\.XPO.S /11 5 3 POLICY RESEARCH WORKING PAPER 1453 Air Pollution and Mortality The relationshiip between particulate air pollution and premature death in Santiago. Results from Santiago, Chile Chile is found to bc very similar to results from Bart Ostro industrial countries Jose Miguel Sanchez Carl!os Aranda Gunnar S. Eskeland The World Bank Policy Research Department Public Economics Division May 1995 POLICY RESEARCH WORKING PAPER 1453 Summary findings Heavy outdoor pollution is found in developing country Multiple regression analysis was used to explain cities such as Jakarta, Katowice, Mexico City, and mortality, with particular attention to the influence of Santiago. But most epidemiological studies of dose- season and temperature. The association persists after response relationships between particulate air pollution controlling for daily minimum temperature and binary (PM10) and premature deaths are from WVestern variables indicating temperature extremes, the day of the industrial nations. This study of such relationships in week, the month, and the year. Additional sensitivity developing countries by Ostro, Sanchez, Aranda, and analysis suggests robust relationships. Eskeland fills an important gap. It is also one of the fcw A change equal to 10-microgram-per-cubic-meter in based on monitored PM10 values, or small particles, daily PM10 (about 9 percent) averaged over three days which is likely to be a more relevant measure of was associated with a 1.1 percent increase in mortality exposure to air pollution than the more traditional (95 percent confidence interval: 0.6 to 1.5 percent). measure of total suspended particulates. Death from respiratory and cardiovascular disease was Over several years, daily measures of ambient PM10 more responsive to changes in PMO10 than total mortality were collected in Santiago. Data were collected for all was. The same holds for mortality among men and deaths, as well as for deaths for all men, all women, and mortality among individuals older than 64. all people over 64. Deaths from respiratory and The results are surprisingly consistent with results cardiovascular disease were recorded separately, and from industrial countries. accidental deaths were excluded. This paper-aproductof the PublicEconomics Division, Policy Research Department-is partof a largereffort in the department to analyze environmental policies. A shorter vcrsion will be published in Journal of Exposure Analysis and Enuironmental Epidemiology. The study was funded by the Bank's Research Support Budget under the r-search project -Air Pollution and Hcalth Effects in Santiago, Chile- (RPO 678-48). Copies of this paper arc availablc frec from the World Bank, 1818 H Strcet NW, Washington, DC 20433. Picase contact CynthiaBernardo, room N10-0SS, extension 37699, or send requests by clectronic mnail (36 pages). May 1995. The Poficy Pesearch Workig Paper Series disseminates the frindings of work in progress to encourge the exchange of idcas about development ssues. An objectwe of the secries is to get the fndings out quickly, even if the presertations are lcss than fully polisbed. The papers carry the names of the authors and should be used and cited accordingly. The findings, intepretations, and conclusions are the authors' owand should not heattrihded to the World Bank its Executive Board of Drtors, or any of its membr countris Produced by the Policy Research Dissemination Center Air Pollution and Mortality: Resultsfrom Santiago, Chile World Bank Bart Ostro Office of Environmental Health Hazard Assessment California Environmental Protection Agency Berkeley, CA Jose Miguel Sanchez C. Programa de Postgrado en Economia LADES-Georgetown University Santiago, Chile Carlos Aranda Colegio Medico de Chile Santiago, Chile Gunnar S. Eskeland The World Bank Washington, D.C. Acknowledgment: A shorter version of this study will be published in Journal of Exposure Analysis and Environmental Epidemiology. Financial support was provided by the World Bank (RPO 678-48). The views expressed are those of the authors and not necessarily those of their affiliations. We are thankful for comments, in particular from Mead Over. 1. Introduction The needfor estimates of environmental benefits in developing countries Increased attention is now being paid to the need to protect the environment in developing countries, especially after "Our Comrnon Future" (T.he Report of the Joint U.N. Commission on Environment and Development, 1987) and the Rio Conference on Environment and Sustainable Development. The concern is not only for the lobs of natural resources and habitat, but also for health problems associated with pollution (See, for instance, The World Bank's "World Development Report 1992", on the environment). Newspaper coverage of health problems in formerly socialist countries, of cholera outbreaks in Latin American and of the Bhopal incident may remind us that problems associated with pollution of air, water and toxics are not limited to western industrialized countries. For air pollution, the topic of this study, new intormation about the high concentrations of particulate matter in Eastern Europe and in large metropolitan areas such as Bangkok, Jakarta, Mexico City and Santiago has motivated an examination of the transferability of results from epidemiological studies conducted primarily in the U-S., Canada and Westem Europe. There is a strong tradition of cost-benefit analysis in evaluation of projects and policy reform in developing countries, but the inclusion of estimates for environmental effects is rare. Thus, the challenge is posed to provide credible estimates of the benefits that can be provided by pollution reductions in developing countries. 3 For air pollution, estimates from industrialized countries has emphasized the benefits of improved public health (See, for instance, OECD, 1989)1. While it is in principle possible to obtain benefit estimates for air pollution reductions directly (say, from survey techniques, or from house values), a more commonly applied approach is to spell out the consequences of reduced air pollution, and then quantify and value these. The present study is a part of a greater effort at the World Bank emphasizing estimation of the health effects of air pollution in developing countries. Ostro (1994) reviewed the literature from developed countries on estimated dose responsefinctions - or inference-based studies of associations between arnbient air pollution and health impacts, such as premature mortality and respiratory illness. In the review, Ostro concentrated on time-series studies. The review concluded with an application in which effects on morbidity and mortality of several air pollutants were estimated for Jakarta, based on assumnptions about transferability of dose response functions. Following Ostro's review, applications of the methodology have been made in many World Bank stadies. While transferability of dose response functions may be a satisfictory approach when localy based inference studies are not available, the present study responds to the need for comparable studies from developing countries. The setting in a developing country may be different in important ways (more time spent out-doors, a younger demographic structure, various aspects of public health status, health services, as well as income related differences, ' We limit our attention here to local air pollution problems. The benefits of reduced emissions of climate gases would be global, and hard to estimate, or even model. A recent reference, on ihe impact on agriculture for the U.S. would be Mendelsohn et al., 1994. 4 such as nutrition and housing), highlighting the need to enrich the literature with studies from these different settings2. The needfor PMIO based studies In 1986, the U.S. Environmental Protection Agency changed its National Ambient Air Quality Standard for particulate matter from one that included all particles (total suspended particulates or TSP) to one that included only those less than 10 microns in diameter (PMlO). Although the standard is based on protecting public health, epidemiologic research on PMIO has been limited by the lack of data from outdoor monitoring stations. In addition, almost all stations that do collect PM1O data operate only every six days, limiting the ability to conduct time-.series anaiysis linking PMIO to various health endpoints, including mortality and morbidity. Thus, in assessing the health effects of particulate matter, many researchers have had to rely on various surrogate measures for PM10. In several recent studies, PM10 data have been available (Pope et al., 1991; Pope et al., 1992; Schwartz, 1993; Dockery et al., 1993), but most have used other pollutant metrics such as TSP (Samet et al., 1981), coefficient of haze (Ostro et al., 1993), fine particulates based on airport visibility (Ostro, 1989), British Smoke (Mazumdar et al., 1982), KM (Shumway et al, 1988) and sulfates (Thurston et al., 1992). 2Regulation of toxic substances is often based on transfer of results from studies of rats under high exposure to assumed effects for humans under low exposure. In this light, the assumption that dose response fimctions may be transferable from. say people in London to people in Calcutta does seem warranted when local research is not available. 5 Santiago, Chile Since 1989, daily measures of ambient PM1O have been collected in Santiago, the capital of Chile. Located in the center of a c:osed basin, Santiago is about 33 degrees south latitude on the western edge of South America. The population of the metropolitan Santiago area is estimated at 4.4 million, roughly one-third of the entire population of Chile (Ministerio de Economia, 1990). The high ambient levels of particulate matter are a result of emissions from motor vehicles, fossil fuel use for energy production, industrial processes and blowing of resuspended dust, coupled with the unique topoclimatology of the region. Mountains, including the coastal range and the Andes, almost completely surround Santiago. The predominant wind direction is from the southwest - coinciding with the one small gap in the surrounding mountains. The city is situated in a zone with fairly stable atmospheric conditions, including low velocity, turbulence and frequency of winds (Comision Especial de Descontaminacion de la Region Metropolitana, 1990). With prevailing anticyclonic conditions throughout the year, an inversion layer typically exists at between 600 and 900 meters above the city. This layer intensifies in the autumn and winter, preventing natural dispersion of pollutants and trapping most particles within 400 meters above the city (Prendez et al., 1991). Thus, between the months of July and August, during the Chilean winter, particulate concentrations in Santiago are among the highest observed in any urban area in the world (300 to 400 jg/m3). These particles include a large proportion less than 2.5 microns in diameter, including sulfates (Sandoval and Martinez, 1990). According to a 1985 analysis based on downtown monitors, diesel sources contibute approximately 74 percent of the ambient PMl0, with gasoline-powered '.'ehicles, industrial, 6 residential, and "other" sources responsible for 6, 6, 2, and 12 percent of the concentrations, respectively.3 Outline Section 2 describes the data, and section 3 presents the methodology. In section 4, we examine the relationship between PMIO and mortality in Santiago. We also report the results of extensive analysis of the sensitivity of the results to alternative regression specifications, methods of controlling for seasonality, functional forms, and health endpoints. Section 5 provides discussion, and Section 6 a comparison with results found elsewhere. 2. Mortality, Air Pollution and Weather Data For the years 1989 through 1991, daily deaths of residents of metropolitan Santiago were extracted from the mortality records of the lnstituto Nacional de Estadisticas. Deaths of Santiago residents that occurred outside of the metropolitan area and deaths from accidents were excluded. In addition to total (all-cause) daily deaths, those due to respiratory disease (ICD 460-519) and cardiovascular disease (ICD 390-448) were tabulated separately. In addition, separate counts of all-cause mortality for males, females, and all people over age 65 were compiled. Daily 24-hour average concentrations of PMlO were collected from five monitoring sites in Santiago for the same years. Four of the sites are located within a few miles of each other in downtown Santiago, while the fifth is located far to the northeast of the central city. Therefore, as measures of outdoor particles, two alternative metrics were examined: the average of the four downtown monitors and the highest reading from the monitors. In order to examine the spatial 3 Sandoval et al., 1985 - more recent analysis of this kind is not known to be available. 7 representativeness of the downtown monitors, available daily historical data on total suspended particulates (TSP) from other monitoring sites were obtained for the previous ten years. Comparisons between one of the downtown monitors and five monitors ringing the city indicated daily correlations of 0.68, 0.73, 0.79, 0.85 and 0.92. Since PM10 is likely to be more evenly distributed than TSP, it appears that, at a minimum, the ambient levels move together throughout the basin. Unfortunately, data on PMI 0 concentrations were not available for every day during this three-year period. Generally, fairly complete data exist for each of the years b_tween April and November, the season of high particulate concentrations. During the rest of the year, missing data are more common. Summary statistics are provided in Table 1. During the three- year period, the mean of the 24-hour average of PM10 concentrations was 115.4 jig/m3, while the average of the highest daily concentration from any monitor was 141.5 pg/m3. This compares to U.S. and Chile standards for annual average concentrations of 50 ig/rm3, and a California standard of 30 ig/rm3. The meai' PM1O (24-hr average) concentration was 76.2 pg/m3 during the summer months, and 141.4 Lg/rm3 during the winter months. The numbers of observations in 1989, 1990 and 1991 were 267, 277 and 246, respectively. 8 Table 1. Descriptive Statistics for Air Pollution, Meteorological and Health Variables. Variable Mean Range PMI0 (plg/m3)m1989) 112.9 32-336 PMI 0( ,ggm3)a(1 990) 119.5 30-367 PMl0("g/m3)-(199l) 113.4 35-308 Maximum PMlOb(pgIm3) (1989) 147.2 35-500 Maximum PMIOb(PgIm3) (1990) 143.5 35-424 Maximum PM,Ob(pgI&m) (1991) 132.9 39-340 Ozone (1-hour maximum, ppb) 52.8 11-264 Nitrogen Dioxide (1-hr max, ppb) 55.6 10-258 Sulfur Dioxide (1-hr max, ppb) 59.9 4-363 Minimum Temperature (CO) 10.51 -0.23-22.7 Maximum Temperature (Co' 22.6 7.37-34.47 Average Daily Humidity 53.1 29.3-93.7 Total Mortality 55.0 22-106 Respiratory Mortality 8.05 0-30 Cardiovascular Mortality 18.0 3-41 Mortality Age 65 and Above 35.6 15-72 Male Mortality 26.8 8-60 Female Mortality 28.1 8-60 a=Daily average of all four monitor readings; b=Daily maximum single monitor reading; Fairly complete daily data on ozone, sulfur dioxide, and nitrogen dioxide were available during this time period, along with daily data on minimum and maximum temperatures and humidity (Table 1). The correlation coefficients for selected variables are displayed in Table 2. 9 Particles are positively correlated witlh oxides of sulfur and nitrogen, and inversely correlated with ozone and temperature. Table 2. Correlation Coefficients for Selected Pollutant and Meteorologic Variables. Max PM1O PMI0 I NO2 S02 Ozone Minimum Temp Maximum Temp MaxPMlO 1.00 PM1O 0.97 1.0D NO2 0.70 0.73 1.00 SO2 0.65 0.64 0.60 1.00 Ozone -0.23 -0.23 -0.06 0.00 1.00 Minimum Temp -0.44 -0.45 -0.36 -0.34 0.42 1.00 Maximum_ Temp -0.30 -0.31 -0.15 -0.13 0.68 0.73 1.00 3. Methodology MIortality counts were first examined to see if they were normally distributed. Figure 1 displays the distribution of counts for total mortality. For these, the normality assUMption was not rejected using the Kolmogorov statistic. Therefore, ordinary least squares regression techniques and parametric tests were used in examining the association between air pollution and total mortality. For the other mortality endpoints (subsets of total mortality), the application of a poisson distribution is more appropriate since the normality assumption is less appropriate for events with a lower daily count. For total mortality, to ensure comparability, results using both distributions are reported. 10 Figure 1: Daily Mortality Santiago, Chile, 1989-91 Number of clays 300 . 150 100 *~~~~~~~~~~~~~~~~~~~~~~~~~~."I ('I'~~~~~~~~~~~~~~~~~~~~~~~~~~~~~''h O 1. 4' 1 JqI s.. 0 *.~'q *i..; * I E id'It *0 AV4 P4 4 0 4'V e~ 4b" C Deaths per day Obviously, air pollution is only one of many factors affecting daily mortality. Therefore, it is imoortant to examine potential confounders, including temperature, month, season, and day of the week. Our analytic approach started with a basic model that included only PMlO as an explanatory variable. We explored both a linear and a semi-log form of PMlO, using both the daily mean and the maximum monitor concentrations. Then, in turn, we examined more complex models as additional potential explanatory variables were added to the regression in a hierarchical fashion. At each stage, both the magnitude and significance of the air pollution variables and the additional explanatory power of the new variables were evaluated Substantial seasonal pattems in mortality exist in Santiago, peaking during the winter months of July and August. Figure 2 displays both the crude mortality counts and a locaLly weighted, smoothed plot of the seasonal pattem. Control for weather and seasonal cycles in the 11 model is important, since several recent mortality studies have indicated the influence of temperature on mortality (Kalkstein, 1991). As a result, temperature was first added to the basic model, in terms of minimum, average, and maximum daily levels. Contemporaneous temperature and 1-, 2-, and 3-day lags were individually examined. Temperature was entered as a continuous variable and in binary form to indicate extremes. Thus, days in the lowest or highest 25 percent, 10 percent, 5 percent and 1 percent of either maximum or minimum temperature readings were examined. In the next model specification, season was modeled by use of dichotomous variables for each quarter. Additional controls for season were tested by adding binary variables for month and a time trend was incorporated through a variable indicating the year of study. Day of week was also added to the model with binary variables. The basic model also was rerun for each separate year. Finally, a model was run that included a separate binary variable for each month in each year, a total of 35 additional terms (using the first month for reference), to allow for differences by month within each year. Figure 2: Crude, Unadjusted Total Mortality Verus Day of Study o. I,X , *- C>,, .. Day . . C3~ ~ ~~~~~1 0 C'J 0 200 400 600 800 1000 Day Subsequent to this analysis we considered dose response functions for other mortality endpoints (i.e., subsets of total mortality: respiratory mortality, cardiovascular mortality, mortality for those aged 65 and above, as well as by gender) using a poisson distribution because of the lower mean levels of these counts. Second, in sensitivity analysis, additional controls for the cyclical nature of the total mortality counts and the effects of temperature were explored through several techniques. The data were stratified first by season and then by year and reanalyzed. Next, the basic regression model wvas rerun after the coldest 1, 5 and 10 percent of the days were deleted, in turn, to reduce the potentially confounding effect of weather (mortality and particles were both higher during the colder seasons). Another method involved the use of a Fourier series comprising sine and cosine terms as covariates. The series included five terms which incorporate harmonic periods of one year and 6, 4, 3, and 2.4 months. These covariates provide good control for both tht short and long waves in the data. The third method to control for cyclical patterns involves the pre- filtering of the data by subtracting multi-day moving averages of 15-days as suggested by previous studies (Kinney and Ozkayrak, 1991). Finally, a generalized additive model was developed using the statistical package S-PLUS (StatSci, 1993) to incorporate non-linear (and non-monotonic) patterns in the temperature-mortality relationship. In this model, the temperature-mortality relationship was smoothed using a nonparametric technique that fits a function in a flexible data-driven manner. Then this fimction is entered as an explanatory variable in the basic regression model. 13 In the third sensitivity analysis, other pollutants. including daily maximlum 1-hour concentrations of ozone, sulfur dioxide and nitrogen dioxide, were examined with and without PMI0 in the model. In the fourth sensitivity analysis, the effect of extreme or unduly influential observations were examined. Approaches included use of robust regression such as M-estimates and least trimmed squares (LTS) regression (Statsci; 1993). As described by Schwartz (1993), the M- estimation iteratively weights each data point, with increasingly less weight for those observations with larger standardized residuals. The LTS regression model minimizes the sum of the portion of the data with the smallest squared residuals and provides robust estimates with small bias even when the data are highly impacted by outliers. The M-estimates regression reduces the impact of responses that are outliers while the LTS regression reduced the impact of both outlier responses and high leverage points; that is, data points that have different x-values relative to most of the observations. Finally, in the last sensitivity analysis, the impact of alternative lag structures for PM10 was investigated. Single-period models using PMI0 lags of zero to 3 days were run. In addition, models using a 3- and 4-day moving average and a polynomial distributed lag of PM10 were examined. Since mortality is likely to be serially correlated, autocorrelation corrections were applied to all of the ordinary least squares models using the AUTOREG procedure in SAS. 14 4. Results Sensitivity to Specification Table 3. Regression Results for Total Mortality Using Alternative Model Specifications Variables in Model T Beta s.C. f RR | 95%C1 MaxPMI O 0.046 0.006 1.13 1.10,1.16 PMIO 0.056 0.007 1.13 1.10, 1.16 Log (maxPMlO) 7.82 0.86 1.10 1.06,1.11 Log (PMIO) 7.62) 0.88 1.16 1.14,1.18 Above +Mintemp (-1)* 5.64 0.78 1.10 1 08, 1.11 AbovcIColdlO+HotlO 5.62 0.B4 1.10 1.08, 1.11 Above+Year 5.78 0.83 1.11 1.09, 1.12 Above+Quarter 4.00 0.92 1.08 1.10,1.13 Above+Day of Weekl 6.31 0.84 1.12 1.10, 1.13 Above-Quarter+Month 2.39 0.94 1.05 1.01,1.08 MlaxPM1O=daily maximum monitor, PM1O=daily average of all monitors,* one-day lag of inimnum daily temperature; ColdlO=coldest 10 percent of the days; Hotl0=hottest 10 percent of the days; Year-binary variable for each year, Quarter-binary variable for each quarter, Month=binaiy variable for each month. Relative Risk is the ratio of predicted mortality when the pollution variable is set at 1-5 times the mean versus .5 times the mean, respectively- In models with MaxPMI0, the mean is of 140 pm3 (so the values are 210 vs 70) and in models with PM1O the mean is 115.4. ln all models PM1O is significant at p