
Recent studies have reported that particulate pollution in ambient air is associated with increased mortality (1-3) and morbidity (4,5). Studies have been done in cities where the primary source of particulate pollution is combustion products. Studies of areas with high industrial particulate pollution show increases in asthma symptoms and hospital admissions. The present study examined the association between particulate pollution and the incidence of acute respiratory diseases as measured by outpatient visits for specific respiratory diagnoses in an area without significant industrial pollution, Anchorage, Alaska. Anchorage is a city of 240,000 people located in a "bowl" surrounded by mountains and sea coast. Wood smoke is not a major contributor to particulate pollution in this area because wood is not commonly used as fuel due to its high cost. Electric power plants are fueled by natural gas. The main sources of particulate pollution are unpaved roads, road sanding, vehicular traffic, and ashfall from volcanic eruptions.
On 18 August 1992, during the period of this study, Mt. Spurr, 60 miles west of Anchorage, erupted and rained ash on the city. Hourly measurements of particulate matter with aerodynamic diameter less than 10 
m (PM10) reached a maximum level of 3000 g/m3. The 24-hr average concentration was 565 g/m3 on the day after the eruption. Computer-controlled scanning electron microscopy (CCSEM) was used to determine the composition of particles from 10 random samples taken before and after the volcano erupted (6). Over 80% of the particle mass was between 2.5 and 10 m. The composition was mainly silica and silica-aluminum. CCSEM showed that less than 5% by weight of the filter mass was carbonaceous particles. This is consistent with source apportionment studies by chemical mass balance done 7 years previously (7), which concluded that more than 85% of the total suspended particulates (TSP) in Anchorage was earth crustal material. Size-fractionated mass measurements below 15 m were made in Anchorage by the U.S. EPA in the early 1980s using dichotomous samplers. Historic data collected by EPA during 1983 also suggest a high coarse-particle mass fraction. Average fine [aerodynamic diameter (da) < 2.5 m] to coarse (2.5 m < da < 15 m) particle mass ratios in the summer of 1983 were 0.14 
0.05. The overall median ratio of fine to coarse PM15 was calculated to be 0.26. This is in distinct contrast with the 0.4 to 0.7 PM2.5/ PM15 ratios reported for 6 communities in the lower 48 states by Spengler and Thurston in 1983 (8). We investigated the relationship between respiratory illness treated on an outpatient basis and ambient particulate PM10 pollution using a health insurance database. Working people and their dependents, generally considered a healthy group, are the sample population used in this analysis. The sample size was approximately 6% of the population of the city of Anchorage.
Methods
Database
Particulates are measured daily as 24-hr PM10 samples using an Anderson head sampler at a central location in Anchorage, the Gambell site. Measurements are made intermittently at two other sites within the city. The Pearson correlation coefficient between sites ranges from 0.76 to 0.81. The Gambell site is located close to a major highway, and PM10 concentrations are 43-76% higher than concentrations measured at other sites. Other than PM10, few pollutants are routinely monitored in Anchorage. Carbon monoxide monitoring is conducted hourly and daily between October and March. CO is routinely monitored at five locations in the Anchorage area. Only two sites exceeded the 9 ppm CO standard during the period of this study: the Seward Highway site, located four blocks from the Gambell PM10 monitoring site, and the Garden site, a residential area about 2 miles from the Gambell PM10 monitoring site. Measurement of criteria pollutants, sulfur dioxide, nitrogen dioxide, and ozone are only done occasionally as they remain quite low.
Available 8-hr maximum CO concentrations measured at each of the five CO monitoring sites in Anchorage between 1 October 1992 and 31 March 1993 and between 1 October 1993 and 31 March 1994 were obtained from the Municipality of Anchorage Environmental Services Division. These data were processed to generate the average 8-hr maximum CO concentration for each day, which was then used in the analysis.
Daily claims made for outpatient visits for respiratory illness were obtained from Aetna Insurance Company, which processes the health insurance claims for both employees of the State of Alaska and employees of the Municipality of Anchorage. Both groups have comprehensive health insurance with low deductibles for employees and dependents. We analyzed data from a 22-month period from 1 May 1992 to 1 March 1994. All outpatient visits that were submitted to insurance, whether they occurred in doctors' offices or in emergency rooms, were captured by this method. The diagnosis code recorded for the visit was based on the International Classification of Diseases 9th Revision (ICD-9) coding. ICD-9 codes were grouped to identify upper respiratory problems such as sore throat, earaches, sinusitis, rhinitis, and other nonspecific upper airway problems. This whole group of illnesses is referred to as upper respiratory illness (URI). The second group, referred to as bronchitis, includes lower airway diseases such as bronchitis, tracheitis, and nonspecific cough. Pneumonia was not included, as it is frequently treated on an inpatient basis. The third respiratory category, referred to as asthma, included all reactive airway disease, bronchospasm, and asthma ICD-9 codes. Diarrhea, a common diagnosis presumably unrelated to air pollution, was recorded as a control diagnosis. The ICD-9 codes used were: for asthma, 519.1, 493.9, 493.0, 495; for bronchitis, 466.0, 490, 490.0, 491.0, 491.1, 786.2; for chronic obstructive pulmonary disease (COPD), 491.2, 491.9, 492.0, 492.8, 496, 506.4; for congestive heart failure (CHF), 428.0, 428.1, 402.01, 402.11, 402.91, 440.9, 398.91, 429.1, 429.4, 429.9; for diarrhea, 558.9; and for upper respiratory illness (URI), 077.2, 460, 461, 461.0, 461.1, 461.2, 461.3, 461.8, 461.9, 462, 465, 465.0, 465.9, 472, 472.0, 472.1, 472.2, 473.0, 473.1, 473.2, 473.3, 477, 477.1, 477.9, 478.2, 478.8.
Reiterations of the insurance data collection were done until we were confident of a stable claims report. Only visits where both patient and provider had an Anchorage zip code were included in the analysis. There were approximately 11,000 State of Alaska and 3000 municipal employees and dependents eligible for health insurance in Anchorage during the time of the study.
Analytical Methods
Daily outpatient visits, temperature, and PM10 series exhibit seasonal cycles, some of which are common. Unless adjusted for long-term cycles, shared seasonal or monthly cycles among outpatient visits and environmental variables could confound results. Adopting the technique used in Kinney and Özkaynak (8), a weighted 19-day moving average filter developed by Shumway (9) was used to detrend the pollution and meteorological series. The method involves subtracting the weighted moving average of each variable (Xt) from itself on each observation. In other words, the Xt on day t = i is filtered as:
(1)
where wi is the filter weights shown in Shumway (9). This process of filtering removes the long-term cycles but not the short-term cycles (i.e., high frequencies). When a linear filter such as this is applied to both the predictor and predicted variables before regression analysis, linear regression relationships among variables are preserved and can be estimated without bias. In addition, this filter efficiently removes the autocorrelation in the pollution and the outpatient visit series. Autocorrelation functions were examined to detect any remaining temporal structure in the filtered data, and none was found.
We computed descriptive statistics for all the filtered and unfiltered data. We used a generalized linear model procedure to test statistical differences in the daily outpatient visits by day of the week. Cross-correlations between filtered outpatient visits (e.g., for asthma) and filtered PM10 were calculated to determine the importance of the relationship between doctors visits and same-day (or lag 0), previous-day (or lag 1) and 2-days prior (or lag 2) PM10 measurements. We analyzed the daily outpatient visit (OV) counts and pollution data using time-series and regression modeling techniques implemented with SAS software (SAS Institute Inc., Cary, North Carolina).
Because of low daily counts for some categories of doctors visits (e.g., asthma, bronchitis), we examined two different methods of modeling the pollution-health effect relationships. Both ordinary and Poisson regression models were fitted to filtered outpatient visit, temperature, and pollution data. Consistency of results and normality of model residuals were examined. In all cases, results from Poisson and multiple regression models were almost identical. Moreover, residuals from the multiple regression models were very nearly normally distributed. Consequently, for technical and practical reasons, we chose multiple regression modeling framework in the analysis. Basic analysis involved fitting multiple regression models to four filtered morbidity variables (i.e., doctors visits diagnosed as asthma, bronchitis, diarrhea, and upper respiratory infections ) using filtered same-day or previous day PM10 and temperature as explanatory variables. The diarrhea category was selected for analysis as a control category. The form of the basic regression model (model I) was:
(2)
where OV_F is the filtered daily outpatient visits, X_Fi is the filtered same-day or previous-day daily temperature and PM10 measurements, and E is the error term. Other models were also done. Model 2 added a weekend/weekday indicator variable (W_D) as an additional explanatory variable to Equation 2. Model 3 was a regression specification using as the dependent variable outpatient visits that were both filtered and weekend/weekday adjusted. Specifically, model 3 was written as:
(3)
where:
(4)
In models 1-3, same-day temperature and same-day PM10 were included. We also ran models with different lags of temperature and PM10. We present results from one of these, Model 4, where previous day's PM10 (or lag 1 PM10) instead of same-day PM10 is included in the specification. The models were run for all ages combined and separately for three age groups (<10 years, 11-45 years, and 46+). Due to sample size limitations, male and female outpatient visits were combined.
Finally, we also examined potential statistical confounding of results due to other pollutants of health concern and the influence of variations in the PM10 composition over time. We included the available wintertime CO measurements independently, as well as jointly with PM10 data, in the regression models tested. Potential changes in the seasonal composition of PM10 and the influence of the volcanic eruption that occurred on 18 August 1992 were also modeled using nested regression modeling methods. In this case, we estimated separate PM10 slopes for winter versus summer seasons and periods strongly influenced by volcanic eruption (18 August 1992-31 December 1992) versus the remaining period less influenced by volcano ash (i.e., 1 May 1992-17 August 17 1992; 1 January 1993-1 March 1994).
Results
 |
 |
|
Figure 1. Measurements of PM10 (particulate matter <10 m in diameter) in Anchorage, Alaska (Gambell site). (A) Original data; (B) filtered data. |
Figure 2. Daily outpatient visits for upper respiratory infections in Anchorage, Alaska. (A) Original data; (B) filtered data. |
 |
 |
| Figure 3. Daily outpatient visits for bronchitis in Anchorage, Alaska. (A) Original data; (B) filtered data. |
Figure 4. Daily outpatient visits for asthma in Anchorage, Alaska. (A) Original data; (B) filtered data. |
Figure 1 displays the daily PM10 measurements collected at the Gambell site in Anchorage. Both the original and filtered series are shown. The influence of volcanic eruptions on PM10 levels during the fall of 1992 are clear. After detrending, long-term cycles and seasonal patterns are no longer apparent. Daily counts for outpatient visits for asthma, bronchitis, and URI are shown in Figures 2-4. Again, the 19-day moving average filter detrends the observations for respiratory illness visits. Because of the relatively young age of the sample population, CHF and COPD visits were infrequent, and no analysis of these was done. Table 1 presents the summary statistics for the analysis variables: temperature, PM10, CO, visits for illnesses of asthma, bronchitis, diarrhea, and URI. Correlation of 8-hr maximum CO measurements among the five different sites was quite high (
0.8). Consequently, using the data from all monitors, we calculated the 8-hr maximum CO value in the Anchorage area for each day data were collected. The correlation between daily PM10 and daily average maximum CO was found to be small ( 0.15). Table 1 also provides a breakdown of the statistics by different age categories. Clearly, most of the visits are recorded in the largest age category, 11-45 year olds. Because only active employee insurance records were analyzed, most of the population at risk were under 65 years of age.
Ordinary regression models were run for all outpatient visit categories. Table 2 presents the results for the basic model (model 1) for the three respiratory illness categories: asthma, bronchitis, and URI. All of the estimated PM10 regression coefficients were significant for these illness categories. However, a generalized linear model analysis indicated substantial weekend/ weekday differences in the recorded outpatient visits. Because most doctors' offices are closed on the weekends, typical weekend visit counts for all causes of illness were five times lower than during the weekdays. Moreover, there was also a slight difference, though not statistically significant, in the PM10 concentrations during weekdays versus weekends. Average PM10 concentration on Saturdays and Sundays was about 37 g/m3, whereas during weekdays it was around 48 g/m3. We suspected that less traffic over the weekend results in slightly lower PM10 levels in the city. Consequently, we attempted to control for the differences in the outpatient visits during weekends in two different ways. One way was to add a weekend-versus-weekday dummy or indicator variable to the basic regression model (Eq. 2), which we called model 2. The other way was to adjust for day-of-the-week effect on outpatient visits first, and then run the regression on the residuals with either same-day PM10 (model 3) or previous-day PM10 value (model 4). The results from these alternative model fits are shown in Table 2. Model 2 results with the weekend/weekday dummy variable do not indicate a significant association between PM10 and doctors visits. The reason for nonsignificant findings under model 2 is the induced statistical collinearity between the estimated PM10 coefficients and the weekend/weekday (W_D) coefficients. Because the underlying reason for reduced outpatient visits over weekend days are different and much more pronounced than those affecting the differences in weekend/weekday PM10 concentrations, a separate adjustment of outpatient visits was considered more appropriate. Therefore, a two-stage regression analysis was considered to be the most reliable method with this data set. Models 3 and 4 were both developed as two-stage regression analyses using filtered weekday/weekend adjusted outpatient visits as the dependent variable.
Statistically significant associations were found between both same-day and previous-day PM10 (lag 0, lag 1) and asthma visits and between same-day PM10 and URI diagnosed outpatient visits based on model 3 and 4 specifications. The statistical association found between lag 1 PM10 and visits for asthma was stronger and more significant than the association found between same-day PM10 and visits for asthma (Tables 2 and 4). Other lags (i.e., lag 2,3) of PM10 were also studied but not found to be significant in the models tested. Using the coefficients from model 3 and model 4, the magnitude of the projected PM10 effect on outpatient visits for each 10 g/m3 increase in PM10 is 2.5-3.5% excess outpatient visits for asthma and 1.2% excess outpatient visits for URI (Table 2).
Next we examined the age dependence of the results by repeating the model 3 analysis for daily visits recorded separately under the three age categories: <10 years, between 11 and 45 years, and >46 years (but typically less than 65 years). Table 3 presents these regression results. Due to the small number of daily counts, some of the age-specific regression estimates were not significant. For asthma visits, the effect estimate for the 11- to 45-years age group (2.6% excess visits) was not significant (p = 0.14) but similar in magnitude to one previously found for the all ages combined. However, statistically significant associations were found between PM10 and URI-related outpatient visits for children under 10 years of age and adults over 46 years of age. The predicted PM10 effect on URI visits associated with an increase of 10 g/m3 PM10 was 1.9% and 1.2%, respectively, in these two age categories. Outpatient visits for diarrhea were not significant either in models 2 or 3.
We examined the association between daily CO and outpatient visits using the regression models (i.e., models 3 and 4). Table 4 presents the estimated regression coefficients for CO from models of outpatient visits for asthma, bronchitis, and URI. These results are based on model 3 specifications. Models with lag 1, 2, or 3 CO variables did not result in statistically significant coefficients. Same-day CO was highly significantly associated with outpatient visits for bronchitis and URI using the available CO series, obtained during fall/winter of 1992-1993. In comparison to the estimated PM10 effect, the magnitude of the estimated CO effect on URI and bronchitis outpatient visits seems to be greater. For an increase of 1 ppm (8-hr maximum) CO, it is estimated that doctor visits for bronchitis and URI will rise by 10% and 13%, respectively. The significant associations found between CO and bronchitis and URI are not influenced or confounded by PM10. Models in which both PM10 and CO variables were included produced results essentially the same as the single pollutant regression models.
Temperature and Volcano Effects
The temperature coefficient was significant in only one of the models. The estimated coefficient for the filtered temperature variable from the model of CO and temperature on URI was 0.24 (p<0.04). We examined further the temperature and PM10, and temperature and CO relationships, and found those to be weak. We re-ran the PM10 regression models with lag 1 temperature instead of the same-day temperature and obtained identical results. We assume that the 19-day weighted Shumway filter adequately removes not only the seasonal trends in the data, but multiday variations in the temperature observations that may influence respiratory diseases and symptoms more than the day-to-day variations in temperature. Variations in temperature may have less effect on health in a young, working population than on a more vulnerable population.
We also examined whether the estimated PM10 coefficients were influenced by seasonal or other compositional factors. We ran nested regression models to estimate separate PM10 slopes for summer (April- October) versus winter (November- March) seasons and also for the period influenced by the volcano eruption (18 August 1992-31 December 1992) versus the period not expected to be influenced by the Mt. Spurr volcano (1 May 1992-17 August 1992; 1 January 1993-1 March 1994). Table 5 presents the estimated PM10 and (lag 1) PM10 coefficients obtained from these models. The association found with PM10 and asthma does not seem to be influenced much by season. A significant (lag 1) PM10 coefficient is estimated for the winter season from models of asthma-related outpatient visits. For URI, because of sample size limitations, PM10 coefficient loses its significance when the data set is split by summer and winter season. However, the magnitude of the estimated summer and winter coefficients remain similar to the PM10 coefficient estimated from the full data set (see Tables 4 and 5). Interestingly, the period immediately after the volcanic eruption resulted in nonsignificant PM10 coefficients in models of asthma and URI outpatient visits. In contrast, the period not affected by the volcanic eruption resulted in statistically significant PM10 coefficients. Average PM10 concentration during the period influenced by the volcano was around 70 g/m3 , whereas the period not affected by the volcano had an average PM10 of 40 g/m3. The magnitude of the estimated same-day or previous-day PM10 effect on doctors visits for asthma, during the period not influenced by the volcano, was about 6%, corresponding to an increase of 10 g/m3 PM10. Likewise, doctors visits for URI are expected to increase by about 3% corresponding to increase of 10 g/m3 PM10 during the period not influenced by volcanic activity.
Discussion
We analyzed 22 months of daily PM10, temperature, and daily cause-specific outpatient visit data from Anchorage, Alaska, to study the acute relationship between PM10 and respiratory illnesses. The health data were obtained from a large health insurance provider to state and municipal employees in Anchorage. Even though the coverage was only partial (80-90%) and records may have included repeat visits to a doctor by the same individuals, the data set is considered to be representative. Furthermore, we applied conservative statistical methods to control for potential seasonal, weekly, and daily confounders of PM10 health effects. In particular, we controlled for potential influences of temperature on daily outpatient visits.
In Anchorage, continuous records for other pollutants such as ozone (O3), sulfur dioxide (SO2), and nitrogen dioxide (NO2) were not available for the period of analysis. It is unlikely that these omitted variables could confound potential associations between PM10 and outpatient visits. Limited monitoring data available for these pollutants indicate very low levels
Last Update: May 16, 1997
