JOURNAL
An
OF ENVIRONSIENTAL
Empirical
ECONOMICS
Analysis
AND
of Air
Pollution
ROBERT
Economics
Department,
SIANAGEMENT
8,
8!%106
( 1979)
Dose-Response
Curves1
MENDELSOHN
University
of Washington,
Seattle, Washington
98195
AND
GUY Economics
Department,
ORCUTT
Yale University,
Received January
New Haven, Connecticut
06520
16, 1978; revised July 24, 1978
The authors combine information from 2 million death certificates and 2 million observations from the Public Use Sample. With several strategies for controlling extraneous variation, the authors explore these data in order to measure the chronic effects of several air pollutants on white mortality rates. In the United States, approximately 140,006 deaths a year (9% of all deaths) may be associated with air pollution. The size of this effect increases dramatically with age, with children displaying no detectable associations. Some pollutants, especially sulfate, are closely associated with many deaths, whereas other pollutants, especially ozone and nitrogen dioxide, have no apparent effect on expected lifetimes.
INTRODUCTION Our society currently enforces a number of environmental policies designed to reduce damage caused by air pollution. In 1973, these regulations prompted industry aIone to spend 2.4 billion dollars on air pollution control, The overall cost to society of air pollution regulations is substantially larger than these expenditures alone if one considers transportation policies, lost production opportunities, and the uncertainty of future regulation. What then, is society achieving with these environmental policies? One expected benefit of these expenditures is reduced population mortality rates. The following study explores the association between mortality rates and ambient concentrations of air pollutants in order to verify and quantify these expected effects. The effects on humans 0; long-term low-level exposure to air pollutants are very complex and subtle phenomena. Conceivably, the effects vary with the climate, population characteristics, and timing of the exposures. Also, the magnitude of the effects may be small and thus easily overshadowed by other unrelated events. In order to isolate the effects of pollution and to explore possible interactions among pollution variables, other environmental factors, and personal variables, one would like to carry out a large program of well-planned 1 This work has been supported by ISPS and the Pierce Foundation. We would especially like to thank Jan Stolwijk, Brian Leaderer, Eric Hanushek, Susan Rose-Ackerman, and Eugene Scskin for their helpful criticism. 85 0095~0696/79/020085-22$02.00/O All
Copyright @ 1979 by Academic Press Inc. rights of reproduction in any form reserved.
86
MENDELSOHN
AND
ORCUTT
experimentation. Moral standards and cost prohibit such experimentation, however, because it would involve purposefully exposing large numbers of people to potentially harmful substances. As an alternative, epidemiologists have searched for natural experiments in which, inadvertently, all the relevant variables have similar values, except for the particular pollutants. It is dificult, though, to find environments whose only difference is the desired treatment variable.2 Most natural experiments involve an unbalanced &sign in which several environmental factors vary simultaneously. Although this means that natural experiments are complex and difficult to interpret, with sufficient data and careful analysis they can provide important information about our environment which is otherwise not attainable, Past studies concentrated frequently upon removing the confounding influences of basic personal characteristics such as age, race, and sex. Often, large samples, from only a few sites, were carefully selected so that these basic personal characteristics were similar across all sites. Although this method may remove the most influential impacts of variations in personal characteristics, other sources of variation may continue to plague any analysis of the data. With only a handful of sites, unwanted variations across sites cannot be removed even if there are only a few unwanted variables. In order to detect and take account of unwanted sources of variation, one needs detailed information about each site and as many sites as possible. Previous data bases lack detailed information about both personal and area-wide characteristics3 Also, previous studies did not take advantage of the natural variation in pollution concentrations across the country and often used two pollutants to represent the entrie spectrum of air pollutants.4 It is quite likely, given the limited available information, that at least part of the chronic mortality effect attributed to pollution in past studies is really caused by other area-wide factors. The data in this study provide an improved opportunity to disentangle the interactions among extraneous factors and the adverse health effects of air pollutants. The Public Use Sample, a rich account of over 2 million individuals, has not been helpful to past researchers because it lacks health information such as mortality rates. This study overcomes this deficiency by carefully matching death certificates with the Census data by county groups and personal characteristics. Additional data sets on pollution concentrations are also matched by county group. This matching of large data sets (which were collected for disparate pur~See, for example, Warner and Stevens [Zl], who criticize Shy [17] for neglecting the effects of nitric acid vapor. 3 Traditional epidemiological studies resort to examining three or four sites in one metropolitan area (see Finklea [5], National Academy of Sciences [14, 151, or Shy [17]). The few studies which examine severa sites generally have only limited information about both persona1 and area-wide characteristics (see Lave and Seskin [B, 9, 111). By disaggregating their analysis by age, sex, and race cells, Lave and Seskin [ 121 make an important improvement over their previous analyses. 4 Since urban and rural citizens are often exposed to very different levels of pollution, a study across such sites has the advantage of widely varying pollution dosages. It is hoped that the inclusion of many undesired variables reduces the unwanted variations which are associated with widely differing sites. The inclusion of many pollutants is also an important improvement if one is concerned about which pollutants are really causing damage. It is particularly important to include secondary pollutants because they are less likely to be associated with their emission sources since they are formed in the atmosphere hours after being emitted.
AIR
POLLUTION
poses) provides a substantially ships, and at a modest cost.
DOSE-RESPONSE
improved
DATA
opportunity
CURVES
sy
to explore desired relation-
BASE
Cur data base is a composite of four large United States data sets which, together, provide somewhat comprehensive information about individuals, the areas they live in, their climate, and the ambient levels of several pollutants in their area. About 2 million recorded death certificates filed in 1970 are organized into 24 race, sex, and age cells by county group. Another 2 million observations from the 1970 Public Use Sample are matched to the race, sex, and age cells for each of the 404 county groups in the study.” This matching provides not only cell-specific mortality rates but also socioeconomic variables such as mean education, income, and age for each cell in the county group. The effects of the mean education level of white males aged 25-44, for example, can thus be separated from the effects of the mean schooling level of the entire county group. Variables describing the entire county group are computed from both the Census Public Use Sample and the 1970 City and County Data Book. Finally, statistics on pallution variables in 1974, from EPA publications, are matched to the appropriate county groupsG Despite our efforts to include the best available pollutant measurements and other data in the analysis, there are still many problems with the available information. Some of the pollutant measurements are not accurate, measurements taken in one location often have to represent large geographical areas, and historical pollution measurements are virtually nonexistent, Also, studies by Amdur [1, 21 suggest that we would like to have more information about the size and type of particles in the air. Additional problems with the data are measurement errors in the reporting of cause of death and possible biases from population mobility. It is always possible that our pollution variables merely serve as inadcquately measured proxies of some nonmeasured pollutants or other variables.? Z The U.S. Census has aggregated the 3000-odd counties in the United States into 408 county groups of over 250,000 persons each. Hawaii, Alaska. and Washington, D.C., arc cxeluded from our study. ‘; Sulfate and nitrate statistics were made available through personal correspondence with Martha Abernathy, National Air Data Bank, U.S. EPA, Research Triangle Park, North Carolina The remaining statistics were collected from “The National Air Monitoring Program: Air Quality and Emissions Trend Annual Report” (1973) and “Air Quality Data-1974 Annual Statistics” (1976), by the U.S. EPA, Research Triangle Park, North Carolina. Pollutant measurements for both 1974 and 1970 were tried in the regressions. The coefficients for the two regressions are similar, but the 1974 coefficients are slightly more significant. This is probably due to the fact that there are more numerous and accurate observations for 1974. The values for 1974 are used in the paper. The benefit of having more observations is that they provide a closer matching between the air pollution data and the county groups. Two methods of matching these data were tried. EPA air quality regions were assigned to Census county groups through which the data were combined. The second method was to determine the exact location of each monitoring station and assign its data to the relevant county group. The two methods produce similar results. County groups which have no monitoring stations are assumed to have the air quality of their neighbors. These air measurement problems probably cause some measurement errors in the independent variables. As a result, the pollution coefficients may be biased toward zero. 7 Omitted variables will not necessarily lead to biased coefficients. For instance, the absrnct of smoking habits in this study may not bias the pollution coefficients since there is no evidence
88
MENDELSOHN
THE
MODEL
AND
ORCUTT
AND
In order to test the effects of pollution regression of the following form: 8
RESULTS on
mortality
rates,
developed
we
a
Di = Bo + BlXli + BzXz, +. . . + GlZl + GzZz+. . . + Ei. (1) Whether the individual died over the year (Di: a zero-one dummy) is equated to his set of personal characteristics( Xii, XZi, . . .), the variables for his county group (21, 22, . . .), and an error term (Et). In order to estimate this relationship, the individual observations are aggregated by cell (age, race, and sex) to the county group level.
+ NGlZl + NG,Zz +. . . + C Ei.
(2)
Dividing Eq. (2) by N, the number of observations in each cell and each county group, yields the regression model we estimated for the probability of death cpD>* p*=~Di/N=Be+Bl~XliiN+Bz~X2i!N+,.. 1
1
1
+ G,Zl + GzZz+. . . + g El/N.
(3)
1
In order to correct for observed heteroscedasticity in the error terms, we use weighted least squares, multiplying each observation by the square root of the appropriate population sizes9 Ideally, one would like to choose data in which the only independent variables which changed across observations were the pollutant variables. Unfortunately, with complex phenomena such as the environmental causes of death, almost all data will have unwanted variation from sources other than pollutants. If this unwanted variation is not addressed in the regression model, it will reduce the precision of the estimates of the dose-response curves. In addition, if these factors are spatially correlated with pollution, their omission from a regression equation could lead to serious biases in the pollution coefficients. Because we analyze individual age, race, and sex cells separately, we already dramatically reduce effects due to the distribution of the population across county groups. To the extent that we accurately model the effects of other unwanted factors, we can also dampen the magnitude of the problems due to these remaining influences. that pollution concentrations across the country are correlated with smoking habits. Further, studies which do measure smoking habits still detect an additive effect from pollutants (see National Academy of Sciences [ 141 or Finklea [5]). Also, the absence of hydrocarbons or oxidants should not lead to serious biases in the remaining pollutant coefficients. Analyses of subsamples of the data which include either of these pollutant types indicate they are generally insignificant. 8 This model was developed by Orcutt et al. [16]. @ Heteroscedasticity due to the variance of the probability of death has been substantially reduced by stratifying the sample by age.
AIR
POLLUTION
DOSE-RESPONSE
CURVES
SC3
a first approximation, we have assumed that these extraneous factors have additive linear effects on mortality. Several categories of these unwanted factors have been analyzed in this study. The first category describes the characteristics of the particular age-race-sex cell such as the average age, family income, and percentage divorced. County group variables make up a second set of variables. These variables are intended to be proxies of area-wide factors such as the level of public hygiene, urbanization, and industrialization. Because of the important interactions between meteorological variables and pollutant concentrations, proxies for the local climate are also included. Finally, dummies for regions have been included in some of the regressions (all regions but North Central). The means, standard deviations, and units of all the variables used in the regressions are available in the Appendix. In an effort to understand how extraneous factors impact on the pollution COefficients, we have run several regressions for each cell (the results for white adults aged 45 to 64 are shown in Table I). lo The first regression includes only a constant term and the pollutants as independent variables. The second equation includes several characteristics of the cell being studied. The third equation includes a host of variables which, one hopes, represent the most important areawide factors. In the fourth equation, we control the variations among regions in order to examine the variation within regions.ll This regional technique controls for unmeasured region-wide factors which may confound the other regressions. The parasimonious regressions include only unwanted variables which are highly significant across cells. The final regression drops all pollution variables except sulfate. The introduction of extraneous variables, either personal or areawide factors, reduces the magnitudes of the coefficients for all the primary polluants. Concentrations of primary pollutants are spatially associated with the emission sources. Because many emission sources are located downtown or in industrial centers, there are many factors which could be correlated with primary pollutants. The introduction of these other factors in the analysis clearly diminishes the impact of primary pollutants. Ozone, sulfate, and nitrate, on the other hand, are products of atmospheric reactions. Their peak ambient concentrations are often many miles from the original source, The introduction of extraneous factors in the regressions understandably has no consistent effect on these secondary products. National spatial cross-sectional data are particularly suitable for detecting the health effects of secondary products. The likelihood that the pollutant IeveIs consistently act as a proxy measure for an unmeasured factor is small. The strong pattern of sulfate coefficients across regressions is reassuring evidence that this pollutant, at least, does not act as a proxy for other unwanted variables. As
10The series of regressionsfor whites of all ages is available front the authors. “extraneous factors” refers to causes of death which do not alter the consequences of tion
The tern) air polln-
exposures on premature deaths. 11 Inclusion of regional dummies is functionally equivalent to taking the difference from the regional mean for each observation, This technique parallels work on time series in which one transforms one’s data into the difference between the observation and a moving average of several years. Some of the variations among regions are captured through the climatic variables. Relative humidity instead of water vapor was also tried in the regressions. The coefficients of these two measures of water in the air are quite similar. See Lave and Seskin [12] for an examination of other meteorological variables,
county group Age Education Relative income No. of children % Divorced 7O Never married Density Unemployment Y0 Owner occupied % w/o plumbing % Overcrowded Net migration Centrrtl city
Personal characteristics Age Education Relative income 70 Divorced 7O Never married
OIOIW
Pollutants Sulfate Nitrate Particulate8 Carbon monoxide Sulfur dioxide Nitrogen dioxide
1.142 -2.722 0.043 0.555 0.164 -0.038 0.056
Crude
(4.28) (3.25) (0.74) (0.80) (2.36) (0.71) (0.79)
Multiple
0.919 (5.00) -0.580 (4.23) -0.285 (1.12) 0.392 (8.66) 0.113 (3.24)
1.46 (6.32) - 1.99 (2.96) -0.018 (0.39) 0.081 (0.15) 0.201 (3.55) -0.024 (0.55) -0.019 (0.35)
Cell characteristics
Regression
-0.026 -0.897 -1.154 -2.191 0.297 0.241 -0.044 0.033 -0.001 -0.013 -0.114 0.010 0.918
0.404 0.240 0.097 0.021 -0.053
(0.41) (2.26) (1.83) (3.92) (3.72) (3.83) (2.83) (0.59) (0.08) (0.66) (2.57) (1.97) (3.50)
(1.96) (0.83) (0.18) (0.35) (1.03)
(5.55) (2.83) (0.42) (0.79) (2.99) (1.48) (2.19)
Area-wide
of Illness-Related
1.22 -1.61 -0.020 0.360 0.160 -0.050 -0.100
Analysis
0.526 -0.521 0.877 0.082 -0.040
0.864 -0.410 0.036 0.554 0.100 -0.C95 -0.110
(0.72) (0.83) (0.28) (4.45) (2.48) (3.84) (3.72) (2.46) (1.24) (0.40) (1.97) (1.85) (4.12)
(2.53) (1.61) (1.56) (1.37) (0.76)
(3.31) (0.59) (0.87) (1.14) (1.81) (2.45) (2.48)
Regional
Death
Ia
0.046 0.372 0.188 2.51 0.296 0.246 -0.058 0.181 -0.016 -0.008 -0.102 0.010 1.10
TABLE Rates
(4.59) (3.48) (4.90) (3.12) (1.89) (1.71) -0.129 (3.61) 0.008 (1.60) 1.01 (3.84)
-2.15 0.275 0.201 -0.052 0.136 -0.020
(3.50) (2.22) (1.36) (1.39)
(1.91) (1.69)
0.014 -0.072
0.588 -0.291 0.629 0.083
(2.53) (0.46) (0.84)
White,
0.787 -0.358 0.034
Squared
: Male,
45-64a
(3.57) (2.49) (1.52) (1.34)
(2.01) (1.03)
-0.137 -0.077 1.01
(3.78) (1.50) (3.84)
-2.142 (4.52) 0.274 (3.41) 0.206 (4.86) -0.054 (3.43) 0.152 (2.14) -0.020 (1.73)
0.578 -0.333 0.702 0.079
0.126 -0X46
0.336 (3.04) -0.290 (0.41) O.C24 (0.57)
Dispersion
Aged
-2.02 0.556
0.173
1.67 -1.C6 -0.C98 1.35 0.153 -0.058 -0.068
(14.68) (15.29)
(0.46)
(8.16) (1.59) (2.53) (2.92) (2.96) (1.51) (1.45)
-1.88 0.531
0.206
1.54
(14.48) (14.86)
(6.38)
(8.04)
Parsimonious kxllfak only)
=i
2
AIR POLLUTION DOSE-RESPONSE CURVES
County group Age Education Relative income No. of children 70 Divorced To Never married Density Unemployment YO Owner occupied yo w/o ptumbing ?70Overcrowded Net migration Centml city
Personal characteristics Age Education Relative income Y0 Divorced Y. Never married
Pollutants Sulfate Nitrate Partieulates Carbon monoxide Sulfur dioxide Nitrogen dioxide OZOll.2
0.478 -0.145 0.017 0.684 0.143 0.001 -0.028
Crude
(4.16) (0.42) (0.71) (2.38) (5.05) (0.05) (0.97)
Multiple
0.264 (3.22) 0.026 (00.40) -0.399 (3.79) 0.126 (7.17) 0.078 (4.71)
0.649 (5.68) -0.135 (0.43) -o.om (0.00) 0.584 (2.34) 0.132 (5.21) -0.015 (0.79) -0.077 (3.10)
Cell chsrrtcteristios
0.005 -0.090 -0.532 -0.371 0.082 0.104 -0.017 0.136 - 0.0005 -0.038 0.005 -0.0066 0.331
-0.022 0.044 -0.415 0.009 -0.002
0.730 -0.360 -0.010 0.480 0.130 -0.030 -0.070
(6.48) (1.23) (0.72) (2.10) (4.96) (1.69) (3.66)
(0.17) (0.45) (1.68) (1.04) (2.16) (3.36) (2.21) (4.64) (0.09) (3.64) (0.26) (0.23) (2.46)
(-0.22) (0.27) (1.28) (0.32) (0.07)
Area-wide
Ib
0.011 0.187 -0.491 -0.896 0.114 0.080 -0.021 0.103 -0.001 -0.018 0.0156 0.0007 0.429
-0.004 -0.131 -0.110 0.020 -0.009
(0.36) (0.90) 11.43) (2.52) (2.94) (2.63) (2.79) (2.79) (0.27) (1.60) (0.61) (0.25) (3.14)
(0.04) (0.80) (0.34) (0.74) (0.34)
(4.07) (0.13) (0.17) (2.05) (3,08) (1.90) (3.27)
Regional
(1.44)
-0.014
0.420 (3.40)
(2.55) (7.78) (4.72) (3.55) (1.52) (2.26)
0.343
-0.018
-0.567 -0.926 0.128 0.075 -0.017 0.093
0.076 -0.005
0.118 (3.15) -0.017 (0.84)
-0.541 -0.891 0.119 0.(169 -0.011 0.082
0.517 -0.116 0.014
(2.74)
(1.85)
(2.72) (8.07) (5.04) (3.91) (2.53) (2.60)
(2.42) (0.26)
(3.85) (0.34) (0.69)
Dispersion
(2.82) (0.49) (0.25)
0.425 0.181 -0.005
Squared
Death Rates: Female, White, Aged 45-64”
0.544 -0.646 -0.003 0.496 0.084 0.037 -0.074
TABLE
Regression Analysis of Illness-Related
-0.553 0.207
(7.72) (11.41)
(3.39)
(2.24)
-0.053
0.053
(7.78) (1.09) (0.77) (3.65) (4.97) (2.54)
0.799 -0.368 -0.015 0.851 0.128 -0.049
-0.428 0.204
0.075
0.816
(6.34) (11.13)
(4.51)
(8.22)
Parsimonious (sulfate only)
4
z
*
zz
-
Mts.
0 The figures
Constant K’ SEE
in parentheses
squared terms Sulfate Nitrate Particulate5 *Sulfur dioxide Nitrogen dioxide
lhpersion t~erma Sulfate Mitrate Particulate8 Sulfur dioxide Nitrogen dioxide
Rocky Pacific
Deep South South West Plains
Regions New England Middle Atlantic South East
Climate January temp. July temp. Water vapor in Jan. Water vapor in July
are t stati,+&.
5.65 (28.33) 0.902 0.802
Crude
-7.73 0.932 0.720
(1.71)
Cell cheracterintics
10.07 0.948 0.604
(1.88)
.Irea-wide
TABLE
7.92 0.955 0.:175
0.027 0.313 -0.470 -0.341 -0.555 -0.570 -0.666 -0.450
0.011 0.021 0.051 -0.044
(1.13)
(0.15) (2.48) (2.64) (1.47) (2.69) (3.72) (2.44) (1.89)
(0.78) (1.46) (1.06) (2.37)
Ib-continued
8.73 0.950 0.577
0.032 -0.00s -0.029 -0.017 -0.048
0.292 -0.367 -0.354 -0.460 -0.570 -0.576 -0.296
(8.28)
(1.21) (0.17) (1.15) (0.50) (1.31)
(2.73) (2.23) (1.62) (2.49) (3.99) (2.29) (1.32)
0.017 (1.59) 0.077 (1.81) -0.040 (2.35)
8.89 0.953 0.574 --___I
0.011 0.512 -0.004 0.012 -0.028
0.236 -0.407 -0.381 -0.550 -0.583 -0.443 -0.293
0.011 0.053 -0.033
(8.44)
(0.20) (1.82) (0.81) (0.98) (2.11)
(2.13) (2.43) (1.70) (2.86) (4.16) (1.75) (1.33)
(1.05) (1.22) (1.96)
Dispersion
4.73 0.940 0.681
-0.251
0.505
(7.90)
(1.64)
(5.02)
Parsimonious
3.57 0.929 0.734
-0.143
0.708
(5.86)
(1.13)
(7.75)
Parsimonious (sulfate only)
6,
2! s
94
MENDELSOHN
AND
ORCUTT
Controlling for between-region variance results in several consistent changes in the pollution coefficients. The effects on mortality of sulfate and sulfur dioxide are smaller for variations within regions than they are for variations between regions across all adult cells. One conceivable explanation of this phenomenon is that the sulfur compounds in regions with high concentrations of ambient sulfur are somehow more potent than the sulfur compounds in the other regions. Further investigation of the size and chemical makeup of ambient sulfur compounds in different regions could help explain these regional differences. After the introduction of regional variables, nitrates are no longer statistically significant. Withinregion variations of nitrates have no observable consistent effects on mortality rates. Because biological dose-response curves often assume an S shape, it is possible that a probit or a logit analysis is more appropriate than the linear model used here. However, laboratory experiments with animals suggest that the lethal dosage of these pollutants for 5070 of the population is about three orders ‘of magnitude above the normal daily dosage tested here. I2 It is apparent that the range of doses we examine is too low to trace out an S shape, but it may exhibit some nonlinear effects. Examination of the residuals of each regression indicates they are randomly distributed around the predicted values. At least within each age and sex cell, the overall linear assumption of the model seems reasonable. Also, squared terms for each pollutant, when introduced into the regressions, are almost always insignificant (the F tests indicate the coefficients as a group are not significantly different from zero). The linear dose-response curve seems to be a reasonable approximation for the doses and the effects actually measured. Another linear assumpti,on ,of the model is that the annual dose is merely the average of the daily pollution levels over the year. That is, daily variations around the annual mean pollutant level are assumed to have no effect. Of course, this conflicts with interpretations of daily mortality data which suggest that daily extremes cause many deaths. 13 In order to test this hypothesis, a measure of daily dispersion around the annual mean is included for each pollutant. If daily variations really have an acute effect, this measure should have a positive significant coefficient. In fact, the coefficients are often negative, and, as a group, they are not significantly different from zero (see Table I). Although episodes ‘of extremely poor air quality result in higher daily mortality rates, their only effect on annual rates is to raise the average annual dose of that pollutant. It is difficult to know which of the regressions in Table I are most correct. Luckily, the choice is not critical with this data set since most of the significant COefficients are relatively stable across the more complete specifications. In order to conserve space, only the parsimonious regressions for each age and sex cell for whites are shown in Table II. As can be seen from the summary statistics of Table II, the independent variables explain between 14 and 93% of the variation within each cell. Because 12 See, for example, National Academy of Sciences 114, Vol. 2, pp. 197-1991 on nitrogen dioxide experiments and National Academy of Sciences [15, Chap. 31 on sulfur dioxide experiments. Detectable changes in the physiology of laboratory animals begin at about 100 times the normal daily dosage. Many animals are exposed to up to 1000 times the normal average and still survive. Furthermore, man appears to be among the species which are more highly resistant to these pollutants. 13 See Lave and Seskin [lo], Goldsmith [6], or U.S. Public Health Service [18-201.
AIR POLLUTION
DOSE-RESPONSE
CURVES
!$,i
the series of regressions also captures the variation between the age and sex cells, the analysis accounts for over 987’L of the total variation of the mortality of the white population across county group cells. The models fit most cells so closely that they probably capture almost all of the sizable causes of death which var! across large areas. Another interesting result shown in Table II is that the coefficients for each cell are quite different. The differences across age groups are the most striking; the coefficients generally increase in size with age. There are no apparent patterns in the pollution coefficients of younger age groups (under 18). Air pollution appears to affect the mortality rate of adults only. Men and women also have different coefficients. Even with the level differences between the sexes controlled, the remaining male and female coefficients are different at 1% significance for all age groups except the elderly. Because of these marked differences among the age and sex cells, the pollution coefficients of each segment of the population arc examined individually. One of our most important findings is the significant difference in the effects of each pollutant. Sulfates have positive, sizable, and significant coefficients for all adult whites.14 Sulfur dioxide and carbon monoxide also have positive coefficients for almost all adult cells, but they are less accurate and smaller than the sulfate coefficients. The coefficients of the remaining pollutants are both positive and negative and generally insignificantly different from zero. Only ozone has consistently negative coefficients, and they are almost always insignificant, In order to gain a further perspective on the importance of the size of these coefficients, we have calculated the effect of existing levels of pollutants on white adults in the United States. Using the county group population, local ambient concentrations of each pollutant, and the dose response curves, we have estimated the effect of each pollutant on each cell in every county group. The national amma effects, the sum of the county group deaths, are presented in Table III. A best-point estimate and a confidence interval arc shown for each pollutant and cell. The dose-response curves of the upper- and lower-limit effects assume the components are statistically independent of each other. Assuming that we have correctly specified the model, we can be 95c/r certain that the true final effect of each pollutant lies within the range of these limits. One of the most striking features of Table III is the magnitude of the sulfatt effect. Just across white adult cells, sulfate accounts for 187,686 deaths per year. If one lowers these estimates by two standard deviations, sulfate still accounts for over 150,000 deaths per year. Even with several alternative specifications of the regression equation, the sulfate effect is large and significant. This paper presents striking evidence that sulfates are deleterious to human health. Both sulfur dioxide and carbon monoxide also appear to be lethal. In seven of the eight cells, the effects of both pollutants arc positive, and, in at least half of these cells, the positive effects are statistically significant. The numbers of annual expected deaths due to these two pollutants, 23,756 from sulfur dioxide and 21,403 from carbon monoxide, are still relatively small compared to the number of expected deaths due to sulfate. Thcrc is no evidence that the remaining pollutants shorten expected lifetimes.
0 The
SEE
R=
Constant
figures
Mid-Atlantic Pacific
Region
in parentheses
(3.15)
(1.61) (1.04)
are t statistics.
0.386 0.332 0.137
-0.033 -0.031
(0.11) (4.83) (1.82)
0.404 0.468 0.059
-0.027 0.024
0.046 -0.065 -0.003
(7.42)
(3.04) (1.78)
(3.15) (10.83) (1.59)
(0.24) (1.43)
0.334 0.728 0.043
-0.024 0.037
0.003 -0.066 0.003
0.002 -0.003
-0.003 0.034 -0.001
0.002 0.000
(8.49)
(0.37) (0.39)
(1.98) (15.19) (11.16)
(1.85) (2.08)
(2.29) (2.31) (0.41)
(0.24) (0.01)
-8.54 0.740 0.661
0.223 -0.450
0.336 0.027 0.042
0.025 -0.034
0.033 -0.012 -0.004
0.201 -0.177
(14.41)
(2.16) (3.12)
(21.22) (0.41) (2.32)
(1.38) (1.52)
(1.72) (0.05) (0.14)
(1.96) (0.55)
from
0.792 0.855 0.242
0.082 -0.195
0.027 -0.264 0.069
0.022 -0.008
0.003 0.254 0.023
0.203 -0.229
25-44
Illness
(3.52)
(2.23) (3.63)
(4.74) (10.75) (10.60)
(3.21) (0.93)
(0.37) (3.07) (2.42)
(5.49) (1.92)
: White
45-64
10.9 0.930 1.35
0.611 -0.884
0.173 -2.02 0.556
-0.058 -0.068
-0.098 1.35 0.153
-1.061.67
Malesa
(8.97)
(3.04) (2.92)
(5.46) (14.68) (15.29)
(1.51) (1.45)
(2.53) (2.92) (2.96)
(8.16) (1.59)
82.21 0.920 6.92
3.45 0.978
-0.506 -1.41 0.426
-0.796 -0.289
0.059 2.22 0.263
8.22 -2.15
(14.44)
(3.32) (0.64)
(3.53) (1.98) (2.32)
(4.05) (1.24)
(0.30) (0.91) (1.01)
(7.96) (0.64)
‘55-t
z!
E g
!2
5:
2
2
%
-0.000 -0.067 0.007
-0.000 -0.003
(0.61) (1.16) (1.37)
(3.54) (0.37)
18-24
on Deaths
Age of children No. Percentage divorced
(1.16) (0.31)
0.001 -0.023 -0.003
-0.032 -0.011
5-17
Regressions
f3
0.005 0.001
(2.19) (0.88) (0.34)
(0.83) (2.00)
l-4
Multiple
IIa
Group
County
dioxide
0.00s -0.041 0.002
Particulates monoxide Carbon Sulfur dioxide
Nitrogen Ozone
-0.017 0.134
Sulfate Nitrate
Pollutants
O-l
Parsimonious
TABLE
AIR POLLUTION
DOSE-RESPONSE
CURVES
163,587 187,686 211,784
3,069 12,304 16,560 20,817 49,007 63,857 78,707
1,962 761 1,915
Loo0 1,481
65,854 82,400
31,591 39,333 49,307
-42 2,042 4,125 2,788 4,386 5,983 23,848
-34,757 -4,193 26,371
-261 392 1,045 -1,735 101 1,937 -6,241 -1,735 2,771 -15,370 3,374 22,118
-392 101 593 -627 473 1,573 -6,310 -2,226 1,858 -1,460 12,447 26,354
-24,765 -2,081 20,603
24,495
- 18,105 3,195
-363 2,231 4,825 -2,040 463 2,966 -21,869 -12,214 -2,558
10,599
-62 -13,217 -5,854 1,509 -20,575 -4,988
-1,499
-2,480 -535 1,410 -3,060
III
-81,739 -63,936 -46,132
-685 -206 273 -1,254 266 1,786 -8,526 -4,770 1,014 - 40,569 -27,503 -14,437
-22,101 -29,764 -37,427
7,416 21,403 35,390
105 379 652 262 908 1,554 1,990 4,403 6,816 -4,987 3,538 12,063
479 1,374 2,269 2,012 6,386 10,763 -5,323 4,444 14,211
764 2,028 3,292
- 12,029 -5,175 1,679
-31 1,209
-1,271
1,188 2,910
-533
Carbon monoxide
-34,638 -20,220 -5,802
-6,402 2,490
-15,294
-452 -175 102 -691 -60 572 -5,529 -2,921 -313
461 530 -8,156 -3,428 1,300 - 15,994 -6,121 3,751
-
- 1,452
-1,509 -652 206
Ozone
87,570 142,415 197,260
1,052 2,236 3,419 1,066 4,139 7,211 4,485 14,783 25,081 21,752 55,648 89,543
-464 17,189 34,842 -348 36,992 74,332
11,133
-454 4,163 8,780 3,397 7,265
Net effect of air pollutants
Levels of Air Pollutants”
5.4 8.9 12.3
1,603,OOO
among components.
4.1 10.6 17.0
10.8 18.4
3.7 14.5 25.3 3.3
16.9
-6.8 13.8 6.3 12.7
52.6 6.5 13.9 21.2
24.9
% of all deaths s.ssoc. w/pollution
526,400
136,500
28,500
6,233
584,300
251,600
52,400
16,706
Total deaths from all causes
from the coefficients. Totals of these ranges assume independence
7,730 23,756 39,782
17 264 511 -114 536 1,186 3,270 5,472 7,674 -1,486 6,337 14,160
-1,222 -80 1,062 169 974 1,778 1,907 5,881 9,854 -8,764 4,372 17,508
Nitrogen dioxide
Whites From Local Exposure
Sulfur dioxide
Deaths of Adult
Partieulates
National
Nitrate
Annual
a The lower and upper dose-response curves are two standard deviations
18-24
Lower Expected Upper 25-44 Lower Expected Upper 45-64 Lower Expected Upper Lower 65+ Expected Upper Female age 18-24 Lower Expected Upper 25-44 Lower Expected Upper 45-64 Lower Expected Upper 65+ Lower Expected Upper Total effect Lower Expected Upper
Male age
Sulfate
Predicted
TABLE
z
3
zx b
Is z
3
s m
%
AIR
POLLUTION
DOSE-RESPONSE
CURVES
99
Nitrate, nitrogen dioxide, and particulates have both positive and negative coefficients which are rarely statistically different from zero. Ozone has only negative coefficients but these, too, are insignificant. While local air quality is a nationwide concern, the worst air quality is concentrated in the North Central and Northeastern regions. As can be seen in Fig. I, the probability of dying from air pollution in these two regions is about twice X5 high as that in the rest of the country. The maximum damage in the nation appears to occur in eastern Ohio, a region with close to the maximum concentration of sulfates. The best air quality in the country is broadly shared by most states west of the Great Plains. Searching for insight into the causal link between pollution and mortality rates, we examine whether the independent variables correlate with specific causes of death. Presumably, if a pollutant is particularly harmful to an organ (the lungs, for example), then high-pollution areas should have higher proportions of deaths from diseases of that organ. One can test these hypotheses with regressions on the absolute numbers and also the proportions of deaths caused by cancer, heart, lung, infections, and other illnesses, respectively. These regressions on the causes of death produce several patten1s.l” Sulfate is correlated with increases in the death rate from all causes, but it increases heart failure relatively more and cancer relatively less than the other causes. Particulates and carbon monoxide appear to be strongly correlated with heart failure-they raise the frequency of deaths from heart failure across all cells. Sulfur dioxide, like sulfate, is associated with increases in every cause of death, but it increases lung failure relatively more. Finally, nitrogen dioxide is negatively associated with all causes of death, though it decreases deaths from cancer relatively less. In summary, there appears to be a definite association between pollution and deaths from heart and circulatory failure. CONCLUSION Considerable effort has been undertaken in this study to remove variations of mortality rates across county groups which are not caused by pollution. Age, race, and sex are carefully controlled in the analysis. Other personal characteristics and county group variables are included to remove as many area-wide phenomena unrelated to pollution as possible. Despite all these efforts to explain observed variations in the mortality rates with other factors, the pollution variables are nonetheless significant and positive. Along with the corroborating evidtlnce of laboratory studies, these findings suggest that pollutants have a serious c,fFcct on mortality rates. lF The results also suggest that some pollutants are far more damaging to humans than others. Sulfates and any unmeasured ambient compounds spatially correlated with sulfates have quite sizable impacts on mortality rates. Carbon monoxide, sulfur dioxide, and unmeasured compounds spatially associated with either of these pollutants appear to have relatively 12The results of the regressions on cause of death are presented. in R. Mendelsohn and G. orcntt, “Pollution Dose Response Curves: A Microanalytic Study,” 1SPS Working Paper No. 78.5, Y’alc University, 1077. 16; Several of these important toxicological findings have Iwzn discovered by Amdnr [I, 21. Colucci [4], Morrow [13], and the U.S. Public Health %-vice [ 18-201 providr helpful rwicws of thcb touicr)lllgical literature.
MENDELSOHN
AND
ORCUTT
AIR POLLUTION
DOSE-RESPONSE
CURVES
101
smaller, if any, effects on mortality rates. Finally, we found no evidence that nitrogen dioxide or ozone has harmful effects on mortality rates. Other hypotheses about pollutants are also examined. The dose-response curves, even for the relatively low measurement levels in the sample, appear to be linear. For all practical purposes, there is no evidence of a lower threshold effect for any of the pollutants. If society wants to eliminate air pollution effects, no one can be exposed to the lowest existing level of any of the harmful air pollutants. Another important finding is that daily air quality extremes seem to have no effect on annual mortality rates. This result suggests that regulating 24-hr concentrations of harmful pollutants is not nearly as important as controlling annual average concentrations. Also, the magnitudes of the pollution effects are quite different for the different age and sex groups of white adults. Finally, areas with lower air quality have higher incidences of death from heart disease. It is interesting to note that many of these conclusions are also reached by Lave and Seskin [12]. They, too, find that the dose-effect curves appear to be linear; the size of the effects increases with age; and the pollutants from stationary sources appear to be more dangerous, in general, than those from mobile sources. Lave and Seskin, in contrast, find that the combination of all particulates is far more important than sulfate particles alone. This is surprising in the light of toxicological evidence which points to the greater hazard of respirable sulfates relative to that of particles in general (see Amdur ]I, 21). Another different feature of the Lave and Seskin results is that all their pollution coefficients differ across specifications. This may be due to the lack of independent variation across just-urban observations. There are several exciting possibilities for further improvement of this work. The unit of observation could be reduced from the county group to the county itself in order to provide greater spatial detail. One could examine whether the observed relationship are stable over time or whether mortality is sensitive to annual or seasonal variations in the independent variables by analyzing additional years of data and different seasons. The effects of poorly measured pollutants could be analyzed through subsets of the data which contain particularly good measurements. Finally, additional explanatory variables such as water pollution, occupation, and individual mobility could be introduced into the regressions. The data explored in this paper should serve as a very promising base for future studies of the effects of air pollution and the environment, in general, upon mortality.
102
MENDELSOHN
APPENDIX:
MEANS,
AND
ORCUTT
STANDARD DEVIATIONS, OF ALL VARIABLES TABLE Pollution
AND
UNITS
AI Variablesa Mean
Standard deviation
Mean ambient concentrations Sulfate Nitrate Sulfur dioxide Nitrogen dioxide Carbon monoxide (10 mg/m”) Total suspended particulates Ozone
1.05 0.31 2.11 4.96 0.26 6.99 2.77
0.38 0.16 1.83 2.51 0.16 1.91 1.49
Dispersion concentrationsb Sulfate Nitrate Sulfur dioxidec Nitrogen dioxide Carbon monoxide (10 mg/m3) Total suspended particulates” Ozone
1.06 0.30 3.30 4.99 0.46 7.63 8.72
0.65 0.15 3.33 4.02 0.29 6.76 3.53
a Variables in 10 pg/m3 units unless otherwise marked. b Differences between upper 95% concentration and mean concentration in 10 pg/m3 units; 9570 concentration refers to the distribution of daily measurements, Only 5’% of the measurements exceeded this value. ~These dispersion values refer to the daily standard deviations around the annual mean for each measurement site.
AIR POLLUTION
DOSE-RESPONSE TABLE
Area-Wide
CURVES
103
-411 Variables hIean
County group-adults Average age (years) Average schooling (years) Relative income Number of children (at home) Percentage divorced Percentage never married Neighborhood characteristics Density (persons/square mile) Unemployment rate Percentage owner occupied) Percentage without plumbing Overcrowding (index of people/room) Yet migration rate” Regional variablesb New England Middle Atlantic Southeast Deep South Southwest Great Plains Rocky Mountains Far West Climatic variables January temperature (Fahrenheit) July-January temperature (Fahrenheit) Water vapor pressure in January (millibars) Water vapor pressure in July (millibars)
31.7 12.5 2.5 3.6 0.06 0.09 247.8 4.4 64.2 5.3 7.8 5.5 0.06 0.22 0.09
0.10 0.09 0.09 0.03 0.13 34.9
40.7 5.5 20.9
Standard deviation 2.1 0.9 0.3 0.6 0.02 0.03 771.8 1.4 11.9 6.0 3.1 17.4 0.23 0.41 0.28 0.20 0.28 0.28 0.17 0.33 11.2 9.4 2.7 -1.6
” Net migration rate is defined as the number of inmigrants minus out,migrants between 1960 1970 divided by the 1960 population. * The means of the regional variables give the distribution of the population across regions. Including t,he North Central region, these figures add up to I .OO. and
104
MENDELSOHN
AND
TABLE Cell-Specific
ORCUTT
AI11
Mortality
Ratesa
Cause of death
White Male
Female
l-4 Years old All illnesses y. Heart and circulatory y0 Lung and respiratory ye Cancer y0 Infectious diseases
0.241 (0.08) 5.3 (4.1) 44.4 (19.0) 12.0 (7.0) 3.1 (3.3)
0.06 10.5 8.9 4.8 49.4
5-17 Years old All illnesses $& Heart and circulatory ‘% Lung and respiratory y0 Cancer $& Infectious diseases
0.299 15 19 20 0.5
0.031 (0.01) 52.2 (25.1) 0.8 (2.3) 8.7 (7.5) 3.8 (4.9)
18-24 Years old All illnesses $& Heart and circulatory $& Lung and respiratory To Cancer $‘J Infectious diseases
2.83 19.4 7 49.5 0.03
25-44 Years old All illnesses y0 Heart and circulatory y0 Lung and respiratory To Cancer y0 Infectious diseases
1.5 42 4 25 1
(0.06) (7)
(6) 6-o (0.1)
0) (8.1) (3.3) (17.1) (0.01) (0.33) (7) (3)
(‘3) (1)
0.77 10.1 4.7 5.1 4.5 1.0 23 5 41 2
(0.03) (12.3) (11.5)
(8.0) (32.2)
(0.20) (3.1) (2.3) (1.8)
(2.1) (0.20) (7) (3)
(8) (2)
45-64 Years old All illnesses y0 Heart and circulatory y0 Lung and respiratory y0 Cancer $& Infectious diseases
13.5 55 5 25 1
(1.9)
65+ Years old All illnesses y0 Heart and circulatory y0 Lung and respiratory y0 Cancer y0 Infectious diseases
69.0 64 7 19 0.6
(7.4) (3) (1)
(4) (1) (3) (0.4)
(2) (0.3)
Q Rates in deaths per thousand persons. Standard deviations deaths are presented as percentages of deaths from all illnesses.
6.6 38 4 41 1 48.0 69 4 17 9.5
(1.11) 00)
(2) w.4 0) (5.14) (3) (1)
are in parentheses.
(2) (0.2) Causes of
AIR
POLLUTION
DOSE-RESPONSE TABLE
Cell-Specific
1O.i
CURVES
AIV
Individual
Charactcristirs”
JJ’hite
1-4
Years Average Average Relative Percentage Percentage
old age (years) schooling (years) income6 divorced never married
2.5 0.0 2.9 0.0 100.0
(0.09) (0.0) (0.56) (0.0) (0.06)
2.5 0.0 2.9 0.0 100.0
(0.97) (0.0) (0.56) (0.0) (0.09)
5-17 Years old Average age (years) Average schooling (years) Relative incomeb Percentage divorced Percentage never married
10.9 6.2 3.2 0.01 99.6
(0.30) (0.25) (0.58) (0.09) (0.3%)
I I .o 6.3 3.2 0.09 98.9
(0.20) (0.25) (0.59) (0.16) (0.79)
18-24 Years old Average age (years) Average schooling (years) Relative income* Percentage divorced Percentage never married
20.8 14.1 3.0 1.8 65.7
((J.21) (O.=lQ) (0.87) (1.01) (7.X)
20.9 l-4. I 3.3 3.6 17.-l
(0. IQ) (0.41) (0.69) (1.58) (9.83)
25-44 Years old Average age (years) Average schooling (years) Relative incomeh Percentage divorced Perrentage never married
3-4.2 l-1.1 3.8 4.0 11.0
(0.14) (0.81) (0.55) (2.0) (5.0)
34.3 13.7 3.7 7.0 7.0
(O.-u) (0.65) (0.53) (2.0) (-1.0)
45-64 Years old Average age (years) Average schooling (years) Relative incomeb Percentage divorced Pcrcent,agc never married
53.7 12.7 4.4 5.0 6.0
(0.53) (0.97) (0.61) (2.0) (3.0)
53.8 12.6 1.2 7.0 6.0
(0.54) (0.78) (0.61) (2.0) (3.0)
65+ Years old Average age (years) Average schooling (years) Relative income6 Percentage divorced Percentage never married
73.1 10.5 2.8 1.0 7.0
(0.60) (0.95) (0.60) (Lo, (1.0)
73.8 10.8 2.5 4.0 8.0
(0.61) (0.90) (0.54) (20) (3.0)
a Standard deviations in parentheses. b Relative income is family money family’s charactcri&x
income
divided
by t.hc oliickrl
pouirty
score
nt,tacshing
to a
106
MENDELSOHN
AND
ORCUTT
REFERENCES I. M. 0. Amdur, Toxicological appraisal of particulate matter, oxides of sulfur, and sulfuric acid, J. Amer. Pollut. Cow. Assoc. 9, 638-644 ( 1969). 2. M. 0. Amdur and M. Corn, The irritant Ipotency of zinc ammonium sulfate of different particle size, Ind. Hyg. I., 326-333 (1963). 3. R. W. Buechley, W. B. Riggan, V. Husselbad, and J. B. Van Beuggen, SO, levels and perturbations in mortality, Arch. E. Health 27, 134-137 (1973). 4. A. V. Colucci, “Sulfur Oxides: Current Status of Knowledge,” EPRI EA-316, Palo Alto, Calif. (1976). 5. J. F. Finklea, et al., “Health Consequences of Sulfur Oxides: A Report From CHESS,” EPA-650/l-74-004, Washington, DC. ( 1974). 6. J. R. Goldsmith, Health hazards from power plant emission, in “Energy, the Environment, and Human Health” (A. J. Finkel, Ed.), Publishing Sciences Group, Acton, Mass. ( 1974). 7. E. Landau, R. Smith, and D. B. Lynn, Carbon mon’oxide and lead-An environmental appraisal, JAPCA 19, 684-87 ( 1969). 8. L. Lave and E. Seskin, Air pollution and human health, Science 169, 723-33 (1976). 9. L. Lave and E. Seskin, Air pollution, climate, and home heating: Their effects on U.S. mortality rates, Amer. J. Public Health 62, 909-916 (1972). 10. L. Lave and E. Seskin, Acute relationships among daily mortality, air pollution, and climate, in “Economic Analysis of Environmental Problems” (E. Mills, Ed.), Nat. Bur. Econ. Res., New York (1975). 11. L. Lave and E. Seskin, Health and air pollution: The effects of occupation mix, Swed.
J. Econ. 73, 76-95 (1971). 12. L. Lave and E. Seskin, “Air Pollution and Human Health,” Johns Hopkins Press, Baltimore (1977). 13. P. Morrow, An exaluation of recent NOx toxicity data and an attempt to derive an ambient air standard for NOx by established toxicological procedures, Enoiron. Res. 10, 92-112 (1975). 14. National Academy of Sciences, “Air Quality and Automobile Emission Control,” for U.S. Senate Committee on Public Works ( 1974). 15. National Academy of Sciences, “Air Quality and Stationary Source Emission Control,” for U.S. Senate Committee on Public Works ( 1975). 16. G. Orcutt, S, Franklin, R. Mendelsohn, and J. Smith, Does your probability of death depend on your environment: A micro analytic study, Amer. Econ. Reu. 67, 260-64 (1977). 17. C. M. Shy, et al., The Chattanooga school children study: The effects of community exposure to nitrogen dioxide, JAPCA 20, 539-45, 582-88 ( 1970). 18. U.S. Public Health Service, “Air Quality Criteria for Nitrogen Oxides,” Pub. No. AP-84, National Air Pollution Control Admin., U.S. Govt. Printing Office, Washington, D.C. (1971). 19. U.S. Public Health Service, “Air Quality Criteria for Particulates,” Pub. No. AP-49, NaD.C. tional Air Pollution Control Admin., U.S. Govt. Printing Office, Washington, (1971). 26. U.S. Public Health Service, “Air Quality Criteria for Sulfur Oxides,” Pub. No. AP-50, National Air Pollution Control Admin., U.S. Govt. Printing Office, Washington, D.C. (1970). 21. P. 0. Warner and L. Stevens, Reevaluation of the “Chattanooga School Children Study” in light of other contemporary governmental studies, J. Amer. PO&~. Cont. Assoc. 23, 769-72 ( 1973 ) .