External Effects of Diesel Trucks Circulating Inside the São Paulo Megacity

External Effects of Diesel Trucks Circulating Inside the São Paulo Megacity Abstract The medical literature documents adverse health effects of acute exposure to diesel exhaust, yet quasi-experimental evidence of a policy intervention sustained over months at the scale of a metropolis is lacking. Exploiting the inauguration of a beltway that removed 20,000 cargo trucks passing daily through inner-city roads in São Paulo, we examine the spatially differentiated impacts on the megacity's traffic, air quality and public health. We combine rich panel data on road congestion, ambient NOx concentrations (as a signature of diesel exhaust), and hospital admissions and deaths. The policy reduced congestion, pollution, and hospitalizations, with effects attenuating at increasing distances from a key inner-city corridor used by the transit trucks prior to the beltway opening. The change in congestion was transient, as gasoline–ethanol passenger cars responded by filling the space the diesel trucks left behind. Effects on air and health persisted thanks to the compositional change in road users. We use 2SLS regression, taking policy-induced variation in NOx to instrument for measured pollution, to quantify about one annual hospitalization for every 10–20 trucks—and one annual death for every 100–200 trucks—using inner-city roads. Policymakers in megacities where humans and diesel vehicles reside and transit in close proximity may learn from São Paulo's experience. 1. Introduction Road vehicles are a major source of localized air pollution in the world's megacities, right by where people circulate and live (Molina and Molina 2004). Although a sizable literature examines the public health impacts of urban traffic flows, there is much less research on the health impact of the fleet composition associated with these aggregate traffic quantities, heavy-duty diesel vehicles in particular (Brunekreef et al. 1997; Beatty and Shimshack 2011). Due to data limitations, the literature typically lumps vehicle emissions—whether from diesel or gasoline engines, heavy versus light duty, aging or new—together as a single category of environmental bads.1 Moreover, to the best of our knowledge, there is no observational study of a fleet-composition intervention on public health, air quality and road congestion at the scale of a real-world megacity.2 A recent cross-disciplinary review calls for more quasi-experimental evidence, based on identifiable and credibly exogenous policy shifts, to complement observational studies that document statistical associations among human health outcomes and airborne pollutants (Dominici et al. 2014). In the previous context, one such policy would shift the diesel–gasoline urban transportation fuel mix while holding the usage of urban roads fixed. Cities in rich Europe, where the penetration of diesel-fueled vehicles is high relative to North America, as well as in the developing world, such as Delhi and Calcutta, are considering whether to restrict the use of diesel (Sharman 2015; Economist 2016; Mitral 2016).3 Seeking to fill this gap, this study exploits a rare large-scale quasi-experiment in the subtropical megacity of São Paulo, of population 20 million. The metropolis lies on busy commercial routes that connect the Atlantic port of Santos, Latin America's largest and 100 km away, to the wider state of São Paulo, which alone accounts for one-third of Brazil's Gross Domestic Product. Until the recent opening of a beltway along mostly undeveloped land at a 25-km radius from the city center, each day about 20,000 heavy vehicles transiting between points other than São Paulo city (including Santos port) had to pass through congested inner-city roads. Described by the state government as “the most important infrastructure project for the state”, the beltway's southern section, inaugurated in April 2010, immediately enabled the passing commercial vehicles to bypass the densely populated city, in particular, a key inner-city corridor, Avenida dos Bandeirantes, en route to and from the port. Alleviating traffic congestion was the policy's stated goal: “The project will prevent these (19,000) heavy-duty vehicles, as well as passenger cars, from transiting through the city, causing severe traffic congestion” (São Paulo State Government 2008). Indeed, Appendix Figure A.1 confirms that about 20,000 heavy vehicles (3 axles or more) paid tolls per day in 2011 along the beltway's southern section. On top of this, in the immediate wake of the beltway, the São Paulo city government tightened restrictions on trucks circulating in the inner city. Since 2008, also with the goal of relieving congestion, restrictions had applied to a proportion of the truck fleet, certain roads and times of the day (Mayor's Office 2008). Over the third quarter of 2010, these restrictions became more stringent, for example, along Av. dos Bandeirantes, during weekday daytime hours, and for larger trucks in particular (Diário Oficial da Cidade 2010).4 We document three unintended consequences of road congestion policies to abate truck flows inside the São Paulo metropolis—which we hereafter refer to as “truck abatement policy” or simply “policy”. First, examining over 5,100 traffic-monitored road segments across the city, we find that the truck policy did instantly relieve road congestion, by about 20% emanating from the key inner-city corridor used previously by the passing trucks, but this effect on traffic volume was short-lived, as passenger cars substituted into the scarce inner-city road space that the trucks left behind. The behavioral response by households in a city with gridlocked roads was such that within a couple of years the traffic relief brought about by the policy was largely undone. Our study provides a clean and vivid demonstration of the “fundamental law of road congestion”.5 Second, the policy achieved abatement in diesel pollution, as proxied by ambient NOx (nitrogen oxides, NO + NO2) concentrations (Anenberg et al. 2017), again emanating from the key inner-city corridor. This second effect on air was not an original policy motivation—it is in this sense that this effect was “unintended”. Importantly, the drop in NOx levels survived the return of traffic volumes to pre-policy levels, as gasoline–ethanol passenger cars increased their share of road space. Our finding that pollution levels remained lower even after traffic congestion rebounded demonstrates the critical influence of fleet composition beyond fleet size. The policy's third unintended consequence was a long-lasting reduction in public health damage, also due to the compositional change in traffic, away from heavy-duty diesel vehicles toward gasoline–ethanol light vehicles. With a study period between 2008 and 2013, and 3.3 million individual public hospital records aggregated up to the level of month by patient's residential zip code by age group, the health effect we examine is for changes in exposure sustained over months, not from daily variation in diesel exhaust.6 Consistent with the policy's effect on air, we find that cardiovascular and respiratory hospitalizations abated in increasing proximity to the key inner-city corridor, and particularly in the vulnerable age groups. Estimates amount to 886 (standard error, s.e., of 141) less public hospital admissions (down 8% from total) and 116 (s.e. 37) less in-hospital deaths (down 9%) per year per one million residents per 10 ppb (parts per billion) abatement in NOx equivalent units of diesel pollution. We quantify one annual hospitalization for every 11–23 diesel trucks—and one annual death for every 86–172 trucks—using roads with high human exposure. We test for the potentially confounding “sorting” hypothesis, by which wealthier, more educated and healthier households would have moved in response to the policy, by shortening the sample period further, as well as by including admissions that are unlikely to be triggered by pollution as additional controls in our health equations (wealthier households are presumably healthier across the disease range). Our goal is not to provide benefit-cost analysis of alternative policies to abate diesel pollution. In particular, our intent is not to evaluate the net benefits associated with the beltway and tightened truck circulation, or how a beltway on undeveloped land today might affect urban sprawl and economic growth (and thus emissions and human exposure) in the future, questions we leave for subsequent research.7 What we do is use these interventions, irrespective of whether they are sound, to evaluate the health damage—today—caused by human exposure to diesel exhaust. Our contribution is to quantify external effects of an actual, not hypothetical, diesel truck fleet observed transiting in the densely urbanized areas of a real-world megacity, under existing operating/maintenance conditions, in close proximity to human populations. By sharing the São Paulo experience, we hope to alert policymakers in urban areas with similar levels of immediate exposure, in developing and rich nations alike—such as Europe, Singapore and India.8 2. Data and Reduced-Form Econometric Models of Environmental Outcomes 2.1. Summary Statistics We combine spatial data on the different outcomes of interest, namely traffic volume, exposure to diesel pollution, and hospital admissions, along with data on meteorology and fuel prices, variables that may influence traffic, pollution and/or health. The traffic, air, and hospital data are not only of interest in their own right but, importantly, brought together they tell a consistent, credible story of how different urban road users (firms, households) responded to the policy, and what their behavioral response reveals about the external damage from the use of diesel. The study period is November 1, 2008 to May 31, 2013. Table 1 provides summary statistics for our datasets. Table 1. Summary statistics for road congestion, NOx concentration (as a proxy for exposure to diesel exhaust), hospital admissions, meteorology, and other control variables. Variable  N  Mean  Std.Dev.  Min  Max  Road congestion   Citywide congestion, weekdays only (km)  1,070  83.2  33.9  0.1  226.7   Citywide congestion (km)  1,673  55.8  46.1  0.0  226.7   Road segment congestion, weekday only (a prop.)  5.5 m  0.15  0.25  0.00  1.00   Road segment congestion (a proportion)  8.6 m  0.10  0.21  0.00  1.00  Exposure to diesel exhaust   NOx concentration, weekdays only (ppb)  10,364  64.0  49.9  0.4  490.2   NOx concentration (ppb)  16,152  56.0  46.1  0.0  490.2  Hospital admissions rate   Cardiovascular (per 100,000 residents per month)  2,365  54.2  19.8  10.7  166.4   Respiratory (per 100,000 residents per month)  2,365  46.5  20.8  5.0  130.0   Trauma (per 100,000 residents per month)  2,365  6.6  3.3  0.0  26.0  Meteorology and thermal inversion   Ground temperature (°C)  1,673  20.8  3.4  10.5  29.6   Relative humidity (%)  1,669  77.3  10.4  29.6  98.0   Solar radiation (7 a.m. to 8 p.m., Wm−2)  1,672  299.2  113.7  25.5  648.3   Atmospheric pressure (hPa)  1,673  932.2  4.8  912.1  945.2   Wind speed (ms−1)  1,673  1.4  0.5  0.3  2.9   Wind blows from N-E (a prop., all 24 h = 1, none = 0)  6,692  0.15  0.17  0.00  1.00   Wind blows from S-E (a prop., all 24 h = 1, none = 0)  6,692  0.44  0.35  0.00  1.00   Wind blows from S-W (a prop., all 24 h = 1, none = 0)  6,692  0.08  0.16  0.00  1.00   Wind blows from N-W (a prop., all 24 h = 1, none = 0)  6,692  0.18  0.22  0.00  1.00   Precipitation (mm/h)  1,673  0.21  0.49  0.00  4.37   Some precipitation on day (yes = 1)  1,673  0.45  0.50  0.00  1.00   Thermal inversion at 9 a.m. or 9 p.m. (yes = 1)  1,673  0.49  0.50  0.00  1.00  Other control variables   Gasoline price (BRL$/l, October 2008 CPI)  1,673  2.24  0.09  2.06  2.44   Ethanol price (BRL$/l, October 2008 CPI)  1,673  1.44  0.17  1.06  1.93   Diesel price index (October 2008 CPI = 1)  1,673  0.86  0.06  0.78  1.01   Ozone concentration, 24-h mean (μg/m3)  18,803  35.2  17.0  0.1  114.2   Ozone concentration, maximum 8-h mean (μg/m3)  18,803  64.7  32.0  0.4  233.4  Variable  N  Mean  Std.Dev.  Min  Max  Road congestion   Citywide congestion, weekdays only (km)  1,070  83.2  33.9  0.1  226.7   Citywide congestion (km)  1,673  55.8  46.1  0.0  226.7   Road segment congestion, weekday only (a prop.)  5.5 m  0.15  0.25  0.00  1.00   Road segment congestion (a proportion)  8.6 m  0.10  0.21  0.00  1.00  Exposure to diesel exhaust   NOx concentration, weekdays only (ppb)  10,364  64.0  49.9  0.4  490.2   NOx concentration (ppb)  16,152  56.0  46.1  0.0  490.2  Hospital admissions rate   Cardiovascular (per 100,000 residents per month)  2,365  54.2  19.8  10.7  166.4   Respiratory (per 100,000 residents per month)  2,365  46.5  20.8  5.0  130.0   Trauma (per 100,000 residents per month)  2,365  6.6  3.3  0.0  26.0  Meteorology and thermal inversion   Ground temperature (°C)  1,673  20.8  3.4  10.5  29.6   Relative humidity (%)  1,669  77.3  10.4  29.6  98.0   Solar radiation (7 a.m. to 8 p.m., Wm−2)  1,672  299.2  113.7  25.5  648.3   Atmospheric pressure (hPa)  1,673  932.2  4.8  912.1  945.2   Wind speed (ms−1)  1,673  1.4  0.5  0.3  2.9   Wind blows from N-E (a prop., all 24 h = 1, none = 0)  6,692  0.15  0.17  0.00  1.00   Wind blows from S-E (a prop., all 24 h = 1, none = 0)  6,692  0.44  0.35  0.00  1.00   Wind blows from S-W (a prop., all 24 h = 1, none = 0)  6,692  0.08  0.16  0.00  1.00   Wind blows from N-W (a prop., all 24 h = 1, none = 0)  6,692  0.18  0.22  0.00  1.00   Precipitation (mm/h)  1,673  0.21  0.49  0.00  4.37   Some precipitation on day (yes = 1)  1,673  0.45  0.50  0.00  1.00   Thermal inversion at 9 a.m. or 9 p.m. (yes = 1)  1,673  0.49  0.50  0.00  1.00  Other control variables   Gasoline price (BRL$/l, October 2008 CPI)  1,673  2.24  0.09  2.06  2.44   Ethanol price (BRL$/l, October 2008 CPI)  1,673  1.44  0.17  1.06  1.93   Diesel price index (October 2008 CPI = 1)  1,673  0.86  0.06  0.78  1.01   Ozone concentration, 24-h mean (μg/m3)  18,803  35.2  17.0  0.1  114.2   Ozone concentration, maximum 8-h mean (μg/m3)  18,803  64.7  32.0  0.4  233.4  Notes: The sample period is November 1, 2008 to May 31, 2013, that is, 1,673 days or 55 months. An observation is either a day (citywide congestion, meteorology except wind direction, fuel price), a day by monitoring site pair (NOx, wind direction, ozone), a road segment by day (road segment congestion), or a 3-digit residential zip code by month pair (hospital admissions), in the São Paulo metropolis. Citywide congestion is the mean extension during afternoon commute hours (5 p.m. to 8 p.m.). Road segment congestion is the proportion of afternoon commute hours that a road segment exhibits congestion (per traffic authority definition). NOx and meteorology are daily (24 h) means unless noted otherwise. Source: ANP, CET, CETESB, DATASUS, FAB, IBGE, INMET. View Large Table 1. Summary statistics for road congestion, NOx concentration (as a proxy for exposure to diesel exhaust), hospital admissions, meteorology, and other control variables. Variable  N  Mean  Std.Dev.  Min  Max  Road congestion   Citywide congestion, weekdays only (km)  1,070  83.2  33.9  0.1  226.7   Citywide congestion (km)  1,673  55.8  46.1  0.0  226.7   Road segment congestion, weekday only (a prop.)  5.5 m  0.15  0.25  0.00  1.00   Road segment congestion (a proportion)  8.6 m  0.10  0.21  0.00  1.00  Exposure to diesel exhaust   NOx concentration, weekdays only (ppb)  10,364  64.0  49.9  0.4  490.2   NOx concentration (ppb)  16,152  56.0  46.1  0.0  490.2  Hospital admissions rate   Cardiovascular (per 100,000 residents per month)  2,365  54.2  19.8  10.7  166.4   Respiratory (per 100,000 residents per month)  2,365  46.5  20.8  5.0  130.0   Trauma (per 100,000 residents per month)  2,365  6.6  3.3  0.0  26.0  Meteorology and thermal inversion   Ground temperature (°C)  1,673  20.8  3.4  10.5  29.6   Relative humidity (%)  1,669  77.3  10.4  29.6  98.0   Solar radiation (7 a.m. to 8 p.m., Wm−2)  1,672  299.2  113.7  25.5  648.3   Atmospheric pressure (hPa)  1,673  932.2  4.8  912.1  945.2   Wind speed (ms−1)  1,673  1.4  0.5  0.3  2.9   Wind blows from N-E (a prop., all 24 h = 1, none = 0)  6,692  0.15  0.17  0.00  1.00   Wind blows from S-E (a prop., all 24 h = 1, none = 0)  6,692  0.44  0.35  0.00  1.00   Wind blows from S-W (a prop., all 24 h = 1, none = 0)  6,692  0.08  0.16  0.00  1.00   Wind blows from N-W (a prop., all 24 h = 1, none = 0)  6,692  0.18  0.22  0.00  1.00   Precipitation (mm/h)  1,673  0.21  0.49  0.00  4.37   Some precipitation on day (yes = 1)  1,673  0.45  0.50  0.00  1.00   Thermal inversion at 9 a.m. or 9 p.m. (yes = 1)  1,673  0.49  0.50  0.00  1.00  Other control variables   Gasoline price (BRL$/l, October 2008 CPI)  1,673  2.24  0.09  2.06  2.44   Ethanol price (BRL$/l, October 2008 CPI)  1,673  1.44  0.17  1.06  1.93   Diesel price index (October 2008 CPI = 1)  1,673  0.86  0.06  0.78  1.01   Ozone concentration, 24-h mean (μg/m3)  18,803  35.2  17.0  0.1  114.2   Ozone concentration, maximum 8-h mean (μg/m3)  18,803  64.7  32.0  0.4  233.4  Variable  N  Mean  Std.Dev.  Min  Max  Road congestion   Citywide congestion, weekdays only (km)  1,070  83.2  33.9  0.1  226.7   Citywide congestion (km)  1,673  55.8  46.1  0.0  226.7   Road segment congestion, weekday only (a prop.)  5.5 m  0.15  0.25  0.00  1.00   Road segment congestion (a proportion)  8.6 m  0.10  0.21  0.00  1.00  Exposure to diesel exhaust   NOx concentration, weekdays only (ppb)  10,364  64.0  49.9  0.4  490.2   NOx concentration (ppb)  16,152  56.0  46.1  0.0  490.2  Hospital admissions rate   Cardiovascular (per 100,000 residents per month)  2,365  54.2  19.8  10.7  166.4   Respiratory (per 100,000 residents per month)  2,365  46.5  20.8  5.0  130.0   Trauma (per 100,000 residents per month)  2,365  6.6  3.3  0.0  26.0  Meteorology and thermal inversion   Ground temperature (°C)  1,673  20.8  3.4  10.5  29.6   Relative humidity (%)  1,669  77.3  10.4  29.6  98.0   Solar radiation (7 a.m. to 8 p.m., Wm−2)  1,672  299.2  113.7  25.5  648.3   Atmospheric pressure (hPa)  1,673  932.2  4.8  912.1  945.2   Wind speed (ms−1)  1,673  1.4  0.5  0.3  2.9   Wind blows from N-E (a prop., all 24 h = 1, none = 0)  6,692  0.15  0.17  0.00  1.00   Wind blows from S-E (a prop., all 24 h = 1, none = 0)  6,692  0.44  0.35  0.00  1.00   Wind blows from S-W (a prop., all 24 h = 1, none = 0)  6,692  0.08  0.16  0.00  1.00   Wind blows from N-W (a prop., all 24 h = 1, none = 0)  6,692  0.18  0.22  0.00  1.00   Precipitation (mm/h)  1,673  0.21  0.49  0.00  4.37   Some precipitation on day (yes = 1)  1,673  0.45  0.50  0.00  1.00   Thermal inversion at 9 a.m. or 9 p.m. (yes = 1)  1,673  0.49  0.50  0.00  1.00  Other control variables   Gasoline price (BRL$/l, October 2008 CPI)  1,673  2.24  0.09  2.06  2.44   Ethanol price (BRL$/l, October 2008 CPI)  1,673  1.44  0.17  1.06  1.93   Diesel price index (October 2008 CPI = 1)  1,673  0.86  0.06  0.78  1.01   Ozone concentration, 24-h mean (μg/m3)  18,803  35.2  17.0  0.1  114.2   Ozone concentration, maximum 8-h mean (μg/m3)  18,803  64.7  32.0  0.4  233.4  Notes: The sample period is November 1, 2008 to May 31, 2013, that is, 1,673 days or 55 months. An observation is either a day (citywide congestion, meteorology except wind direction, fuel price), a day by monitoring site pair (NOx, wind direction, ozone), a road segment by day (road segment congestion), or a 3-digit residential zip code by month pair (hospital admissions), in the São Paulo metropolis. Citywide congestion is the mean extension during afternoon commute hours (5 p.m. to 8 p.m.). Road segment congestion is the proportion of afternoon commute hours that a road segment exhibits congestion (per traffic authority definition). NOx and meteorology are daily (24 h) means unless noted otherwise. Source: ANP, CET, CETESB, DATASUS, FAB, IBGE, INMET. View Large We proxy for traffic volume using hourly observations on road congestion at the road segment level. The traffic authority monitors a grid of main roads comprising 5,133 segments across the city, with a total extension of 406 km (Figure 1(c)). Thus, the unit of observation measures 80 m on average. For each road segment by hour pair, we observe “congestion” or “no congestion”, per the traffic authority's definition. There were 1,070 regular weekdays (i.e., workdays excluding public holidays and a vacation fortnight around New Year) in the study period; on these weekdays’ afternoon commute hours, between 5 p.m. and 8 p.m., the extent of congestion throughout the city averages 83 km. Further reflecting the state of gridlock, the proportion of the afternoon commute that a road segment exhibits congestion (from 0 to 1, if congested continuously from 5 p.m. to 8 p.m.) averages 0.15 across 5.5 million road segment by weekday pairs. Figure 1. View largeDownload slide The beltway around the São Paulo metropolis that removed trucks from urban roads, and locations in the combined datasets. (a, b) Southeast of São Paulo state and the São Paulo metropolis, respectively. Circles (filled in red, labeled according to the environmental authority's identifier, N = 15) indicate 15 NOx monitoring sites. Balloons (filled in gray, N = 43) indicate the 3-digit zip codes (district centroids averaged by population) contained in the residential addresses of patients admitted to hospital. The beltway (new route) and Avenida dos Bandeirantes (key link in the original route) are shown. Arrows indicate 7 (out of ten) major highways connecting farms and industry in the state, via the metropolis, to the port of Santos. (c) Road segments monitored for traffic congestion in the city of São Paulo. Circles (N = 5,133) indicate the midpoints of monitored road segments. Source: CET, CETESB, DATASUS. Figure 1. View largeDownload slide The beltway around the São Paulo metropolis that removed trucks from urban roads, and locations in the combined datasets. (a, b) Southeast of São Paulo state and the São Paulo metropolis, respectively. Circles (filled in red, labeled according to the environmental authority's identifier, N = 15) indicate 15 NOx monitoring sites. Balloons (filled in gray, N = 43) indicate the 3-digit zip codes (district centroids averaged by population) contained in the residential addresses of patients admitted to hospital. The beltway (new route) and Avenida dos Bandeirantes (key link in the original route) are shown. Arrows indicate 7 (out of ten) major highways connecting farms and industry in the state, via the metropolis, to the port of Santos. (c) Road segments monitored for traffic congestion in the city of São Paulo. Circles (N = 5,133) indicate the midpoints of monitored road segments. Source: CET, CETESB, DATASUS. As a signature of diesel exhaust (Pérez-Martínez et al. 2014, 2015), we observe hourly ambient NOx concentrations at each of 11 monitoring sites across the São Paulo metropolis, all located inside of the beltway and mostly in the inner city (Figure 1(b), sites marked with circles). The 24-h mean concentration averages 64 ppb across 10,364 site by weekday pairs and 56 ppb if all days including weekends are considered.9 We also observe hourly NOx concentrations at monitoring sites located in three cities 50–100 km away from the metropolis (Figure 1(a)). NOx levels at these sites outside the metropolis are comparable to levels in the metropolis excluding sites that are close to large city roads. We note that routine monitoring of 2.5-μm-diameter particulate matter (PM2.5) began only in January 2011, nine months after the beltway opened, and at a single site in the metropolis (CETESB 2008–2013, several years). “Submicron” particles, with diameter smaller than 1 μm including gas-like “nanoparticles” with diameter measured in nanometers10 that recent toxicological research suggests are highly damaging to health (Donaldson et al. 2005; Nel 2005; Choi et al. 2010; Smita et al. 2012), are not officially monitored. Similarly, the authorities do not monitor black carbon, a species of particulate matter that is understood to be more toxic compared to inorganic compounds such as nitrates, sulfates, and crustal material (Lippmann et al. 2013; Lelieveld et al. 2015). For the later part of the study period, Figure 2 reports the tight correlation between NOx, PM2.5, submicron particle and black carbon levels in São Paulo measured by the authorities or researchers concurrently in time and space.11 The scatterplots are consistent with diesel combustion being a major if not dominant source across these pollutants, including ultrafine and carbonaceous particulates, and underscore the role of routinely monitored NOx as a marker of diesel exhaust. We provide technical references on diesel exhaust and diesel particulate matter (DE and DPM) in what follows. Figure 2. View largeDownload slide NOx as a marker for diesel exhaust and diesel particulate matter, including ultrafines and carbonaceous. (a) Daily (24 h) mean PM2.5 (μg/m3) against daily mean NOx (ppb) concentrations at the three sites in the São Paulo metropolis that concurrently monitored PM2.5 and NOx after policy implementation (PM2.5 starting in: January 2011 for site 8 in Figure 1, August 2011 for site 31, January 2012 for site 27; through the end of the sample period in May 2013). (b) Daily mean submicron 7–800 nm (“PM 0.007–0.8”) number concentration (cm−3), (c) daily mean ultrafine 7–100 nm (“PM 0.007–0.1”) number concentration (cm−3), or (d) daily mean black carbon mass concentration (μg/m3), against NOx at the University of São Paulo (at or very close to site 31 in Figure 1) after policy implementation (October 2010–September 2011 for submicron and ultrafines, or to July 2012 for black carbon). An observation is a site by day in (a) or a day in (b–d). The scatters show deseasoned concentrations, that is, residuals fitted after separately regressing each pollutant concentration on a set of month-of-year dummies. Scatters plotting raw concentrations show similar (and even tighter) correlation. Source: CETESB, field campaigns in Salvo et al. (2017). Figure 2. View largeDownload slide NOx as a marker for diesel exhaust and diesel particulate matter, including ultrafines and carbonaceous. (a) Daily (24 h) mean PM2.5 (μg/m3) against daily mean NOx (ppb) concentrations at the three sites in the São Paulo metropolis that concurrently monitored PM2.5 and NOx after policy implementation (PM2.5 starting in: January 2011 for site 8 in Figure 1, August 2011 for site 31, January 2012 for site 27; through the end of the sample period in May 2013). (b) Daily mean submicron 7–800 nm (“PM 0.007–0.8”) number concentration (cm−3), (c) daily mean ultrafine 7–100 nm (“PM 0.007–0.1”) number concentration (cm−3), or (d) daily mean black carbon mass concentration (μg/m3), against NOx at the University of São Paulo (at or very close to site 31 in Figure 1) after policy implementation (October 2010–September 2011 for submicron and ultrafines, or to July 2012 for black carbon). An observation is a site by day in (a) or a day in (b–d). The scatters show deseasoned concentrations, that is, residuals fitted after separately regressing each pollutant concentration on a set of month-of-year dummies. Scatters plotting raw concentrations show similar (and even tighter) correlation. Source: CETESB, field campaigns in Salvo et al. (2017). We obtained admissions at all public hospitals in the metropolis, comprising 3.3 million individual records over the study period. We observe the patient's dates of admission and discharge (including death), her condition on admission (per the International Classification of Disease, ICD-10), the admitting hospital, and the patient's residential zip code and age. Importantly, we use the patient's 3-digit zip code, not the hospital's, as a proxy for the site of main exposure to diesel pollution (Figure 1(b), zip codes marked with balloons). To compute incidence, we obtained resident population over time by age group by district (a subset of a zip code; see Online Appendix B.1). Across the 55 month by 43 zip code (i.e., 2,365) pairs in the sample, mean cardiovascular and respiratory admissions rates are each about 50 per month per 100,000 residents. Admissions due to these conditions that ultimately result in death are an order of magnitude lower (not shown in Table 1 for brevity). We do not observe mortality outside of public hospitals but note that in São Paulo the majority of deaths occur in (public) hospitals (Bravo et al. 2016). Mean monthly admissions rates due to trauma, excluding those related to traffic accidents, are 7 per 100,000 residents. Monthly admissions due to all causes are about 400 per 100,000 residents (again not shown), so admissions that are recorded as cardiorespiratory (cardiovascular + respiratory) account for about one-quarter of the total. 2.2. Reduced-Form Econometric Models To let the data speak, and prior to developing a structural model of the policy affecting public health through its impact on pollution, we first present reduced-form evidence. We use ordinary least squares (OLS) regression to estimate the following reduced-form model separately for each localized outcome variable, denoted by ylt, namely traffic congestion, diesel pollution and hospital admissions, where l and t index location and time, respectively,   \begin{equation} {y_{lt}} = {\alpha _1}I{\left( {{\rm{truck}}} \right)_t}\, +\, {\alpha _2}{d_l} + {\alpha _3}I{\left( {{\rm{truck}}} \right)_t}{d_l} + {x_{lt}}^{\prime}{\alpha _{4l}} + {\phi _l} + {\delta _t} + {\varepsilon _{lt}}{} \end{equation} (1) The traffic outcome is the proportion (frequency) of afternoon hours between 5 p.m. and 8 p.m. that a road segment l was congested on day t. The diesel pollution outcome is the 24-h mean NOx concentration at a monitoring site l on day t. Hospital outcomes are, separately, cardiovascular and respiratory admissions rates for the resident population in a zip code l in month t. We allow each outcome variable to shift on March 30, 2010, when the beltway opened (and just before inner-city truck circulation was restricted further), indicated by the dummy variable I(truck)t. Put simply, the dummy is set to one for all time periods t on or after this date, and zero otherwise. To capture the spatially differentiated policy effects, emanating from the key inner-city corridor—Av. dos Bandeirantes—that prior to the beltway was heavily used by trucks, we interact the time dummy I(truck)t with location l's distance (in logarithms) to the nearest point on this key inner-city corridor, dl. The main coefficients of interest to be estimated are thus α1 and α3. The core Av. dos Bandeirantes, labeled “original truck route”, is shown in Figure 1(b). Figure 3 shows how levels of traffic congestion in the city systematically fall as one moves away from this key inner-city corridor, an aspect that the policy sought to address.12 Figure 3. View largeDownload slide Levels of traffic congestion in the city of São Paulo systematically fall in the distance from the key inner-city corridor, Avenida dos Bandeirantes. Each observation in the scatterplot represents a bin, of length 100 m, at increasing radial distance from the original truck route, that is, 0–100 m away, 100–200 m away, and so forth. Plotted is the mean proportion of the weekday afternoon commute (5 p.m. to 8 p.m.) that all road segments in a given distance bin exhibit congestion (per traffic authority definition, as in Table 1). We average segment-level proportions over afternoon commute hours within each weekday, aggregate in space weighting by segment length (within each bin), then average over all weekdays in the November 1, 2008 to May 31, 2013 sample. Source: CET. Figure 3. View largeDownload slide Levels of traffic congestion in the city of São Paulo systematically fall in the distance from the key inner-city corridor, Avenida dos Bandeirantes. Each observation in the scatterplot represents a bin, of length 100 m, at increasing radial distance from the original truck route, that is, 0–100 m away, 100–200 m away, and so forth. Plotted is the mean proportion of the weekday afternoon commute (5 p.m. to 8 p.m.) that all road segments in a given distance bin exhibit congestion (per traffic authority definition, as in Table 1). We average segment-level proportions over afternoon commute hours within each weekday, aggregate in space weighting by segment length (within each bin), then average over all weekdays in the November 1, 2008 to May 31, 2013 sample. Source: CET. To account for spatial heterogeneity, we specify location fixed effects, φl, that is, separate intercepts by road corridor in the traffic model, by NOx monitoring site in the diesel pollution model, or by residential zip code in the hospital admissions model. In the traffic model, there are 178 road corridors comprising the 5,133 traffic-monitored road segments, such that one road corridor consists of about 30 road segments.13 In the traffic and pollution models, for which observations are at the day level, day-of-week and week-of-year fixed effects, δt, account for systematic weekly and annual cycles. In the hospital admissions model, with more aggregated temporal variation at the monthly level, we use month-of-year fixed effects instead. Except in the traffic model with its numerous locations, for added flexibility we interact seasonal controls (e.g., day-of-week, week-of-year) with location fixed effects—for simplicity, notation in (1) does not reflect this. We further include in δt a linear time trend, common across locations, to capture any long-run changes over the 2008–2013 sample period.14 Vector x$${lt}$$ includes extensive controls to account for other potential determinants of traffic, air quality and health outcomes, namely meteorology, the occurrence of atmospheric thermal inversions within 500 m of the ground, and consumer prices for fuels used in the heavy-vehicle (diesel) and light-vehicle (gasoline, ethanol) fleets (see Table 1). The table of results that we present after spelling out our hypotheses summarizes the covariates in each reduced-form model. The idiosyncratic disturbance term in each reduced-form equation is denoted by ε$${lt}$$. 2.3. Hypotheses Reduced-form model (1) estimates the environmental effects of abating truck flows emanating from the key original truck route, and allows the effects at locations to differ based on their distance from this route. We hypothesize that the policy relieved road congestion and, unintentionally, reduced ambient NOx levels (among other unobserved contaminants associated with diesel exhaust) and hospital visits due to cardiovascular and respiratory disease, that is, α1 < 0 for all three environmental externalities. Moreover, we expect the environmental impacts of abating diesel truck flows to attenuate as we move away from the original truck route, that is, α3 > 0. With trucks being induced out of the inner city's gridlocked roads, we further expect that the road space they left behind was gradually filled by light vehicles, consistent with the law of road congestion (Downs 1962; Vickrey 1969; Duranton and Turner 2011; Hsu and Zhang 2014). Figure 4(a) indeed shows a large drop in recorded congestion (deseasoned) along the key inner-city corridor when the policy was introduced, and that the change in congestion was short-lived. To formally test for this phenomenon, on applying model (1) to traffic congestion we additionally specify a time trend that begins only on the date the beltway opened, and interact this “trend after policy implementation” with dl, road segment l's (log) distance from the original truck route.15 We hypothesize that the truck policy's effect on inner-city traffic volume—which was the original policy motivation—to be undone in the medium run as light vehicles selected into newly available road space. (For brevity, only Table 2, but not expression (1), reflects these additional terms.) In contrast, the spatial effects of the policy on diesel pollution and hospitalizations were long lasting, as the policy did change traffic composition even if it did not meet the goal of relieving congestion—see Figure 4(b) and (c) and the analysis that follows. Figure 4. View largeDownload slide Road congestion, diesel pollution, and hospitalizations near and far and over time. (a) Proportions of the city's road segments that are congested: along Av. dos Bandeirantes (black, solid); and between 10 and 20 km from this key inner-city corridor (red, dotted). (b) Ambient NOx concentrations: at a site only 20 m from Av. dos Bandeirantes (black solid); and at four sites 50–100 km away from the São Paulo metropolis (red, dotted). (c) Cardiovascular admission rates among residents: at the 10 zip codes (25%) that are closer to the Av. dos Bandeirantes (black, solid, left axis); at the 10 zip codes that are furthest from this key inner-city corridor (red, dotted, right axis); and the difference (blue, thick, right axis). In (a), we average segment-level proportions over afternoon commute hours within each weekday, aggregate in space weighting by segment length, deseason (regress on week-of-year and day-of-week), and plot mean residuals by month. In (b), we average monitor-level concentrations over 24 h within each weekday, deseason, and plot mean residuals by month (and across sites outside the metropolis). In (c), we sum admissions over all days within each quarter and across residents of each group of 10 zip codes, and divide by the resident population. The vertical line indicates the opening of the beltway. Figure 4. View largeDownload slide Road congestion, diesel pollution, and hospitalizations near and far and over time. (a) Proportions of the city's road segments that are congested: along Av. dos Bandeirantes (black, solid); and between 10 and 20 km from this key inner-city corridor (red, dotted). (b) Ambient NOx concentrations: at a site only 20 m from Av. dos Bandeirantes (black solid); and at four sites 50–100 km away from the São Paulo metropolis (red, dotted). (c) Cardiovascular admission rates among residents: at the 10 zip codes (25%) that are closer to the Av. dos Bandeirantes (black, solid, left axis); at the 10 zip codes that are furthest from this key inner-city corridor (red, dotted, right axis); and the difference (blue, thick, right axis). In (a), we average segment-level proportions over afternoon commute hours within each weekday, aggregate in space weighting by segment length, deseason (regress on week-of-year and day-of-week), and plot mean residuals by month. In (b), we average monitor-level concentrations over 24 h within each weekday, deseason, and plot mean residuals by month (and across sites outside the metropolis). In (c), we sum admissions over all days within each quarter and across residents of each group of 10 zip codes, and divide by the resident population. The vertical line indicates the opening of the beltway. Table 2. Reduced-form analysis of traffic volume, diesel pollution, and hospitalizations: effects of the truck abatement policy on different outcome variables across the São Paulo metropolis.   (1)  (2)  (3)  (4)  Dependent variable  Road segment congestion (% afternoon commute hours)  Diesel exhaust marker (24-h NOx, ppb)  Cardiovascular (all ages, per 100,000 per month)  Respiratory (all ages, per 100,000 per month)  Truck policy implemented (yes = 1)  −0.049***  −9.93***  −5.80***  −3.30**    (0.004)  (2.24)  (1.76)  (1.67)  ln(distance) from key original truck route  −0.042***  Subsumed in  Subsumed in  Subsumed in    (0.001)  site FEs  zip FEs  zip FEs  Truck policy implemented (yes = 1)  0.018***  8.78***  3.24***  1.61***   × ln(distance) from orig. truck route  (0.002)  (0.49)  (0.59)  (0.58)  Trend after policy implementation  0.066***  –  –  –    (0.019)  –  –  –  Trend after policy implementation  −0.013***  –  –  –   × ln(distance) from orig. truck route  (0.002)  –  –  –  Road corridor fixed effects (FEs)  Yes (178)  –  –  –  NOx monitoring site FEs  –  Yes (11)  –  –  3-digit residential zip code FEs  –  –  Yes (43)  Yes (43)  Week-of-year FEs  Yes  –  –  –  Week-of-year by location FEs  –  Yes  –  –  Day-of-week FEs  Yes  –  –  –  Day-of-week by location FEs  –  Yes  –  –  Month-of-year by location FEs  –  –  Yes  Yes  Meteorology and thermal inversion  Yes  Yes  Yes  Yes  Time trend over the entire sample period  Yes  Yes  Yes  Yes  Fuel prices (gasoline, ethanol, diesel)  Yes  Yes  Yes  Yes  May–September 2009 swine flu epidemic (yes = 1)  –  –  Yes  Yes  Number of observations  8,561,844  16,106  2,365  2,365  Number of regressors  259  699  576  576  Number of locations  5,133 segments  11 sites  43 zips  43 zips  R2  0.296  0.757  0.902  0.912  Mean value of dependent variable  0.102  55.92  54.21  46.52    (1)  (2)  (3)  (4)  Dependent variable  Road segment congestion (% afternoon commute hours)  Diesel exhaust marker (24-h NOx, ppb)  Cardiovascular (all ages, per 100,000 per month)  Respiratory (all ages, per 100,000 per month)  Truck policy implemented (yes = 1)  −0.049***  −9.93***  −5.80***  −3.30**    (0.004)  (2.24)  (1.76)  (1.67)  ln(distance) from key original truck route  −0.042***  Subsumed in  Subsumed in  Subsumed in    (0.001)  site FEs  zip FEs  zip FEs  Truck policy implemented (yes = 1)  0.018***  8.78***  3.24***  1.61***   × ln(distance) from orig. truck route  (0.002)  (0.49)  (0.59)  (0.58)  Trend after policy implementation  0.066***  –  –  –    (0.019)  –  –  –  Trend after policy implementation  −0.013***  –  –  –   × ln(distance) from orig. truck route  (0.002)  –  –  –  Road corridor fixed effects (FEs)  Yes (178)  –  –  –  NOx monitoring site FEs  –  Yes (11)  –  –  3-digit residential zip code FEs  –  –  Yes (43)  Yes (43)  Week-of-year FEs  Yes  –  –  –  Week-of-year by location FEs  –  Yes  –  –  Day-of-week FEs  Yes  –  –  –  Day-of-week by location FEs  –  Yes  –  –  Month-of-year by location FEs  –  –  Yes  Yes  Meteorology and thermal inversion  Yes  Yes  Yes  Yes  Time trend over the entire sample period  Yes  Yes  Yes  Yes  Fuel prices (gasoline, ethanol, diesel)  Yes  Yes  Yes  Yes  May–September 2009 swine flu epidemic (yes = 1)  –  –  Yes  Yes  Number of observations  8,561,844  16,106  2,365  2,365  Number of regressors  259  699  576  576  Number of locations  5,133 segments  11 sites  43 zips  43 zips  R2  0.296  0.757  0.902  0.912  Mean value of dependent variable  0.102  55.92  54.21  46.52  Notes: An observation is: (1) a day by road segment, (2) a day by monitoring site pair, or (3, 4) a month by 3-digit residential zip code. The sample period is November 2008 to May 2013. All models control for: location fixed effects (road corridor, NOx monitoring site, or zip code); location by seasonality fixed effects (week-of-year and day-of-week, or month-of-year); meteorology and thermal inversion; linear time trend; and road transportation fuel prices. Policy implemented indicates an observation: (1, 2) on or after March 30, 2010, or (3, 4) on or after April 2010. OLS regressions. Standard errors, in parentheses, are: (1, 2) one-way clustered by week-of-sample, or (3, 4) robust. Clustering by month-of-sample in (1, 2), or two-way clustering by week-of-sample and segment in (1), yields slightly higher standard errors, but the same significance levels. Allowing for spatial and serial correlation (cutoffs at 5 km and up to 2 months) yields similar precision. *Significant at 0.1; **significant at 0.05; ***significant at 0.01. View Large Table 2. Reduced-form analysis of traffic volume, diesel pollution, and hospitalizations: effects of the truck abatement policy on different outcome variables across the São Paulo metropolis.   (1)  (2)  (3)  (4)  Dependent variable  Road segment congestion (% afternoon commute hours)  Diesel exhaust marker (24-h NOx, ppb)  Cardiovascular (all ages, per 100,000 per month)  Respiratory (all ages, per 100,000 per month)  Truck policy implemented (yes = 1)  −0.049***  −9.93***  −5.80***  −3.30**    (0.004)  (2.24)  (1.76)  (1.67)  ln(distance) from key original truck route  −0.042***  Subsumed in  Subsumed in  Subsumed in    (0.001)  site FEs  zip FEs  zip FEs  Truck policy implemented (yes = 1)  0.018***  8.78***  3.24***  1.61***   × ln(distance) from orig. truck route  (0.002)  (0.49)  (0.59)  (0.58)  Trend after policy implementation  0.066***  –  –  –    (0.019)  –  –  –  Trend after policy implementation  −0.013***  –  –  –   × ln(distance) from orig. truck route  (0.002)  –  –  –  Road corridor fixed effects (FEs)  Yes (178)  –  –  –  NOx monitoring site FEs  –  Yes (11)  –  –  3-digit residential zip code FEs  –  –  Yes (43)  Yes (43)  Week-of-year FEs  Yes  –  –  –  Week-of-year by location FEs  –  Yes  –  –  Day-of-week FEs  Yes  –  –  –  Day-of-week by location FEs  –  Yes  –  –  Month-of-year by location FEs  –  –  Yes  Yes  Meteorology and thermal inversion  Yes  Yes  Yes  Yes  Time trend over the entire sample period  Yes  Yes  Yes  Yes  Fuel prices (gasoline, ethanol, diesel)  Yes  Yes  Yes  Yes  May–September 2009 swine flu epidemic (yes = 1)  –  –  Yes  Yes  Number of observations  8,561,844  16,106  2,365  2,365  Number of regressors  259  699  576  576  Number of locations  5,133 segments  11 sites  43 zips  43 zips  R2  0.296  0.757  0.902  0.912  Mean value of dependent variable  0.102  55.92  54.21  46.52    (1)  (2)  (3)  (4)  Dependent variable  Road segment congestion (% afternoon commute hours)  Diesel exhaust marker (24-h NOx, ppb)  Cardiovascular (all ages, per 100,000 per month)  Respiratory (all ages, per 100,000 per month)  Truck policy implemented (yes = 1)  −0.049***  −9.93***  −5.80***  −3.30**    (0.004)  (2.24)  (1.76)  (1.67)  ln(distance) from key original truck route  −0.042***  Subsumed in  Subsumed in  Subsumed in    (0.001)  site FEs  zip FEs  zip FEs  Truck policy implemented (yes = 1)  0.018***  8.78***  3.24***  1.61***   × ln(distance) from orig. truck route  (0.002)  (0.49)  (0.59)  (0.58)  Trend after policy implementation  0.066***  –  –  –    (0.019)  –  –  –  Trend after policy implementation  −0.013***  –  –  –   × ln(distance) from orig. truck route  (0.002)  –  –  –  Road corridor fixed effects (FEs)  Yes (178)  –  –  –  NOx monitoring site FEs  –  Yes (11)  –  –  3-digit residential zip code FEs  –  –  Yes (43)  Yes (43)  Week-of-year FEs  Yes  –  –  –  Week-of-year by location FEs  –  Yes  –  –  Day-of-week FEs  Yes  –  –  –  Day-of-week by location FEs  –  Yes  –  –  Month-of-year by location FEs  –  –  Yes  Yes  Meteorology and thermal inversion  Yes  Yes  Yes  Yes  Time trend over the entire sample period  Yes  Yes  Yes  Yes  Fuel prices (gasoline, ethanol, diesel)  Yes  Yes  Yes  Yes  May–September 2009 swine flu epidemic (yes = 1)  –  –  Yes  Yes  Number of observations  8,561,844  16,106  2,365  2,365  Number of regressors  259  699  576  576  Number of locations  5,133 segments  11 sites  43 zips  43 zips  R2  0.296  0.757  0.902  0.912  Mean value of dependent variable  0.102  55.92  54.21  46.52  Notes: An observation is: (1) a day by road segment, (2) a day by monitoring site pair, or (3, 4) a month by 3-digit residential zip code. The sample period is November 2008 to May 2013. All models control for: location fixed effects (road corridor, NOx monitoring site, or zip code); location by seasonality fixed effects (week-of-year and day-of-week, or month-of-year); meteorology and thermal inversion; linear time trend; and road transportation fuel prices. Policy implemented indicates an observation: (1, 2) on or after March 30, 2010, or (3, 4) on or after April 2010. OLS regressions. Standard errors, in parentheses, are: (1, 2) one-way clustered by week-of-sample, or (3, 4) robust. Clustering by month-of-sample in (1, 2), or two-way clustering by week-of-sample and segment in (1), yields slightly higher standard errors, but the same significance levels. Allowing for spatial and serial correlation (cutoffs at 5 km and up to 2 months) yields similar precision. *Significant at 0.1; **significant at 0.05; ***significant at 0.01. View Large 2.4. Unintended Policy Consequences Table 2 reports estimates of model (1) applied to traffic volume (column (1)), diesel pollution (column (2)), and hospitalizations (columns (3) and (4)) across different locations in the metropolis as the truck abatement policy was implemented. All of our hypotheses are borne out in the data.16 First, road congestion, ambient NOx levels and cardiovascular and respiratory hospitalizations all fell (with significance at the 1% level) as truck flows in the inner city abated. Second, the reduction in all of these environmental outcomes was less pronounced (again, with significance at the 1% level) the more distant the site of exposure—road segment, air monitor and residential zip code—from the key original truck route. Third, as Figure 4(a) suggests, we find evidence of reversion in the temporary traffic relief that the abatement of truck flows brought about.17 To help interpret Table 2 estimates, Table 3 fits the different models to varying distances from the key inner-city corridor (50 m, 1, 2, 5, 10 km from Av. dos Bandeirantes) at varying points in time (April 2009, 2010, 2012). The proportion of the afternoon commute that a road segment suffered congestion right by the original truck route (50 m) one year prior to policy implementation, in April 2009, was 0.48 (s.e. 0.01)—see panel (1). As fitted by the model, congestion at this location fell sharply by 20% right after policy implementation, in April 2010, to 0.38 (s.e. 0.01). Only two years after policy implementation, however, the effect on congestion at this location from removing trucks was 50% undone, with the predicted congestion frequency rising back to 0.43 (s.e. 0.01). In the face of widespread gridlock on the city's roads and repressed demand for travel among city dwellers, passenger cars increased their share of road space (for a simple theoretical economic model of this mechanism, see Appendix  A). The policy's mild effect on traffic volume only a few years from implementation was its first unintended consequence. Table 3. Fitted road congestion, diesel pollution and health outcomes one year before, upon (April 2010), and two years after policy implementation. Outcome  Distance from key original truck route  One year before policy implemented (April 2009)  Upon policy implementation (April 2010)  Two years after policy implemented (April 2012)  (1) Road segment congestion (% afternoon commute hours, 5 p.m. to 8 p.m.)  50 m away  0.482 (0.005)  0.376 (0.010)  0.429 (0.012)    1 km away  0.355 (0.003)  0.302 (0.009)  0.331 (0.012)    2 km away  0.326 (0.003)  0.284 (0.009)  0.308 (0.012)    5 km away  0.287 (0.004)  0.261 (0.009)  0.278 (0.012)    10 km away  0.258 (0.005)  0.244 (0.009)  0.256 (0.012)  (2) Diesel exhaust marker (24-h ambient NOx levels, ppb)  50 m away  148.1 (3.3)  111.0 (4.2)  109.4 (4.8)    1 km away  125.2 (0.4)  114.4 (2.3)  112.8 (3.3)    2 km away  119.9 (0.9)  115.2 (2.3)  113.6 (3.3)    5 km away  112.9 (1.8)  116.3 (2.7)  114.6 (3.5)    10 km away  107.6 (2.6)  117.1 (3.1)  115.4 (3.9)  (3) Cardiovascular admissions (all ages, per 100,000 residents per month)  50 m away  Out of sample  Out of sample  Out of sample    1 km away  49.4 (0.2)  45.1 (2.1)  44.8 (2.2)    2 km away  49.2 (1.4)  47.2 (2.3)  46.9 (2.4)    5 km away  49.1 (3.2)  50.0 (3.5)  49.7 (3.6)    10 km away  49.0 (4.5)  52.1 (4.7)  51.8 (4.7)  (4) Respiratory admissions (all ages, per 100,000 residents per month)  50 m away  Out of sample  Out of sample  Out of sample    1 km away  41.3 (0.2)  38.2 (1.9)  38.5 (2.1)    2 km away  42.9 (1.2)  41.0 (2.0)  41.3 (2.1)    5 km away  45.1 (2.7)  44.7 (2.9)  45.0 (3.0)    10 km away  46.8 (3.8)  47.5 (3.9)  47.8 (4.0)  Outcome  Distance from key original truck route  One year before policy implemented (April 2009)  Upon policy implementation (April 2010)  Two years after policy implemented (April 2012)  (1) Road segment congestion (% afternoon commute hours, 5 p.m. to 8 p.m.)  50 m away  0.482 (0.005)  0.376 (0.010)  0.429 (0.012)    1 km away  0.355 (0.003)  0.302 (0.009)  0.331 (0.012)    2 km away  0.326 (0.003)  0.284 (0.009)  0.308 (0.012)    5 km away  0.287 (0.004)  0.261 (0.009)  0.278 (0.012)    10 km away  0.258 (0.005)  0.244 (0.009)  0.256 (0.012)  (2) Diesel exhaust marker (24-h ambient NOx levels, ppb)  50 m away  148.1 (3.3)  111.0 (4.2)  109.4 (4.8)    1 km away  125.2 (0.4)  114.4 (2.3)  112.8 (3.3)    2 km away  119.9 (0.9)  115.2 (2.3)  113.6 (3.3)    5 km away  112.9 (1.8)  116.3 (2.7)  114.6 (3.5)    10 km away  107.6 (2.6)  117.1 (3.1)  115.4 (3.9)  (3) Cardiovascular admissions (all ages, per 100,000 residents per month)  50 m away  Out of sample  Out of sample  Out of sample    1 km away  49.4 (0.2)  45.1 (2.1)  44.8 (2.2)    2 km away  49.2 (1.4)  47.2 (2.3)  46.9 (2.4)    5 km away  49.1 (3.2)  50.0 (3.5)  49.7 (3.6)    10 km away  49.0 (4.5)  52.1 (4.7)  51.8 (4.7)  (4) Respiratory admissions (all ages, per 100,000 residents per month)  50 m away  Out of sample  Out of sample  Out of sample    1 km away  41.3 (0.2)  38.2 (1.9)  38.5 (2.1)    2 km away  42.9 (1.2)  41.0 (2.0)  41.3 (2.1)    5 km away  45.1 (2.7)  44.7 (2.9)  45.0 (3.0)    10 km away  46.8 (3.8)  47.5 (3.9)  47.8 (4.0)  Notes: Point estimates and standard errors, in parentheses, are based on, or more restrictive versions of, the reduced-form models reported in Table 2. For traffic outcome (1), to the sample mean for road segment congestion on the original truck route, we apply the estimated temporal by spatial variation (respectively, pre and post policy, 50 m to 10 km away), according to estimates in the first 5 rows of column (1), Table 2 and the time trend. For diesel pollution outcome (2), to the sample mean for NOx at one site 20 m from the original truck route, we apply variation estimated in a more restrictive model than in Table 2, without monitor fixed effects, in order to identify the coefficient on (log) distance. For hospital outcomes (3, 4), to the sample mean for hospitalizations among residents of three zip codes closest to the original truck route, we apply variation estimated in a more restrictive model than in Table 2, without zip code fixed effects, in order to identify the coefficient on (log) distance. The closest zip code—according to the population-weighted average centroid of a zip code's districts—is 2 km away. View Large Table 3. Fitted road congestion, diesel pollution and health outcomes one year before, upon (April 2010), and two years after policy implementation. Outcome  Distance from key original truck route  One year before policy implemented (April 2009)  Upon policy implementation (April 2010)  Two years after policy implemented (April 2012)  (1) Road segment congestion (% afternoon commute hours, 5 p.m. to 8 p.m.)  50 m away  0.482 (0.005)  0.376 (0.010)  0.429 (0.012)    1 km away  0.355 (0.003)  0.302 (0.009)  0.331 (0.012)    2 km away  0.326 (0.003)  0.284 (0.009)  0.308 (0.012)    5 km away  0.287 (0.004)  0.261 (0.009)  0.278 (0.012)    10 km away  0.258 (0.005)  0.244 (0.009)  0.256 (0.012)  (2) Diesel exhaust marker (24-h ambient NOx levels, ppb)  50 m away  148.1 (3.3)  111.0 (4.2)  109.4 (4.8)    1 km away  125.2 (0.4)  114.4 (2.3)  112.8 (3.3)    2 km away  119.9 (0.9)  115.2 (2.3)  113.6 (3.3)    5 km away  112.9 (1.8)  116.3 (2.7)  114.6 (3.5)    10 km away  107.6 (2.6)  117.1 (3.1)  115.4 (3.9)  (3) Cardiovascular admissions (all ages, per 100,000 residents per month)  50 m away  Out of sample  Out of sample  Out of sample    1 km away  49.4 (0.2)  45.1 (2.1)  44.8 (2.2)    2 km away  49.2 (1.4)  47.2 (2.3)  46.9 (2.4)    5 km away  49.1 (3.2)  50.0 (3.5)  49.7 (3.6)    10 km away  49.0 (4.5)  52.1 (4.7)  51.8 (4.7)  (4) Respiratory admissions (all ages, per 100,000 residents per month)  50 m away  Out of sample  Out of sample  Out of sample    1 km away  41.3 (0.2)  38.2 (1.9)  38.5 (2.1)    2 km away  42.9 (1.2)  41.0 (2.0)  41.3 (2.1)    5 km away  45.1 (2.7)  44.7 (2.9)  45.0 (3.0)    10 km away  46.8 (3.8)  47.5 (3.9)  47.8 (4.0)  Outcome  Distance from key original truck route  One year before policy implemented (April 2009)  Upon policy implementation (April 2010)  Two years after policy implemented (April 2012)  (1) Road segment congestion (% afternoon commute hours, 5 p.m. to 8 p.m.)  50 m away  0.482 (0.005)  0.376 (0.010)  0.429 (0.012)    1 km away  0.355 (0.003)  0.302 (0.009)  0.331 (0.012)    2 km away  0.326 (0.003)  0.284 (0.009)  0.308 (0.012)    5 km away  0.287 (0.004)  0.261 (0.009)  0.278 (0.012)    10 km away  0.258 (0.005)  0.244 (0.009)  0.256 (0.012)  (2) Diesel exhaust marker (24-h ambient NOx levels, ppb)  50 m away  148.1 (3.3)  111.0 (4.2)  109.4 (4.8)    1 km away  125.2 (0.4)  114.4 (2.3)  112.8 (3.3)    2 km away  119.9 (0.9)  115.2 (2.3)  113.6 (3.3)    5 km away  112.9 (1.8)  116.3 (2.7)  114.6 (3.5)    10 km away  107.6 (2.6)  117.1 (3.1)  115.4 (3.9)  (3) Cardiovascular admissions (all ages, per 100,000 residents per month)  50 m away  Out of sample  Out of sample  Out of sample    1 km away  49.4 (0.2)  45.1 (2.1)  44.8 (2.2)    2 km away  49.2 (1.4)  47.2 (2.3)  46.9 (2.4)    5 km away  49.1 (3.2)  50.0 (3.5)  49.7 (3.6)    10 km away  49.0 (4.5)  52.1 (4.7)  51.8 (4.7)  (4) Respiratory admissions (all ages, per 100,000 residents per month)  50 m away  Out of sample  Out of sample  Out of sample    1 km away  41.3 (0.2)  38.2 (1.9)  38.5 (2.1)    2 km away  42.9 (1.2)  41.0 (2.0)  41.3 (2.1)    5 km away  45.1 (2.7)  44.7 (2.9)  45.0 (3.0)    10 km away  46.8 (3.8)  47.5 (3.9)  47.8 (4.0)  Notes: Point estimates and standard errors, in parentheses, are based on, or more restrictive versions of, the reduced-form models reported in Table 2. For traffic outcome (1), to the sample mean for road segment congestion on the original truck route, we apply the estimated temporal by spatial variation (respectively, pre and post policy, 50 m to 10 km away), according to estimates in the first 5 rows of column (1), Table 2 and the time trend. For diesel pollution outcome (2), to the sample mean for NOx at one site 20 m from the original truck route, we apply variation estimated in a more restrictive model than in Table 2, without monitor fixed effects, in order to identify the coefficient on (log) distance. For hospital outcomes (3, 4), to the sample mean for hospitalizations among residents of three zip codes closest to the original truck route, we apply variation estimated in a more restrictive model than in Table 2, without zip code fixed effects, in order to identify the coefficient on (log) distance. The closest zip code—according to the population-weighted average centroid of a zip code's districts—is 2 km away. View Large As Table 3, panel (1) further reports, the model's estimate that the policy hardly affected traffic congestion at locations distant from the key inner-city corridor is reassuring. For example, 5 km from Av. dos Bandeirantes congestion frequency was 0.29 in April 2009, slightly down to 0.26 in April 2010, and back up to 0.28 in April 2011. Figure 5(a) provides stark visualization of the immediate policy effect, that is, in April 2010, estimated by the reduced-form traffic model. To prepare the scatterplot, we use a more flexible variant than in Table 2: instead of constraining the effect to vary in proportion to log distance, we interact the intervention indicator I(truck)t with 5,133 segment fixed effects and estimate 5,133 coefficients, plotting these against distance. Figure 5. View largeDownload slide Estimated localized policy effects on road congestion, diesel pollution and cardiovascular and respiratory hospitalizations. An observation in each scatterplot marks a specific location in the São Paulo metropolis, at a distance from the original truck route: (a) 5,133 road segments, (b) 11 NOx monitoring sites, or (c, d) 43 3-digit residential zip codes. Scatters in (a) and (c, d) are estimates based on reduced-form models as in Table 2, except that we interact the policy implementation indicator with a full set of location fixed effects (N = 5,133 or 43). This is more flexible than interacting policy implementation with (log) distance from the original truck route, as Table 2 reports for ease of exposition. Moreover, the fitted line and 95% confidence intervals overlaid on these three scatters are obtained from the more restrictive models used in Table 3, in which policy implementation is interacted with (log) distance and detailed location fixed effects are excluded in order to identify the coefficient on (log) distance. In (b), 95% CI are from Table 4, specification 1 (Appendix Table B.4 details individual effects, N = 11). Figure 5. View largeDownload slide Estimated localized policy effects on road congestion, diesel pollution and cardiovascular and respiratory hospitalizations. An observation in each scatterplot marks a specific location in the São Paulo metropolis, at a distance from the original truck route: (a) 5,133 road segments, (b) 11 NOx monitoring sites, or (c, d) 43 3-digit residential zip codes. Scatters in (a) and (c, d) are estimates based on reduced-form models as in Table 2, except that we interact the policy implementation indicator with a full set of location fixed effects (N = 5,133 or 43). This is more flexible than interacting policy implementation with (log) distance from the original truck route, as Table 2 reports for ease of exposition. Moreover, the fitted line and 95% confidence intervals overlaid on these three scatters are obtained from the more restrictive models used in Table 3, in which policy implementation is interacted with (log) distance and detailed location fixed effects are excluded in order to identify the coefficient on (log) distance. In (b), 95% CI are from Table 4, specification 1 (Appendix Table B.4 details individual effects, N = 11). The second unintended consequence of the policy to induce trucks off the megacity's roads was to abate diesel pollution, particularly right by the densely populated inner-city corridor (Table 2, column (2)). The reduced-form model applied to NOx levels predicts that 50 m from the key-inner city corridor, a 24-h mean of 148 ppb (s.e. 3 ppb) in April 2009, a year before policy implementation, dropped by 25% to 111 ppb (s.e. 4 ppb) on the month that immediately followed the beltway opening, with little variation thereafter (Table 3, panel (2)). For example, two years after implementation, NOx levels 50 m from the core corridor remained at 109 ppb. Again, the predicted abatement in ambient NOx is decreasing in the distance from this core road. For example, the fitted model suggests that 5 km from Av. dos Bandeirantes, NOx levels did not fall upon policy implementation (116 ppb in April 2010 vs. 113 ppb a year earlier). Emissions inventories published by the environmental authority estimate that heavy-duty (diesel) vehicles account for over three-quarters of vehicular NOx emissions in the metropolis (CETESB 2008–2013, several years). The passenger car and motorcycle fleets run almost entirely on mixtures of gasoline and ethanol (Salvo and Geiger 2014). Power generation in southeastern Brazil is mostly hydroelectric, and mild winters imply minimal residential heating (Negri 2010).18 Together these features imply that ambient NOx levels in the metropolis serve more broadly as a marker for diesel exhaust, which contains or forms many toxins beyond NOx that are not routinely monitored by the environmental authority, including submicron particles and black carbon shown in Figure 2 (Zhu et al. 2002; Kittelson et al. 2004; Burtscher 2005; Gentner et al. 2012; Liu et al. 2012; Rollins et al. 2012; Brito et al. 2013; Vara-Vela et al. 2016. In Online Appendix B.2, we use approximate emission factors and other assumptions to show that diverting 20,000 trucks to the beltway, and replacing each inner-city truck-km by twice as many light-vehicle-km, may have abated the equivalent of 17% and 18%, respectively, of NOx and PM2.5 emissions in the metropolis (against a 2% increase in aggregate CO emissions).19 Even at the aggregate metropolis level, this was a large-scale intervention, shifting the urban transportation fuel mix and associated pollutants away from diesel.20 Reduced-form model estimates of the spatially differentiated removal of trucks on health damage parallel those for diesel pollution. Such joint evidence is important since the dependent variables—hospitalization rates and ambient NOx levels—come from distinct data sources, respectively, the health and environmental authorities. The last two columns of Table 2 present estimates for cardiovascular and respiratory hospital admissions for patients of all ages. Hospitalizations for patients living in zip codes near to the key inner-city corridor dropped upon policy implementation and this drop attenuates as we move away from the original truck route. Findings are similar if we condition on vulnerable age groups, such as cardiovascular disease in the elderly subpopulation or respiratory disease among children aged 0–4 years. Table 3 predictions, based on a less-flexible reduced-form model without zip code fixed effects—only for the purpose of identifying a distance coefficient to describe the variation—suggest that monthly hospitalization rates for residences 1 km from the key inner-city corridor dropped by 4 (cardiovascular) and 3 (respiratory) per 100,000, but did not fall 10 km away. Instead, 10 km away, they rose by 3 and 1 per 100,000, respectively.21 Estimates are similar in Figure 5(c) and (d) when we use a more flexible model with the policy indicator I(truck)t interacted with a full set of zip code fixed effects (rather than an interaction with log distance as in Table 2). Data plotted in Figure 4(c) shows that monthly cardiovascular rates per 100,000 residents in the 10 zip codes closest versus the 10 zip codes furthest from the core Av. dos Bandeirantes were higher prior to policy implementation (e.g., due to high exposure to diesel pollution) but similar after the intervention. Figure 1(b) shows that the 10 zip codes furthest from the core road tend to lie to the northeast of the city, in contrast to the beltway, southwest of it and even further away. Importantly, all the patterns reported up to here come directly from the data; the structural model we turn to next additionally allows us to quantify the public health impact of an actual fleet of diesel-fueled trucks crossing densely populated neighborhoods of a megacity. 3. Estimating the Effect of Diesel Pollution on Health in a Real-World Megacity 3.1. Structural Econometric Model The reduced-form analysis separately and significantly linked spatial environmental outcomes of interest—road congestion, ambient NOx and hospitalizations—to the intervention. The empirical strategy we now propose maps month-by-zip code level diesel pollution directly to hospitalizations, by way of a dose-response function—label this the structural health equation.22 One approach is to fit the health equation by OLS regression. An alternative approach, and our preferred approach, is to fit the health equation using two-stage least squares (2SLS) regression. The 2SLS estimator requires the use of an instrumental variable that is correlated with the measured dose variable (pollution exposure) and that does not drive the response variable (hospitalizations) directly. The use of 2SLS alleviates concern with regard to possible unobserved determinants of hospitalizations that correlate with diesel pollution—but not with the instrument—that under OLS generate “omitted variable bias”, or concern with measurement error in pollution exposure—measurement error that is uncorrelated with the instrument. Measurement error in the measured dose can lead to “attenuation bias”, that is, an OLS estimate of the dose-response slope that is biased toward zero. 2SLS is standard in environmental economics research into the health response to air pollution (Moretti and Neidell 2011; Schlenker and Walker 2015). Policy interventions typically function as instruments, to the extent that they affect health outcomes only indirectly through their effect on emissions (Isen et al. 2017). Our 2SLS estimates use policy-induced variation in NOx as an instrument for pollution variation that is included in the health equation. One takeaway from the reduced-form analysis is heightened confidence that localized exposure to diesel pollution shifted with the truck policy, I(truck)t. The identifying assumption is that, in our sample, the policy to abate truck flows affected health in the metropolis only by shifting local air pollutant composition, away from diesel exhaust and to gasoline–ethanol emissions.23 We report not only 2SLS but also OLS estimates of the health effect of exposure to diesel pollution; to preview our results, we obtain larger magnitudes under 2SLS (which tends to be the case in the cited literature). The health equation is   \begin{equation} {\mathit {healt}}{\mathit h_{lt}} = {\beta _{1l}}{\mathit {diese}}{l_{lt}} + {x_{lt}}^{\prime}{\beta _2} + {\mu _l} + {\xi _t} + {\epsilon _{lt}} \end{equation} (2) Health outcomes of interest are hospitalizations, including admissions that terminate in death, by age group and by condition—cardiovascular, respiratory, and as a placebo, nontraffic related trauma. Similar to the reduced form, an observation is a residential zip code l by month t pair (43 zips in the metropolis, 55 months). We proxy for exposure to diesel exhaust, diesel$${lt}$$, using monthly mean ambient NOx at the closest monitor—but not to exceed 5 km of a zip code's district centroid or, for robustness, 10 km (Bravo et al. 2016). Covariates (μl, ξt, x$${lt}$$) control for potentially confounding spatial and temporal variation including residential zip code fixed effects, seasonality (month-of-year by zip code fixed effects), month-of-sample fixed effects (more flexible than a time trend), and meteorology (subsumed by month-of-sample, if not interacted with zip code). In a robustness test, we introduce local road congestion as an additional control in the health equation (since congestion can be an additional channel through which hospitalizations are affected, for example, Peters et al. (2004)). The econometric error is ε$${lt}$$.24 To generate the instrument $${\widehat {\mathit {diesel}}_{lt}}$$, we predict variation in diesel (NOx) pollution that with confidence we attribute to the truck policy, by fitting a model that is spatially more flexible than the reduced-form pollution model (1):   \begin{equation} {\mathit {diese}}{l_{lt}} = {\alpha _1}I{\left( {{\rm{truck}}} \right)_t} + {\alpha _{2l,l \in {\mathit {metro}}}}I{\left( {{\rm{truck}}} \right)_t} + {x_{lt}}^{\prime}{\alpha _{3l}} + {\phi _l} + {\delta _t} + {\varepsilon _{lt}}. \end{equation} (3) As in the reduced form, an observation is a NOx monitoring site l by day t pair, except that the sample now includes all sites in cities 50–100 km away from the São Paulo metropolis, in addition to the 11 sites in the metropolis (Figure 1). Model (3) is a “difference in difference” estimator, the “after” versus “before” policy—that is, I(truck)t equal to 1 versus 0—variation in localized NOx in the metropolis (site l ∈ metro) corrected for any common time-varying NOx determinants outside the metropolis (l ∉ metro) that are omitted. Supporting this assumption, Figure 4(b) shows deseasoned ambient NOx at these “control sites” and at the site in which “treatment” was most severe.25 As we show in a robustness test, our findings are qualitatively similar if we base our policy instrument only on variation within the metropolis instead. As we predict $${\widehat {\mathit {diesel}}_{lt}}$$ at the site by day level, we average these fitted values across days within each month; again $${\widehat {\mathit {diesel}}_{lt}}\ $$by monitoring site is assigned to zip code of residence based on proximity, not to exceed a distance threshold (5 km in the base case). Model (2) is then estimated by 2SLS, with $${\widehat {\mathit {diesel}}_{lt}}$$ instrumenting for diesel$${lt}$$. Through model (3) we restrict variation in our diesel exhaust measure in health equation (2) to that explained by the policy, flexibly by location (β1l). To emphasize, our empirical design, including the use of policy to instrument for spatially differentiated changes in pollution, follows other examples in the environmental economics literature (Chay and Greenstone 2005; Schlenker and Walker 2015; Isen et al. 2017).26 3.2. Effect of Truck Policy on Diesel Pollution, by Monitor Table 4 reports estimates of pollution model (3) across different specifications, to assess sensitivity. In column (1), 24-h NOx levels at a site only 20 m from the original truck route (site 8) were, on average, 59 ppb (s.e. 3 ppb) lower after the policy was implemented compared to before, relative to average changes over time at sites 50–100 km away (Figure 4(b)). The mean policy effect at four sites 2–6 km from the original truck route is estimated to be an order of magnitude lower, at −6 ppb (s.e. 1 ppb; Table 1 sample mean of 56 ppb). For brevity, Table 4 reports mean effects across sites grouped according to distance from the original truck route, but Figure 5(b) and Online Appendix Table B.4 show the individual effect estimated for each of the 11 sites in the metropolis, α2l, l∈metro. Table 4. The spatial effects of the truck abatement policy on localized diesel pollution, using ambient NOx as a marker.   (1)  (2)  (3)  (4)  (5)  (6)  Model specification  Base  Trend, not year-mo.  Site × meteor.  36 months  Daytime 7 a.m.–8 p.m.  Metro only  Dependent variable (ppb)  24 h NOx  24 h NOx  24 h NOx  24 h NOx  14 h NOx  24 h NOx  Policy implemented (yes = 1)  × siteFEs  One site 20 m from original truck route  −59.2***  −59.0***  −59.5***  −51.8***  −74.6***  −56.6***   (site 8 in Figure 1)  (3.1)  (3.1)  (2.6)  (3.7)  (3.0)  (3.2)  Sites 2–6 km from original truck route  −6.0***  −5.9***  −6.7***  −4.7***  −6.7***  −3.2***   (mean sites 5, 10, 27, 31 in Figure 1)  (1.3)  (1.3)  (1.1)  (1.5)  (1.1)  (1.1)  Sites 6–10 km from original truck route  −2.0  −2.1*  −2.3**  0.7  −3.4***  0.6   (mean sites 1, 7, 20 in Figure 1)  (1.2)  (1.2)  (1.1)  (1.5)  (1.0)  (0.9)  Sites 10–20 km from original truck route  −2.7**  −2.9**  −2.9***  0.3  −4.0***  –    (mean sites 17, 22, 29 in Figure 1)  (1.2)  (1.2)  (1.0)  (1.4)  (1.3)  –  Policy implemented (yes = 1)  6.1***  5.3***  7.4***  4.5**  6.8***  4.9*    (2.1)  (1.5)  (2.2)  (2.3)  (1.9)  (2.8)  Site fixed effects (FEs)  Yes  Yes  Yes  Yes  Yes  Yes  Week-of-year, day-of-week by site FEs  Yes  Yes  Yes  Yes  Yes  Yes  Meteorology and thermal inversion  Yes  Yes  Yes  Yes  Yes  Yes  Month-of-sample FEs  Yes  –  Yes  Yes  Yes  Yes  Quadratic time trend  –  Yes  –  –  –  –  Fuel prices (gasoline, ethanol, diesel)  –  Yes  –  –  –  –  Number of observations  22,168  22,168  22,168  14,674  22,244  16,106  Number of regressors  1,007  958  1,175  986  1,007  756  R2  0.791  0.786  0.847  0.803  0.805  0.767  Mean value of dep. var. (for site 8)  110.3  110.3  110.3  121.1  113.7  110.3    (1)  (2)  (3)  (4)  (5)  (6)  Model specification  Base  Trend, not year-mo.  Site × meteor.  36 months  Daytime 7 a.m.–8 p.m.  Metro only  Dependent variable (ppb)  24 h NOx  24 h NOx  24 h NOx  24 h NOx  14 h NOx  24 h NOx  Policy implemented (yes = 1)  × siteFEs  One site 20 m from original truck route  −59.2***  −59.0***  −59.5***  −51.8***  −74.6***  −56.6***   (site 8 in Figure 1)  (3.1)  (3.1)  (2.6)  (3.7)  (3.0)  (3.2)  Sites 2–6 km from original truck route  −6.0***  −5.9***  −6.7***  −4.7***  −6.7***  −3.2***   (mean sites 5, 10, 27, 31 in Figure 1)  (1.3)  (1.3)  (1.1)  (1.5)  (1.1)  (1.1)  Sites 6–10 km from original truck route  −2.0  −2.1*  −2.3**  0.7  −3.4***  0.6   (mean sites 1, 7, 20 in Figure 1)  (1.2)  (1.2)  (1.1)  (1.5)  (1.0)  (0.9)  Sites 10–20 km from original truck route  −2.7**  −2.9**  −2.9***  0.3  −4.0***  –    (mean sites 17, 22, 29 in Figure 1)  (1.2)  (1.2)  (1.0)  (1.4)  (1.3)  –  Policy implemented (yes = 1)  6.1***  5.3***  7.4***  4.5**  6.8***  4.9*    (2.1)  (1.5)  (2.2)  (2.3)  (1.9)  (2.8)  Site fixed effects (FEs)  Yes  Yes  Yes  Yes  Yes  Yes  Week-of-year, day-of-week by site FEs  Yes  Yes  Yes  Yes  Yes  Yes  Meteorology and thermal inversion  Yes  Yes  Yes  Yes  Yes  Yes  Month-of-sample FEs  Yes  –  Yes  Yes  Yes  Yes  Quadratic time trend  –  Yes  –  –  –  –  Fuel prices (gasoline, ethanol, diesel)  –  Yes  –  –  –  –  Number of observations  22,168  22,168  22,168  14,674  22,244  16,106  Number of regressors  1,007  958  1,175  986  1,007  756  R2  0.791  0.786  0.847  0.803  0.805  0.767  Mean value of dep. var. (for site 8)  110.3  110.3  110.3  121.1  113.7  110.3  Notes: An observation is a day by NOx monitoring site pair. The dependent variable is the mean NOx concentration (in ppb) over 24 h (all specifications but (5)) or between 7 a.m. and 8 p.m. (specification (5)). Specifications (1)–(5) are “difference in difference” models, with the second difference corresponding to sites in the metropolis versus sites 50–100 km away. (6) excludes sites 50–100 km away. We allow policy effects to vary by site but in the table show means across sites grouped according to distance from the original truck route (Appendix Table B.4 shows individual effects). The sample period is November 2008 to May 2013. We control for: site fixed effects; seasonality (week-of-year and day-of-week) by site fixed effects, month-of-sample fixed effects; and meteorology and thermal inversion. (2) replaces month-of-sample fixed effects by a quadratic time trend and fuel prices. (3) allows the effects of meteorology to vary by site. (4) restricts the sample to 36 months from January 2009 to December 2011. OLS regressions. Standard errors, in parentheses, are one-way clustered by week-of-sample. *Significant at 0.1; **significant at 0.05; ***significant at 0.01. View Large Table 4. The spatial effects of the truck abatement policy on localized diesel pollution, using ambient NOx as a marker.   (1)  (2)  (3)  (4)  (5)  (6)  Model specification  Base  Trend, not year-mo.  Site × meteor.  36 months  Daytime 7 a.m.–8 p.m.  Metro only  Dependent variable (ppb)  24 h NOx  24 h NOx  24 h NOx  24 h NOx  14 h NOx  24 h NOx  Policy implemented (yes = 1)  × siteFEs  One site 20 m from original truck route  −59.2***  −59.0***  −59.5***  −51.8***  −74.6***  −56.6***   (site 8 in Figure 1)  (3.1)  (3.1)  (2.6)  (3.7)  (3.0)  (3.2)  Sites 2–6 km from original truck route  −6.0***  −5.9***  −6.7***  −4.7***  −6.7***  −3.2***   (mean sites 5, 10, 27, 31 in Figure 1)  (1.3)  (1.3)  (1.1)  (1.5)  (1.1)  (1.1)  Sites 6–10 km from original truck route  −2.0  −2.1*  −2.3**  0.7  −3.4***  0.6   (mean sites 1, 7, 20 in Figure 1)  (1.2)  (1.2)  (1.1)  (1.5)  (1.0)  (0.9)  Sites 10–20 km from original truck route  −2.7**  −2.9**  −2.9***  0.3  −4.0***  –    (mean sites 17, 22, 29 in Figure 1)  (1.2)  (1.2)  (1.0)  (1.4)  (1.3)  –  Policy implemented (yes = 1)  6.1***  5.3***  7.4***  4.5**  6.8***  4.9*    (2.1)  (1.5)  (2.2)  (2.3)  (1.9)  (2.8)  Site fixed effects (FEs)  Yes  Yes  Yes  Yes  Yes  Yes  Week-of-year, day-of-week by site FEs  Yes  Yes  Yes  Yes  Yes  Yes  Meteorology and thermal inversion  Yes  Yes  Yes  Yes  Yes  Yes  Month-of-sample FEs  Yes  –  Yes  Yes  Yes  Yes  Quadratic time trend  –  Yes  –  –  –  –  Fuel prices (gasoline, ethanol, diesel)  –  Yes  –  –  –  –  Number of observations  22,168  22,168  22,168  14,674  22,244  16,106  Number of regressors  1,007  958  1,175  986  1,007  756  R2  0.791  0.786  0.847  0.803  0.805  0.767  Mean value of dep. var. (for site 8)  110.3  110.3  110.3  121.1  113.7  110.3    (1)  (2)  (3)  (4)  (5)  (6)  Model specification  Base  Trend, not year-mo.  Site × meteor.  36 months  Daytime 7 a.m.–8 p.m.  Metro only  Dependent variable (ppb)  24 h NOx  24 h NOx  24 h NOx  24 h NOx  14 h NOx  24 h NOx  Policy implemented (yes = 1)  × siteFEs  One site 20 m from original truck route  −59.2***  −59.0***  −59.5***  −51.8***  −74.6***  −56.6***   (site 8 in Figure 1)  (3.1)  (3.1)  (2.6)  (3.7)  (3.0)  (3.2)  Sites 2–6 km from original truck route  −6.0***  −5.9***  −6.7***  −4.7***  −6.7***  −3.2***   (mean sites 5, 10, 27, 31 in Figure 1)  (1.3)  (1.3)  (1.1)  (1.5)  (1.1)  (1.1)  Sites 6–10 km from original truck route  −2.0  −2.1*  −2.3**  0.7  −3.4***  0.6   (mean sites 1, 7, 20 in Figure 1)  (1.2)  (1.2)  (1.1)  (1.5)  (1.0)  (0.9)  Sites 10–20 km from original truck route  −2.7**  −2.9**  −2.9***  0.3  −4.0***  –    (mean sites 17, 22, 29 in Figure 1)  (1.2)  (1.2)  (1.0)  (1.4)  (1.3)  –  Policy implemented (yes = 1)  6.1***  5.3***  7.4***  4.5**  6.8***  4.9*    (2.1)  (1.5)  (2.2)  (2.3)  (1.9)  (2.8)  Site fixed effects (FEs)  Yes  Yes  Yes  Yes  Yes  Yes  Week-of-year, day-of-week by site FEs  Yes  Yes  Yes  Yes  Yes  Yes  Meteorology and thermal inversion  Yes  Yes  Yes  Yes  Yes  Yes  Month-of-sample FEs  Yes  –  Yes  Yes  Yes  Yes  Quadratic time trend  –  Yes  –  –  –  –  Fuel prices (gasoline, ethanol, diesel)  –  Yes  –  –  –  –  Number of observations  22,168  22,168  22,168  14,674  22,244  16,106  Number of regressors  1,007  958  1,175  986  1,007  756  R2  0.791  0.786  0.847  0.803  0.805  0.767  Mean value of dep. var. (for site 8)  110.3  110.3  110.3  121.1  113.7  110.3  Notes: An observation is a day by NOx monitoring site pair. The dependent variable is the mean NOx concentration (in ppb) over 24 h (all specifications but (5)) or between 7 a.m. and 8 p.m. (specification (5)). Specifications (1)–(5) are “difference in difference” models, with the second difference corresponding to sites in the metropolis versus sites 50–100 km away. (6) excludes sites 50–100 km away. We allow policy effects to vary by site but in the table show means across sites grouped according to distance from the original truck route (Appendix Table B.4 shows individual effects). The sample period is November 2008 to May 2013. We control for: site fixed effects; seasonality (week-of-year and day-of-week) by site fixed effects, month-of-sample fixed effects; and meteorology and thermal inversion. (2) replaces month-of-sample fixed effects by a quadratic time trend and fuel prices. (3) allows the effects of meteorology to vary by site. (4) restricts the sample to 36 months from January 2009 to December 2011. OLS regressions. Standard errors, in parentheses, are one-way clustered by week-of-sample. *Significant at 0.1; **significant at 0.05; ***significant at 0.01. View Large Individual effects at NOx monitors are quite informative. The three largest reductions occurred at precisely the three sites with the highest levels of ambient NOx from surrounding traffic: besides site 8 (down 59 s.e. 3 ppb), these are sites 10 and 17, with estimated reductions of 17 (s.e. 2) and 12 (s.e. 2) ppb, respectively.27 Site 17 is located en route to a heavily congested exit to a key highway (Castello Branco), on the inside of the beltway, which trucks bound from and to Santos port were able to bypass after April 2010 (Figure 1), explaining the 12 ppb NOx abatement associated with the policy. For perspective, such reductions, with point estimates of 59, 17, and 12 ppb, are large compared to pre-policy average NOx levels of 146, 75, and 106 ppb at sites 8, 10, and 17, respectively (proportionate changes of −40%, −23%, and −11%). Further, we average the immediate change in congestion frequency—road segment level estimates shown in Figure 5(a)—over all road segments within 1 km of each NOx monitor, and compare this with the estimated change in NOx. The (transient) reduction in congestion brought about by the policy is strongly associated with the sustained change in diesel pollution, with a correlation coefficient of 0.91.28 Estimated policy effects on localized diesel pollution are robust across specifications. In column (2), we replace month-of-sample fixed effects by a less-flexible quadratic trend and now control for fuel prices (month-of-sample previously accounted for fuel prices). In column (3), we allow the effects of meteorology to vary by site—notice that precision grows. In column (4), we shorten the sample to three years (2009–2011)—notice that estimated magnitudes shrink somewhat, as does precision. In column (5), dependent and independent variables are means over daytime hours, 7 a.m. to 8 p.m., when traffic flows are higher, rather than 24-h (0 a.m. to 11 p.m.) means—estimated policy effects grow. We also confirm that our findings are robust to alternatively specifying the daily maximum NOx level, with a sample mean of 227 ppb compared to 110 ppb for the daily mean (not reported in Table 4 for brevity, but see Online Appendix Table B.3, column (2). Another robustness test we perform is to control for both citywide road congestion and congestion within a 2 km radius of a site, interacting such controls with site fixed effects for flexibility (again not shown). Table 4 further indicates that ambient NOx at sites 50–100 km from the São Paulo metropolis were about 6 ppb higher after the policy implementation date compared to before (see the row labeled “Policy implemented”, showing α1 estimates across alternative specifications). In column (6), we exclude sites outside the metropolis from the estimation sample and instead use sites 10–20 km from the original truck route as controls. 3.3. Effects of Diesel Pollution on Morbidity and Mortality Table 5 reports the mean estimated coefficient of interest in health equation (2), β1l, across the 25 zip codes for which NOx is monitored at a distance of 5 km or less.29 By age group and health outcome, we report both: (1) 2SLS estimates—where measured NOx (diesel$${lt}$$) is instrumented with the component of NOx variation that is explained by the policy change ($${\widehat {\mathit {diesel}}_{lt}}$$, fitted per specification 1, Table 4); and (ii) OLS estimates—where diesel$${lt}$$ serves as its own instrument. Table 5. Effects of diesel pollution on hospital admission and death rates (structural analysis of health). Dependent variable: hospitalization rate per 100,000 residents per month  All ages  0–4 years  5–19 years  20–64 years  65+ years  Cardiovascular admissions  Diesel exhaust (in 10 ppb NOx,  4.2***  –  –  2.7***  20.8***  mean over zip)  (0.7)      (0.7)  (5.9)    … compare 2SLS estimate to  0.6*  –  –  0.3  3.1  OLS estimate:  (0.4)      (0.4)  (3.0)  Mean of dependent variable  53.1  –  –  42.9  275.6  Respiratory admissions  Diesel exhaust (in 10 ppb NOx,  3.2***  22.9***  3.7***  0.9*  11.5***  mean over zip)  (0.7)  (6.4)  (1.1)  (0.5)  (3.9)    … compare 2SLS estimate to  1.0***  9.5***  0.6  0.5**  0.7  OLS estimate:  (0.4)  (3.4)  (0.6)  (0.2)  (2.0)  Mean of dependent variable  43.2  251.5  29.9  18.3  120.5  Cardiovascular/respiratory death  Diesel exhaust (in 10 ppb NOx,  1.0***  –  –  0.5*  6.7**  mean over zip)  (0.3)      (0.3)  (3.0)    … compare 2SLS estimate to  0.2  –  –  0.0  1.2  OLS estimate:  (0.2)      (0.1)  (1.7)  Mean of dependent variable  10.2  –  –  5.4  74.9  Placebo: Trauma admissions (excl. traffic)  Diesel exhaust (in 10 ppb NOx,  0.1  –  –  0.3  0.8  mean over zip)  (0.2)      (0.3)  (1.1)    … compare 2SLS estimate to  0.0  –  –  0.1  0.5  OLS estimate:  (0.1)  –  –  (0.1)  (0.5)  Mean of dependent variable  6.9  –  –  6.7  14.0  Number of observations  1,343  1,343  1,343  1,343  1,343  Number of zip codes (with nearby NOx monitor)  25  25  25  25  25  Dependent variable: hospitalization rate per 100,000 residents per month  All ages  0–4 years  5–19 years  20–64 years  65+ years  Cardiovascular admissions  Diesel exhaust (in 10 ppb NOx,  4.2***  –  –  2.7***  20.8***  mean over zip)  (0.7)      (0.7)  (5.9)    … compare 2SLS estimate to  0.6*  –  –  0.3  3.1  OLS estimate:  (0.4)      (0.4)  (3.0)  Mean of dependent variable  53.1  –  –  42.9  275.6  Respiratory admissions  Diesel exhaust (in 10 ppb NOx,  3.2***  22.9***  3.7***  0.9*  11.5***  mean over zip)  (0.7)  (6.4)  (1.1)  (0.5)  (3.9)    … compare 2SLS estimate to  1.0***  9.5***  0.6  0.5**  0.7  OLS estimate:  (0.4)  (3.4)  (0.6)  (0.2)  (2.0)  Mean of dependent variable  43.2  251.5  29.9  18.3  120.5  Cardiovascular/respiratory death  Diesel exhaust (in 10 ppb NOx,  1.0***  –  –  0.5*  6.7**  mean over zip)  (0.3)      (0.3)  (3.0)    … compare 2SLS estimate to  0.2  –  –  0.0  1.2  OLS estimate:  (0.2)      (0.1)  (1.7)  Mean of dependent variable  10.2  –  –  5.4  74.9  Placebo: Trauma admissions (excl. traffic)  Diesel exhaust (in 10 ppb NOx,  0.1  –  –  0.3  0.8  mean over zip)  (0.2)      (0.3)  (1.1)    … compare 2SLS estimate to  0.0  –  –  0.1  0.5  OLS estimate:  (0.1)  –  –  (0.1)  (0.5)  Mean of dependent variable  6.9  –  –  6.7  14.0  Number of observations  1,343  1,343  1,343  1,343  1,343  Number of zip codes (with nearby NOx monitor)  25  25  25  25  25  Notes: An observation is a 3-digit residential zip code by month pair. The dependent variable is the hospitalization rate for a given disease-age group pair (except pairs with very low incidence, for example, cardiovascular among children aged 0–4 years). For each of 14 disease-age pairs, we estimate health equation (2) separately by 2SLS regression and by OLS regression; we thus report mean β1l, averaged across all zip codes l, for each of 14 × 2 = 28 regressions. 2SLS estimates use fitted NOx per the specification reported in Table 4, column (1) as an instrument for measured NOx. We assign each district within a zip code (there is a ratio of 103 districts to 43 zip codes) to the nearest NOx monitor from the district centroid; we then average NOx across districts within each zip code. Across the multiple first-stage equations in the 2SLS implementation, the first-stage F-statistic for the excluded instruments averages 93 (range 6–258; endogenous regressors are measured NOx interacted with zip code fixed effects and excluded instruments are fitted NOx interacted with zip code fixed effects). The sample period is November 2008 to May 2013. We control for: zip code fixed effects; seasonality (month-of-year) by zip code fixed effects; and month-of-sample fixed effects (these subsume meteorology and the May to September 2009 swine flu epidemic). Robust standard errors are reported in parentheses. *Significant at 0.1; **significant at 0.05; ***significant at 0.01. View Large Table 5. Effects of diesel pollution on hospital admission and death rates (structural analysis of health). Dependent variable: hospitalization rate per 100,000 residents per month  All ages  0–4 years  5–19 years  20–64 years  65+ years  Cardiovascular admissions  Diesel exhaust (in 10 ppb NOx,  4.2***  –  –  2.7***  20.8***  mean over zip)  (0.7)      (0.7)  (5.9)    … compare 2SLS estimate to  0.6*  –  –  0.3  3.1  OLS estimate:  (0.4)      (0.4)  (3.0)  Mean of dependent variable  53.1  –  –  42.9  275.6  Respiratory admissions  Diesel exhaust (in 10 ppb NOx,  3.2***  22.9***  3.7***  0.9*  11.5***  mean over zip)  (0.7)  (6.4)  (1.1)  (0.5)  (3.9)    … compare 2SLS estimate to  1.0***  9.5***  0.6  0.5**  0.7  OLS estimate:  (0.4)  (3.4)  (0.6)  (0.2)  (2.0)  Mean of dependent variable  43.2  251.5  29.9  18.3  120.5  Cardiovascular/respiratory death  Diesel exhaust (in 10 ppb NOx,  1.0***  –  –  0.5*  6.7**  mean over zip)  (0.3)      (0.3)  (3.0)    … compare 2SLS estimate to  0.2  –  –  0.0  1.2  OLS estimate:  (0.2)      (0.1)  (1.7)  Mean of dependent variable  10.2  –  –  5.4  74.9  Placebo: Trauma admissions (excl. traffic)  Diesel exhaust (in 10 ppb NOx,  0.1  –  –  0.3  0.8  mean over zip)  (0.2)      (0.3)  (1.1)    … compare 2SLS estimate to  0.0  –  –  0.1  0.5  OLS estimate:  (0.1)  –  –  (0.1)  (0.5)  Mean of dependent variable  6.9  –  –  6.7  14.0  Number of observations  1,343  1,343  1,343  1,343  1,343  Number of zip codes (with nearby NOx monitor)  25  25  25  25  25  Dependent variable: hospitalization rate per 100,000 residents per month  All ages  0–4 years  5–19 years  20–64 years  65+ years  Cardiovascular admissions  Diesel exhaust (in 10 ppb NOx,  4.2***  –  –  2.7***  20.8***  mean over zip)  (0.7)      (0.7)  (5.9)    … compare 2SLS estimate to  0.6*  –  –  0.3  3.1  OLS estimate:  (0.4)      (0.4)  (3.0)  Mean of dependent variable  53.1  –  –  42.9  275.6  Respiratory admissions  Diesel exhaust (in 10 ppb NOx,  3.2***  22.9***  3.7***  0.9*  11.5***  mean over zip)  (0.7)  (6.4)  (1.1)  (0.5)  (3.9)    … compare 2SLS estimate to  1.0***  9.5***  0.6  0.5**  0.7  OLS estimate:  (0.4)  (3.4)  (0.6)  (0.2)  (2.0)  Mean of dependent variable  43.2  251.5  29.9  18.3  120.5  Cardiovascular/respiratory death  Diesel exhaust (in 10 ppb NOx,  1.0***  –  –  0.5*  6.7**  mean over zip)  (0.3)      (0.3)  (3.0)    … compare 2SLS estimate to  0.2  –  –  0.0  1.2  OLS estimate:  (0.2)      (0.1)  (1.7)  Mean of dependent variable  10.2  –  –  5.4  74.9  Placebo: Trauma admissions (excl. traffic)  Diesel exhaust (in 10 ppb NOx,  0.1  –  –  0.3  0.8  mean over zip)  (0.2)      (0.3)  (1.1)    … compare 2SLS estimate to  0.0  –  –  0.1  0.5  OLS estimate:  (0.1)  –  –  (0.1)  (0.5)  Mean of dependent variable  6.9  –  –  6.7  14.0  Number of observations  1,343  1,343  1,343  1,343  1,343  Number of zip codes (with nearby NOx monitor)  25  25  25  25  25  Notes: An observation is a 3-digit residential zip code by month pair. The dependent variable is the hospitalization rate for a given disease-age group pair (except pairs with very low incidence, for example, cardiovascular among children aged 0–4 years). For each of 14 disease-age pairs, we estimate health equation (2) separately by 2SLS regression and by OLS regression; we thus report mean β1l, averaged across all zip codes l, for each of 14 × 2 = 28 regressions. 2SLS estimates use fitted NOx per the specification reported in Table 4, column (1) as an instrument for measured NOx. We assign each district within a zip code (there is a ratio of 103 districts to 43 zip codes) to the nearest NOx monitor from the district centroid; we then average NOx across districts within each zip code. Across the multiple first-stage equations in the 2SLS implementation, the first-stage F-statistic for the excluded instruments averages 93 (range 6–258; endogenous regressors are measured NOx interacted with zip code fixed effects and excluded instruments are fitted NOx interacted with zip code fixed effects). The sample period is November 2008 to May 2013. We control for: zip code fixed effects; seasonality (month-of-year) by zip code fixed effects; and month-of-sample fixed effects (these subsume meteorology and the May to September 2009 swine flu epidemic). Robust standard errors are reported in parentheses. *Significant at 0.1; **significant at 0.05; ***significant at 0.01. View Large Coefficients on diesel exposure, by zip code, estimated by 2SLS tend to be 2–4 times those estimated by OLS. 2SLS estimates indicate that a 10 ppb abatement in the diesel exhaust marker generates monthly morbidity benefits per one million residents of: (i) 42 fewer cardiovascular admissions (s.e. 7; sample mean of 531); and (ii) 32 fewer respiratory admissions (s.e. 7; sample mean of 432). Conditional on specific age group-disease pairs, monthly admissions rates fall by as much as 208 (age 65+, cardiovascular, s.e. 59) and 229 (age 0–4, respiratory, s.e. 64) per one million residents in the age group. Estimated benefits of diesel control on the lower respiratory disease subcategory are also significant (e.g., pneumonia, not reported). We also find a significant effect on mortality: a 10 ppb drop in NOx equivalent units of diesel pollution causes 10 fewer monthly in-hospital deaths per one million residents across age groups and cardiorespiratory conditions (s.e. 3, sample mean of 102). Moreover, estimated impacts of diesel exposure on the placebo—nontraffic related trauma—are not significantly different from zero. Evaluated at the 2SLS point estimate reported in Table 5, the implied elasticity for all-age cardiorespiratory deaths with respect to diesel pollution is (+10/102)$$/$$(10/56) ≈ 0.55, evaluated at the sample means reported in Tables 5 and 1 for the dependent variable and key regressor, respectively. Similarly, the elasticity among persons aged 65+ years is (+67/749)$$/$$(10/56) ≈ 0.50. These elasticities are three times those implied by a recent study in the United States of all-cause mortality among Medicare seniors and different subgroups30 from long-term (i.e., annual average) exposure to PM2.5 (Di et al. 2017). Mean PM2.5 in the US sample (11 μg/m3) is about half the level in São Paulo (20 μg/m3). In contrast, Chen et al. (2013) obtain higher elasticities of ∼0.9 for cardiorespiratory mortality responses to coal combustion due to a sustained winter heating policy in China.31 Differences in the chemical and physical composition of emissions likely explain at least in part the different findings across studies; Lelieveld et al. (2015), for example, assumes “carbonaceous PM2.5 is five times more toxic than inorganic particles” (p. 367). Examining traffic-induced pollution variation—overall vehicle exhaust, not specific to diesel or gasoline, and week to week—find “a 1 unit decrease in PM10 saves roughly 10 lives per 100,000 live births, an elasticity of approximately 1” (p. 350). Currie and Walker (2011) find that a localized intervention that reportedly lowered road congestion and (overall) vehicle emissions reduced the incidence of low weight and prematurity at birth by about 10%.32 A rare study examining an (assumed yet credible) shift in diesel exposure from a school bus retrofit program in Washington state finds “that adopter districts experienced 23% fewer children's bronchitis and asthma cases… relative to a control group… (and) 37% fewer children's pneumonia cases” (Beatty and Shimshack 2011, p. 997). OLS estimates, although smaller in magnitude, also indicate statistically significant public health benefits from abating diesel trucks. In what follows, we provide sensitivity analysis showing that the preferred 2SLS estimates are robust to alternative modeling choices. These include the assignment of monitor-level diesel pollution to patients’ residential zip code, constructing the instrumental variable according to alternative Table 4 specifications such as dropping NOx monitors 50–100 km away, and adding local traffic volume controls in the health equation. A key assumption in our analysis is that shifts in local demographic composition in response to the policy—say wealthier, more educated and healthier households moving to zip codes near the key inner-city corridor—were not significant within our relatively short study period. We tried to obtain district-level demographic data to test the assumption, but the available data are collected once every ten years (the last Census occurring the year the beltway opened). Our relatively short sample of 4.5 years, including the before and after intervention periods, suggests that sorting is unlikely to be confounding our estimates. The robustness test reported in what follows that shortens the sample period even further, to three years, is consistent with this. In two recent studies that used a design based on spatially differentiated pollution abatement over time, sample periods of similar length were specified, of five and six years. The studies did not detect short-run demographic shifts or sorting in response to localized abatement policy—namely the introduction of electronic collection at US highway toll plazas, and the closure of a refinery within Mexico City (Currie and Walker 2011; Hanna and Oliva 2015).33 Moreover, it is plausible that the elderly are less likely to shift residence in response to the policy,34 as well as less likely to commute35 and suffer from traffic congestion other than through the pollution it produces (Peters et al. 2004). Thus, the significant estimates we obtain for the elderly subpopulation provide further reassurance that we are picking up a direct health response from exposure to air contaminants. As further evidence against the sorting hypothesis, a robustness test we report includes nontraffic related trauma admissions as a control in the cardiorespiratory health equations. Presumably, were they to be moving near to the inner-city corridor in response to the policy, wealthier and healthier households would require less medical attention on account not only of cardiorespiratory symptoms but also symptoms that are less likely triggered by pollution, such as trauma. As an alternative, we include the three-quarters of all admissions that are not cardiorespiratory or trauma related as controls in the cardiorespiratory equations: wealthier and healthier households would require less medical attention across the disease range. This test is likely to err on the side of caution, as other symptoms caused by pollution, or cardiorespiratory symptoms that were misdiagnosed or misclassified, would be included in this control. Our statistically significant estimates of the effect of diesel pollution on hospital admission and death rates survive. We show estimates for these and other tests next. 3.4. Robustness Tests Table 6 reports estimates for alternative specifications of the health model (2). Across all specifications, estimated effects of diesel pollution on hospitalizations and in-hospital death rates are of similar magnitude to those in Table 5. Table 6. Effects of diesel pollution on hospital admission and death rates: robustness tests of the structural analysis. Dependent variable: hospitalization rate per 100,000 residents per month  Number of observations  Number of zip codes  All ages  0–4 years  5–19 years  20–64 years  65+ years  Cardiovascular admissions  Preferred specification (Table 5, 2SLS)  1,343  25  4.2*** (0.7)  –  –  2.7*** (0.7)  20.8*** (5.9)   Closest monitor within 10 km  1,830  34  4.3*** (0.8)  –  –  3.1*** (0.8)  21.6*** (6.2)   Mean of monitors within 5 km  1,343  25  3.7*** (0.7)  –  –  2.4*** (0.7)  18.5*** (5.4)   Mean of monitors within 10 km  1,839  34  3.9*** (0.7)  –  –  3.2*** (0.7)  16.9*** (5.6)   Site-specific meteorology effects (IV)  1,343  25  2.7*** (0.6)  –  –  1.7*** (0.7)  13.4** (5.4)   Include policy–wind interactions (IV)  1,343  25  4.2*** (0.8)  –  –  2.7*** (0.7)  21.1*** (6.0)   Incl. site-specific policy–wind interact. (IV)  1,343  25  2.8*** (0.6)  –  –  1.7** (0.7)  15.2*** (5.3)   Drop control sites 50–100 km away (IV)  1,343  25  4.8*** (1.0)  –  –  3.3*** (1.0)  17.9** (8.1)   Policy implemented May 31, 2010 (IV)  1,343  25  4.2*** (0.7)  –  –  3.1*** (0.7)  19.0*** (5.7)   Shorter period of 36 months  879  25  2.1** (0.9)  –  –  1.0 (0.9)  11.9* (7.0)   Local road congestion as health control  1,343  25  4.1*** (0.7)  –  –  2.8*** (0.7)  21.4*** (5.6)   Repeat w/ IV based on policy and congestion  1,343  25  3.7*** (0.7)  –  –  2.5*** (0.7)  18.8*** (5.2)   Trauma admissions as health control  1,343  25  4.1*** (0.8)  –  –  2.5*** (0.7)  21.0*** (5.9)   All other admissions as health control  1,343  25  2.7*** (0.7)  –  –  1.7** (0.7)  17.8*** (6.0)   Distance from patient to hospital as control  1,343  25  4.0*** (0.7)  –  –  2.5*** (0.7)  20.5*** (5.9)   Dist. from patient to hospital × zip code FE  1,343  25  3.6*** (0.8)  –  –  2.6*** (0.8)  18.1*** (5.8)   Quadratic trend  1,343  25  3.8*** (0.8)  –  –  2.6*** (0.8)  15.5*** (5.7)   Quadratic trend plus fuel prices  1,343  25  4.6*** (0.8)  –  –  3.0*** (0.8)  23.7*** (5.7)   Weighted 2SLS regression (sq.rt.pop.)  1,343  25  3.3*** (0.7)  –  –  2.0*** (0.7)  16.4*** (5.5)   Population-weighted average for β1l  1,343  25  3.3*** (0.8)  –  –  1.8** (0.8)  19.3*** (5.1)  Respiratory admissions  Preferred specification (Table 5, 2SLS)  1,343  25  3.2*** (0.7)  22.9*** (6.4)  3.7*** (1.1)  0.9* (0.5)  11.5*** (3.9)   Closest monitor within 10 km  1,830  34  3.0*** (0.7)  22.8*** (6.7)  3.3*** (1.0)  0.8* (0.5)  14.7*** (3.9)   Mean of monitors within 5 km  1,343  25  2.4*** (0.6)  19.7*** (5.8)  3.3*** (1.0)  0.5 (0.4)  8.6** (3.6)   Mean of monitors within 10 km  1,839  34  1.4** (0.7)  9.9* (5.9)  1.4 (0.9)  0.6 (0.4)  11.6*** (3.6)   Site-specific meteorology effects (IV)  1,343  25  2.5*** (0.6)  25.2*** (5.8)  1.6* (1.0)  0.7* (0.4)  3.1 (3.6)   Include policy–wind interactions (IV)  1,343  25  3.2*** (0.7)  23.7*** (6.5)  3.9*** (1.1)  0.9* (0.5)  11.8*** (3.9)   Incl. site-specific policy–wind interact. (IV)  1,343  25  2.6*** (0.6)  23.3*** (5.6)  1.6* (1.0)  0.8** (0.4)  4.1 (3.6)   Drop control sites 50–100 km away (IV)  1,343  25  4.2*** (1.0)  28.3*** (9.3)  6.3*** (1.5)  1.3** (0.6)  11.2** (4.8)   Policy implemented May 31, 2010 (IV)  1,343  25  3.3*** (0.7)  27.1*** (6.8)  4.0*** (1.0)  0.9** (0.5)  12.1*** (3.8)   Shorter period of 36 months  879  25  2.4*** (0.8)  24.0*** (8.1)  0.8 (1.2)  0.3 (0.5)  8.7* (4.9)   Local road congestion as health control  1,343  25  3.4*** (0.7)  24.2*** (6.2)  3.6*** (1.0)  1.1** (0.5)  12.6*** (3.7)   Repeat w/ IV based on policy and congestion  1,343  25  3.3*** (0.6)  23.3*** (5.6)  3.3*** (1.0)  1.2*** (0.4)  11.3*** (3.4)   Trauma admissions as health control  1,343  25  3.0*** (0.7)  22.6*** (6.5)  3.8*** (1.1)  0.8* (0.5)  11.5*** (3.9)   All other admissions as health control  1,343  25  1.9*** (0.7)  13.7** (6.3)  2.0* (1.1)  0.5 (0.5)  10.7*** (3.9)   Distance from patient to hospital as control  1,343  25  2.9*** (0.7)  22.4*** (6.3)  3.5*** (1.1)  0.8* (0.5)  10.8*** (3.7)   Dist. from patient to hospital × zip code FE  1,343  25  2.7*** (0.7)  20.9*** (6.1)  3.0*** (1.1)  0.8 (0.5)  11.4*** (3.8)   Quadratic trend  1,343  25  2.9*** (0.8)  24.2*** (7.3)  1.7 (1.1)  0.8 (0.5)  12.4*** (4.1)   Quadratic trend plus fuel prices  1,343  25  3.8*** (0.7)  30.9*** (6.9)  3.5*** (1.1)  1.1** (0.5)  11.7*** (3.7)   Weighted 2SLS regression (sq.rt.pop.)  1,343  25  2.5*** (0.6)  18.5*** (5.8)  2.5*** (0.9)  0.7* (0.4)  10.1*** (3.6)   Population-weighted average for β1l  1,343  25  2.9*** (0.7)  20.6*** (7.3)  3.4*** (1.2)  1.0** (0.5)  9.9*** (3.4)  Cardiovascular/respiratory death  Preferred specification (Table 5, 2SLS)  1,343  25  1.0*** (0.3)  –  –  0.5* (0.3)  6.7** (3.0)   Closest monitor within 10 km  1,830  34  1.3*** (0.3)  –  –  0.8*** (0.3)  8.5*** (2.9)   Mean of monitors within 5 km  1,343  25  0.8*** (0.3)  –  –  0.4* (0.3)  5.4* (2.8)   Mean of monitors within 10 km  1,839  34  0.7*** (0.2)  –  –  0.3 (0.2)  5.4** (2.6)   Site-specific meteorology effects (IV)  1,343  25  0.7*** (0.3)  –  –  0.4* (0.2)  3.6 (2.9)   Include policy–wind interactions (IV)  1,343  25  1.0*** (0.3)  –  –  0.5* (0.3)  7.0** (3.1)   Incl. site-specific policy–wind interact. (IV)  1,343  25  0.8*** (0.3)  –  –  0.4 (0.2)  4.6 (2.8)   Drop control sites 50–100 km away (IV)  1,343  25  1.1** (0.4)  –  –  0.6 (0.4)  7.4* (3.9)   Policy implemented May 31, 2010 (IV)  1,343  25  0.7** (0.3)  –  –  0.3 (0.3)  5.1* (2.9)   Shorter period of 36 months  879  25  0.8** (0.4)  –  –  0.5 (0.3)  6.0* (3.5)   Local road congestion as health control  1,343  25  1.0*** (0.3)  –  –  0.4* (0.3)  8.2*** (2.9)   Repeat w/ IV based on policy and congestion  1,343  25  0.9*** (0.3)  –  –  0.4 (0.2)  7.4*** (2.7)   Trauma admissions as health control  1,343  25  0.9*** (0.3)  –  –  0.5* (0.3)  6.6** (3.0)   All other admissions as health control  1,343  25  0.7** (0.3)  –  –  0.4 (0.3)  6.0** (3.0)   Distance from patient to hospital as control  1,343  25  0.9*** (0.3)  –  –  0.5* (0.3)  6.4** (3.0)   Dist. from patient to hospital × zip code FE  1,343  25  0.8** (0.3)  –  –  0.4 (0.3)  5.0 (3.1)   Quadratic trend  1,343  25  1.0*** (0.3)  –  –  0.6** (0.3)  6.6** (3.1)   Quadratic trend plus fuel prices  1,343  25  1.1*** (0.3)  –  –  0.7*** (0.3)  7.6** (3.0)   Weighted 2SLS regression (sq.rt.pop.)  1,343  25  0.6** (0.3)  –  –  0.2 (0.2)  3.9 (2.7)   Population-weighted average for β1l  1,343  25  0.8** (0.3)  –  –  0.4 (0.3)  6.3** (2.7)  Placebo: Trauma admissions (excl. traffic)  Preferred specification (Table 5, 2SLS)  1,343  25  0.1 (0.2)  –  –  0.3 (0.3)  0.8 (1.1)   Closest monitor within 10 km  1,830  34  −0.1 (0.2)  –  –  0.1 (0.3)  0.2 (1.1)   Mean of monitors within 5 km  1,343  25  0.0 (0.2)  –  –  0.2 (0.3)  1.1 (1.0)   Mean of monitors within 10 km  1,839  34  −0.2 (0.2)  –  –  0.0 (0.2)  −0.5 (0.9)   Site-specific meteorology effects (IV)  1,343  25  −0.2 (0.2)  –  –  0.2 (0.3)  −0.5 (1.0)   Include policy–wind interactions (IV)  1,343  25  0.1 (0.2)  –  –  0.3 (0.3)  0.9 (1.1)   Incl. site-specific policy–wind interact. (IV)  1,343  25  −0.1 (0.2)  –  –  0.3 (0.3)  −0.7 (1.0)   Drop control sites 50–100 km away (IV)  1,343  25  0.2 (0.3)  –  –  0.2 (0.4)  1.1 (1.4)   Policy implemented May 31, 2010 (IV)  1,343  25  0.2 (0.2)  –  –  0.3 (0.3)  0.9 (1.1)   Shorter period of 36 months  879  25  0.0 (0.3)  –  –  −0.1 (0.4)  1.6 (1.3)   Local road congestion as health control  1,343  25  0.1 (0.2)  –  –  0.2 (0.3)  0.6 (1.0)   Repeat w/ IV based on policy and congestion  1,343  25  0.0 (0.2)  –  –  0.1 (0.3)  0.5 (1.0)   Distance from patient to hospital as control  1,343  25  0.1 (0.2)  –  –  0.2 (0.3)  0.9 (1.1)   Dist. from patient to hospital × zip code FE  1,343  25  −0.2 (0.3)  –  –  0.1 (0.3)  0.7 (1.1)   Quadratic trend  1,343  25  0.4 (0.2)  –  –  0.5* (0.3)  1.0 (1.1)   Quadratic trend plus fuel prices  1,343  25  0.2 (0.2)  –  –  0.3 (0.3)  1.4 (1.1)   Weighted 2SLS regression (sq.rt.pop.)  1,343  25  0.2 (0.2)  –  –  0.3 (0.3)  0.9 (0.9)   Population-weighted average for β1l  1,343  25  0.1 (0.2)  –  –  0.4 (0.3)  0.6 (1.0)  Dependent variable: hospitalization rate per 100,000 residents per month  Number of observations  Number of zip codes  All ages  0–4 years  5–19 years  20–64 years  65+ years  Cardiovascular admissions  Preferred specification (Table 5, 2SLS)  1,343  25  4.2*** (0.7)  –  –  2.7*** (0.7)  20.8*** (5.9)   Closest monitor within 10 km  1,830  34  4.3*** (0.8)  –  –  3.1*** (0.8)  21.6*** (6.2)   Mean of monitors within 5 km  1,343  25  3.7*** (0.7)  –  –  2.4*** (0.7)  18.5*** (5.4)   Mean of monitors within 10 km  1,839  34  3.9*** (0.7)  –  –  3.2*** (0.7)  16.9*** (5.6)   Site-specific meteorology effects (IV)  1,343  25  2.7*** (0.6)  –  –  1.7*** (0.7)  13.4** (5.4)   Include policy–wind interactions (IV)  1,343  25  4.2*** (0.8)  –  –  2.7*** (0.7)  21.1*** (6.0)   Incl. site-specific policy–wind interact. (IV)  1,343  25  2.8*** (0.6)  –  –  1.7** (0.7)  15.2*** (5.3)   Drop control sites 50–100 km away (IV)  1,343  25  4.8*** (1.0)  –  –  3.3*** (1.0)  17.9** (8.1)   Policy implemented May 31, 2010 (IV)  1,343  25  4.2*** (0.7)  –  –  3.1*** (0.7)  19.0*** (5.7)   Shorter period of 36 months  879  25  2.1** (0.9)  –  –  1.0 (0.9)  11.9* (7.0)   Local road congestion as health control  1,343  25  4.1*** (0.7)  –  –  2.8*** (0.7)  21.4*** (5.6)   Repeat w/ IV based on policy and congestion  1,343  25  3.7*** (0.7)  –  –  2.5*** (0.7)  18.8*** (5.2)   Trauma admissions as health control  1,343  25  4.1*** (0.8)  –  –  2.5*** (0.7)  21.0*** (5.9)   All other admissions as health control  1,343  25  2.7*** (0.7)  –  –  1.7** (0.7)  17.8*** (6.0)   Distance from patient to hospital as control  1,343  25  4.0*** (0.7)  –  –  2.5*** (0.7)  20.5*** (5.9)   Dist. from patient to hospital × zip code FE  1,343  25  3.6*** (0.8)  –  –  2.6*** (0.8)  18.1*** (5.8)   Quadratic trend  1,343  25  3.8*** (0.8)  –  –  2.6*** (0.8)  15.5*** (5.7)   Quadratic trend plus fuel prices  1,343  25  4.6*** (0.8)  –  –  3.0*** (0.8)  23.7*** (5.7)   Weighted 2SLS regression (sq.rt.pop.)  1,343  25  3.3*** (0.7)  –  –  2.0*** (0.7)  16.4*** (5.5)   Population-weighted average for β1l  1,343  25  3.3*** (0.8)  –  –  1.8** (0.8)  19.3*** (5.1)  Respiratory admissions  Preferred specification (Table 5, 2SLS)  1,343  25  3.2*** (0.7)  22.9*** (6.4)  3.7*** (1.1)  0.9* (0.5)  11.5*** (3.9)   Closest monitor within 10 km  1,830  34  3.0*** (0.7)  22.8*** (6.7)  3.3*** (1.0)  0.8* (0.5)  14.7*** (3.9)   Mean of monitors within 5 km  1,343  25  2.4*** (0.6)  19.7*** (5.8)  3.3*** (1.0)  0.5 (0.4)  8.6** (3.6)   Mean of monitors within 10 km  1,839  34  1.4** (0.7)  9.9* (5.9)  1.4 (0.9)  0.6 (0.4)  11.6*** (3.6)   Site-specific meteorology effects (IV)  1,343  25  2.5*** (0.6)  25.2*** (5.8)  1.6* (1.0)  0.7* (0.4)  3.1 (3.6)   Include policy–wind interactions (IV)  1,343  25  3.2*** (0.7)  23.7*** (6.5)  3.9*** (1.1)  0.9* (0.5)  11.8*** (3.9)   Incl. site-specific policy–wind interact. (IV)  1,343  25  2.6*** (0.6)  23.3*** (5.6)  1.6* (1.0)  0.8** (0.4)  4.1 (3.6)   Drop control sites 50–100 km away (IV)  1,343  25  4.2*** (1.0)  28.3*** (9.3)  6.3*** (1.5)  1.3** (0.6)  11.2** (4.8)   Policy implemented May 31, 2010 (IV)  1,343  25  3.3*** (0.7)  27.1*** (6.8)  4.0*** (1.0)  0.9** (0.5)  12.1*** (3.8)   Shorter period of 36 months  879  25  2.4*** (0.8)  24.0*** (8.1)  0.8 (1.2)  0.3 (0.5)  8.7* (4.9)   Local road congestion as health control  1,343  25  3.4*** (0.7)  24.2*** (6.2)  3.6*** (1.0)  1.1** (0.5)  12.6*** (3.7)   Repeat w/ IV based on policy and congestion  1,343  25  3.3*** (0.6)  23.3*** (5.6)  3.3*** (1.0)  1.2*** (0.4)  11.3*** (3.4)   Trauma admissions as health control  1,343  25  3.0*** (0.7)  22.6*** (6.5)  3.8*** (1.1)  0.8* (0.5)  11.5*** (3.9)   All other admissions as health control  1,343  25  1.9*** (0.7)  13.7** (6.3)  2.0* (1.1)  0.5 (0.5)  10.7*** (3.9)   Distance from patient to hospital as control  1,343  25  2.9*** (0.7)  22.4*** (6.3)  3.5*** (1.1)  0.8* (0.5)  10.8*** (3.7)   Dist. from patient to hospital × zip code FE  1,343  25  2.7*** (0.7)  20.9*** (6.1)  3.0*** (1.1)  0.8 (0.5)  11.4*** (3.8)   Quadratic trend  1,343  25  2.9*** (0.8)  24.2*** (7.3)  1.7 (1.1)  0.8 (0.5)  12.4*** (4.1)   Quadratic trend plus fuel prices  1,343  25  3.8*** (0.7)  30.9*** (6.9)  3.5*** (1.1)  1.1** (0.5)  11.7*** (3.7)   Weighted 2SLS regression (sq.rt.pop.)  1,343  25  2.5*** (0.6)  18.5*** (5.8)  2.5*** (0.9)  0.7* (0.4)  10.1*** (3.6)   Population-weighted average for β1l  1,343  25  2.9*** (0.7)  20.6*** (7.3)  3.4*** (1.2)  1.0** (0.5)  9.9*** (3.4)  Cardiovascular/respiratory death  Preferred specification (Table 5, 2SLS)  1,343  25  1.0*** (0.3)  –  –  0.5* (0.3)  6.7** (3.0)   Closest monitor within 10 km  1,830  34  1.3*** (0.3)  –  –  0.8*** (0.3)  8.5*** (2.9)   Mean of monitors within 5 km  1,343  25  0.8*** (0.3)  –  –  0.4* (0.3)  5.4* (2.8)   Mean of monitors within 10 km  1,839  34  0.7*** (0.2)  –  –  0.3 (0.2)  5.4** (2.6)   Site-specific meteorology effects (IV)  1,343  25  0.7*** (0.3)  –  –  0.4* (0.2)  3.6 (2.9)   Include policy–wind interactions (IV)  1,343  25  1.0*** (0.3)  –  –  0.5* (0.3)  7.0** (3.1)   Incl. site-specific policy–wind interact. (IV)  1,343  25  0.8*** (0.3)  –  –  0.4 (0.2)  4.6 (2.8)   Drop control sites 50–100 km away (IV)  1,343  25  1.1** (0.4)  –  –  0.6 (0.4)  7.4* (3.9)   Policy implemented May 31, 2010 (IV)  1,343  25  0.7** (0.3)  –  –  0.3 (0.3)  5.1* (2.9)   Shorter period of 36 months  879  25  0.8** (0.4)  –  –  0.5 (0.3)  6.0* (3.5)   Local road congestion as health control  1,343  25  1.0*** (0.3)  –  –  0.4* (0.3)  8.2*** (2.9)   Repeat w/ IV based on policy and congestion  1,343  25  0.9*** (0.3)  –  –  0.4 (0.2)  7.4*** (2.7)   Trauma admissions as health control  1,343  25  0.9*** (0.3)  –  –  0.5* (0.3)  6.6** (3.0)   All other admissions as health control  1,343  25  0.7** (0.3)  –  –  0.4 (0.3)  6.0** (3.0)   Distance from patient to hospital as control  1,343  25  0.9*** (0.3)  –  –  0.5* (0.3)  6.4** (3.0)   Dist. from patient to hospital × zip code FE  1,343  25  0.8** (0.3)  –  –  0.4 (0.3)  5.0 (3.1)   Quadratic trend  1,343  25  1.0*** (0.3)  –  –  0.6** (0.3)  6.6** (3.1)   Quadratic trend plus fuel prices  1,343  25  1.1*** (0.3)  –  –  0.7*** (0.3)  7.6** (3.0)   Weighted 2SLS regression (sq.rt.pop.)  1,343  25  0.6** (0.3)  –  –  0.2 (0.2)  3.9 (2.7)   Population-weighted average for β1l  1,343  25  0.8** (0.3)  –  –  0.4 (0.3)  6.3** (2.7)  Placebo: Trauma admissions (excl. traffic)  Preferred specification (Table 5, 2SLS)  1,343  25  0.1 (0.2)  –  –  0.3 (0.3)  0.8 (1.1)   Closest monitor within 10 km  1,830  34  −0.1 (0.2)  –  –  0.1 (0.3)  0.2 (1.1)   Mean of monitors within 5 km  1,343  25  0.0 (0.2)  –  –  0.2 (0.3)  1.1 (1.0)   Mean of monitors within 10 km  1,839  34  −0.2 (0.2)  –  –  0.0 (0.2)  −0.5 (0.9)   Site-specific meteorology effects (IV)  1,343  25  −0.2 (0.2)  –  –  0.2 (0.3)  −0.5 (1.0)   Include policy–wind interactions (IV)  1,343  25  0.1 (0.2)  –  –  0.3 (0.3)  0.9 (1.1)   Incl. site-specific policy–wind interact. (IV)  1,343  25  −0.1 (0.2)  –  –  0.3 (0.3)  −0.7 (1.0)   Drop control sites 50–100 km away (IV)  1,343  25  0.2 (0.3)  –  –  0.2 (0.4)  1.1 (1.4)   Policy implemented May 31, 2010 (IV)  1,343  25  0.2 (0.2)  –  –  0.3 (0.3)  0.9 (1.1)   Shorter period of 36 months  879  25  0.0 (0.3)  –  –  −0.1 (0.4)  1.6 (1.3)   Local road congestion as health control  1,343  25  0.1 (0.2)  –  –  0.2 (0.3)  0.6 (1.0)   Repeat w/ IV based on policy and congestion  1,343  25  0.0 (0.2)  –  –  0.1 (0.3)  0.5 (1.0)   Distance from patient to hospital as control  1,343  25  0.1 (0.2)  –  –  0.2 (0.3)  0.9 (1.1)   Dist. from patient to hospital × zip code FE  1,343  25  −0.2 (0.3)  –  –  0.1 (0.3)  0.7 (1.1)   Quadratic trend  1,343  25  0.4 (0.2)  –  –  0.5* (0.3)  1.0 (1.1)   Quadratic trend plus fuel prices  1,343  25  0.2 (0.2)  –  –  0.3 (0.3)  1.4 (1.1)   Weighted 2SLS regression (sq.rt.pop.)  1,343  25  0.2 (0.2)  –  –  0.3 (0.3)  0.9 (0.9)   Population-weighted average for β1l  1,343  25  0.1 (0.2)  –  –  0.4 (0.3)  0.6 (1.0)  Notes: The first row within each disease category/death reproduces estimates for our preferred specification (Table 5). Each row presents a different robustness test (with separate 2SLS regressions by age group in the columns). As in Table 5, we report β1l, averaged across all zip codes l, for a 10 ppb increase in NOx equivalent units of diesel pollution. See Table 5 notes. IV denotes instrumental variable. Robust standard errors are reported in parentheses. *Significant at 0.1; **significant at 0.05; ***significant at 0.01. View Large Table 6. Effects of diesel pollution on hospital admission and death rates: robustness tests of the structural analysis. Dependent variable: hospitalization rate per 100,000 residents per month  Number of observations  Number of zip codes  All ages  0–4 years  5–19 years  20–64 years  65+ years  Cardiovascular admissions  Preferred specification (Table 5, 2SLS)  1,343  25  4.2*** (0.7)  –  –  2.7*** (0.7)  20.8*** (5.9)   Closest monitor within 10 km  1,830  34  4.3*** (0.8)  –  –  3.1*** (0.8)  21.6*** (6.2)   Mean of monitors within 5 km  1,343  25  3.7*** (0.7)  –  –  2.4*** (0.7)  18.5*** (5.4)   Mean of monitors within 10 km  1,839  34  3.9*** (0.7)  –  –  3.2*** (0.7)  16.9*** (5.6)   Site-specific meteorology effects (IV)  1,343  25  2.7*** (0.6)  –  –  1.7*** (0.7)  13.4** (5.4)   Include policy–wind interactions (IV)  1,343  25  4.2*** (0.8)  –  –  2.7*** (0.7)  21.1*** (6.0)   Incl. site-specific policy–wind interact. (IV)  1,343  25  2.8*** (0.6)  –  –  1.7** (0.7)  15.2*** (5.3)   Drop control sites 50–100 km away (IV)  1,343  25  4.8*** (1.0)  –  –  3.3*** (1.0)  17.9** (8.1)   Policy implemented May 31, 2010 (IV)  1,343  25  4.2*** (0.7)  –  –  3.1*** (0.7)  19.0*** (5.7)   Shorter period of 36 months  879  25  2.1** (0.9)  –  –  1.0 (0.9)  11.9* (7.0)   Local road congestion as health control  1,343  25  4.1*** (0.7)  –  –  2.8*** (0.7)  21.4*** (5.6)   Repeat w/ IV based on policy and congestion  1,343  25  3.7*** (0.7)  –  –  2.5*** (0.7)  18.8*** (5.2)   Trauma admissions as health control  1,343  25  4.1*** (0.8)  –  –  2.5*** (0.7)  21.0*** (5.9)   All other admissions as health control  1,343  25  2.7*** (0.7)  –  –  1.7** (0.7)  17.8*** (6.0)   Distance from patient to hospital as control  1,343  25  4.0*** (0.7)  –  –  2.5*** (0.7)  20.5*** (5.9)   Dist. from patient to hospital × zip code FE  1,343  25  3.6*** (0.8)  –  –  2.6*** (0.8)  18.1*** (5.8)   Quadratic trend  1,343  25  3.8*** (0.8)  –  –  2.6*** (0.8)  15.5*** (5.7)   Quadratic trend plus fuel prices  1,343  25  4.6*** (0.8)  –  –  3.0*** (0.8)  23.7*** (5.7)   Weighted 2SLS regression (sq.rt.pop.)  1,343  25  3.3*** (0.7)  –  –  2.0*** (0.7)  16.4*** (5.5)   Population-weighted average for β1l  1,343  25  3.3*** (0.8)  –  –  1.8** (0.8)  19.3*** (5.1)  Respiratory admissions  Preferred specification (Table 5, 2SLS)  1,343  25  3.2*** (0.7)  22.9*** (6.4)  3.7*** (1.1)  0.9* (0.5)  11.5*** (3.9)   Closest monitor within 10 km  1,830  34  3.0*** (0.7)  22.8*** (6.7)  3.3*** (1.0)  0.8* (0.5)  14.7*** (3.9)   Mean of monitors within 5 km  1,343  25  2.4*** (0.6)  19.7*** (5.8)  3.3*** (1.0)  0.5 (0.4)  8.6** (3.6)   Mean of monitors within 10 km  1,839  34  1.4** (0.7)  9.9* (5.9)  1.4 (0.9)  0.6 (0.4)  11.6*** (3.6)   Site-specific meteorology effects (IV)  1,343  25  2.5*** (0.6)  25.2*** (5.8)  1.6* (1.0)  0.7* (0.4)  3.1 (3.6)   Include policy–wind interactions (IV)  1,343  25  3.2*** (0.7)  23.7*** (6.5)  3.9*** (1.1)  0.9* (0.5)  11.8*** (3.9)   Incl. site-specific policy–wind interact. (IV)  1,343  25  2.6*** (0.6)  23.3*** (5.6)  1.6* (1.0)  0.8** (0.4)  4.1 (3.6)   Drop control sites 50–100 km away (IV)  1,343  25  4.2*** (1.0)  28.3*** (9.3)  6.3*** (1.5)  1.3** (0.6)  11.2** (4.8)   Policy implemented May 31, 2010 (IV)  1,343  25  3.3*** (0.7)  27.1*** (6.8)  4.0*** (1.0)  0.9** (0.5)  12.1*** (3.8)   Shorter period of 36 months  879  25  2.4*** (0.8)  24.0*** (8.1)  0.8 (1.2)  0.3 (0.5)  8.7* (4.9)   Local road congestion as health control  1,343  25  3.4*** (0.7)  24.2*** (6.2)  3.6*** (1.0)  1.1** (0.5)  12.6*** (3.7)   Repeat w/ IV based on policy and congestion  1,343  25  3.3*** (0.6)  23.3*** (5.6)  3.3*** (1.0)  1.2*** (0.4)  11.3*** (3.4)   Trauma admissions as health control  1,343  25  3.0*** (0.7)  22.6*** (6.5)  3.8*** (1.1)  0.8* (0.5)  11.5*** (3.9)   All other admissions as health control  1,343  25  1.9*** (0.7)  13.7** (6.3)  2.0* (1.1)  0.5 (0.5)  10.7*** (3.9)   Distance from patient to hospital as control  1,343  25  2.9*** (0.7)  22.4*** (6.3)  3.5*** (1.1)  0.8* (0.5)  10.8*** (3.7)   Dist. from patient to hospital × zip code FE  1,343  25  2.7*** (0.7)  20.9*** (6.1)  3.0*** (1.1)  0.8 (0.5)  11.4*** (3.8)   Quadratic trend  1,343  25  2.9*** (0.8)  24.2*** (7.3)  1.7 (1.1)  0.8 (0.5)  12.4*** (4.1)   Quadratic trend plus fuel prices  1,343  25  3.8*** (0.7)  30.9*** (6.9)  3.5*** (1.1)  1.1** (0.5)  11.7*** (3.7)   Weighted 2SLS regression (sq.rt.pop.)  1,343  25  2.5*** (0.6)  18.5*** (5.8)  2.5*** (0.9)  0.7* (0.4)  10.1*** (3.6)   Population-weighted average for β1l  1,343  25  2.9*** (0.7)  20.6*** (7.3)  3.4*** (1.2)  1.0** (0.5)  9.9*** (3.4)  Cardiovascular/respiratory death  Preferred specification (Table 5, 2SLS)  1,343  25  1.0*** (0.3)  –  –  0.5* (0.3)  6.7** (3.0)   Closest monitor within 10 km  1,830  34  1.3*** (0.3)  –  –  0.8*** (0.3)  8.5*** (2.9)   Mean of monitors within 5 km  1,343  25  0.8*** (0.3)  –  –  0.4* (0.3)  5.4* (2.8)   Mean of monitors within 10 km  1,839  34  0.7*** (0.2)  –  –  0.3 (0.2)  5.4** (2.6)   Site-specific meteorology effects (IV)  1,343  25  0.7*** (0.3)  –  –  0.4* (0.2)  3.6 (2.9)   Include policy–wind interactions (IV)  1,343  25  1.0*** (0.3)  –  –  0.5* (0.3)  7.0** (3.1)   Incl. site-specific policy–wind interact. (IV)  1,343  25  0.8*** (0.3)  –  –  0.4 (0.2)  4.6 (2.8)   Drop control sites 50–100 km away (IV)  1,343  25  1.1** (0.4)  –  –  0.6 (0.4)  7.4* (3.9)   Policy implemented May 31, 2010 (IV)  1,343  25  0.7** (0.3)  –  –  0.3 (0.3)  5.1* (2.9)   Shorter period of 36 months  879  25  0.8** (0.4)  –  –  0.5 (0.3)  6.0* (3.5)   Local road congestion as health control  1,343  25  1.0*** (0.3)  –  –  0.4* (0.3)  8.2*** (2.9)   Repeat w/ IV based on policy and congestion  1,343  25  0.9*** (0.3)  –  –  0.4 (0.2)  7.4*** (2.7)   Trauma admissions as health control  1,343  25  0.9*** (0.3)  –  –  0.5* (0.3)  6.6** (3.0)   All other admissions as health control  1,343  25  0.7** (0.3)  –  –  0.4 (0.3)  6.0** (3.0)   Distance from patient to hospital as control  1,343  25  0.9*** (0.3)  –  –  0.5* (0.3)  6.4** (3.0)   Dist. from patient to hospital × zip code FE  1,343  25  0.8** (0.3)  –  –  0.4 (0.3)  5.0 (3.1)   Quadratic trend  1,343  25  1.0*** (0.3)  –  –  0.6** (0.3)  6.6** (3.1)   Quadratic trend plus fuel prices  1,343  25  1.1*** (0.3)  –  –  0.7*** (0.3)  7.6** (3.0)   Weighted 2SLS regression (sq.rt.pop.)  1,343  25  0.6** (0.3)  –  –  0.2 (0.2)  3.9 (2.7)   Population-weighted average for β1l  1,343  25  0.8** (0.3)  –  –  0.4 (0.3)  6.3** (2.7)  Placebo: Trauma admissions (excl. traffic)  Preferred specification (Table 5, 2SLS)  1,343  25  0.1 (0.2)  –  –  0.3 (0.3)  0.8 (1.1)   Closest monitor within 10 km  1,830  34  −0.1 (0.2)  –  –  0.1 (0.3)  0.2 (1.1)   Mean of monitors within 5 km  1,343  25  0.0 (0.2)  –  –  0.2 (0.3)  1.1 (1.0)   Mean of monitors within 10 km  1,839  34  −0.2 (0.2)  –  –  0.0 (0.2)  −0.5 (0.9)   Site-specific meteorology effects (IV)  1,343  25  −0.2 (0.2)  –  –  0.2 (0.3)  −0.5 (1.0)   Include policy–wind interactions (IV)  1,343  25  0.1 (0.2)  –  –  0.3 (0.3)  0.9 (1.1)   Incl. site-specific policy–wind interact. (IV)  1,343  25  −0.1 (0.2)  –  –  0.3 (0.3)  −0.7 (1.0)   Drop control sites 50–100 km away (IV)  1,343  25  0.2 (0.3)  –  –  0.2 (0.4)  1.1 (1.4)   Policy implemented May 31, 2010 (IV)  1,343  25  0.2 (0.2)  –  –  0.3 (0.3)  0.9 (1.1)   Shorter period of 36 months  879  25  0.0 (0.3)  –  –  −0.1 (0.4)  1.6 (1.3)   Local road congestion as health control  1,343  25  0.1 (0.2)  –  –  0.2 (0.3)  0.6 (1.0)   Repeat w/ IV based on policy and congestion  1,343  25  0.0 (0.2)  –  –  0.1 (0.3)  0.5 (1.0)   Distance from patient to hospital as control  1,343  25  0.1 (0.2)  –  –  0.2 (0.3)  0.9 (1.1)   Dist. from patient to hospital × zip code FE  1,343  25  −0.2 (0.3)  –  –  0.1 (0.3)  0.7 (1.1)   Quadratic trend  1,343  25  0.4 (0.2)  –  –  0.5* (0.3)  1.0 (1.1)   Quadratic trend plus fuel prices  1,343  25  0.2 (0.2)  –  –  0.3 (0.3)  1.4 (1.1)   Weighted 2SLS regression (sq.rt.pop.)  1,343  25  0.2 (0.2)  –  –  0.3 (0.3)  0.9 (0.9)   Population-weighted average for β1l  1,343  25  0.1 (0.2)  –  –  0.4 (0.3)  0.6 (1.0)  Dependent variable: hospitalization rate per 100,000 residents per month  Number of observations  Number of zip codes  All ages  0–4 years  5–19 years  20–64 years  65+ years  Cardiovascular admissions  Preferred specification (Table 5, 2SLS)  1,343  25  4.2*** (0.7)  –  –  2.7*** (0.7)  20.8*** (5.9)   Closest monitor within 10 km  1,830  34  4.3*** (0.8)  –  –  3.1*** (0.8)  21.6*** (6.2)   Mean of monitors within 5 km  1,343  25  3.7*** (0.7)  –  –  2.4*** (0.7)  18.5*** (5.4)   Mean of monitors within 10 km  1,839  34  3.9*** (0.7)  –  –  3.2*** (0.7)  16.9*** (5.6)   Site-specific meteorology effects (IV)  1,343  25  2.7*** (0.6)  –  –  1.7*** (0.7)  13.4** (5.4)   Include policy–wind interactions (IV)  1,343  25  4.2*** (0.8)  –  –  2.7*** (0.7)  21.1*** (6.0)   Incl. site-specific policy–wind interact. (IV)  1,343  25  2.8*** (0.6)  –  –  1.7** (0.7)  15.2*** (5.3)   Drop control sites 50–100 km away (IV)  1,343  25  4.8*** (1.0)  –  –  3.3*** (1.0)  17.9** (8.1)   Policy implemented May 31, 2010 (IV)  1,343  25  4.2*** (0.7)  –  –  3.1*** (0.7)  19.0*** (5.7)   Shorter period of 36 months  879  25  2.1** (0.9)  –  –  1.0 (0.9)  11.9* (7.0)   Local road congestion as health control  1,343  25  4.1*** (0.7)  –  –  2.8*** (0.7)  21.4*** (5.6)   Repeat w/ IV based on policy and congestion  1,343  25  3.7*** (0.7)  –  –  2.5*** (0.7)  18.8*** (5.2)   Trauma admissions as health control  1,343  25  4.1*** (0.8)  –  –  2.5*** (0.7)  21.0*** (5.9)   All other admissions as health control  1,343  25  2.7*** (0.7)  –  –  1.7** (0.7)  17.8*** (6.0)   Distance from patient to hospital as control  1,343  25  4.0*** (0.7)  –  –  2.5*** (0.7)  20.5*** (5.9)   Dist. from patient to hospital × zip code FE  1,343  25  3.6*** (0.8)  –  –  2.6*** (0.8)  18.1*** (5.8)   Quadratic trend  1,343  25  3.8*** (0.8)  –  –  2.6*** (0.8)  15.5*** (5.7)   Quadratic trend plus fuel prices  1,343  25  4.6*** (0.8)  –  –  3.0*** (0.8)  23.7*** (5.7)   Weighted 2SLS regression (sq.rt.pop.)  1,343  25  3.3*** (0.7)  –  –  2.0*** (0.7)  16.4*** (5.5)   Population-weighted average for β1l  1,343  25  3.3*** (0.8)  –  –  1.8** (0.8)  19.3*** (5.1)  Respiratory admissions  Preferred specification (Table 5, 2SLS)  1,343  25  3.2*** (0.7)  22.9*** (6.4)  3.7*** (1.1)  0.9* (0.5)  11.5*** (3.9)   Closest monitor within 10 km  1,830  34  3.0*** (0.7)  22.8*** (6.7)  3.3*** (1.0)  0.8* (0.5)  14.7*** (3.9)   Mean of monitors within 5 km  1,343  25  2.4*** (0.6)  19.7*** (5.8)  3.3*** (1.0)  0.5 (0.4)  8.6** (3.6)   Mean of monitors within 10 km  1,839  34  1.4** (0.7)  9.9* (5.9)  1.4 (0.9)  0.6 (0.4)  11.6*** (3.6)   Site-specific meteorology effects (IV)  1,343  25  2.5*** (0.6)  25.2*** (5.8)  1.6* (1.0)  0.7* (0.4)  3.1 (3.6)   Include policy–wind interactions (IV)  1,343  25  3.2*** (0.7)  23.7*** (6.5)  3.9*** (1.1)  0.9* (0.5)  11.8*** (3.9)   Incl. site-specific policy–wind interact. (IV)  1,343  25  2.6*** (0.6)  23.3*** (5.6)  1.6* (1.0)  0.8** (0.4)  4.1 (3.6)   Drop control sites 50–100 km away (IV)  1,343  25  4.2*** (1.0)  28.3*** (9.3)  6.3*** (1.5)  1.3** (0.6)  11.2** (4.8)   Policy implemented May 31, 2010 (IV)  1,343  25  3.3*** (0.7)  27.1*** (6.8)  4.0*** (1.0)  0.9** (0.5)  12.1*** (3.8)   Shorter period of 36 months  879  25  2.4*** (0.8)  24.0*** (8.1)  0.8 (1.2)  0.3 (0.5)  8.7* (4.9)   Local road congestion as health control  1,343  25  3.4*** (0.7)  24.2*** (6.2)  3.6*** (1.0)  1.1** (0.5)  12.6*** (3.7)   Repeat w/ IV based on policy and congestion  1,343  25  3.3*** (0.6)  23.3*** (5.6)  3.3*** (1.0)  1.2*** (0.4)  11.3*** (3.4)   Trauma admissions as health control  1,343  25  3.0*** (0.7)  22.6*** (6.5)  3.8*** (1.1)  0.8* (0.5)  11.5*** (3.9)   All other admissions as health control  1,343  25  1.9*** (0.7)  13.7** (6.3)  2.0* (1.1)  0.5 (0.5)  10.7*** (3.9)   Distance from patient to hospital as control  1,343  25  2.9*** (0.7)  22.4*** (6.3)  3.5*** (1.1)  0.8* (0.5)  10.8*** (3.7)   Dist. from patient to hospital × zip code FE  1,343  25  2.7*** (0.7)  20.9*** (6.1)  3.0*** (1.1)  0.8 (0.5)  11.4*** (3.8)   Quadratic trend  1,343  25  2.9*** (0.8)  24.2*** (7.3)  1.7 (1.1)  0.8 (0.5)  12.4*** (4.1)   Quadratic trend plus fuel prices  1,343  25  3.8*** (0.7)  30.9*** (6.9)  3.5*** (1.1)  1.1** (0.5)  11.7*** (3.7)   Weighted 2SLS regression (sq.rt.pop.)  1,343  25  2.5*** (0.6)  18.5*** (5.8)  2.5*** (0.9)  0.7* (0.4)  10.1*** (3.6)   Population-weighted average for β1l  1,343  25  2.9*** (0.7)  20.6*** (7.3)  3.4*** (1.2)  1.0** (0.5)  9.9*** (3.4)  Cardiovascular/respiratory death  Preferred specification (Table 5, 2SLS)  1,343  25  1.0*** (0.3)  –  –  0.5* (0.3)  6.7** (3.0)   Closest monitor within 10 km  1,830  34  1.3*** (0.3)  –  –  0.8*** (0.3)  8.5*** (2.9)   Mean of monitors within 5 km  1,343  25  0.8*** (0.3)  –  –  0.4* (0.3)  5.4* (2.8)   Mean of monitors within 10 km  1,839  34  0.7*** (0.2)  –  –  0.3 (0.2)  5.4** (2.6)   Site-specific meteorology effects (IV)  1,343  25  0.7*** (0.3)  –  –  0.4* (0.2)  3.6 (2.9)   Include policy–wind interactions (IV)  1,343  25  1.0*** (0.3)  –  –  0.5* (0.3)  7.0** (3.1)   Incl. site-specific policy–wind interact. (IV)  1,343  25  0.8*** (0.3)  –  –  0.4 (0.2)  4.6 (2.8)   Drop control sites 50–100 km away (IV)  1,343  25  1.1** (0.4)  –  –  0.6 (0.4)  7.4* (3.9)   Policy implemented May 31, 2010 (IV)  1,343  25  0.7** (0.3)  –  –  0.3 (0.3)  5.1* (2.9)   Shorter period of 36 months  879  25  0.8** (0.4)  –  –  0.5 (0.3)  6.0* (3.5)   Local road congestion as health control  1,343  25  1.0*** (0.3)  –  –  0.4* (0.3)  8.2*** (2.9)   Repeat w/ IV based on policy and congestion  1,343  25  0.9*** (0.3)  –  –  0.4 (0.2)  7.4*** (2.7)   Trauma admissions as health control  1,343  25  0.9*** (0.3)  –  –  0.5* (0.3)  6.6** (3.0)   All other admissions as health control  1,343  25  0.7** (0.3)  –  –  0.4 (0.3)  6.0** (3.0)   Distance from patient to hospital as control  1,343  25  0.9*** (0.3)  –  –  0.5* (0.3)  6.4** (3.0)   Dist. from patient to hospital × zip code FE  1,343  25  0.8** (0.3)  –  –  0.4 (0.3)  5.0 (3.1)   Quadratic trend  1,343  25  1.0*** (0.3)  –  –  0.6** (0.3)  6.6** (3.1)   Quadratic trend plus fuel prices  1,343  25  1.1*** (0.3)  –  –  0.7*** (0.3)  7.6** (3.0)   Weighted 2SLS regression (sq.rt.pop.)  1,343  25  0.6** (0.3)  –  –  0.2 (0.2)  3.9 (2.7)   Population-weighted average for β1l  1,343  25  0.8** (0.3)  –  –  0.4 (0.3)  6.3** (2.7)  Placebo: Trauma admissions (excl. traffic)  Preferred specification (Table 5, 2SLS)  1,343  25  0.1 (0.2)  –  –  0.3 (0.3)  0.8 (1.1)   Closest monitor within 10 km  1,830  34  −0.1 (0.2)  –  –  0.1 (0.3)  0.2 (1.1)   Mean of monitors within 5 km  1,343  25  0.0 (0.2)  –  –  0.2 (0.3)  1.1 (1.0)   Mean of monitors within 10 km  1,839  34  −0.2 (0.2)  –  –  0.0 (0.2)  −0.5 (0.9)   Site-specific meteorology effects (IV)  1,343  25  −0.2 (0.2)  –  –  0.2 (0.3)  −0.5 (1.0)   Include policy–wind interactions (IV)  1,343  25  0.1 (0.2)  –  –  0.3 (0.3)  0.9 (1.1)   Incl. site-specific policy–wind interact. (IV)  1,343  25  −0.1 (0.2)  –  –  0.3 (0.3)  −0.7 (1.0)   Drop control sites 50–100 km away (IV)  1,343  25  0.2 (0.3)  –  –  0.2 (0.4)  1.1 (1.4)   Policy implemented May 31, 2010 (IV)  1,343  25  0.2 (0.2)  –  –  0.3 (0.3)  0.9 (1.1)   Shorter period of 36 months  879  25  0.0 (0.3)  –  –  −0.1 (0.4)  1.6 (1.3)   Local road congestion as health control  1,343  25  0.1 (0.2)  –  –  0.2 (0.3)  0.6 (1.0)   Repeat w/ IV based on policy and congestion  1,343  25  0.0 (0.2)  –  –  0.1 (0.3)  0.5 (1.0)   Distance from patient to hospital as control  1,343  25  0.1 (0.2)  –  –  0.2 (0.3)  0.9 (1.1)   Dist. from patient to hospital × zip code FE  1,343  25  −0.2 (0.3)  –  –  0.1 (0.3)  0.7 (1.1)   Quadratic trend  1,343  25  0.4 (0.2)  –  –  0.5* (0.3)  1.0 (1.1)   Quadratic trend plus fuel prices  1,343  25  0.2 (0.2)  –  –  0.3 (0.3)  1.4 (1.1)   Weighted 2SLS regression (sq.rt.pop.)  1,343  25  0.2 (0.2)  –  –  0.3 (0.3)  0.9 (0.9)   Population-weighted average for β1l  1,343  25  0.1 (0.2)  –  –  0.4 (0.3)  0.6 (1.0)  Notes: The first row within each disease category/death reproduces estimates for our preferred specification (Table 5). Each row presents a different robustness test (with separate 2SLS regressions by age group in the columns). As in Table 5, we report β1l, averaged across all zip codes l, for a 10 ppb increase in NOx equivalent units of diesel pollution. See Table 5 notes. IV denotes instrumental variable. Robust standard errors are reported in parentheses. *Significant at 0.1; **significant at 0.05; ***significant at 0.01. View Large In the order presented in the rows of Table 6: to assign diesel pollution to zip code of residence, we specify the closest NOx monitor within a 10 km, rather than a 5 km, distance of each zip code's district centroid. As another alternative, we specify the mean concentration across all NOx monitors within 5 km (alternatively, 10 km) of a district centroid, rather than the concentration at the closest NOx monitor, and accordingly modify the instrument (NOx variation fitted by the policy variation). To generate the instrument $${\widehat {\mathit {diesel}}_{lt}}$$, we alternatively use the specifications reported in columns (3) and (6) of Table 4, respectively allowing the effects of meteorology to vary by site,36 and excluding sites outside the metropolis from the estimation sample and instead using sites 10–20 km from the original truck route as controls. In forming the instrument, we alternatively specify the policy implementation date as May 31, rather than March 30, 2010. As an alternative to our preferred November 2008 to May 2013 study period, we shorten the sample to January 2009 to December 2011. Instrument $${\widehat {\mathit {diesel}}_{lt}}$$ is then constructed according to specification 4, Table 4, which also uses the shorter sample. We specify additional controls in the structural health equation. We include local road congestion (summed across road segments within 2 km of the NOx monitor closest to each district, as explained, and during daytime hours, 7 a.m. to 8 p.m., when traffic flows are higher) interacted with zip code fixed effects for flexibility.37 As a test of the sorting hypothesis, as discussed previously, we include traffic-unrelated trauma admissions, and as an “acid test” we include all admissions other than cardiorespiratory and traffic-unrelated trauma (the mean value in the sample is 305 monthly admissions per 100,000 residents, to be compared to 53 for cardiovascular and 43 for respiratory). We include mean distance between the patient and the admitting hospital to capture, for example, patients seeking better hospitals further from home in response to the temporary traffic relief. An alternative to specifying month-of-sample fixed effects in the health equation is to use a less-flexible quadratic trend. Instrument $${\widehat {\mathit {diesel}}_{lt}}$$ is then constructed according to specification 2, Table 4, with month-of-sample fixed effects—which subsume fuel prices—replaced by a quadratic time trend and fuel prices. We show estimates for a health equation that includes a quadratic trend but not fuel prices (in which case fuel prices provide instruments in addition to the policy change), and estimates for a specification with both a quadratic trend and fuel prices. We weigh the 2SLS regression specifying as weights the square root of the zip code population in the given month by age group. As an alternative to reporting the arithmetic mean for β1l across all zip codes l, in the last row within each disease category/death of Table 6 we report a population-weighted average for β1l across all zip codes l.38 Finally, Online Appendix Table B.5, panel A considers the logarithm of the admissions or death rate, by age group-disease pair, as the dependent variable. Effects are consistent with (or in some cases slightly higher) than those in Table 5. Point estimates for respiratory are higher in the vulnerable age groups. Panel B considers the (within-month average of) daily maximum NOx, rather than daily mean NOx, as an alternative measure of the severity of diesel exposure, modifying the instrument accordingly (policy-induced maximum NOx variation). We repeated the analysis, now further stratifying by gender, and find some evidence that respiratory admissions respond to diesel exhaust more for boys than for girls aged 0–4 years, but is similar by age 5–19 years (Clougherty 2010).39 3.5. A Multi-Pollutant Model The interpretation we offer for our adopted single-pollutant approach is that of NOx as a wider “indicator” (Dominici et al. 2010) for variation in atmospheric toxins both (i) emitted by diesel combustion—for example, particles of varying chemical and physical composition such as black carbon and PM2.5 to ultrafine—as well as (ii) formed from or that react with these tailpipe emissions. As noted, the bulk of primary and secondary pollutants associated with diesel exhaust, numbering in the hundreds or thousands of compounds, are unmonitored. One co-pollutant that is monitored, and is depleted by vehicular NO in proximity to roads, is ground-level ozone. Indeed, plots of deseasoned ozone measured at sites 5, 27, and 31, the three sites nearest to the original truck route,40 show an increase over the fall of 2010, after the beltway opened. However, rising ozone was observed elsewhere in the metropolis and in cities 50–100 km away. This was due in large part to unseasonal rising temperatures that contribute to atmospheric ozone production, illustrating why correcting for meteorology is key (plots omitted for brevity). Taking the single-pollutant 2SLS model in Tables 4 and 5 as our point of departure, we estimate a multipollutant NOx-ozone health model in Tables 7 and 8 (and Table 4). Table 7 reports estimates for an ozone pollution model similar to Table 4, but where the dependent variable is the daily ozone concentration either averaged over 24 h or the maximum 8-h average.41 We now require a second exclusion restriction, available by virtue of price-induced shifts in the gasoline–ethanol mix used in the light-vehicle fleet (Salvo and Geiger 2014). This additional source of identifying variation is temporal, as fluctuations in the price of ethanol relative to gasoline were similar across São Paulo state, so our chosen specification controls for long-run ozone drifts using a quadratic trend as in Table 4, column (2) for NOx. Month-of-sample fixed effects would subsume most of the variation in fuel prices (Table 4, column (1)). As seen in Table 7, ozone levels fall in the ethanol-to-gasoline price ratio.42 Consistent with NO's depletion of ozone, when the truck policy was introduced maximum 8-h ozone rose by 4 μg/m3 (2 ppb) at the sites closest to the key inner-city corridor, against a mean of 65 μg/m3 for the metropolis. Table 7. The effects of the (spatially differentiated) truck abatement policy and the (spatially invariant) light-vehicle gasoline–ethanol fuel mix on localized ozone pollution. Dependent variable (μg/m3)  Ozone, daily 24-h mean  Ozone, daily maximum 8-h mean    (1)  (2)  (3)  (4)  (5)  (6)  Model specification  Year  Quadratic  Metro  Year  Quadratic  Metro    FE  trend  only  FE  trend  only  Truck policy in metro (yes = 1)  × Site FEs  Sites 2–6 km from original truck  1.6**  1.5**  3.2***  4.3***  4.2***  4.5***   route (mean three sites: 5, 27, 31)  (0.7)  (0.7)  (0.8)  (1.4)  (1.4)  (1.2)  Sites 6–10 km from original truck route  −0.0  −0.1  1.7***  2.0  1.9  2.2**  (mean four sites: 15, 7, 1, 3)  (0.8)  (0.8)  (0.6)  (1.5)  (1.5)  (1.0)  Sites 10–20 km from original truck route  −1.5**  1.7**  –  −0.0  −0.1  –   (mean five sites: 2, 6, 18, 29, 22)  (0.7)  (0.7)  –  (1.3)  (1.3)  –  Policy implemented (yes = 1)  2.8  1.8  0.6  5.1*  2.6  3.1    (1.7)  (1.5)  (1.6)  (2.7)  (2.4)  (2.6)  Ethanol-to-gasoline price ratio × Site FEs  −27.3***  −29.5***  −30.9***  −35.8**  −46.2***  −50.0***   (mean 12 sites in metropolis, 4 outside)  (8.7)  (7.6)  (7.7)  (14.2)  (12.3)  (13.0)  Site fixed effects (FEs)  Yes  Yes  Yes  Yes  Yes  Yes  Week-of-year, day-of-week by site FEs  Yes  Yes  Yes  Yes  Yes  Yes  Meteorology and thermal inversion  Yes  Yes  Yes  Yes  Yes  Yes  Diesel price (ethanol–gasoline ratio given previously)  Yes  Yes  Yes  Yes  Yes  Yes  Year FEs  Yes  –  –  Yes  –  –  Quadratic time trend  –  Yes  Yes  –  Yes  Yes  Number of observations  25,183  25,183  18,753  25,183  25,183  18,753  Number of regressors  1,038  1,035  778  1,038  1,035  778  R2  0.663  0.660  0.650  0.700  0.699  0.703  Mean value of dep. var.  36.4  36.4  35.2  67.0  67.0  64.8  Dependent variable (μg/m3)  Ozone, daily 24-h mean  Ozone, daily maximum 8-h mean    (1)  (2)  (3)  (4)  (5)  (6)  Model specification  Year  Quadratic  Metro  Year  Quadratic  Metro    FE  trend  only  FE  trend  only  Truck policy in metro (yes = 1)  × Site FEs  Sites 2–6 km from original truck  1.6**  1.5**  3.2***  4.3***  4.2***  4.5***   route (mean three sites: 5, 27, 31)  (0.7)  (0.7)  (0.8)  (1.4)  (1.4)  (1.2)  Sites 6–10 km from original truck route  −0.0  −0.1  1.7***  2.0  1.9  2.2**  (mean four sites: 15, 7, 1, 3)  (0.8)  (0.8)  (0.6)  (1.5)  (1.5)  (1.0)  Sites 10–20 km from original truck route  −1.5**  1.7**  –  −0.0  −0.1  –   (mean five sites: 2, 6, 18, 29, 22)  (0.7)  (0.7)  –  (1.3)  (1.3)  –  Policy implemented (yes = 1)  2.8  1.8  0.6  5.1*  2.6  3.1    (1.7)  (1.5)  (1.6)  (2.7)  (2.4)  (2.6)  Ethanol-to-gasoline price ratio × Site FEs  −27.3***  −29.5***  −30.9***  −35.8**  −46.2***  −50.0***   (mean 12 sites in metropolis, 4 outside)  (8.7)  (7.6)  (7.7)  (14.2)  (12.3)  (13.0)  Site fixed effects (FEs)  Yes  Yes  Yes  Yes  Yes  Yes  Week-of-year, day-of-week by site FEs  Yes  Yes  Yes  Yes  Yes  Yes  Meteorology and thermal inversion  Yes  Yes  Yes  Yes  Yes  Yes  Diesel price (ethanol–gasoline ratio given previously)  Yes  Yes  Yes  Yes  Yes  Yes  Year FEs  Yes  –  –  Yes  –  –  Quadratic time trend  –  Yes  Yes  –  Yes  Yes  Number of observations  25,183  25,183  18,753  25,183  25,183  18,753  Number of regressors  1,038  1,035  778  1,038  1,035  778  R2  0.663  0.660  0.650  0.700  0.699  0.703  Mean value of dep. var.  36.4  36.4  35.2  67.0  67.0  64.8  Notes: An observation is a day by ozone monitoring site pair. The dependent variable is the daily ozone concentration (in μg/m3) averaged over 24 h (specifications (1)–(3)) or the maximum 8-h average (specifications (4)–(6)). Specifications (1, 2) and (4, 5) are “difference in difference” models, with the second difference corresponding to sites in the metropolis, potentially impacted by the truck policy, versus sites 50–100 km away. (3) and (6) exclude sites 50–100 km away. We allow the effects of both the truck policy and the light-vehicle gasoline–ethanol mix to vary by site but the table show means across sites grouped according to distance from the original truck route (site-specific coefficients on truck policy) or over all sites (site-specific coefficients on per-liter price ratio for ethanol to gasoline). The sample period is November 2008 to May 2013. We control for: site fixed effects; seasonality (week-of-year and day-of-week) by site fixed effects; meteorology and thermal inversion; and diesel prices. (1) and (4) include year fixed effects, whereas all other specifications include a quadratic time trend. OLS regressions. Standard errors, in parentheses, are one-way clustered by week-of-sample. *Significant at 0.1; **significant at 0.05; ***significant at 0.01. View Large Table 7. The effects of the (spatially differentiated) truck abatement policy and the (spatially invariant) light-vehicle gasoline–ethanol fuel mix on localized ozone pollution. Dependent variable (μg/m3)  Ozone, daily 24-h mean  Ozone, daily maximum 8-h mean    (1)  (2)  (3)  (4)  (5)  (6)  Model specification  Year  Quadratic  Metro  Year  Quadratic  Metro    FE  trend  only  FE  trend  only  Truck policy in metro (yes = 1)  × Site FEs  Sites 2–6 km from original truck  1.6**  1.5**  3.2***  4.3***  4.2***  4.5***   route (mean three sites: 5, 27, 31)  (0.7)  (0.7)  (0.8)  (1.4)  (1.4)  (1.2)  Sites 6–10 km from original truck route  −0.0  −0.1  1.7***  2.0  1.9  2.2**  (mean four sites: 15, 7, 1, 3)  (0.8)  (0.8)  (0.6)  (1.5)  (1.5)  (1.0)  Sites 10–20 km from original truck route  −1.5**  1.7**  –  −0.0  −0.1  –   (mean five sites: 2, 6, 18, 29, 22)  (0.7)  (0.7)  –  (1.3)  (1.3)  –  Policy implemented (yes = 1)  2.8  1.8  0.6  5.1*  2.6  3.1    (1.7)  (1.5)  (1.6)  (2.7)  (2.4)  (2.6)  Ethanol-to-gasoline price ratio × Site FEs  −27.3***  −29.5***  −30.9***  −35.8**  −46.2***  −50.0***   (mean 12 sites in metropolis, 4 outside)  (8.7)  (7.6)  (7.7)  (14.2)  (12.3)  (13.0)  Site fixed effects (FEs)  Yes  Yes  Yes  Yes  Yes  Yes  Week-of-year, day-of-week by site FEs  Yes  Yes  Yes  Yes  Yes  Yes  Meteorology and thermal inversion  Yes  Yes  Yes  Yes  Yes  Yes  Diesel price (ethanol–gasoline ratio given previously)  Yes  Yes  Yes  Yes  Yes  Yes  Year FEs  Yes  –  –  Yes  –  –  Quadratic time trend  –  Yes  Yes  –  Yes  Yes  Number of observations  25,183  25,183  18,753  25,183  25,183  18,753  Number of regressors  1,038  1,035  778  1,038  1,035  778  R2  0.663  0.660  0.650  0.700  0.699  0.703  Mean value of dep. var.  36.4  36.4  35.2  67.0  67.0  64.8  Dependent variable (μg/m3)  Ozone, daily 24-h mean  Ozone, daily maximum 8-h mean    (1)  (2)  (3)  (4)  (5)  (6)  Model specification  Year  Quadratic  Metro  Year  Quadratic  Metro    FE  trend  only  FE  trend  only  Truck policy in metro (yes = 1)  × Site FEs  Sites 2–6 km from original truck  1.6**  1.5**  3.2***  4.3***  4.2***  4.5***   route (mean three sites: 5, 27, 31)  (0.7)  (0.7)  (0.8)  (1.4)  (1.4)  (1.2)  Sites 6–10 km from original truck route  −0.0  −0.1  1.7***  2.0  1.9  2.2**  (mean four sites: 15, 7, 1, 3)  (0.8)  (0.8)  (0.6)  (1.5)  (1.5)  (1.0)  Sites 10–20 km from original truck route  −1.5**  1.7**  –  −0.0  −0.1  –   (mean five sites: 2, 6, 18, 29, 22)  (0.7)  (0.7)  –  (1.3)  (1.3)  –  Policy implemented (yes = 1)  2.8  1.8  0.6  5.1*  2.6  3.1    (1.7)  (1.5)  (1.6)  (2.7)  (2.4)  (2.6)  Ethanol-to-gasoline price ratio × Site FEs  −27.3***  −29.5***  −30.9***  −35.8**  −46.2***  −50.0***   (mean 12 sites in metropolis, 4 outside)  (8.7)  (7.6)  (7.7)  (14.2)  (12.3)  (13.0)  Site fixed effects (FEs)  Yes  Yes  Yes  Yes  Yes  Yes  Week-of-year, day-of-week by site FEs  Yes  Yes  Yes  Yes  Yes  Yes  Meteorology and thermal inversion  Yes  Yes  Yes  Yes  Yes  Yes  Diesel price (ethanol–gasoline ratio given previously)  Yes  Yes  Yes  Yes  Yes  Yes  Year FEs  Yes  –  –  Yes  –  –  Quadratic time trend  –  Yes  Yes  –  Yes  Yes  Number of observations  25,183  25,183  18,753  25,183  25,183  18,753  Number of regressors  1,038  1,035  778  1,038  1,035  778  R2  0.663  0.660  0.650  0.700  0.699  0.703  Mean value of dep. var.  36.4  36.4  35.2  67.0  67.0  64.8  Notes: An observation is a day by ozone monitoring site pair. The dependent variable is the daily ozone concentration (in μg/m3) averaged over 24 h (specifications (1)–(3)) or the maximum 8-h average (specifications (4)–(6)). Specifications (1, 2) and (4, 5) are “difference in difference” models, with the second difference corresponding to sites in the metropolis, potentially impacted by the truck policy, versus sites 50–100 km away. (3) and (6) exclude sites 50–100 km away. We allow the effects of both the truck policy and the light-vehicle gasoline–ethanol mix to vary by site but the table show means across sites grouped according to distance from the original truck route (site-specific coefficients on truck policy) or over all sites (site-specific coefficients on per-liter price ratio for ethanol to gasoline). The sample period is November 2008 to May 2013. We control for: site fixed effects; seasonality (week-of-year and day-of-week) by site fixed effects; meteorology and thermal inversion; and diesel prices. (1) and (4) include year fixed effects, whereas all other specifications include a quadratic time trend. OLS regressions. Standard errors, in parentheses, are one-way clustered by week-of-sample. *Significant at 0.1; **significant at 0.05; ***significant at 0.01. View Large Table 8. Effects of diesel pollution on hospital admission and death rates: multipollutant model with ozone. Dependent variable: all-age hospitalization rate per 100,000 residents per month  Fitted pollut. as IV, from  Number of obs.  No. of zip codes  NOx (+10 ppb)  Ozone (+10 μg/m3)  Cardiovascular admissions  Preferred single-pollutant model (Table 5)  Table 4(1)  1,343  25  4.2*** (0.7)  –  Include 24-h ozone  Tables 4(2) and 7(2)  1,126  21  6.3*** (2.0)  −2.8 (2.1)  Include maximum 8-h ozone  Tables 4(2) and 7(5)  1,126  21  6.7*** (1.8)  −1.3 (1.0)  Include max. 8-h ozone (within 10 km)  Tables 4(2) and 7(5)  1,771  33  2.8** (1.1)  −1.1 (0.7)  Respiratory admissions  Preferred single-pollutant model (Table 5)  Table 4(1)  1,343  25  3.2*** (0.7)  –  Include 24-h ozone  Tables 4(2) and 7(2)  1,126  21  4.9** (2.0)  −1.7 (1.9)  Include maximum 8-h ozone  Tables 4(2) and 7(5)  1,126  21  5.1*** (1.8)  −0.3 (0.9)  Include max. 8-h ozone (within 10 km)  Tables 4(2) and 7(5)  1,771  33  3.3*** (1.2)  −0.6 (0.7)  Cardiovascular/respiratory death  Preferred single-pollutant model (Table 5)  Table 4(1)  1,343  25  1.0*** (0.3)  –  Include 24-h ozone  Tables 4(2) and 7(2)  1,126  21  1.9** (0.8)  −0.6 (0.8)  Include maximum 8-h ozone  Tables 4(2) and 7(5)  1,126  21  2.0*** (0.7)  −0.4 (0.4)  Include max. 8-h ozone (within 10 km)  Tables 4(2) and 7(5)  1,771  33  1.1** (0.4)  −0.2 (0.3)  Placebo: Trauma admissions (excl. traffic)  Preferred single-pollutant model (Table 5)  Table 4(1)  1,343  25  0.1 (0.2)  –  Include 24-h ozone  Tables 4(2) and 7(2)  1,126  21  0.8 (0.6)  −0.5 (0.7)  Include maximum 8-h ozone  Tables 4(2) and 7(5)  1,126  21  0.9 (0.6)  −0.2 (0.3)  Include max. 8-h ozone (within 10 km)  Tables 4(2) and 7(5)  1,771  33  0.6 (0.4)  −0.2 (0.2)  Dependent variable: all-age hospitalization rate per 100,000 residents per month  Fitted pollut. as IV, from  Number of obs.  No. of zip codes  NOx (+10 ppb)  Ozone (+10 μg/m3)  Cardiovascular admissions  Preferred single-pollutant model (Table 5)  Table 4(1)  1,343  25  4.2*** (0.7)  –  Include 24-h ozone  Tables 4(2) and 7(2)  1,126  21  6.3*** (2.0)  −2.8 (2.1)  Include maximum 8-h ozone  Tables 4(2) and 7(5)  1,126  21  6.7*** (1.8)  −1.3 (1.0)  Include max. 8-h ozone (within 10 km)  Tables 4(2) and 7(5)  1,771  33  2.8** (1.1)  −1.1 (0.7)  Respiratory admissions  Preferred single-pollutant model (Table 5)  Table 4(1)  1,343  25  3.2*** (0.7)  –  Include 24-h ozone  Tables 4(2) and 7(2)  1,126  21  4.9** (2.0)  −1.7 (1.9)  Include maximum 8-h ozone  Tables 4(2) and 7(5)  1,126  21  5.1*** (1.8)  −0.3 (0.9)  Include max. 8-h ozone (within 10 km)  Tables 4(2) and 7(5)  1,771  33  3.3*** (1.2)  −0.6 (0.7)  Cardiovascular/respiratory death  Preferred single-pollutant model (Table 5)  Table 4(1)  1,343  25  1.0*** (0.3)  –  Include 24-h ozone  Tables 4(2) and 7(2)  1,126  21  1.9** (0.8)  −0.6 (0.8)  Include maximum 8-h ozone  Tables 4(2) and 7(5)  1,126  21  2.0*** (0.7)  −0.4 (0.4)  Include max. 8-h ozone (within 10 km)  Tables 4(2) and 7(5)  1,771  33  1.1** (0.4)  −0.2 (0.3)  Placebo: Trauma admissions (excl. traffic)  Preferred single-pollutant model (Table 5)  Table 4(1)  1,343  25  0.1 (0.2)  –  Include 24-h ozone  Tables 4(2) and 7(2)  1,126  21  0.8 (0.6)  −0.5 (0.7)  Include maximum 8-h ozone  Tables 4(2) and 7(5)  1,126  21  0.9 (0.6)  −0.2 (0.3)  Include max. 8-h ozone (within 10 km)  Tables 4(2) and 7(5)  1,771  33  0.6 (0.4)  −0.2 (0.2)  Notes: An observation is a 3-digit residential zip code by month pair. The first row within each disease category/death reproduces estimates for our preferred single-pollutant specification, using NOx as a marker for diesel exhaust (Table 5). The following rows each represent a different 2SLS regression, where the dependent variable is the all-age hospitalization rate for the given disease category/death, and where ozone exposure is proxied by either the daily 24-h or maximum 8-h concentration, as indicated. We instrument for NOx and ozone using fitted NOx and fitted ozone per the specifications reported, respectively, in Table 4(2) and Table 7 (column as indicated), and plotted in Appendix Figure A.2. Similar to Table 5, we report $$\beta _{1l}^{\rm {NOx}}$$ and $$\beta _{1l}^{\rm {ozone}}$$, the coefficients on diesel$${lt}$$ and ozone$${lt}$$, respectively, averaged across all zip codes l, for increases of 10 ppb in (24-h) NOx-equivalent units of diesel pollution and 10 μg/m3 in (24 h or maximum 8 h) ozone. We assign each district within a zip code to the nearest NOx monitor and nearest ozone monitor from the district centroid, not to exceed 5 km (or, in the last row, 10 km); we then average NOx and ozone across districts within each zip code. The first-stage F-statistic for the excluded instruments averages 29. The sample period is November 2008 to May 2013. We control for: zip code fixed effects; seasonality (month-of-year) by zip code fixed effects; a quadratic time trend; and meteorology and thermal inversion. Robust standard errors are reported in parentheses. *Significant at 0.1; **significant at 0.05; ***significant at 0.01. View Large Table 8. Effects of diesel pollution on hospital admission and death rates: multipollutant model with ozone. Dependent variable: all-age hospitalization rate per 100,000 residents per month  Fitted pollut. as IV, from  Number of obs.  No. of zip codes  NOx (+10 ppb)  Ozone (+10 μg/m3)  Cardiovascular admissions  Preferred single-pollutant model (Table 5)  Table 4(1)  1,343  25  4.2*** (0.7)  –  Include 24-h ozone  Tables 4(2) and 7(2)  1,126  21  6.3*** (2.0)  −2.8 (2.1)  Include maximum 8-h ozone  Tables 4(2) and 7(5)  1,126  21  6.7*** (1.8)  −1.3 (1.0)  Include max. 8-h ozone (within 10 km)  Tables 4(2) and 7(5)  1,771  33  2.8** (1.1)  −1.1 (0.7)  Respiratory admissions  Preferred single-pollutant model (Table 5)  Table 4(1)  1,343  25  3.2*** (0.7)  –  Include 24-h ozone  Tables 4(2) and 7(2)  1,126  21  4.9** (2.0)  −1.7 (1.9)  Include maximum 8-h ozone  Tables 4(2) and 7(5)  1,126  21  5.1*** (1.8)  −0.3 (0.9)  Include max. 8-h ozone (within 10 km)  Tables 4(2) and 7(5)  1,771  33  3.3*** (1.2)  −0.6 (0.7)  Cardiovascular/respiratory death  Preferred single-pollutant model (Table 5)  Table 4(1)  1,343  25  1.0*** (0.3)  –  Include 24-h ozone  Tables 4(2) and 7(2)  1,126  21  1.9** (0.8)  −0.6 (0.8)  Include maximum 8-h ozone  Tables 4(2) and 7(5)  1,126  21  2.0*** (0.7)  −0.4 (0.4)  Include max. 8-h ozone (within 10 km)  Tables 4(2) and 7(5)  1,771  33  1.1** (0.4)  −0.2 (0.3)  Placebo: Trauma admissions (excl. traffic)  Preferred single-pollutant model (Table 5)  Table 4(1)  1,343  25  0.1 (0.2)  –  Include 24-h ozone  Tables 4(2) and 7(2)  1,126  21  0.8 (0.6)  −0.5 (0.7)  Include maximum 8-h ozone  Tables 4(2) and 7(5)  1,126  21  0.9 (0.6)  −0.2 (0.3)  Include max. 8-h ozone (within 10 km)  Tables 4(2) and 7(5)  1,771  33  0.6 (0.4)  −0.2 (0.2)  Dependent variable: all-age hospitalization rate per 100,000 residents per month  Fitted pollut. as IV, from  Number of obs.  No. of zip codes  NOx (+10 ppb)  Ozone (+10 μg/m3)  Cardiovascular admissions  Preferred single-pollutant model (Table 5)  Table 4(1)  1,343  25  4.2*** (0.7)  –  Include 24-h ozone  Tables 4(2) and 7(2)  1,126  21  6.3*** (2.0)  −2.8 (2.1)  Include maximum 8-h ozone  Tables 4(2) and 7(5)  1,126  21  6.7*** (1.8)  −1.3 (1.0)  Include max. 8-h ozone (within 10 km)  Tables 4(2) and 7(5)  1,771  33  2.8** (1.1)  −1.1 (0.7)  Respiratory admissions  Preferred single-pollutant model (Table 5)  Table 4(1)  1,343  25  3.2*** (0.7)  –  Include 24-h ozone  Tables 4(2) and 7(2)  1,126  21  4.9** (2.0)  −1.7 (1.9)  Include maximum 8-h ozone  Tables 4(2) and 7(5)  1,126  21  5.1*** (1.8)  −0.3 (0.9)  Include max. 8-h ozone (within 10 km)  Tables 4(2) and 7(5)  1,771  33  3.3*** (1.2)  −0.6 (0.7)  Cardiovascular/respiratory death  Preferred single-pollutant model (Table 5)  Table 4(1)  1,343  25  1.0*** (0.3)  –  Include 24-h ozone  Tables 4(2) and 7(2)  1,126  21  1.9** (0.8)  −0.6 (0.8)  Include maximum 8-h ozone  Tables 4(2) and 7(5)  1,126  21  2.0*** (0.7)  −0.4 (0.4)  Include max. 8-h ozone (within 10 km)  Tables 4(2) and 7(5)  1,771  33  1.1** (0.4)  −0.2 (0.3)  Placebo: Trauma admissions (excl. traffic)  Preferred single-pollutant model (Table 5)  Table 4(1)  1,343  25  0.1 (0.2)  –  Include 24-h ozone  Tables 4(2) and 7(2)  1,126  21  0.8 (0.6)  −0.5 (0.7)  Include maximum 8-h ozone  Tables 4(2) and 7(5)  1,126  21  0.9 (0.6)  −0.2 (0.3)  Include max. 8-h ozone (within 10 km)  Tables 4(2) and 7(5)  1,771  33  0.6 (0.4)  −0.2 (0.2)  Notes: An observation is a 3-digit residential zip code by month pair. The first row within each disease category/death reproduces estimates for our preferred single-pollutant specification, using NOx as a marker for diesel exhaust (Table 5). The following rows each represent a different 2SLS regression, where the dependent variable is the all-age hospitalization rate for the given disease category/death, and where ozone exposure is proxied by either the daily 24-h or maximum 8-h concentration, as indicated. We instrument for NOx and ozone using fitted NOx and fitted ozone per the specifications reported, respectively, in Table 4(2) and Table 7 (column as indicated), and plotted in Appendix Figure A.2. Similar to Table 5, we report $$\beta _{1l}^{\rm {NOx}}$$ and $$\beta _{1l}^{\rm {ozone}}$$, the coefficients on diesel$${lt}$$ and ozone$${lt}$$, respectively, averaged across all zip codes l, for increases of 10 ppb in (24-h) NOx-equivalent units of diesel pollution and 10 μg/m3 in (24 h or maximum 8 h) ozone. We assign each district within a zip code to the nearest NOx monitor and nearest ozone monitor from the district centroid, not to exceed 5 km (or, in the last row, 10 km); we then average NOx and ozone across districts within each zip code. The first-stage F-statistic for the excluded instruments averages 29. The sample period is November 2008 to May 2013. We control for: zip code fixed effects; seasonality (month-of-year) by zip code fixed effects; a quadratic time trend; and meteorology and thermal inversion. Robust standard errors are reported in parentheses. *Significant at 0.1; **significant at 0.05; ***significant at 0.01. View Large Table 8 reports estimates when we include 24-h or maximum 8-h ozone$${lt}$$, in addition to diesel$${lt}$$ (i.e., 24-h NOx), in health equation (2), now with a quadratic trend instead of month-of-sample fixed effects of Table 5 to correct for long-run unobserved health determinants. We include NOx and ozone variation induced by the truck policy and ethanol–gasoline prices (fitted NOx and ozone using corresponding specifications in Tables 4 and 7) in the vector of excluded instruments.43 Although point estimates on the NOx coefficients (means over 21 or 33 zip codes44) are larger or smaller depending on whether we proxy for multipollutant exposure using the closest monitor up to 5 or 10 km from a zip code's district centroid, results are consistent in magnitude and significance with those reported in the single pollutant analysis of Table 5. Although the epidemiological/economics literature (Gouveia and Fletcher 2000; Arceo et al. 2016) finds collinearity in multi-pollutant analysis of health challenging, the key feature of our pollution data is the large drop in NOx emanating from the inner city coinciding with the truck policy.45 4. Concluding Remarks We find that abating diesel truck traffic in the urban core of the São Paulo megacity resulted in sustained reductions in ambient NOx—in its “indicator role” for diesel exhaust known to contain many toxic substances not routinely monitored46—of magnitude 5–50 ppb, depending on proximity of roadside exposure, with quantifiable morbidity and mortality benefits. The research focus is not whether traffic pollution harms health, but to identify the types of traffic that cause large damage. Summing cardiovascular and respiratory conditions, we estimate reductions of 886 (s.e. 141) public hospital admissions and 116 (s.e. 37) in-hospital deaths, or 8% and 9% of total, per year per one million residents per 10 ppb abatement in NOx equivalent units of diesel pollution (e.g., Table 5, (4.2 + 3.2) × 12 months/year × 1 million/100,000). Considering: (i) a 1 to 2 million inner-city population that benefited from a 10 ppb-equivalent diesel pollution abatement, and (ii) a 20,000 abatement in trucks passing daily, we arrive at a back-of-the-envelope annual impact of one hospitalization for every 11–23 diesel trucks, and one death for every 86–172 diesel trucks, driving daily through the inner city (e.g., 20,000$$/$$(886 × 2) to 20,000$$/$$886). Studies in toxicology and epidemiology have established causal links between exposure to diesel exhaust and human health damage (Ohtoshi et al. 1998; Kilburn 2000; Castranova et al. 2001; Baulig et al. 2003; Gauderman et al. 2005; McCreanor et al. 2007; Krivoshto et al. 2008; Jalava et al. 2010; Marks et al. 2010; Patel et al. 2011). Yet, to our knowledge, this is the first observational study to combine evidence at the scale of a gridlocked megacity, exploiting a large, abrupt and identifiable policy-induced shift in the composition of a real-world circulating fleet. The evidence is based on spatially and temporally resolved observations of hospitalizations, ambient diesel pollution, traffic volume, meteorology, and other controls. With measurements from different sources, the joint distribution of variables is compelling. Upon implementation, the beltway-access cum tightened-truck-circulation policy abated road congestion, NOx levels, and cardiorespiratory hospitalizations, particularly in the road segments, air monitoring sites, and residential zip codes that were most exposed to the passing truck traffic. Offering a uniquely clean demonstration of the fundamental law of road congestion (Duranton and Turner 2011), the traffic volume relief brought about by the intervention—its stated motivation—was temporary. However, the resulting shift in road user composition unintentionally improved air quality and public health, enabling us to quantify the health effect of urban exposure to diesel exhaust. More broadly, other world megacities might stand to gain similarly by curbing the circulation of aging heavy-duty diesel fleets at times and locations of high human exposure, including investing in fleet renovation. Notes The editor in charge of this paper was Claudio Michelacci. Acknowledgments We gratefully acknowledge numerous people from ANP, CET, CETESB, INMET, and SPMar who facilitated data sharing. In particular, we thank Wagner Baptista, Cristina Costa, Alaor Dall’Antonia Jr, Marcos Abreu Fonseca, Dirce Maria Franco, Masayuki Kuromoto, Carlos Lacava, Clarice Muramoto, Alexandre Romualdo, Marcelo Safadi and Roseni dos Santos. We thank Sam Ritchey for coding the GPS coordinates of road segment endpoints and Lucas Soriano Martins with compiling other data. We thank seminar audiences at ASSA, Columbia, Duke-NUS, International Society for Environmental Epidemiology conference, Melbourne, NUS, Regional Air Quality Management Workshop at HKUST, Singapore Health Economics Association conference, Singapore Public Health and Occupational Medicine conference, Society for Risk Analysis World Congress, and WCERE, as well as Reshad Ahsan, Antonio Bento, Pat Kinney, Matt Neidell, Ivan Png, and Chiu Yu Ko, for helpful comments. A.S. acknowledges support from the Initiative for Sustainability and Energy, General Motors Research Center for Strategy in Management, the Center for Research in Technology & Innovation, and the Zell Center for Risk Research, all at Northwestern University, as well as Singapore's Ministry of Education Academic Research Fund Tier 1 (FY2013-FRC3-003). The data and code replicating all of the results in this article are available at https://goo.gl/GG2pgh. Appendix A: Adding Composition Effects to the Law of Road Congestion To help fix ideas, we lay out a simple conceptual framework. Consider an open-access resource such as urban (inner-city) road space. A user's cost to access this resource (including time spent commuting), denoted t(q), is common across users and is increasing and convex in the aggregate number of users, q ≥ 0 (Small and Verhoef 2007). The nominal (“boilerplate”) capacity of the resource is of measure k. We thus specify t(q) > 0, t΄(q) ≥ 0 and t΄΄(q) ≥ 0, with limq→kt(q) → ∞. On the demand side, there are two groups of users, l and h, of mass (“market size”) ml and mh, each user type with heterogeneous preferences over using the resource distributed according to the cumulative density functions Fl and Fh, with domain $${\mathbb{R}^ + }$$ (Figure A.3(a)–(c)). These functions, along with user mass, describe both aggregate demand for the open-access resource as well as the composition of demand. Thus, for example, the number of users of type l with value up to $$v$$ is given by mlFl($$v$$). In our setting, l and h represent light- and heavy-duty vehicles, or low- and high-diesel-exhaust vehicles, respectively. Given our setting, we make the following assumption Figure A.1. View largeDownload slide Daily number of southern beltway users, by vehicle type. Data, based on toll-paying vehicles, are daily means within calendar month since toll operator SPMar's contract began in August 2011. “Heavy commercial vehicles” are commercial vehicles with at least three axles. Source: SPMar. Figure A.1. View largeDownload slide Daily number of southern beltway users, by vehicle type. Data, based on toll-paying vehicles, are daily means within calendar month since toll operator SPMar's contract began in August 2011. “Heavy commercial vehicles” are commercial vehicles with at least three axles. Source: SPMar. Figure A.2. View largeDownload slide Ozone versus NOx concentrations in the São Paulo metropolis. (a) Daily (24 h) mean ozone (μg/m3), or (b) daily maximum 8-h mean ozone (μg/m3), against daily mean NOx (ppb). Seven sites in the metropolis concurrently monitor both NOx and ozone. An observation is a site by month-of-sample. We show (site-specific) means over days within month-of-sample. (c) Fitted 24-h ozone (μg/m3), or (d) fitted maximum 8-h ozone (μg/m3), against fitted 24-h NOx (ppb), per the excluded instruments in the 2SLS multi-pollutant health models of Table 8. An observation is a 3-digit residential zip code by month-of-sample. Source: CETESB, Tables 4(2), 7(2), and 7(5). Figure A.2. View largeDownload slide Ozone versus NOx concentrations in the São Paulo metropolis. (a) Daily (24 h) mean ozone (μg/m3), or (b) daily maximum 8-h mean ozone (μg/m3), against daily mean NOx (ppb). Seven sites in the metropolis concurrently monitor both NOx and ozone. An observation is a site by month-of-sample. We show (site-specific) means over days within month-of-sample. (c) Fitted 24-h ozone (μg/m3), or (d) fitted maximum 8-h ozone (μg/m3), against fitted 24-h NOx (ppb), per the excluded instruments in the 2SLS multi-pollutant health models of Table 8. An observation is a 3-digit residential zip code by month-of-sample. Source: CETESB, Tables 4(2), 7(2), and 7(5). Figure A.3. View largeDownload slide Adding composition effects to the Law of Road Congestion. The top panels (a)–(c) indicate the marginal user of each type, and the aggregate number of users and individual cost in equilibrium. In the middle panels (d)–(f), the opening of a beltway raises the value of the outside option to urban road space for type-h users. In the bottom panels (g)–(i), exogenous demand growth shifts the demand curve to the right. Figure A.3. View largeDownload slide Adding composition effects to the Law of Road Congestion. The top panels (a)–(c) indicate the marginal user of each type, and the aggregate number of users and individual cost in equilibrium. In the middle panels (d)–(f), the opening of a beltway raises the value of the outside option to urban road space for type-h users. In the bottom panels (g)–(i), exogenous demand growth shifts the demand curve to the right. A$$\scriptstyle{\rm SSUMPTION}$$ 1 (Megacity). ml + mh ≫ k. User i's utility from accessing the resource is thus given by $$u$$i = $$v$$i  −  t(q), with the second term capturing the congestion externality. Normalize the utility from the outside option, u0, to be equal across user types and users, at 0, that is, $$u$$l0 = $$u$$h0  =  0. User i's selection into road space is then given by   \begin{equation*} {\rm{max}}\left( {0,\ {u_i}\left( q \right)} \right) \end{equation*} In equilibrium, the marginal users of each type, with values ($$v$$l, $$v$$h), are defined implicitly (assuming interior solutions) by the system:   \begin{equation*} {v_l} - t\ \left( {{m_l}\left( {1 - {F_l}\left( {{v_l}} \right)} \right) + {m_h}\left( {1 - {F_h}\left( {{v_h}} \right)} \right)} \right) = \ 0, \end{equation*}   \begin{equation*} {v_h} - t\ \left( {{m_l}\left( {1 - {F_l}\left( {{v_l}} \right)} \right) + {m_h}\left( {1 - {F_h}\left( {{v_h}} \right)} \right)} \right) = \ 0, \end{equation*} where the aggregate number of users, q*, is ml(1 − Fl($$v$$l)) + mh(1 − Fh($$v$$h)). We now briefly consider how the composition of road users changes under different shocks. As a summary measure, let sh denote the type-h user share among all users, that is,   \begin{equation*} \ {s_h} = \frac{{{m_h}\left( {1 - {F_h}\left( {{v_h}} \right)} \right)}}{{{m_l}\left( {1 - {F_l}\left( {{v_l}} \right)} \right) + {m_h}\left( {1 - {F_h}\left( {{v_h}} \right)} \right)}}. \end{equation*} Shock 1 (Opening of a beltway around the urban area). Suppose duh0 > 0. The “immediate” effect of raising the value of an alternative to urban road space for type-h users is that the mass of type-h users falls by mh(Fh($$v$$h + duh0) − Fh($$v$$h))—see shift (i) in Figure A.3(d)–(f). As urban road space becomes available, and the cost of accessing the resource drops, marginal users of both types select into road space—see shift (ii), Figure A.3(d)–(f). As illustrated, the higher the density around Fl($$v$$l) (i.e., the probability density function dFl($$v$$)$$/$$d$$v$$ evaluated at $$v$$l) relative to the density at Fh($$v$$h + duh0), the more will the type-h share fall after the beltway opens, dsh$$/$$d$$u$$h0 < 0. Stated differently, the steeper is Fl($$v$$l) relative to Fh($$v$$h + d$$u$$h0)—Δuh0 is large, say, as in the figure—the more will the type-h share fall. The aggregate number of users falls by a measure, dq*$$/$$duh0 < 0, that depends also on the curvature of t(q). Shock 2 (Investment in light rail). Suppose dul0 > 0. This shift in the value of the outside option to type-l users is analogous to Shock 1. The ensuing increase in the type-h share, dsh$$/$$dul0 > 0, depends on the density around Fh($$v$$h) relative to the density at Fl($$v$$l + dul0). Also, dq*$$/$$dul0 < 0. Shock 3 (Exogenous demand growth). Consider a proportionate shift in population such that dml$$/$$ml =  dmh$$/$$mh > 0, with no change to the preference distributions Fl and Fh. It is clear that this shock will raise the aggregate number of users, q*, while not necessarily changing the composition of demand—see shift (iii), Figure A.3(g)–(i). It is intuitive that a combination of Shock 1 (or Shock 2) followed by population growth may actually result in an increased aggregate number of users. Footnotes 1 Small and Kazimi (1995) compute the external harm per vehicle-mile for the 1992 Los Angeles fleet as 3.3 cents for a gasoline car and 52.7 cents for a heavy-duty diesel truck (1992 US dollars, only health damage was estimated). Auffhammer (2017) writes: “Big trucks are largely powered by diesels… The externalities from these trucks are likely significant in terms of pollution, congestion, and accidents. But I am aware of next to no papers in the economics literature that have attempted to quantify these externalities”. 2 Some work has measured the impact of traffic composition on airborne pollutant concentrations. Kinney et al. (2000) and Lena et al. (2002) associate black carbon concentrations with concurrent traffic counts for diesel trucks and buses along New York city streets. Wolff (2014) studies the effect on particle pollution (PM10) in some German cities that implemented “low-emission zones” restricting the circulation of vehicles according to their age (Euro 1–4) and fuel (diesel vs. gasoline). 3 Economist (2017) writes: “NOx emissions cause the premature deaths of an estimated 72,000 Europeans a year… This week the city of Oslo used new powers to ban diesel cars temporarily in order to improve air quality. Paris, Madrid, and Athens are set to ban diesels altogether by 2025”. Singapore's latest budget selectively introduces a volume-based duty on diesel to reduce its use (Lam 2017). 4 Restrictions on the circulation of light vehicles, based on registration plate and similar to those in Beijing, Bogotá, and Mexico City (Eskeland and Feyzioglu 1997; Davis 2008; Gallego et al. 2013; Viard and Fu 2015; Zhang et al. 2017), have been in place since the late 1990s. 5 We provide references on this phenomenon in what follows. Appendix  A provides a simple model of this mechanism. Relatedly, Gallego et al. (2013) document the fast response by households to transport policy shocks on car use. 6 At the same time, we do not examine the effect of chronic exposure to diesel exhaust, that is, sustained over years and decades (USEPA 2002; Garshick et al. 2004; Chen et al. 2015). We subsequently provide references to the medical literature on health impacts from acute exposure to diesel combustion, and do not repeat them here. 7 For example, logistics or retail operations might locate along the beltway (Souza 2009). Anecdotal evidence suggests that land development was not immediate, perhaps in part due to the distance (a 25-km radius) from the city center, or land use restrictions. For example, by February 2011, ten months after inauguration, no fuel retailer had located, or been allowed to locate, along the beltway's southern section (Transporta Brasil 2011). 8 For example, Kumar et al. (2011) estimated that heavy-duty diesel vehicles contributed 65% of nanoparticle number emissions in Delhi in 2010, but only 4% of vehicle kilometers traveled. In a source contribution study in rich Singapore, Engling et al. (2014) find that diesel exhaust contributes a dominant 62% to suspended particulate matter on typical days. 9 56 ppb is about 105 μg/m3, a level that can double at roadside monitors, and lies between 1998 means for New York at 79 μg/m3 and Mexico City at 130 μg/m3 (Molina and Molina 2004, Table 2). Based on remote sensing, Beirle et al. (2003, Figure 1) reports hotspots for global tropospheric NO2 in São Paulo, Mexico City, Johannesburg, Jakarta, East Asia, Los Angeles, Eastern United States, Western Europe, and selected Middle Eastern cities. 10 1 nm = 10−3 μm = 10−9 m. A micron is 1 μm or 1,000 nm. 11 We obtained hourly number concentrations for submicron particles (PM 7-100 nm and PM 7-800 nm, i.e., PM 0.007-0.1 and PM 0.007-0.8), and mass concentrations for black carbon, sampled during field campaigns beginning October 2010 and lasting about one year at a site near to NOx monitoring site 31 in Figure 1 (Salvo et al. 2017). 12 For clarity, Figure 3 does not report a policy impact. The figure shows how central Av. dos Bandeirantes is to mobility in São Paulo (a level). Policy impacts across outcomes of interest, including localized short-run effects on road congestion, are illustrated, for example, in Figure 5. 13 An alternative to specifying 178 corridor fixed effects would be to include 5,133 road segment fixed effects, in which case the latter would subsume the term α2dl, since distance varies within corridor but not within segment, the spatial unit of observation. Results are similar. 14 Online Appendix Table B.3 specifies year fixed effects. Several time series for the wider economy (omitted for brevity) all paint a picture of economic stability over the 2008–2013 period. These include the size of the economically active population, the mean real wage, public bus ridership, and city airport activity in the São Paulo metropolis, as well as industrial activity and wholesale diesel quantities (including the highway market) for São Paulo state. 15 In the alternative specification of Online Appendix Table B.3, we use, besides a full set of year fixed effects, year fixed effects beginning 2011 × ln(distance) from the original truck route. 16 See Online Appendix Table B.3 for variations on the reduced-form analysis presented in Table 2. 17 São Paulo's roads are most congested during the afternoon rush, between 5 p.m. and 8 p.m., thus our focus on these hours (see, e.g., Table 2 in Salvo and Wang 2017). Applying Table 2, specification 1 to the proportion of congestion time during the earlier daytime window from 7 a.m. to 4 p.m., with data also at the road segment by day level, a similar policy impact in space and over time is observed, though it is less pronounced (while equally statistically significant—see Online Appendix Table B.3, column (1)). 18 Regarding potential shifts in averting behavior that could make our estimated health effects conservative, we note here that São Paulo homes are typically naturally ventilated, and that the use of face masks is rare. 19 Anenberg et al. (2017) cites studies to support the statement that around the world “current diesel vehicles emit far more NOx under real-world operating conditions than during laboratory certification testing” (p. 467). 20 For perspective, some other studies in the environmental economics literature exploit interventions of this magnitude, for example, a 35% drop in NOx emissions due to a cap-and-trade policy in northeastern United States (Deschenes et al. 2017); a 20% drop in SO2 pollution due to the closure of a refinery in Mexico City (Hanna and Oliva 2015). 21 Figure 1(b) shows 10 km is still well inside of the beltway. Recall that the beltway lies on mostly undeveloped land at a 25-km radius from the center. 22 We note that we identify the particular change induced by the natural experiment, not a continuous dosage. 23 A threat to instrument validity would be the possibility that the localized truck abatement policies might have boosted economic activity in the surrounding areas. For example, workers in the inner city would now spend less time commuting and thus accept lower wages, benefiting local business; higher business income would then somehow spill over to the local population, helping explain improved health. However, we note from origin-destination surveys reported in Office of Metropolitan Transport (2013) that most workers live far from their workplace, suggesting that local rent sharing is of lesser concern. Median travel time on public transport is over 2 h/day, and the distribution of travel time was invariant between 2007 and 2012, on either side of the 2010 policy. Over this period, Office of Metropolitan Transport (2013) further reports that growth in both the number of jobs as well as trips by point of origin have been somewhat less pronounced in the center of the metropolis (defined as 16% of its total land area) than around the center (84% of the area). Jobs in an even tighter central area that is crossed by the Av. dos Bandeirantes and is home to 3.5 million people (24% of the population in the 43 zip codes listed in Online Appendix Table B.1) grew at exactly the same rate as in the metropolis, +1.5% per year from 2007 to 2012. Moreover, the assumption that the policy intervention benefited inner-city business has been questioned by industry representatives (Martins 2012). We verify that our inference is robust to relaxing the exclusion restriction, that is, for a range of direct policy effects on health, likely due in part to the strength of our instruments (Conley et al. 2012). 24 Our linear form for the concentration-response is supported by the extant health literature, at least over the range of particle concentrations that São Paulo's population is exposed to (WHO 2006; Mills et al. 2015; Pope et al. 2015; Di et al. 2017). 25 Inspection of Figure 4 suggests that the policy impact may have occurred over a few months, rather than instantly when the beltway opened. We caution that Figure 4 describes data, without accounting for meteorology. 26 To be clear, the estimation routine implements a first stage in which monthly zip code-specific measured NOx levels are projected on fitted NOx (fitted from (3) and averaged across days within a month) and nonpollution covariates included in health equation (2). Isen et al. (2017), for example, similarly instrument for pollution using fitted pollution imputed from a policy intervention. 27 A plot similar to site 8's Figure 4(b) shows a drop in (deseasoned) NOx across these sites 10 and 17, of over 10 ppb in April 2010 (not shown for brevity). 28 Seven NOx monitors have road segments monitored for congestion within a 1 km radius (Figures 1(b) and (c)). If we consider an alternative radius of 0.5 km from each NOx monitor, the correlation coefficient is 0.92 (N = 5). 29 See the table notes and Online Appendix B.3 for details on how we assign monitor-level NOx to zip codes, and the robustness tests in what follows for similar estimates using different assignment rules (e.g., 34 zip codes for which NOx is monitored within 10 km). 30 For perspective, blacks exhibit the highest PM2.5-mortality elasticity, with a point estimate of 0.24. 31 Averaging across estimates for three alternative 2SLS specifications for the (log) cardiorespiratory mortality rate (Chen et al. 2013, Table 3), +0.20/(100/453.2) ≈ 0.91 (+100 μg/m3 of total suspended particulates, with sample mean of 453 μg/m3). OLS estimates imply a similar elasticity. 32 Currie and Walker (2011) lack traffic and pollution data across toll plazas, but they state (see their references) “a crude estimate is that E-ZPass reduced NO2 emissions from traffic by about 6.8%…” (p. 70), which would imply elasticities above 1. The study does not discuss the fuel mix (diesel vs. gasoline in trucks and cars). 33 Office of Metropolitan Transport (2013, Figure 5) shows almost identical distributions for travel time in the São Paulo metropolis, whether on public or private transport, on comparing two years (2007, 2012) on either side of the policy intervention (2010). At least from this angle, horizon and at this level of aggregation, we do not see spatial shifts in the population of residences relative to jobs (about 45% of total daily trips), schools (32%), and so forth. 34 Anderson (2016, footnote 7) reports that the median individual aged over 65 in a Los Angeles cross-section has lived at the current location for 25 years. 35 Office of Metropolitan Transport (2013, Figure 19) reports an inverted-U relationship between commuter age and commuting frequency, with adults aged 30–39 traveling (trips/day) more than twice as often as persons aged 60+ years, and, similarly, compared to children under 4 years. 36 Additional tests reported in Table 6 include interactions of the policy indicator I(truck)t with wind conditions (speed and direction, per Table 1) in generating the pollution instrument; we do this either by restricting interactions to have common effects across sites in the metropolis, or allow these to vary by site. Thus, exogenous site-specific policy effects on diesel exposure shift with observed wind. We note that winds speeds in the metropolis are about one half those in Chicago and Los Angeles (Herrnstadt and Muehlegger 2015; Anderson 2016), and wind patterns have been stable over the sample period. 37 We implement this robustness test in two ways, instrumenting for measured NOx with the component of NOx variation that is explained by: (i) the policy change (NOx fitted per specification 1, Table 4) and, on top of this, (ii) local and citywide road congestion. 38 We also estimated a quadratic version of health equation (2), that is, with exposure terms $${\beta _{11l}}{\mathit {diese}}{l_{lt}} + {\beta _{12l}}{\mathit {diesel}}_{lt}^2$$, adding $$\widehat {{\mathit {diesel}}_{lt}^2}$$ to the vector of instruments (fitted NOx and the square of fitted NOx interacted with zip code fixed effects). The evidence suggests that the linear model is not overly restrictive. Point estimates indicate slight, and statistically insignificant, concavity, or supralinearity, around the mean NOx in the sample, 56 ppb. 39 Estimates are available on request. Although São Paulo's child population is almost evenly distributed across genders (and slightly more male), by middle age women begin to outlive men, reaching a ratio of 65% to 35% among those aged 75+ years. We find that the response of cardiorespiratory deaths to diesel pollution is significant among elderly women, and for men it is imprecisely estimated, perhaps due in part to the decline of men in that subpopulation. 40 We deseason ozone as in Figure 4(b) for NOx at site 8, which does not monitor ozone. 41 São Paulo state's 8-h ozone standard is 140 μg/m3, or about 71 ppb, comparable to the US EPA's 70 ppb (Salvo and Wang 2017). 42 Salvo and Huse (2013) find that the ethanol share among consumers equipped with “flexible fuel” gasoline–ethanol vehicles falls roughly linearly in the ethanol-to-gasoline price ratio. Salvo and Geiger (2014) document lower ozone levels—and unchanged traffic congestion, speeds, and public transport use—in the metropolis in response to higher relative ethanol prices that induce shifts to gasoline in the light-vehicle fuel mix. 43 The instruments are different but the chemistry that underlies the NOx and ozone variation generated by the instruments is the same, highlighting why estimating separate health coefficients can be challenging. Moreover, ethanol–gasoline price variation is only monthly. Recent research also suggests that ozone variation induced by shifts in the gasoline–ethanol mix may inversely correlate with variation in ultrafine particles (Salvo et al. 2017). 44 NOx and ozone are both monitored within 5 km (respectively, 10 km) for 21 (respectively, 33) zip codes. 45 Dominici et al. (2010) discuss the “vast …challenges” the multi-pollutant approach faces and “caution (that) the results of any regression model become highly unstable when incorporating two or more pollutants that are highly correlated” (p. 188). Appendix Figure A.2 shows negative correlation between NOx and ozone. 46 To support this statement, we not only cite the environmental literature; we show the tight correlation between NOx and particles in urban air, specifically in the currently unmonitored and unregulated health-relevant ultrafine size range and carbonaceous composition. To our knowledge, ours is the first economics study to report such data. References Anderson Michael L. ( 2016). “As the Wind Blows: The Effects of Long-Term Exposure to Air Pollution on Mortality.” NBER Working Paper 21578, National Bureau of Economic Research, Cambridge, MA. Anenberg Susan C., Miller Joshua, Minjares Ray, Du Li, Henze Daven K., Lacey Forrest, Malley Christopher S., Emberson Lisa, Franco Vicente, Klimont Zbigniew, Heyes Chris ( 2017). “ Impacts and Mitigation of Excess Diesel-Related NOx Emissions in 11 Major Vehicle Markets.” Nature , 545, 467– 471. Google Scholar CrossRef Search ADS   Arceo Eva, Hanna Rema, Oliva Paulina ( 2016). “ Does the Effect of Pollution on Infant Mortality Differ Between Developing and Developed Countries? Evidence from Mexico City.” The Economic Journal , 126, 257– 280. Google Scholar CrossRef Search ADS   Auffhammer Maximilian ( 2017). Save the California Waiver! How a “Little” California Vehicle Standard Prevented an Urban “Airmaggedon”.  Energy at Haas, Berkeley. Baulig Augustin, Garlatti Michele, Bonvallot Veronique, Marchand Alexandre, Barouki Robert, Marano Francelyne, Baeza Armelle ( 2003). “ Involvement of Reactive Oxygen Species in the Metabolic Pathways Triggered by Diesel Exhaust Particles in Human Airway Epithelial Cells.” American Journal of Physiology-Lung Cellular and Molecular Physiology , 285, L671– L679. Google Scholar CrossRef Search ADS   Beatty Timothy K. M., Shimshack Jay P. ( 2011). “ School Buses, Diesel Emissions, and Respiratory Health.” Journal of Health Economics , 30, 987– 999. Google Scholar CrossRef Search ADS   Beirle S., Platt U., Wenig M., Wagner T. ( 2003). “ Weekly Cycle of NO2 by GOME Measurements: A Signature of Anthropogenic Sources.” Atmospheric Chemistry and Physics , 3, 2225– 2232. Google Scholar CrossRef Search ADS   Bravo Mercedes A., Son Jiyoung, de Freitas Clarice U., Gouveia Nelson, Bell Michelle L. ( 2016). “ Air Pollution and Mortality in São Paulo, Brazil: Effects of Multiple Pollutants and Analysis of Susceptible Populations.” Journal of Exposure Science and Environmental Epidemiology , 26, 150– 161. Google Scholar CrossRef Search ADS   Brito Joel, Rizzo L. V., Herckes P., Vasconcellos P. C., Caumo S. E. S., Fornaro Adalgiza, Ynoue Rita Y., Artaxo Paulo, Andrade Maria F. ( 2013). “ Physical-Chemical Characterisation of the Particulate Matter Inside Two Road Tunnels in the São Paulo Metropolitan Area.” Atmospheric Chemistry and Physics , 13, 12199– 12213. Google Scholar CrossRef Search ADS   Brunekreef Bert, Janssen Nicole A., de Hartog Jeroen, Harssema Hendrik, Knape Mirjam, van Vliet Patricia ( 1997). “ Air Pollution from Truck Traffic and Lung Function in Children Living near Motorways.” Epidemiology , 8, 298– 303. Google Scholar CrossRef Search ADS   Burtscher H. ( 2005). “ Physical Characterization of Particulate Emissions from Diesel Engines: A Review.” Journal of Aerosol Science , 36, 896– 932. Google Scholar CrossRef Search ADS   Castranova Vincent, Ma Jane Y., Yang Hui-Min, Antonini Jim M., Butterworth Leon, Barger Mark W., Roberts Jenny, Ma Joseph K. H. ( 2001). “ Effect of Exposure to Diesel Exhaust Particles on the Susceptibility of the Lung to Infection.” Environmental Health Perspectives , 109, 609– 612. Google Scholar CrossRef Search ADS   CETESB ( 2008–2013, several years). Relatório Anual sobre a Qualidade do Ar no Estado de São Paulo . Companhia Ambiental do Estado de São Paulo, São Paulo. Chay Kenneth Y., Greenstone Michael ( 2005). “ Does Air Quality Matter? Evidence from the Housing Market.” Journal of Political Economy , 113, 376– 424. Google Scholar CrossRef Search ADS   Chen Szu-Ying, Wu Chang-Fu, Lee Jui-Huan, Hoffmann Barbara, Peters Annette, Brunekreef Bert, Chu Da-Chen, Chan Chang-Chuan ( 2015). “ Associations between Long-Term Air Pollutant Exposures and Blood Pressure in Elderly Residents of Taipei City: A Cross-Sectional Study.” Environmental Health Perspectives , 123, 779– 784. Chen Yuyu, Ebenstein Avraham, Greenstone Michael, Li Hongbin ( 2013). “ Evidence on the Impact of Sustained Exposure to Air Pollution on Life Expectancy from China's Huai River Policy.” Proceedings of the National Academy of Sciences , 110, 12936– 12941. Google Scholar CrossRef Search ADS   Choi Hak Soo, Ashitate Yoshitomo, Lee Jeong Heon, Kim Soon Hee, Matsui Aya, Insin Numpon, Bawendi Moungi G., Semmler-Behnke Manuela, Frangioni John V., Tsuda Akira ( 2010). “ Rapid Translocation of Nanoparticles from the Lung Airspaces to the Body.” Nature Biotechnology , 28, 1300– 1303. Google Scholar CrossRef Search ADS   Clougherty Jane E. ( 2010). “ A Growing Role for Gender Analysis in Air Pollution Epidemiology.” Environmental Health Perspectives , 118, 167– 176. Google Scholar CrossRef Search ADS   Conley Tim G., Hansen Christian B., Rossi Peter E. ( 2012). “ Plausibly Exogenous.” Review of Economics and Statistics , 94, 260– 272. Google Scholar CrossRef Search ADS   Currie Janet, Walker Reed ( 2011). “ Traffic Congestion and Infant Health: Evidence from E-ZPass.” American Economic Journal: Applied Economics , 3, 65– 90. Google Scholar CrossRef Search ADS   Davis Lucas W. ( 2008). “ The Effect of Driving Restrictions on Air Quality in Mexico City.” Journal of Political Economy , 116, 38– 81. Google Scholar CrossRef Search ADS   Deschenes Olivier, Greenstone Michael, Shapiro Joseph S. ( 2017). “ Defensive Investments and the Demand for Air Quality: Evidence from the NOx Budget Program.” American Economic Review , 107( 10), 2958– 2989. Google Scholar CrossRef Search ADS   Di Qian, Wang Yan, Zanobetti Antonella, Wang Yun, Koutrakis Petros, Choirat Christine, Dominici Francesca, Schwartz Joel D. ( 2017). “ Air Pollution and Mortality in the Medicare Population.” New England Journal of Medicine , 376, 2513– 2522. Google Scholar CrossRef Search ADS   Diário Oficial da Cidade ( 2010) Dispõe sobre o trânsito de caminhões na Marginal Pinheiros, Av. dos Bandeirantes e outras vias do Município [Regulates the Circulation of Trucks along Marginal Pinheiros, Av. dos Bandeirantes and Other Corridors of the City] . Portaria n.° 108/10-SMT.GAB, São Paulo. Dominici Francesca, Greenstone Michael, Sunstein Cass R. ( 2014). “ Particulate Matter Matters.” Science , 344, 257– 259. Google Scholar CrossRef Search ADS   Dominici Francesca, Peng Roger D., Barr Christopher D., Bell Michelle L. ( 2010). “ Protecting Human Health from Air Pollution: Shifting from a Single-pollutant to a Multi-pollutant Approach.” Epidemiology , 21, 187– 194. Google Scholar CrossRef Search ADS   Donaldson Ken, Tran Lang, Jimenez Luis A., Duffin Rodger, Newby David E., Mills Nicholas, MacNee William, Stone Vicki ( 2005). “ Combustion-Derived Nanoparticles: A Review of Their Toxicology Following Inhalation Exposure.” Particle and Fibre Toxicology , 2, 10, https://doi.org/10.1186/1743-8977-2-10. Google Scholar CrossRef Search ADS   Downs Anthony ( 1962). “ The Law of Peak-Hour Expressway Congestion.” Traffic Quarterly , 16, 393– 409. Duranton Gilles, Turner Matthew A. ( 2011). “ The Fundamental Law of Road Congestion: Evidence from US Cities.” American Economic Review , 101( 6), 2616– 2652. Google Scholar CrossRef Search ADS   Economist ( 2016). “ Cleaning up the Data: The Dangers of Dirty Air Need to be Made Much more Transparent to City-Dwellers.” The Economist . https://www.economist.com/news/leaders/21702753-dangers-dirty-air-need-be-made-much-more-transparent-city-dwellers-cleaning-up. Economist ( 2017). “ To Stop Carmakers Bending the Rules on Emissions, Europe must get Much Tougher.” The Economist . https://www.economist.com/news/leaders/21714991-europe-should-follow-america-taking-hard-line-regulating-car-emissions-stop-carmakers. Engling G., He J., Betha R., Balasubramanian R. ( 2014). “ Assessing the Regional Impact of Indonesian Biomass Burning Emissions Based on Organic Molecular tracers and Chemical Mass Balance Modeling.” Atmospheric Chemistry and Physics , 14, 8043– 8054. Google Scholar CrossRef Search ADS   Eskeland Gunnar S., Feyzioglu Tarhan ( 1997). “ Rationing Can Backfire: The ‘Day without a Car’ in Mexico City.” The World Bank Economic Review , 11, 383– 408. Google Scholar CrossRef Search ADS   Gallego Francisco, Montero Juan-Pablo, Salas Christian ( 2013). “ The Effect of Transport Policies on Car Use: Evidence from Latin American cities.” Journal of Public Economics , 107, 47– 62. Google Scholar CrossRef Search ADS   Garshick Eric, Laden Francine, Hart Jaime E., Rosner Bernard, Smith Thomas J., Dockery Douglas W., Speizer Frank E. ( 2004). “ Lung Cancer in Railroad Workers Exposed to Diesel Exhaust.” Environmental Health Perspectives , 112, 1539– 1543. Google Scholar CrossRef Search ADS   Gauderman W. James, Avol Edward, Lurmann Fred, Kuenzli Nino, Gilliland Frank, Peters John, McConnell Rob ( 2005). “ Childhood Asthma and Exposure to Traffic and Nitrogen Dioxide.” Epidemiology