The health effects of air pollution in Delhi, India

WPs 1%(Po POLICY RESEARCH WORKING PAPER 1860 The Health Effects of Particulate air poliution has less overall impact on Air Pollution in Delhi, India nontraumatic deaths in Delhi, India, than in U.S. cities. But the deaths occur earlier in life Maureen L. Cropper in Delhi, which could mean a Nathalie B. Simon Anna Alberini larger loss in life-years. P. K. Sharma The World Bank Development Research Group December 1997  POLICY RESEARCH WORKING PAPER 1860 Summary findings Cropper, Simon, Alberini, and Shar-a report the results Although air pollution seems to have less impact on of a time-series study of the impact of particulate air mortality counts in Delhi, the number of life-years saved pollution on daily mortality in Delhi. They find: per death avoided is greater in Delhi than in U.S. cities * A positive, significant relationship between - because the age distribution of impacts in these two particulate pollution and daily nontraumatic deaths as places varies. In the United States particulates have the well as deaths from certain causes (respiratory and greatest influence on daily deaths among persons 65 and cardiovascular problems) and for certain age groups. older. In Delhi, they have the greatest impact in the 15- * In general, these impacts are smaller than those to-44 age group. That means that for each death estimated for other countries, where on average a 100- associated with air pollution, on average more life-years microgram increase in total suspended particulates (TSP) would be saved in Delhi than in the United States. leads to a 6-percent increase in nontraumatic mortality. Large differences in the magnitude of effects do call In Delhi, such an increase in TSP is associated with a 2.3- into question the validity of the "concentration-response percent increase in deaths. transfer" procedure. In that procedure, concentration- * The differences in magnitudes of the effects are response relationships found for industrial countries are most likely explained by differences in distributions of applied to cities in developing countries with little or no age at death and cause of death, as most deaths in Delhi adjustment, to estimate the effects of pollution on daily occur before the age of 65 and are not attributed to mortality. causes with a strong association with air pollution. This paper - a product of Development Economics Research Group - is part of a larger effort in the group to examine the benefits and costs of pollution control. The study was funded by the Bank's Research Support Budget under research project "Measuring the Health Effects of Air Pollution in Developing Countries: The Case of Delhi, India" (RPO 679-96). Copies of the paper are available free from the World Bank, 1818 H Street NW, Washington, DC 20433. Please contact Anna Marie Maranon, room N10-031, telephone 202-473-9074, fax 202-522-3230, Internet address amaranon@worldbank.org. December 1997. (40 pages) The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the view of the World Bank, its Executive Directors, or the countries they represent. Produced by the Policy Research Dissemination Center  The Health Effects of Air Pollution in Delhi, India by Maureen L. Cropper, Nathalie B. Simon, Anna Alberini, and P.K. Sharma Maureen L. Cropper is a principal economist at the World Bank (DECRG). Nathalie B. Simon is a consultant to the World Bank (DECRG) and a graduate student at the University of Maryland. Anna Alberini is an assistant professor at the University of Colorado. P. K. Sharma is the Assistant Registrar of Births and Deaths for the New Delhi Municipal Committee.  I I I  The Health Effects of Air Pollution in Delhi, India by Maureen L. Cropper, Nathalie B. Simon, Anna Alberini, and P.K. Sharma 1. INTRODUCTION An important reason for controlling air pollutants such as particulate matter or sulfur dioxide is the damaging effects they have on human health. These effects include premature death, as well as increases in the incidence of chronic heart and lung disease. Estimates of the health damages associated with air pollution are important because they can provide both an impetus for environmental controls and a means of evaluating the benefits of specific pollution control policies. To estimate the health damages associated with air pollution in developing countries, policy makers are often forced to extrapolate results from studies conducted in industrialized countries. These extrapolations, however, may be inappropriate for two reasons. First, it is not clear that the relationships found between pollution and health at the relatively low levels of pollution experienced in industrialized countries hold for the extremely high pollution levels witnessed in developing countries. Levels of particulate matter, for instance, are often three to four times higher in developing countries than in industrialized ones. Second, in developing countries, people die at younger ages and from different causes than in industrialized countries, implying that extrapolations of the impacts of air pollution on mortality may be especially misleading. 1  This paper reports the results of a study relating levels of particulate matter to daily deaths in Delhi, India between 1991 and 1994. We focus on Delhi, the capital of India, because it is one of the world's most polluted cities. During the study period, the average total suspended particulate (TSP) level in Delhi was 378 micrograms per cubic meter--approximately five times the World Health Organization's (WHO) annual average standard. Furthermore, TSP levels in Delhi during this time period exceeded the WHO 24-hour standard on 97 percent of all days on which readings were taken. In addition, the distributions of deaths by age and by cause in Delhi are very different from those in the U.S. In the U.S., over 70 percent of all deaths occur after the age of 65. In Delhi, over 70 percent of all deaths occur before the age of 65, with over 20 percent occurring before the age of five. Furthermore, 46 percent of all non-trauma deaths in the U.S. are attributable to cardiovascular disease compared to only 23 percent in Delhi.' Because the main effects of acute exposure to air pollution on daily deaths occur through impacts on cardiovascular and respiratory disease, for which age is a known risk factor, we expect these differences to affect the relationship between pollution and mortality. Our estimates of health damages have policy implications for pollution control in Delhi, and permit us to compare extrapolations from U.S. studies with actual pollution impacts. We find that a given reduction in TSP reduces non-trauma deaths in Delhi by a smaller percentage lStatistics for the U.S. are taken from WHO's World Health Statistics Annual, 1993. Figures for Delhi are reported in Appendix 1, Table A3 and are taken from the National Capital Territory of Delhi's Registration of Births and Deaths Act - 1969: Annual Report 1991. 2  than predicted by U.S.. studies. Indeed, the percentage decrease in deaths corresponding to a 100 microgram reduction in TSP is 2.3 percent--about one-third of the effect found in the U.S. On the other hand, because the age distribution of impacts vary from Delhi to the U.S., so do the number of life-years saved. The largest impact of particulates on daily deaths in the U.S. occurs among persons 65 and older. In Delhi, the largest impact occurs in the 15 to 44 age group, implying that for each death associated with air pollution, more life-years will be saved in Delhi than in the U.S. on average. II. ESTIMATING CONCENTRATION-RESPONSE RELATIONSHIPS FOR DELHI The relationship between air pollution and premature mortality is most often studied using time-series analysis of daily observations of the number of deaths and pollution levels. These studies capture the effects of short-term exposure to pollution on the probability of dying. The underlying assumption is that there is a distribution of susceptibility to the effects of air pollution in any population. People who are in a weakened physical state or who have a history of chronic obstructive pulmonary disease (COPD) or cardio-pulmonary problems are thought to be the most vulnerable. In the case of a sharp rise in pollution, the most vulnerable people are more likely to die. Clearly, this type of analysis does not capture all of the effects of pollution exposure. Long-term exposure to pollution can also reduce life expectancy by altering lung function and making people more susceptible to COPD. However, measuring the effects of chronic exposure requires a long-term prospective study in which a sample population is followed long enough for the chronic effects to manifest themselves. Two such studies have been completed to date: the 3  Harvard Six Cities study (Dockery et al., 1993) which followed participants over the course of 15 years and the American Cancer Society study (Pope et al., 1995) which lasted approximately eight years. Due to cost considerations and time constraints, our work focusses on the acute effects of air pollution. Because time-series studies focus on a given geographic location over a number of years, factors that are often thought to influence the health of the population, such as percentage of smokers, income level, occupational exposure to pollutants, access to medical care and age distribution, do not need to be incorporated into the analysis as they are considered to remain relatively constant within the study area over time. Typically, the only other factors aside from pollution included in these models are weather variables and seasonal controls. In measuring the effects of air pollution, most attention has been focussed on particulates, especially those particles measuring less than 10 microns in diameter (PM 10), which penetrate the lungs more readily. Even though particulate matter tends to be the pollutant most strongly associated with premature mortality, the presence of other pollutants may be important as well. A. Data Mortality data for years 1991 through 1994 were obtained from the New Delhi Municipal Committee (NDMC), one of the three distinct regions which comprise the National Capital Territory.2 Because the NDMC houses a large concentration of Delhi's hospitals, approximately one-fourth of the 60,000 deaths in Delhi each year occur in the NDMC, in spite of the fact that 2 While registration of vital statistics became mandatory in the National Capital Territory in 1957, only the NDMC maintained a sufficiently detailed, computerized data base suitable for time-series analysis of this type, with a full four years worth of data available as of January 1995. The other two regions, the Municipal Corporation of Delhi and the Delhi Cantonment Board, began their computerization efforts several years later and, at the time this study began, did not provide the same level of detail in their data as the NDMC. 4  only 3.6 percent of the population resides there. Although the NDMC data represent only 25 percent of all deaths occurring in Delhi, the geographic distribution of the Delhi residents who died due to non-traumatic causes in the NDMC mirrors the geographic distribution of the population, as shown in Table 1. In the Appendix, we compare the distributions of deaths by cause and by age for the NDMC and the National Capital Territory of Delhi. The distributions of deaths by cause are similar, with between 20 and 25 percent of all medically certified deaths attributable to causes associated with air pollution (respiratory illnesses and cardiovascular disease). Roughly the same percentage of deaths are attributable to infectious diseases and to perinatal causes. The distribution of deaths by age group differs somewhat between the NDMC and the National Capital Territory. Forty-three percent of deaths in the NDMC occur before the age of five, compared to 33 percent of deaths in Delhi. For this reason, we estimate impacts by age group, as well as for specific causes of death (respiratory illness and cardiovascular disease).3 We also estimate the impact of air pollution on all non-trauma deaths. Because of problems associated with pinpointing the precise cause of an individual's death, counts of total non-trauma deaths have often been used as the basis of time-series analysis of this type, even though the links are strongest between air pollution and both cardiac and respiratory disorders. Daily counts of total non-trauma deaths covering the study period are plotted in Figure 1. The 3 Respiratory illness includes ICD8 460-519, excluding 463, 464, and 474 while cardiovascular disease consists of ICD8 390-448. 5  data display a marked seasonal pattern, with the highest number of deaths occurring during the 4 rainy monsoon seasons The Central Pollution Control Board (CPCB) provided daily data on air pollution levels collected at the nine monitoring stations located throughout the city. Six of these monitors have been in operation since 1987 and are operated directly by the CPCB. The other three monitors, added to the monitoring network in 1990, are operated on behalf of the CPCB by the National Environmental Engineering Research Institute (NEERI). Readings of total suspended particulates (TSP), sulfur dioxide (SO2) and nitrogen oxides (NOj) are taken at each station on a rotational basis approximately every three days. The data include average, maximum and minimum daily observations of these pollutants at each monitor. The monitors were not in operation on weekends or holidays during the study period. Daily means of TSP, SO, and NOR, were calculated using all available readings on a given day. Figure 2 shows average daily Total Suspended Particulate (TSP) levels for Delhi over the four years of the study. WHO established a guideline range several years ago of 150 to 230 micrograms per cubic meter per 24-hour period (WHO, 1992). As shown in Figure 2, TSP levels routinely fall well above this 24-hour guideline and, at times, reach as high as six times the guideline levels.5 While two coal-fired power plants, a large motor vehicle fleet and chemical and cement industries located within the city all contribute to the high particulate levels in Delhi, we suspect that much of the TSP is currently attributable to resuspended dust and natural sources, 4With its semi-arid climate, Delhi experiences only three seasons: (l) the dry season from April through June; (2) the monsoon season from July through September and (3) winter from October through March. 5Note that two observations greater than 1200 micrograms per cubic meter were omitted from Figure 2 to make the graph more legible. 6  primarily because sulfur dioxide (SO2) and nitrogen oxide (NOJ) levels remain well below the WHO guidelines (see Figures 3 and 4). Even so, available data on the size composition of particles suggest that the ratio of particles less than or equal to 10 microns in diameter (PM 10) to TSP is about the same as in the United States (0.53). Daily meteorological data covering all four years of the study period were obtained from the National Climatic Data Center in Asheville, NC. These data, collected on a daily basis at Delhi's International Airport located in the southwest of city, include average daily temperature, maximum and minimum temperature, mean daily dew point temperature, rainfall and visual range. The seasonality demonstrated by the data are shown clearly in Figures 5 and 6. Temperatures never fall below freezing and typically range between 60 and 100 degrees Fahrenheit. During the study period, fewer than five percent of the days fell below 55 degrees and fewer than one percent exceeded 100 degrees. Descriptive statistics for all of the variables are provided in Table 2. B. The Econometric Model Daily mortality figures are considered counts of rare events and are therefore often modeled using Poisson regression analysis. We fit Poisson regressions to the Delhi mortality data using the method of maximum likelihood. Formally, the log likelihood function is: T logL E [-XI + ytlogXi - log(yt!)] (1) where y, is the count of deaths occurring on day t, Xt = exp(X, j) is both the mean and the variance of deaths, X, is a matrix of covariates on day t, and P is a vector of regression 7  coefficients. The predicted count of deaths on day t is therefore y, =exp (X, P), implying that, if pollution is entered in a linear fashion in the right-hand side of the model, the percentage change in mortality following a given change in pollution levels is--holding all other factors unchanged-- AC-pc , where C is the pollution concentration variable and Pc is the coefficient associated with this variable. The maximum likelihood estimate of P is obtained by applying the method of iteratively re-weighted least squares (IRLS) to minimize the quadratic form: [Y-'X}Q I[Y-X] (2) where Y is the Tx 1 vector of daily deaths, X =exp(XO) , X is the Tx 1 matrix of covariates, and Q=diag{exp(Xt P)} (Farhmeir and Tutz, 1994). We exploit this equivalence to modify expression (2) to correct for serial correlation. Specifically, we assume that daily deaths follow a serially correlated process of the first order with autocorrelation paramneter p, and implement the generalized estimating equation (GEE) approach devised by Liang and Zeger (1986), replacing the matrix 0 in equation (2) with the matrix A=AAPAY' where A=diag{exp(Xt,p)}, P={rt }, and rtd=pItsI , for t, s = 1, 2, ..., T. Coefficient estimates are obtained from an IRLS step nested within a grid search over values of p ranging between -1 and 1 .6 As is often the case with Poisson models, we are also concerned about the possible presence of over dispersion, which may signal omitted variables and tends to result, when 6Liang and Zeger (1986) and Prentice (1988) show that for large samples the GEE coefficient estimates are normally distributed, and provide the expression for their asymptotic covariance matrix. Had we not accounted for serial correlation, the estimates of 1 from the Poisson model that assume serial independence would remain consistent under general assumptions, but their standard errors would not be reliable. 8  ignored, in unrealistically low standard errors (Agresti, 1996). To avoid such a problem, the Poisson standard errors can be multiplied by a correction factor equal to x2/dof, where x2 is the value of the Pearson chi square goodness of fit statistic (see Agresti, 1996 and McCullagh and Nelder, 1989) and dof is the number of degrees of freedom, i.e., the number of observations minus the number of independent variables. III. MODEL SPECIFICATION A. Development of the Core Model We estimated the autoregressive Poisson model described above for total non-trauma deaths, as well as for deaths by selected cause (respiratory and cardiovascular deaths) and age group. Our modeling strategy was carried out in four steps. We began by accounting for the variability in the number of total non-trauma deaths using seasonal/cyclical terms, a daily time trend and year dummy variables. Then, we added weather variables to control for the impact of temperature and humidity. Next, we added pollution variables to see if they had any additional explanatory power. Because of the complexity of the autoregressive Poisson model, we used a log-linear specification and ordinary least squares to build the models, using the Akaike Information Criterion (AIC) at each step to select additional regressors. Finally, we corrected the model for serial correlation using IRLS as described above. Following Schwartz et al. (1996), we have controlled for the seasonality in the data by including trigonometric terms for cycles ranging from one year to 2.4 months in length. A daily time trend and dummy variables for the year of the study account for population increases and 9  other unobserved factors thought to influence the number of deaths. We considered using indicator variables for both the month of the study and the season to control for the seasonality in the data but found that the trigonometric terms provided the best fit. We accounted for the effects of weather on daily mortality by adding temperature and dew point temperature variables to the model. We considered both maximum and mean temperatures with various lag structures up to seven days in length as well as indicator variables for the hottest and most humid one, five and ten percent of the days during the study period. In the end, we found that contemporaneous mean weather variables and dummy variables for the most extreme 10 percent of the days were the best predictors. With season and weather accounted for, we incorporated particulate pollution into the model. Again, we considered daily maximum, minimum and mean values of TSP as well as values lagged up to three days. While we did not find significant effects of contemporaneous levels of TSP on mortality, we did find that mean TSP lagged two days had a significant effect. Again, using AIC, we found that this specification provided the best fit. The results of the autoregressive Poisson model using our core specification are presented in Table 3. Particulate pollution has a positive and significant impact on daily mortality (P=0.00023, se=0.0001) as does mean daily temperature (P=0.01 1, se=0.0047) and the dummy variable for the most humid days (,B=0.008, se=0.0019). Mean daily dew point temperature also has a significant impact on daily mortality; however, the coefficient is negative (P=-0.002, se=0.0007). This same model is used to estimate the relationship between particulate pollution and the other mortality endpoints (see Tables 4 and 5). 10  B. Sensitivity Analyses To examine the robustness of the core model, we conducted several sensitivity tests. First, we added dummy variables to the model to indicate which monitoring stations were included in the calculation of the TSP variable for a given day. As shown in Tables 4 and 5, the addition of these variables did not affect the significance of the pollution variable nor did it affect the magnitudes of the coefficients dramatically. An F-test of the joint significance of these variables, however, did not allow us to reject the null hypothesis that all the coefficients were equal to zero. Since other studies have shown statistically significant impacts of SO2 on daily mortality, we examined the impacts of SO2 on daily deaths in Delhi. First, we considered the impact of SO2 on mortality alone and then included it in the model with TSP. When considered alone, SO2 was found to have a negative coefficient for every mortality endpoint except cardiovascular disease and was significant for all endpoints except respiratory deaths, cardiovascular deaths and deaths in the 0 to 4 age group. When considered jointly with SO2, the effect of TSP on daily mortality retained its significance in all cases (except total non-trauma deaths) and, in general, did not change in magnitude. While the negative coefficient on SO2 is somewhat puzzling, we do not give much weight to these results. SO2 levels during the study period were very low (see Figure 3) and the correlation coefficient between SO2 and TSP, while positive, is not large (less than 0.3). In studies where SO2 was found to have a significant and positive effect on daily mortality (Xu et al, 1994; Sunyer et al., 1996, Vigotti et al., 1996), SO2 and TSP were much more highly correlated and SO2 levels were in general much higher. We do not report the effects of NO, on 11  daily deaths since levels during the study period were similarly low and the link with daily mortality is not well established in the medical literature. We tested the model's sensitivity to outliers by removing the highest five percent of the TSP values and repeating the regressions. While admittedly the magnitudes of the coefficients were affected in some specifications of the model by the removal of these extreme values, in general, the significance of the TSP variable was not. In some instances, such as cardiovascular deaths, deaths among children aged 5 to 14, and among the elderly, the impact actually increased once the extreme values were removed. Finally, to test for the effects of over dispersion, we estimated models for each endpoint using a serially independent Poisson model and then applied the correction used by McCullagh and Nelder (1989) and Agresti (1996). Correcting for over dispersion in this manner had only slight effects on the magnitudes of the coefficients on TSP and did not affect their significance. IV. DISCUSSION A. Comparison of Results with Other Studies One of the striking features of the results reported in Table 3 is the fact that the impact of TSP, while statistically significant, is small when compared to results for other cities. Table 6 lists the results of daily mortality studies conducted in a number of cities in the U.S., Europe, and other developing countries. All of the studies listed find a significant and positive relationship between particulate pollution and daily mortality, with an average impact on daily mortality 12  (excluding Delhi) of six percent for a 100 gig/m3 increase in TSP.7 By contrast, in Delhi a 100 microgram per cubic meter change in TSP is associated with only a 2.3 percent change in total non-trauma mortality. There are many potential explanations for the seemingly small impact of particulate pollution on mortality in Delhi. One line of reasoning focusses on differences in the particulate matter itself. For example, it could be that the health effects of particulate pollution are lower in Delhi because the size of the particulates is larger than in western cities, thereby impeding penetration into the lungs and causing fewer detrimental effects. In 1992, however, NEERI measured PM 10 at the three monitoring stations that they operate (NEERI, 1994). Results indicate that the ratio of PM1O to TSP at these monitoring stations is only slightly smaller than the average ratio occurring in the U.S. (0.53 compared to 0.55 in U.S. cities). Even after converting the mean TSP value of 378 p.g/m3 to PM1O using the local ratio to account for the slight difference in particle size, we find that, on average, PM 1O over the four year period reached 200 ig/rm3, a value well above even the 24-hour standard set by the U.S. of 150 Vg/M3. It may be however that, because a large proportion of particulate pollution in Delhi arises from natural sources rather than combustion of fossil fuels, the chemical composition of the particulates is such that they pose a lower risk to health. Unfortunately, not enough is known about the chemical composition of particulates in the various cities to make a comparison. 7 Bacharova et al. (1996) recently completed a daily mortality study for the Slovak Republic and find no significant relationship between particulate pollution and daily mortality. In a similar study conducted in Poland, Wojtyniak and Piekarski (1996) find mixed results for four cities. Zmirou et al. (1996), on the other hand, find no significant relationship between particulates and total non-trauma mortality in Lyon, France but do find significant results for respiratory deaths. 13  Another line of reasoning, and in our opinion a more compelling one, focusses on demographic differences. Deaths in Delhi tend to occur at much younger ages than in the U.S. and, also, for much different reasons. To place our results in context, we compare our estimates of the impact of TSP on mortality in Delhi with similar estimates for Philadelphia (Schwartz and Dockery, 1 992a), presented in Table 7. The Philadelphia results are based on an econometric model similar to ours and use TSP as the measure of air pollution. Average TSP levels in Philadelphia during the time of the study were, however, 300 micrograms lower on average than levels in Delhi. The difference in the impact on total non-trauma deaths is no doubt due in part to the continued importance of infectious diseases in Delhi (accounting for close to 20 percent of deaths in our sample) as well as to the smaller role played by cardiovascular and respiratory deaths. The latter account for only a quarter of the deaths in our sample (see Appendix), whereas they account for approximately half of all deaths in Philadelphia. This, however, is not the entire story. Table 7 indicates that the impact of TSP on cardiovascular and respiratory deaths is also smaller in Delhi than in Philadelphia. This result is most likely driven by differences in the nature of these illnesses between the two cities. For example pneumonia, which has a weaker association with particulate matter than COPD, comprises a larger fraction of respiratory deaths in Delhi. The differences in the effects of TSP by age are also striking. In Philadelphia, the impact of TSP on deaths before the age of 65 is not statistically significant. The main impact of TSP is on persons who die after the age of 65. In Delhi the impact of TSP on deaths after the age of 65 is not statistically significant. Peak impact occurs between 15 and 44, with significant positive 14  effects also found for the 5 to 14 and 45 to 64 age groups. A clear implication of this finding is that more life-years are likely to be lost per person due to the impacts of air pollution. B. Policy Implications One of the implications of our findings is that extrapolations of the mortality impacts of air pollution to developing countries from U.S. studies are likely to be misleading. If one applies the coefficients from Schwartz and Dockery's study to the population of Delhi, the number of premature deaths associated with air pollution is over estimated. As shown in Table 7 using vital statistics for 1991, deaths would increase by 1,385 in Delhi if TSP were to increase by 100 micrograms, whereas the Schwartz and Dockery coefficient for total non-trauma deaths predicts an increase of 3,524 deaths8' 9. If one cares about life-years lost, however, the impacts of a 100 microgram increase in TSP in Delhi are more startling. As illustrated in Table 8, weighting each of the 1,385 lives lost by remaining life expectancy implies a loss of 51,403 life-years. To put this number in perspective, we use Table 8 to contrast the impact of increasing TSP by 100 micrograms in Philadelphia and in Delhi. Although the impact of the change in air pollution on total non- trauma deaths is lower in Delhi than in Philadelphia, the same increase in air pollution causes more life-years to be lost in Delhi. This is because more deaths occur at younger ages in Delhi 8 The number of lives lost per 100 jig/m3 increase in TSP is calculated by multiplying the total number of non-trauma deaths in Delhi (52,601) by 0.067, the coefficient on the TSP variable from the Philadelphia study. 9 Brandon and Hommann (1995) estimate the number of deaths that could be avoided in Delhi using the accepted metric (computed from U.S. studies) that a 10 jg/m3 change in PM IO leads to a one percent change in mortality. They estimate, for a 141.16 g/rm3 change in PMIO (a change large enough to reduce levels to the lower end of WHO guideline range), that 7,490 deaths would be avoided. Using our estimates, a change in PM] Oof this magnitude would result in only 3,430 avoided deaths, less than half of what they predict. 15  than in Philadelphia and because the impact of air pollution on deaths is greater at younger ages in Delhi than in Philadelphia. VI. CONCLUSIONS We draw two conclusions from this study: (1) The impact of particulate matter on total non-trauma deaths in Delhi, India is smaller than effects found in the U.S. This is due to the fact that in Delhi a greater proportion of deaths occurs at younger ages and from causes not associated with air pollution than is the case in the U.S. Estimating the number of non-trauma deaths associated with air pollution in a developing country city based on U.S. studies may therefore yield misleading results. (2) The impacts of air pollution on deaths by age group may be very different in developing countries than in the U.S. In the U.S., peak effects occur among people 65 and older. In Delhi, peak effects occur in the 15 to 44 age group, implying that a death associated with air pollution causes more life-years to be lost. We recognize, however, that our results are based on data representing only 25 percent of the deaths that occurred in Delhi during the study period. When data for the remainder of the city become available, this study should be replicated to confirm our findings. 16  TABLE 1: Distribution of the Population of Delhi vs the Distribution of the NDMC Deaths Among the Three Areas of Delhi by Place of Residence Census Population, 1991 NDMC Non-Trauma Deaths, Region 1991-94 Percent of T Percent of Number Total Number Total Municipal Corporation of Delhi -- Urban 8,075,935 95.3 34,455 94.4 New Delhi Municipal Committee 301,297 3.6 1,999 5.5 Delhi Cantonment Board 94,393 1.1 49 0.1 Total 8,471,625 100 36,503 100 Source: National Capital Territory of Delhi, 1991. Current Study. 17  TABLE 2: Descriptive Statistics for Variables Used in Regression Analysis Standard Variable N Mean Deviation Minimum Maximum Weather Variables: -Temperature 1393 76.32 12.73 45.4 102.5 * Dew Point 1392 57.91 12.34 28.3 80.2 Mortality (non-trauma)': * Total Deaths per day 1461 24.98 7.07 6.0 50 * Respiratory Deaths 1461 2.22 1.68 0 14 * Cardiovascular Deaths 1461 3.86 2.10 0 13 *Ages 0 to 4 1461 10.61 4.63 0 31 * Ages 5 to 14 1461 1.41 1.25 0 7 *Ages 15 to 44 1461 4.86 2.45 0 15 Ages 45 to 64 1461 4.78 2.39 0 14 Ages 65 and over 1461 3.34 1.97 0 12 Pollution Measures: Average Daily TSP 961 377.56 139.44 68 1714 Average Daily SO2 974 19.04 8.84 3.2 79.93 Average Daily NO, 974 32.74 11.80 6.2 127.54 'Included deaths occurring among Delhi residents only. 18  TABLE 3: Regression Results for Autoregressive Poisson Model Dependent Variable=Total Non-trauma Deaths Variable Definition Coefficient Standard Error Constant Intercept 2.83* 0.3165 TSPLAG2 Mean daily TSP lagged two days 0.00023* 0.0001 TEMP Mean daily temperature (