Available online 24 June 2014. JEL classification: I12. Q53. J13 .... tion due to a voluntary pollution prevention progr
Journal of Health Economics 37 (2014) 219–231
Contents lists available at ScienceDirect
Journal of Health Economics journal homepage: www.elsevier.com/locate/econbase
Air pollution and infant mortality: A natural experiment from power plant desulfurization夽 Simon Luechinger ∗ University of Lucerne and KOF Swiss Economic Institute, ETH Zurich, Switzerland
a r t i c l e
i n f o
Article history: Received 28 September 2012 Received in revised form 6 June 2014 Accepted 17 June 2014 Available online 24 June 2014 JEL classification: I12 Q53 J13
a b s t r a c t The paper estimates the effect of SO2 pollution on infant mortality in Germany, 1985–2003. To avoid endogeneity problems, I exploit the natural experiment created by the mandated desulfurization at power plants and power plants’ location and prevailing wind directions, which together determine treatment intensity for counties. Estimates translate into an elasticity of 0.07–0.13 and the observed reduction in pollution implies an annual gain of 826–1460 infant lives. There is no evidence for disproportionate effects on neonatal mortality, but for an increase in the number of infants with comparatively low birth weight and length. © 2014 Elsevier B.V. All rights reserved.
Keywords: Health Infants Mortality Infant mortality Air pollution
1. Introduction Health concerns are a primary rationale for air quality regulations such as the U.S. Clean Air Act and the German Federal Immission Control Act. These regulations considerably improved air quality. For example, Fig. 1 depicts the sulfur dioxide (SO2 ) concentration in Germany in 1985–2003. Other developed countries
夽 I thank Wolfgang Bräuniger, Andrea Minkos, and Wolfgang Müller from the German Federal Environmental Agency for the pollution and power plant data, the operating companies for giving confidential information on their generating units, Hiltrud Bayer from the German Youth Institute and the statistical offices of the Länder for the mortality data, and Hans-Peter Mast and Stefan Weil from the Research Data Centers of the Federal Statistical Office and the statistical offices of the Länder for help with remote access to the death and birth certificates. For comments and suggestions, I thank Peter Nilsson, Shinsuke Tanaka, participants of the annual meeting of the Swiss Society of Economics and Statistics 2009, the annual meeting of the European Economic Association 2009, the annual meeting of the American Economic Association 2010, the annual meeting of the German Economic Association 2010 and seminars at the ETH Zurich, the University of Berlin, the University of Fribourg, the University of Mannheim, and the University of Rotterdam as well as the editor Adriana Lleras-Muney and two anonymous referees. ∗ Correspondence to: University of Lucerne, Department of Economics, P.O. Box 4466, 6002 Lucerne, Switzerland. Tel.: +41 41 229 5641. E-mail address:
[email protected] http://dx.doi.org/10.1016/j.jhealeco.2014.06.009 0167-6296/© 2014 Elsevier B.V. All rights reserved.
experienced similar declines in SO2 concentrations. But this general trend masks considerable heterogeneity. Many people are still exposed to high pollution levels and in developing countries air pollution is often getting worse. Even in Europe, SO2 pollution is bound to rise as coal-based power generation experiences a revival. Therefore, knowledge of the health effects of previous air quality regulations and corresponding improvements in air pollution is of considerable interest. Many studies investigate the effect of air pollution on adult mortality. Epidemiologists typically assess acute effects with timeseries analyses and chronic effects of long-term exposure with cross-section and cohort studies. For example, SO2 was significantly associated with mortality in a time-series analysis for London in 1958–1972 (Schwartz and Marcus, 1990), in a crosssection of groups of U.S. counties in 1970 (Mendelsohn and Orcutt, 1979), or in a large group of adult Americans followed over the period 1982–1998 (Pope et al., 2002). However, studies on infant mortality have clearer implications regarding the number of lifeyears lost and suffer less from uncertainty regarding life-time exposure (Chay and Greenstone, 2003b). Omitted variables are an important concern in all these studies since air pollution depends on economic activity and other unobserved factors with independent effects on mortality. The concern
220
S. Luechinger / Journal of Health Economics 37 (2014) 219–231
Fig. 1. SO2 concentration and infant mortality in East and West Germany, 1985–2003. Sources: Federal Statistical Office and Federal Environmental Agency.
is partially addressed in intervention analyses. Evidence of reduced SO2 pollution and mortality following a ban on high sulfur fuel in Hong Kong in 1990 is certainly more convincing than evidence from other time-series analyses, but influences from other concurrent shocks cannot be ruled out without an adequate control group (Hedley et al., 2002). Such concerns sparked a recent interest among economists in natural experiments that allow researchers to identify the effects of air pollution on infant mortality (see Section 2 for a review). This paper estimates the health benefits of an air quality regulation and uses regulation-induced changes in air pollution to identify chronic effects on infant mortality (similar to Chay and Greenstone, 2003a). Specifically, it estimates the effect of the mandated installations of scrubbers at power plants and the resulting reduction in SO2 pollution on infant health with data from Germany in 1985–2003. Thereby, it contributes to the existing literature in two respects. First, the paper provides evidence on infant health effects of air pollution and air quality regulations for another highly developed country than the U.S. Thus, the results help us to understand if and to what extent pollution-mortality relationships found in one context can the transferred to different contexts. While there are no reasons to expect differences in biological dose-response relationships, effects of ambient air pollution also depend on medical care consumption and avoidance behavior, which affects the relationship between ambient air pollution and individual exposure (Graff Zivin and Neidell, 2013). In this regard, Germany and the U.S. may differ along several relevant dimensions. For example, there may be differences in medical care utilization. Pre-natal care can improve infant health and thereby reduce susceptibility to air pollution, good access to medical care after birth allows for timely measures against health problems. Despite a huge increase in Medicaid eligibility of pregnant women in the U.S. in the 1980s, a large share of the population is still uninsured and many eligible women enroll late in their pregnancies (Currie and Gruber, 1996; Gruber, 1997). In contrast, health insurance coverage in Germany is near universal and frequent screenings of infants are statutorily regulated. Similarly, the extent to which individuals can avoid exposure to pollutants may also differ across countries and will depend on such diverse factors as building design with large regional differences in airtightness of houses or typical activity patterns of pregnant women and parents (Ashmore and Dimitroulopoulou, 2009).
Second, the paper analyzes the effect of SO2 pollution. Of course, different pollutants may be correlated (Lleras-Muney, 2010) and the regulation may have affected several pollutants. However, as I explain in more detail in Section 3, the German situation analyzed in this paper is particularly well-suited to specifically isolate exogenous variation in SO2 pollution. A federal regulation mandated the installation of scrubbers at power plants and left local authorities or operating companies little room for discretion. Power plants are the main source of SO2 and the predominant scrubbing technology removes SO2 but not other pollutants. Regulation of TSP was already in place and new regulation of nitrogen oxides (NOx ) was generally not binding, affected different sources, and required different compliance measures. Further, I find that the estimated effect of desulfurization at power plants affected SO2 concentration but not NOx concentration. Therefore, I present not only reduced form effects of the policy but also use it to instrument SO2 pollution. Looking at SO2 is interesting for several reasons. First, toxicity differs across pollutants. Second, pollutants differ in the extent to which ambient air pollution translates into individual exposure. For example, the correlation between outdoor and indoor air pollution is lower for SO2 than for other pollutants (Ashmore and Dimitroulopoulou, 2009). Third, the existing evidence on the effects of SO2 on infant mortality is inconclusive and suffers from omitted variables (see Section 2 for a review). Given that SO2 pollution is the focus of many air quality regulations and that there is considerable evidence for effects of SO2 on adult mortality and on adverse ˇ et al., 2005), the lack pregnancy outcomes (for a review, see Sram of convincing evidence on the effects of SO2 pollution on infant mortality is regrettable and this paper aims to fill this void. The most important finding is that the air quality regulation had beneficial effects on infant health: Infant mortality decreases with predicted reductions in SO2 concentration due to desulfurization at power plants. The sharp and simultaneous drops in SO2 pollution and in infant mortality between 1987 and 1988 in Fig. 1 anticipates this result. Assuming that the air quality regulation only affected infant mortality through its effect on actual SO2 concentrations, I use the regulation-induced changes in SO2 concentration to estimate the effect of SO2 pollution on infant mortality. According to fixed-effects regressions of infant mortality rates on SO2 concentration, 0.026 infant lives (per 1000 live births) are saved for every 1 g/m3 reduction in SO2 concentration. In instrumental variable regressions with the predicted regulation-induced reductions in SO2 concentrations as an instrument, the effect amounts to 0.045 infant lives. Since most of the variation in SO2 concentration is the result of the air quality regulation, the fixed-effects and instrumental variable estimates are similar and both estimates are informative about the health effects of the regulation. The point estimates translate into an elasticity of 0.07–0.13. The results are similar in subperiods and the West German subsample but not the East German subsample. The estimates are robust to controls for local economic and demographic development, weather, TSP pollution, reunification effects, and rural/urban trends. The instrumental variable estimates are also robust to the inclusion of county-specific time trends, the fixed-effects estimates less so. There is no evidence for strongly disproportionate effects of SO2 on neonatal mortality, but evidence for effects on the number of infants with comparatively low birth weight and, in particular, length. Thus, although poor fetal development due to exposure during gestation seems to affect infant health, it is unlikely to be the main biological mechanisms through which SO2 affects infant mortality. The remainder of the paper is organized as follows. Section 2 briefly reviews the related literature. Section 3 introduces the pollution data and the strategy to instrument SO2 concentrations.
S. Luechinger / Journal of Health Economics 37 (2014) 219–231
Section 4 presents the mortality data, the baseline regressions and robustness tests as well as the estimates for mortality for different ages at death and for various birth outcomes. Section 5 concludes. 2. Related literature This paper is closely related to a growing literature in economics that exploits natural experiments to identify the effects of air pollution on infant mortality. Chay and Greenstone (2003a,b) exploit regulation- and recession-induced changes in total suspended particulate (TSP) pollution in U.S. counties in the 1970s and 1980s and find that air pollution increases infant mortality with elasticities between 0.35 and 0.5. Currie and Neidell (2005) and Currie et al. (2009) use within zip-code month variation and find a positive effect of carbon monoxide (CO) with elasticities of 0.09 and 0.04, respectively, but no effects for particulate matter (PM) and ground-level ozone (O3 ). Currie and Schmieder (2009) estimate elasticities for emissions of toxic chemicals ranging from 1.82 for heavy metals to 6.11 for volatile organic compounds and 6.49 for chemicals known to affect child development. Further, there is recent evidence for adverse effects of air pollution on infant survival from developing countries with studies using changes in air pollution due to a voluntary pollution prevention program in Mexico (Foster et al., 2009), wildfires in Indonesia (Jayachandran, 2009), and air quality regulations in China (Tanaka, 2010). In contrast to these well-identified effects of air pollution, the existing evidence on the effects of SO2 on infant mortality suffers from omitted variables. Further, the existing evidence is rather inconclusive. Studies assessing acute effects of SO2 pollution on infant mortality from all causes with time-series analyses consistently find positive effects near a volcano in Japan in 1978–1988 (Shinkura et al., 1999), in Seoul in 1955–1999 (Ha et al., 2003), in São Paulo in 1998–2000 (Lin et al., 2004), and in ten English cities in 1969–1973 (Hajat et al., 2007). The last study estimates an elasticity of 0.03. SO2 pollution has been observed to increase acute deaths from the sudden infant death syndrome (SIDS) in twelve Canadian cities in 1993–2003 (Dales et al., 2004), but not from respiratory causes (Hajat et al., 2007). Results from studies investigating chronic effects from prolonged exposure with crosssection, pooled panel, or case–control studies generally discover no significant effects of SO2 on all-cause infant mortality. This is the case for studies using data across U.S. standard metropolitan statistical areas in 1961–1964 (Hickey et al., 1976), U.S. county groups in 1970 (Medelsohn and Orcutt 1979), counties and boroughs in England and Wales in 1969–1973 (Chinn et al., 1981), Czech districts in 1986–1993 and in 1989–1991 (Bobak and Leon, 1992; 1999), or large U.S. counties in 1999–2002 (Woodruff et al., 2008). An exception is the study by Joyce et al. (1989) with data from U.S. counties in 1970, which finds SO2 pollution to increase neonatal mortality with an elasticity of 0.02–0.05. The results from studies researching the effects of prolonged exposure on mortality from specific causes are conflicting. Chinn et al. (1981) and Woodruff et al. (2008) find no effect on pneumonia or respiratory mortality, while Bobak and Leon (1992; 1999) find positive effects on respiratory mortality. Lipfert et al. (2000) and Woodruff et al. (2008) discover no effect on SIDS mortality. Crocker et al. (1979) estimate mortality from early infant diseases to increase with SO2 in a cross-section of 60 U.S. cities in 1970 with an elasticity of 0.09. 3. Pollution: data, evolution and instrument From the German Federal Environmental Agency (Umweltbundesamt; hereafter UBA for short) I have data on the annual mean SO2
221
concentration measured at air quality monitors for 1985–2003.1,2 There are data for between 196 monitors in 1985 and 416 monitors in 1994, or 553 monitors in total. I interpolate the monitor readings on a grid with cells of 1 km2 . For the interpolation, I use the method of inverse distance weighting with the 9 nearest monitors (without any distance cutoff) and the inverse cubed distance as weights. The UBA determined the parameters on the basis of empirical studies. However, both interpolated values and regression results are very similar for slightly different parameters.3 Following the approach suggested in Currie and Neidell (2005), I evaluate the accuracy of the interpolation procedure by comparing at each monitor actual readings with the concentration level that would be estimated with the interpolation procedure if this particular monitor was not there. The correlation of 0.87 implies that the interpolated values are accurate. The results are very similar if the sample is reduced to county-years with active monitors.4 The number of monitors changes over time and the placement of the monitors may be endogenous. However, if I use pollution measures based on the 64 continuously operating monitors, the results get even stronger but are broadly similar.5 To merge the air pollution data with the mortality data, I aggregate the interpolated pollution values to the county level by taking the average value of all grid cells that fall within the borders of a county.6 Fig. 2 depicts the mean SO2 concentration per county in 1985, 1990, 1995, and 2000. Looking at the pattern and evolution of SO2 pollution, two aspects are worth noting. First, in the mid-1980s, pollution levels were high, especially at three hotspots, the Ruhr area in the west, Northern Hesse in the center, and the area around Leipzig in the east. Back then, these areas were important industrial centers and coal mining areas. Second, air quality diminished substantially after 1985 and 1990 in West Germany and after 1990 in East Germany. These improvements are largely the result of the large combustion plant ordinance enacted in 1983. The ordinance required operating companies to retrofit fossil fuel fired power plants with scrubbers within three to nine years from 1986 on. Retrofitting deadlines were statutorily fixed and depended on a plant’s capacity and emissions. Thus, they were not chosen by regulators or operating companies. The unification treaty of 1990 extended the regulation to East Germany with East German power plants required to install
1 The pollution data and identification strategy have been previously used in Luechinger (2009) to estimate the effect of air pollution on subjective well-being. The description of the data and identification strategy in this section is based on the description of Luechinger (2009). 2 The UBA calculates annual mean concentrations as simple averages based on daily or hourly means. The temporal aggregation follows EU rules which require minimum data capture of 50 percent for annual means (2001/752/EC and email from Andrea Minkos, UBA, June 14, 2013). 3 The baseline estimates reported in column 1 of Table 2 for the cubed distance and the nine nearest monitors are 0.026 (std. err.: 0.004) (fixed-effects estimate) and 0.044 (std. err.: 0.009) (instrumental variable estimate). Using the squared distance and the six nearest monitors, the respective estimates are 0.026 (std. err.: 0.004) and 0.044 (std. err.: 0.009), using the cubed distance and the six nearest monitors 0.030 (std. err.: 0.005) and 0.046 (std. err. 0.010) and using the squared distance and nine nearest monitors 0.027 (std. err.: 0.004) and 0.044 (std. err. 0.009). 4 The baseline estimates reported in column 1 of Table 2 are 0.026 (std. err.: 0.004) (fixed-effects estimate) and 0.044 (std. err.: 0.009) (instrumental variable estimate). For the reduced sample the respective estimates are 0.021 (std. err.: 0.006) and 0.039 (std. err.: 0.012). Thus, despite of a reduction of the sample size by nearly 60 percent, the results are quantitatively similar and still precisely estimated. 5 The baseline estimates reported in column 1 of Table 2 are 0.026 (std. err.: 0.004) (fixed-effects estimate) and 0.044 (std. err.: 0.009) (instrumental variable estimate). The respective estimates with pollution measures based on the 64 continuously operating monitors are 0.046 (std. err.: 0.009) and 0.064 (std. err.: 0.014). 6 County mergers in East Germany caused the number of counties to fall from 543 in 1993 to 439 in 2001. The analysis in this paper is based on the 439 counties at the end of the merging process.
222
S. Luechinger / Journal of Health Economics 37 (2014) 219–231
the product of the operating status, pre-desulfurization or residual emissions, a distance decay function, f(Dcj ), and the frequency county c lies downwind of power plant j, g(Rcj ). The contribution from all power plants is the sum of the contributions from individual plants multiplied by an unknown parameter ˛1 converting pollutant mass into concentration levels. Thus, the SO2 concentration in county c at time t can be described by Eq. (1): Pct = ˛0 + ˛1
1(active)jt · Ej · (1 − ˛2 · 1(scrubber)jt )
j
·f (Dcj ) · g(Rcj ) + c + t + εct .
(1)
Slightly re-arranging terms yields Eq. (2): Pct = ˛0 + ˛1 −˛1 ˛2
1(active)jt · Ej · f (Dcj ) · g(Rcj )
j
(2)
1(active)jt · Ej · 1(scrubber)jt · f (Dcj ) · g(Rcj ) + c + t + εct .
j
The third term in Eq. (2),
1(active)jt · Ej · 1(scrubber)jt · f (Dcj ) ·
j
g(Rcj ), is the estimated effect of desulfurization at power plants on SO2 pollution, which I will use to identify the effects of air pollution. The temporal variation comes from two sources: The retrofitting of power plants with scrubbers and changes in the power plant population. Only the first is exogenous. Changes in power plant population may be related to unobserved factors with independent effects on infant mortality. Therefore, I include the effect of changes in the power plant population captured by the second term in Eq.
1(active)jt · Ej · f (Dcj ) · g(Rcj ), as a control variable. The geo-
(2),
j
Fig. 2. SO2 concentration in German counties; 1985, 1990, 1995 and 2000. Legend:
≤ 20 g/m3 ,
20–40 g/m3 ,
40–60 g/m3 ,
60–80 g/m3 ,
80–100 g/m3 , 100–125 g/m3 , 125–150 g/m3 and > 150 g/m3 ; cities: D = Dortmund in the Ruhr area, K = Kassel in Northern Hesse, L = Leipzig and B = Berlin. Sources: See text.
scrubbers from 1993 on. Again, different retrofitting deadlines for different categories of power plants were statutorily fixed. Below I will use the estimated effect of the mandated installation of scrubbers at power plants to identify the effects of air pollution on infant mortality. This effect is estimated with a simple model of air pollution, information on power plants’ operating status, annual pre-desulfurization SO2 emissions, and retrofitting status, and information on wind directions as well as distances and directions between power plants and counties. The SO2 concentration in county c at time t, Pct , comprises contributions from power plant emissions and background pollution. Background pollution is captured by county effects, c , time effects, t , and a random component, εct . The contribution of emissions from an individual plant j depends on the operating status and the retrofitting status of the power plant, 1(active)jt and 1(scrubber)jt , pre-desulfurization emissions or residual emissions, Ej or Ej · (1 − ˛2 · 1(scrubber)jt ), and the distance and direction between the plant and the county, Dcj and Rcj . 1(active)jt is a dummy with value one if a power plant operates at time t and zero otherwise, 1(scrubber)jt is a dummy with value one if the power plant has installed a scrubber at time t and zero otherwise, and ˛2 is an unknown parameter reflecting average separation efficiency of scrubbers. The contribution of an individual power plant j is
graphical variation also comes from two sources: Distance to power plants and the distribution of wind directions. The estimated effect of the installations of scrubbers on air pollution is the interaction of temporally and geographically varying components and, thus, similar to a difference-in-difference term. However, the term differs from a standard difference-in-difference term in three respects. First, treatment and control group status is a matter of degree rather than one of kind and depends on distance and wind direction frequencies. Second, I have to include all power plants simultaneously. Therefore, my treatment variable is a weighted sum of desulfurization at all plants with predesulfurization emissions as weights. Third, some power plants are newly constructed, others taken offline. For this reason, I control for changes in the power plant population. For over 300 fossil fuel fired generating units with a capacity of 100 MW or more and active between 1985 and 2003, I have information on the starting year, the year the unit was taken offline, the year of desulfurization, capacity, fuel and fuel efficiency. The information comes from the UBA, publications of operating companies and the engineering literature, a questionnaire sent to operating companies, and statutory provisions (see Luechinger, 2009 for details). Panel A of Fig. 3 pictures power plants’ locations. I use published emission factors (Bakkum et al., 1987) and plants physical characteristics to estimate the pre-desulfurization SO2 emissions.7
7 Lacking data on utilization rates, I have to assume full utilization of capacities. The emissions are then simply the product of the emission factor, the capacity, inverse fuel efficiency, and the time period. This calculation may overstate emissions because the assumption of full utilization may not be plausible. However, for my purpose, the absolute level of emissions is immaterial. I am only interested in the relative size of pre-desulfurization emissions from different power plants, which are mainly due to differences in fuels and plant size. Further, using actual
S. Luechinger / Journal of Health Economics 37 (2014) 219–231
Fig. 3. Locations of fossil fuel fired power plants and wind stations. Sources: See text.
I model distance decay with an exponential curve and a characteristic decay distance of 480 km, i.e. g(Dcj ) = exp(−2.1E-6 · Dcj ), as proposed by field estimates (Schwartz, 1989; Summers and Fricke, 1989). Frequencies of wind directions in 12 30-degree sectors at the nearest wind station characterize the wind situation at each power plant (Traup and Kruse, 1996). Panel B of Fig. 3 depicts the 43 wind stations used in this paper. Finally, I use the distance and direction between every power plant and every county to relate plant- and county-level data. To interpret the effect of the mandated installations of scrubbers at power plants as the causal effect of improved air quality, I need to assume that the retrofitting of power plants affected infant health only through its effect on air quality. In this context, it is important to note that the statutory provisions were enacted prior to the sample period. Thus, it is unlikely that the actual installation of scrubbers is a response to concurrent demographic, economic or political developments in faraway upwind or nearby downwind regions. However, one might worry that counties lying downwind in the vicinity of a retrofitted power plant experienced improvements in air quality and infant health before the plant installed the scrubber. In this case, the difference-in-difference term in Eq. (2) that is used as an instrument would partly capture these preexisting trends. In order to assess this issue, Fig. 4, panels A and C, plots for each county the differences in SO2 concentration and infant mortality between the first year of the desulfurization process and the last year before the process against the estimated effect of desulfurization averaged over the whole sample period. The predesulfurization differences relate to the years 1986–1985 for West Germany and the years 1993–1992 for East Germany. Fig. 4, panels B and D, also plots the differences in SO2 concentration and infant mortality between the third and the first year of desulfurization process against averaged values of the estimated effect of desulfurization. The post-desulfurization differences relate to the years 1988–1986 for West Germany and the years 1995–1993 for East Germany. These years mark the time frame during which the most polluting power plants were required to install scrubbers. Fig. 4 plots both actual values and Kernel-weighted local polynomial regression-smoothed values. As we can see from Fig. 4, in the years before the desulfurization process, average values of the estimated effect of desulfurization
utilization rates may be problematic because these rates are endogenous and potentially related to unobserved factors.
223
are not related to changes in SO2 concentration. The correlation () is −0.05 (p = 0.294). In contrast, average values of the estimated effect of desulfurization are strongly related to changes in SO2 concentration after the desulfurization process started ( = −0.49; p < 0.0001). The same pattern holds for infant mortality rates even though it may be less visible to the naked eye. The correlation for pre-desulfurization is 0.05 (p = 0.276), post-desulfurization it is −0.11 (p = 0.018).8 Another worry might be that that the retrofitting of power plants lowered concentrations of other pollutants in addition to SO2 . However, the German situation analyzed in this paper is wellsuited to specifically isolate exogenous variation in SO2 pollution. Air quality regulation in other countries, notably the U.S., often leaves authorities at subnational levels substantial leeway in formulating their own implementation plans and in deciding how pollution targets are reached. Similarly, changes in economic activity are likely to affect several pollutants. In contrast, the large combustion plant ordinance specified not only targets but also measures and it left local authorities or operating companies little room for discretion. The regulation of TSP had a long tradition in Germany before the large combustion plant ordinance. However, in addition to emission limits for SO2 , the large combustion plant ordinance did establish for the first time emission limits for nitrogen oxides (NOx ), though the limits were preliminary and generally not binding. Further, for four reasons my instrument is unlikely to capture changes in NOx pollution. First, power plants are the most important source of SO2 , but not of NOx pollution. For example, in 1990 power plants accounted for 60 percent of all SO2 emissions but only for 20 percent of NOx emissions; NOx emissions are primarily caused by road traffic (UBA, 2012). Second, SO2 emission factors of coal and oil fired power plants are two to five times as large as NOx emission factors. Conversely, gas fired power plants are important emitters of NOx , but not SO2 (Bakkum et al., 1987). Third, 93 percent of the installed scrubber capacity was wet scrubbers with a water-gypsum-limestone/lime-slurry as scrubbing liquid (Mittelbach, 1991). This scrubbing technology is aimed at removing SO2 , not NOx . Fourth, and most importantly, for the years 1990–2003, the years for which NOx data are available, there is a strong first stage for SO2 (t-value: −4.29) but none for NOx (t-value: 0.45). 4. Effect of SO2 pollution on infant mortality 4.1. Data and empirical strategy The state statistical agencies collect data on births and deaths in standardized way in accordance with federal laws and then publish infant mortality rates at the county level in state reports or make them available upon request. The German Youth Institute, an independent children and family research institute, compiled the data for years since 1986 and generously shared the data with me. The 1985 data come directly from the statistical agencies.9 Infant
8 It is important to note that the actual development stacks the deck against finding the instrument to be valid. Even though power plants were not statutorily required to install scrubbers before 1986 and 1993 in West and East Germany, respectively, individual power plants (